Hyperspectral Remote Sensing 6 Data Analysis and Future Challenges by José M

Contents | Zoom in | Zoom outFor navigation instructions please click here Search Issue | Next Page

CALL FOR PAPERS

IEEE Geoscience and Remote Sensing Magazine

This is the second issue of the new IEEE Geoscience and Remote Sensing Magazine, which was approved by the IEEE Technical Activities Board in 2012. This is an important achievement for GRSS since it has never had a publication in the magazine format. The magazine will provide a new venue to publish high quality technical articles that by their very nature do not find a home in journals requiring scientific innovation but that provide relevant information to scientists, engineers, end-users, and students who interact in different ways with the geoscience and remote sensing disciplines.

The magazine will publish tutorial papers and technical papers on geoscience and remote sensing topics, as well as papers that describe relevant applications of and projects based on topics addressed by our society.

The magazine will also publish columns on: - New satellite missions - Standard remote sensing data sets - Education in remote sensing - Women in geoscience and remote sensing - Industrial profiles - University profiles - GRSS Technical Committee activities - GRSS Chapter activities - Conferences and workshops.

The new magazine is published in with an appealing layout, and its articles will be included with an electronic format in the IEEE Xplore online archive. The Magazine content is freely available to GRSS members.

This call for papers is to encourage all readers to prepare and submit articles and technical content for review to be published in the IEEE Geoscience and Remote Sensing Magazine. Contributions for the above-mentioned columns of the magazine are also welcome.

All technical papers will undergo blind review by multiple reviewers. The submission and the review process is managed at the IEEE Manuscript Central as it is already done for the three GRSS journals. Prospective authors are required to submit electronically using the website http://mc.manuscriptcentral.com/grs and selecting the “Geoscience and Remote Sensing Magazine” option from the drop-down list. Instructions for creating new user accounts, if necessary, are available on the login screen. No other manners of submission are accepted. Papers should be submitted in single column, double-spaced format. The review process will assess the technical quality and/or the tutorial value of the contributions.

The magazine will publish also special issues. Readers interested to propose a special issue can contact the Editor In Chief.

For any additional information and for submitting papers contact the Editor In Chief:

Prof. Lorenzo Bruzzone University of Trento, Trento, Italy

E-Mail: [email protected] Phone: +39 0461 28 2056

Digital Object Identifier 10.1109/MGRS.2013.2261353

JUNE 2013 VOLUME 1, NUMBER 2 WWW.GRSS-IEEE.ORG______

FEATURES

Hyperspectral Remote Sensing 6 Data Analysis and Future Challenges by José M. Bioucas-Dias, Antonio Plaza, Gustavo Camps-Valls, Paul Scheunders, Nasser M. Nasrabadi, and Jocelyn Chanussot

Spectroradiometric Field Surveys in 37 Remote Sensing Practice: A Workflow Proposal, from Planning to Analysis by L. Pompilio, P. Villa, M. Boschetti, and M. Pepe © GRAPHICS DYNAMIC

PG. 6

ON THE COVER: Airborne and satellite hyperspectral sensors can acquire images of the Earth surface that allow a fine characterization of materials, objects, and biophysical variables.

Digital Object Identifier 10.1109/MGRS.2013.2260233

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 1

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE EDITORIAL BOARD 2013 GRS OFFICERS Dr. Lorenzo Bruzzone President Editor-in-Chief Dr. Melba M. Crawford COLUMNS & University of Trento Purdue University, USA DEPARTMENTS Department of Information Engineering and Computer Science Executive Vice-President Via Sommarive, 5 Dr. Kamal Sarabandi I-38123 Povo, Trento, ITALY University of Michigan, USA 3 FROM THE EDITOR E-mail: [email protected] Vice-President of Meetings and Symposia Dr. Adriano Camps Dr. William Blackwell Universitat Politecnica de 5 PRESIDENT’S MESSAGE MIT Lincoln Laboratory Catalunya-Barcelona Tech, Spain Lexington, MA 02420-9108, USA E-mail: [email protected]______Vice-President of Publications 52 REMOTE SENSING SATELLITES Dr. William Emery Dr. Kun Shan Chen University of Colorado, USA National Central University 68 TECHNICAL COMMITTEES Vice-President of Technical Activities Chungli, TAIWAN Dr. John Kerekes E-mail: [email protected] ______Rochester Institute of Technology, USA 72 CHAPTERS Dr. Paul Gader Vice-President of Professional Activities CISE Dept., University of Florida Dr. Wooil M. Moon 76 ORGANIZATION PROFILES 301 CSE Bldg. University of Manitoba, Canada Gainesville, FL 32611, USA E-mail: [email protected]______Vice-President of Information Resources 80 EDUCATION Dr. Steven C. Reising Dr. John Kerekes Colorado State University, USA 84 WOMEN IN GRS Rochester Institute of Technology 54 Lomb Memorial Dr. Rochester, NY 14623, USA IEEE PERIODICALS 86 CONFERENCE REPORTS E-mail: [email protected]______MAGAZINES DEPARTMENT Associate Editor Dr. Antonio J. Plaza Laura Ambrosio 92 GRSS MEMBER HIGHLIGHTS Department of Technology of Computers and Communications Senior Art Director Escuela Politecnica de Caceres, Janet Dudar 94 CALENDAR University of Extremadura Assistant Art Director Avda. de la Universidad S/N Gail A. Schnitzer E-10071 Cáceres, SPAIN E-mail: [email protected]______Production Coordinator Theresa L. Smith Dr. Gail Skofronick Jackson NASA Goddard Space Flight Center Business Development Manager Code 612 Susan Schneiderman Greenbelt, MD 20771, USA +1 732 562 3946 [email protected] E-mail: [email protected]______Fax: +1 732 981 1855 Dr. Stephen Volz Advertising Production Manager NASA Earth Science Div. Felicia Spagnoli 300 E St., SW Washington, DC 20546, USA Production Director E-mail: [email protected]______Peter M. Tuohy MISSION STATEMENT Editorial Director The IEEE Geoscience and Remote Sensing Dawn Melley Society of the IEEE seeks to advance science Staff Director, Publishing Operations and technology in geoscience, remote sensing Fran Zappulla and related fields using conferences, education, and other resources.

IEEE Geoscience and Remote Sensing Magazine (ISSN 2168-6831) is pub- the first page, provided the per-copy fee indicated in the code is paid through lished quarterly by The Institute of Electrical and Electronics Engineers, the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923 Inc., IEEE Headquarters: 3 Park Ave., 17th Floor, New York, NY 10016-5997, USA; 2) pre-1978 articles without fee. For all other copying, reprint, or repub- +1 212 419 7900. Responsibility for the contents rests upon the authors and lication information, write to: Copyrights and Permission Department, IEEE not upon the IEEE, the Society, or its members. IEEE Service Center (for Publishing Services, 445 Hoes Lane, Piscataway, NJ 08854 USA. Copyright orders, subscriptions, address changes): 445 Hoes Lane, Piscataway, NJ 08854, © 2013 by the Institute of Electrical and Electronics Engineers, Inc. All rights +1 732 981 0060. Price/Publication Information. Subscriptions: included in reserved. Postmaster: Send address changes to IEEE Geoscience and Remote Society fee for each member of the IEEE Geoscience and Remote Sensing Society. Sensing Magazine, IEEE, 445 Hoes Lane, Piscataway, NJ 08854 USA. Canadian Nonmember subscription prices available on request. Copyright and Reprint GST #125634188 PRINTED IN USA Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limits of U.S. Copyright Law for private IEEE prohibits discrimination, harassment, and bullying. For more information, use of patrons: 1) those post-1977 articles that carry a code at the bottom of visit http://www.ieee.org/web/aboutus/whatis/policies/p9-26.html.

_____

Digital Object Identifier 10.1109/MGRS.2013.2260234

2 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

FROM THE EDITOR

BY LORENZO BRUZZONE

his is the second issue of the IEEE Geoscience and enabling a highly detailed analysis of their properties. TRemote Sensing Magazine, which as most of you know This results in many possible applications. The paper is a new publication of the IEEE Geoscience and Remote presents an overview of the relevant hyperspectral data Sensing Society (GRSS) established this year. The goal analysis methods and algorithms. It contains six main of the magazine is to publish tutorial and technical topics: data fusion, unmixing, classification, target detec- papers on geoscience and remote sensing topics, as well tion, physical parameter retrieval, and fast computing. In as papers that describe relevant projects based on and all topics, the paper describes the state of the art, provides applications of topics addressed by our society. All tuto- illustrative examples, and points to future challenges and rial and technical papers will undergo blind review by research directions. The second contribution addresses multiple reviewers. The review process is managed on the problem of spectral analysis from a different perspec- the IEEE Manuscript Central site, as is already done for tive by considering the issue of spectroradiometric field the three GRSS journals. In addition, the magazine will surveys. It presents a proposal for an operational work- publish regular columns on education in remote sens- flow that merges various aspects of in-situ spectral data ing, remote sensing missions, standard data sets, wom- acquisition, from planning to data analysis, and provides en in geoscience and remote sensing, space agency news, a set of well-established guidelines. These in-situ acquisi- book reviews, etc. You may refer to the call for papers in tions are of fundamental importance for multispectral this issue for more details on the topics of the magazine and hyperspectral remote sensing data analysis. and on the procedure for submitting your contributions The Remote Sensing Satellites column describes the to the magazine. The digital edition of the magazine is recently-approved Cyclone Global Navigation Satellite provided to GRSS members who made subscription to System (CYGNSS), which is a micro-satellite constel- the magazine through a subscriber Qmags page, where- lation designed to enhance future hurricane predic- as the electronic version of the papers is available to all tion. The article describes the motivation for using a GRSS members on the IEEE Xplore online archive. Note micro-satellite constellation, the mission design and the that only the first two issues have been distributed in deployment module. the digital format to all GRSS members. The Technical Committees column presents the activi- This second issue is published shortly before the ties of the Instrumentation and Future Technologies IEEE IGARSS 2013 Symposium that will be held on July (IFT) Technical Committee of the GRSS. After a brief 21–26 in Melbourne, Australia. In this issue you can description of the organization of the technical commit- find more information on IGARSS 2013 and on related tee, consisting of working groups on different topics, the educational activities. article describes the activities of the Remote Sensing Instru- The Features section includes two main contributions. ments and Technologies for Small Satellites working group. The first article is an excellent tutorial paper on hyper- The Chapters column presents two main contribu- spectral systems, which are one of the most important tions. The first is a short note that reminds GRSS Chap- technologies in passive remote sensing. Hyperspectral ters of the very important opportunity to use the GRSS systems allow us to acquire very detailed information Distinguished Lecturer Program. This program provides on the spectral reflectance of various types of land cover, GRSS Chapters with seminars by experts on topics of interest and importance to the geoscience and remote sensing community. The program is structured so that Digital Object Identifier 10.1109/MGRS.2013.2260698 Date of publication: 26 June 2013 Chapters incur little or no cost to use it. Therefore, I

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 3

strongly encourage chapters to take advantage of this very can affect academic and work environments in ways that important initiative. The second article describes the inau- are still being studied and discovered. guration of the new Bangalore Section Chapter of the GRSS, The Conference Reports column contains two contribu- which is one of the most recently formed chapters of GRSS. tions. The first is an article describing the Workshop on The article also describes the workshop on “Mathematical Reflectometry using GNSS and Other Signals of Opportunity Morphology and Pattern Recognition: Theory and Applica- (GNSS+R 2012) held at Purdue University, West Lafayette, tions”, organized on March 26–28, 2013, at the Indian Statis- USA on October 10–11, 2012. The second report describes tical Institute-Bangalore Centre, India. the 9th International Conference of the African Association The Organization Profiles column describes the Wallops of Remote Sensing of the Environment (AARSE), held in El Flight Facility (WFF) of NASA Goddard Space Flight Cen- Jadida, Morocco, from October 29 to November 2, 2012. ter (GSFC). The article describes the Suborbital and Special As a final remark, I encourage you to contribute to the Orbital Projects developed at this facility and describes the success of the Magazine by submitting tutorial, technical, related activities. educational, and organization profiles articles that are of The Education column presents an article that introduces interest to our community. the new Director of Education of IEEE GRSS, Prof. Michael I wish you a relaxing and productive summer, and I am Inggs. The article addresses future plans of GRSS on edu- looking forward to meeting many of you at IGARSS 2013 cation and presents the Remote Sensing Summer School in Melbourne. 2013 (actually a Winter School since it is being held in the Sincerely, Australian winter), which will be held on July 18–19, 2013, Lorenzo Bruzzone shortly before IGARSS 2013 in Melbourne, Australia. Editor-in-Chief______

The Women in Geoscience and Remote Sensing presents [email protected] a short article that discusses the unconscious bias, which GRS

Australian Government Geoscience Australia APPLYING GEOSCIENCE TO AUSTRALIA‘S MOST IMPORTANT CHALLENGES

'!!!! ! ! & " . / !!"!! " ! !! ! !4235 & ". " (43-46"&4235/* " ! $ # !, ! !$!" ! # !! "!!" ! " (! # !("!&$ " ! &*

! ! ! & ". / ! "&#! $ ! !! % ! ( (! ( ! ! !"! "!$ ! #$!! ! " !" "! ! ! #! ! !( (! *

#! ! %! !!$ !"! (# ! ("! !!"! ( ! ( ! ! (! %!! ! ! ! "! ("!( !$ ( " ( ! !!$ 0 " ! & "*

Exciting sponsorship opportunities are still available for interested parties. For more details on how you or your organisation can get involved please visit: http://www.igarss2013.org/SponsorExhibition.asp

Digital Object Identifier 10.1109/MGRS.2013.2261356

4 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

PRESIDENT’S MESSAGE

BY MELBA CRAWFORD

n behalf of the IEEE Geoscience and Remote Director, Mike Inggs, to develop a program with top- OSensing Society, I invite you to participate in ics ranging from analysis of remotely sensed data to IGARSS 2013, our annual International Geoscience applications of remote sensing for carbon accounting and Remote Sensing Symposium. This year’s sympo- and coastal vulnerability. A full day of pre-conference sium will be held in Melbourne, Australia from July tutorials led by internationally recognized experts has 21–26 and will focus on the theme “Building a Sus- also been organized. Social events, tainable Earth Through Remote Sensing”. The theme including a Young Professionals was selected by the local organizing team to empha- Luncheon and the traditional Soc- size the issues that most affect the Earth’s environ- cer Game (the IGARSS Futsal World IGARSS 2013 WILL FOCUS Cup) are also planned. ment and the human impact on the planet. (Please ON THE THEME “BUILDING refer to http://igarss2013.org/ for more information.) Located on the southern coast A SUSTAINABLE EARTH The organizing team, led by general co-chairs Peter of Australia, Melbourne is a beau- THROUGH REMOTE Woodgate and Simon Jones, has developed an out- tiful, multicultural city that is standing conference that includes a plenary session renowned for its international SENSING”. THE THEME with presentations by internationally recognized key- flair and hospitality, providing WAS SELECTED TO note speakers in remote sensing and GIS. The Tech- an outstanding destination for EMPHASIZE THE ISSUES nical Program is the culmination of the combined IGARSS 2013. The Melbourne THAT MOST AFFECT THE contributions of colleagues who participated in the Exhibition and Convention Cen- EARTH’S ENVIRONMENT review and abstract selection process, and the Tech- tre (MECC) is the venue for the AND THE HUMAN IMPACT nical Program Committee that met in Los Angeles, conference. Completed in 2008, ON THE PLANET. California, in March. I congratulate and offer special it is a state-of-the-art facility that thanks to Clive Fraser and Jeff Walker, who are doing was designed to be the world’s first an outstanding job as co-chairs of the IGARSS techni- 6 star green energy rated conven- cal committee. The IGARSS 2013 technical program tion center. The region also offers is now available at http://igarss2013.org/RegularPro-______a wide range of outstanding opportunities for a unique gram.asp.______pre- or post-conference holiday. The local conference IGARSS 2013 will also include activities of spe- team is providing excellent support to help you plan a cial interest to students and young professionals. The great getaway “Down Under”. popular GRSS Remote Sensing School, which was first We look forward to seeing you in Melbourne at organized at IGARSS 2012, is being hosted again as a IGARSS 2013! pre-conference event. Our thanks to Xiuping Jia and Best regards, Kim Lowell, who have collaborated with our Education Melba Crawford President, IEEE GRSS [email protected] Digital Object Identifier 10.1109/MGRS.2013.2260699 ______Date of publication: 26 June 2013 GRS

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 5

JOSÉ M. BIOUCAS-DIAS, ANTONIO PLAZA, GUSTAVO CAMPS-VALLS, PAUL SCHEUNDERS, NASSER M. NASRABADI, AND JOCELYN CHANUSSOT

Abst ract—Hyperspectral remote sensing technology has advanced significantly in the past two decades. Current sensors onboard airborne and spaceborne platforms cover large areas of the Earth surface with unprecedented spectral, spatial, and temporal resolutions. These characteristics enable a myriad of applications requiring fine identification of materials or estimation of physical parameters. Very often, these applications rely on sophisticated and complex data analysis methods. The sources of difficulties are, namely, the high dimensionality and size of the hyperspectral data, the spectral mixing (linear and nonlinear), and the degradation mechanisms associated to the measurement process such as noise and atmospheric effects. This paper presents a tutorial/overview cross section of some relevant hyperspectral data analysis methods and algorithms, organized © DYNAMIC GRAPHICS in six main topics: data fusion, unmixing, classification, target detection, physical parameter retrieval, and fast computing. In all topics, we describe the state-of-the-art, provide illustrative examples, and point to future challenges and research directions.

I. INTRODUCTION yperspectral remote sensing is concerned with the extraction of information from objects or Hscenes lying on the Earth surface, based on their radiance acquired by airborne or spaceborne sensors [1], [2]. Hyperspectral sensing, namely its imaging modality termed hyperspectral imaging, has been increasingly used in applications at lab scale (e.g., food safety, pharmaceutical process monitoring and quality control, biomedical, industrial, biometric, and forensic) using small, commercial, high spatial

Digital Object Identifier 10.1109/MGRS.2013.2244672 Date of publication: 26 June 2013

and spectral resolution instruments (see [3] and refer- correlated and thus lives in a low dimensional manifold. ences therein). This is illustrated at the bottom of Fig. 2, where the spectral Figure 1 gives a partial account of the relevance of vectors of soil, vegetation, and water are represented as Rnb hyperspectral applications, by comparing paper counts dimensional points on a surface. per year in the hyperspectral and radar areas. These results In terms of the geometrical properties of a remote sens- were obtained by searching the SCI-Expanded database of ing imaging system, the spatial resolution of a sensor is the ISI Web-Of-Science with the topics “(hyperspectral) given by its field of view (FOV), and the obtained spectrum and (remote sensing),” in the left hand side, and “(radar) is the average of the material’s reflectances within this FOV. and (remote sensing),” in the right hand side. We conclude The spectral resolution is determined by the bandwidth that the number of items per year in 2011 is similar for the of the spectral bands. When spatially and spectrally sam- hyperspectral and radar areas, with a clear increasing trend pling the information (we will assume that the sampling is in the former and a stabilization or decrease in the latter. performed at the sensors spatial and spectral resolution), ## In hyperspectral imaging, also termed imaging spec- a 3D “hypercube” X ! Rnnn12b is obtained, containing

troscopy [4], the sensor acquires a spectral vector nn= 12# n pixels and nb bands (see Fig. 2). Different forms with hundreds or thousands of elements from of representation can be used for HSIs: every pixel in a given scene. The result is ◗ In the spectral representation, each pixel is defined in the so-called hyperspectral image the spectral space x ! Rnb. Since neighboring spectra cor- (HSI). It should be noted that HSIs respond to similar materials, grouping in this spectral are spectrally smooth and spa- space is commonly applied to characterize materials. tially piece-wise smooth: the This can be done by clustering neighboring spectra, or values in neighboring loca- by supervised classification (see section on Classifica- tions and wavelengths tion). Since the spectral correlation is high, the data are are highly correlated. likely to reside on a very low-dimensional submanifold This can be observed of the spectral space, and projection of the data on a sub-

by extremely non- space of dimension dn% b, using, e.g., principal compo- diagonal covariance nent analysis (PCA) [3], is commonly applied. matrices and wide ◗ In the spatial representation, each image band is a nn12# autocorrelation func- matrix Xi ! R . Because of the high spatial correla- tions [1]. This piece- tion, neighboring pixels are likely to belong to a simi- wise smoothness lar material and spatial grouping (e.g., segmentation) is holds as well in the commonly applied. spatio-spectral direc- ◗ In the spatial-spectral representation spectral processing tion. The characteristics of a pixel is performed taking neighboring pixels into are similar to those of nat- account, while spatial processing of an image band is ural photographic images performed by accounting for the other bands. and videos and, therefore, These representations have been actively exploited, many tools that were developed namely, in dimensionality reduction, feature extraction, for these data can be extended for unmixing, classification, segmentation, and detection [1], HSI analysis. [5]. Still related with the high dimensionality of the spectral An equivalent interpretation of an HSI information, the most recent trend is sparse and redundant is given by the acquisition of a stack of images modeling, which is currently reaching the areas of, e.g., representing the radiance in the respective band restoration, unmixing, classification, segmentation, and (wavelength interval). Due to this interpretation, the HSIs detection (see, e.g., [6], [7] and references therein). are also termed hyperspectral data cubes. These two points Since the output of a hyperspectral sensor provides raw of view are illustrated in the top left hand side of Fig. 2, digital number (DN) values and for quantification pur- where the HSI has nb spectral bands and nn12# pixels. The poses, a conversion to apparent surface reflectance values plots on the top right hand side show the spectra of pixels is required before using advanced information extraction containing soil, vegetation, and water. Owing to the high techniques such as those mentioned above [8]. The char- spectral sampling, the spectral information is often highly acteristics of the sensor itself are described by its transfer

José M. Bioucas-Dias is with the Instituto de Telecomunicações, Instituto Superior Técnico, Lisbon, 1049-1, Portugal (e-mail: [email protected]). Antonio Plaza is with the Hyperspectral Computing Laboratory, Department of Technology of Computers and Communications, Escuela Politécnica de Cáceres, University of Extremadura, 10003 Cáceres, Spain (e-mail: [email protected]).______Gustavo Camps-Valls is with the Image Processing Laboratory, Universitat de València, C/Catedrático Escardino, E-46980 Paterna (València), Spain (e-mail: [email protected]).______Paul Scheunders is with iMinds, Vision Lab, Depart- ment of Physics, University of Antwerp, 2610 Wilrijk, Belgium (e-mail: [email protected]).______Nasser Nasrabadi is with the U.S. Army Research Labo- ratory, Adelphi, MD 20783 USA (e-mail: [email protected]).______Jocelyn Chanussot is with the GIPSA-Lab, Grenoble Institute of Technology, Grenoble, France (e-mail: [email protected]).______This work was supported by the Portuguese Science and Technology Foundation, project PEst-OE/EEI/0008/2013 and by the Spanish Ministry of Economy and Competitiveness (MINECO) under projects TIN2012-38102-C03-01 and AYA2011-29334-C02-02.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 7

Published Items in Each Year

350 400 Hyperspectral Radar 300 350

250 300 250 200 200 150 150 100 100

50 50

0 0 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 (a) (b)

FIGURE 1. Paper counts per year in hyperspectral and radar topics obtained by searching the SCI-Expanded database of the ISI Web-Of- Science with the following topics: (a) hyperspectral and remote sensing; (b) radar and remote sensing. Search done on January 2013.

[9]. Finally, one has to account for the effects of illumination X and viewing angle and the surfaces structural and optical Soil properties, to lead to the surface reflectance values. The interactions of the light with the atmosphere are extremely complex. The radiative transfer theory is often Veg used to derive models for these interactions [10]. Originally n2 developed for simulating TOA radiance for the preparation of future satellite missions, inversion of these models Water nb allows for atmospheric correction. The bidirectional reflec- n1 tance distribution functions (BRDF) describe the reflected (a) light on a surface as a function of the incoming and outgo- ing light directions. A first approximation for the BRDF, R termed ’albedo’, is the ratio between the reflectance and the sun’s irradiance. BRDF’s are usually accompanied by Vegetation Water complex surface structure models such as leaf-to-canopy models. A schematic overview of the spectral characteristics of hyperspectral data is given in Fig. 3. We remark that, Soil besides for the conversion to surface reflection, all these models are relevant as well for quantitative analysis (see Sections III and VI). (b) For illustrative purposes, Table 1 displays spatial and spectral parameters of eight hyperspectral instruments: two airborne (HYDICE1 and AVIRIS2) and six spaceborne FIGURE 2. (a) Hyperspectral imaging concept. (b) Hyperspectral 3 4 5 6 7 vectors represented in a low-dimensional manifold. (HYPERION , EnMAP , PRISMA , CHRIS , HyspIRI and IASI8). From this list, EnMAP, PRISMA and HyspIRI are

function. To account for this, first a radiometric calibration 1 ______http://rsd-www.nrl.navy.mil/hydice of the spectra is generally performed to obtain at-sensor or 2 http://aviris.jpl.nasa.gov______3 top-of-atmosphere (TOA) radiance values. As the reflected ______http://eo1.usgs.gov 4 http://www.enmap.org______sunlight passes through the atmosphere, it gets partially 5 http://www.asi.it/en/flash______en/observing/prisma 6 absorbed and scattered. Since these effects have a huge https://earth.esa.int/web/guest/missions/esa-operational-eo-missions/______proba___ influence on the spectral values, they need to be corrected 7 http://hyspiri.jpl.nasa.gov______8 to obtain the ground-leaving radiance or reflectance values http://smsc.cnes.fr/IASI______

8 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

not yet operational. The spatial resolutions are higher for sensors carried by low altitude platforms and vice-versa. DN The spectral coverage of HYDICE, AVIRIS, HYPERION, Sensor Transfer Sensor Radiometric EnMAP, PRISMA and HyspIRI corresponds to the vis- Function Calibration ible, the near-infrared, and the shortwave infrared spectral bands, whereas CHRIS covers the visible bands and TOA Radiance IASI covers the mid-infrared and the long-infrared bands. Atmospheric The number of bands is approximately 200 for HYDICE, Atmospheric RTM Atmosphere Correction AVIRIS, HYPERION, EnMAP, PRISMA and HyspIRI, with a spectral resolution of the order of 10 nm. The number of Ground-Leaving Reflectance bands for CHRIS is 63, with spectral resolutions of 4 and 12 Viewing Geometry nm (depending on the region of the spectrum) and 8461 BRDF Surface and -1 for IASI, with a resolution of 05cm. . In any case, the reso- Surface Correction lution is very high (offering a huge potential to discrimi- Surface Reflectance nate materials) in the case of the first seven sensors, and to estimate physical parameters (temperature, moisture and FIGURE 3. Spectral characterization of hyperspectral data. trace gases across the atmospheric column), in the case of the IASI sensor. A summary of the characteristics of several hyperspectral imaging instruments currently in operation, Agency (ESA) already flagged up in 2011 that “data rates and under construction, and missions in a planning stage has data volumes produced by payloads continue to increase, while the been recently provided [11]. available downlink bandwidth to ground stations is comparatively Several factors make the analysis of hyperspectral data stable” [14]. In this context, the design of solutions aimed an often complex and hard task calling for sophisticated at taking advantage of the ever increasing dimensionality methods and algorithms. Among these factors, we refer to of remotely sensed hyperspectral images for near real-time spectral mixing (linear and nonlinear), and degradation applications has gained significant relevance and momen- mechanisms associated to the measurement process (e.g., tum during the last decade [15], [16]. noise and atmosphere). Another important issue is the This paper presents a tour over relevant and distinctive extremely high dimensionality and size of the data, result- hyperspectral data analysis themes, organized in six main ing from the improved spatial, spectral and temporal resolu- topics: data fusion, unmixing, classification, target detec- tions provided by hyperspectral instruments. This demands tion, physical parameter retrieval, and fast computing. Most fast computing solutions that can accelerate the interpreta- of the frameworks used in these topics are rooted on signal tion and efficient exploitation of hyperspectral data sets and image processing, statistical inference, and machine in various applications [12]. For example, it has been esti- learning fields. In all topics, we describe the state-of-the-art mated by the NASA’s Jet Propulsion Laboratory (JPL) that and point to the most likely future challenges and research a volume of 4.5 TBytes of data will be daily produced by directions. Illustrative examples with real data are provided HyspIRI (1630 TBytes per year). Similar data volume ratios for some of the topics covered. are expected for EnMAP and PRISMA. Unfortunately, this The remainder of the paper is organized as follows. extraordinary amount of information jeopardizes the use Section II discusses processing techniques aimed at fus- of latest-generation hyperspectral instruments in real-time ing spatial and spectral information from multiple or near real-time applications, due to the prohibitive delays observation and sources. Section III addresses linear and in the delivery of Earth Observation payload data to ground nonlinear hyperspectral mixing and unmixing. Section processing facilities [13]. In this respect, the European Space IV outlines some of the main techniques and challenges

TABLE 1. PARAMETERS OF EIGHT HYPERSPECTRAL INSTRUMENTS.

PARAMETER HYDICE AVIRIS HYPERION EnMAP PRISMA CHRIS HyspIRI IASI Altitude (km) 1.6 20 705 653 614 556 626 817 Spatial resolution (m) 0.75 20 30 30 5–30 36 60 V: 1–2 km H: 25 km Spectral resolution (nm) 7–14 10 10 6.5–10 10 1.3–12 4–12 0.5 cm-1 Coverage (μm) 0.4–2.5 0.4–2.5 0.4–2.5 0.4–2.5 0.4–2.5 0.4–1.0 0.38–2.5 3.62–15.5 and 7.5–12 (645–2760 cm-1) Number of bands 210 224 220 228 238 63 217 8461 Data cube size 200 # 320 512 # 614 660 # 256 1000 # 1000 400 # 880 748 # 748 620 # 512 765 # 120 (sample # lines # bands) # 210 # 224 # 220 # 228 # 238 # 63 # 210 # 8461

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 9

in hyperspectral image classification. Section V addresses B. SPECTRAL DATA FUSION hyperspectral target detection. Section VI reviews the main Here, we discuss the fusion of spectral bands of an HSI, in problems and methods in model inversion and estimation this way removing high spectral redundancy. Since the high of physical parameters, and finally Section VII outlines sev- number of bands causes dimensionality problems, a dimen- eral strategies to accelerate the hyperspectral image compu- sionality reduction of the hyperspectral vectors can highly tations using different hardware architectures. facilitate the analysis afterwards. The goal is to obtain an image of reduced number of bands while trying to preserve II. DATA FUSION the most useful spectral information as possible. The sim- In this section, we will discuss hyperspectral processing plest way is to select a few of the available bands, but it is techniques (image in-image out), that fuse spatial and clear that better performance can be obtained when bands spectral information from one or multiple hyperspectral are fused together. Traditionally, methods based on PCA observations, or a combination of hyperspectral images are applied that decorrelate bands. In many occasions, the and other image sources. We will refer to this processing as dimensionality reduction is applied for an improved clas- data fusion. In Fig. 4, a schematic overview of the different sification afterwards. This topic is treated in Section IV strategies is given. (IV.A.1 and IV.A.2). A specific application of spectral data fusion is the visu- A. RESTORATION alization of HSIs. A user may need to visualize hyperspec- Signal processing techniques can be applied to restore or tral image data for exploration purposes, e.g., for generating improve the signal-to-noise ratio (SNR) and/or the spatial ground reference data. However, an HSI contains far more resolution. In the case of gray-scale images, many denois- image bands than can be displayed on a standard tristimu- ing and deconvolution techniques were developed to restore lus display. By fusion of the spectral bands, an image of lim- SNR and spatial resolution. It is clear that a band-by-band ited number of bands can be generated, e.g., a panchromatic treatment of the restoration problem in HSIs would not image or an RGB image; how to fuse preserving as much benefit from the high spectral redundancy. The traditional information as possible is an issue. In [20], hyperspectral image restoration techniques are extended to account for images are linearly projected onto color matching basis this spectral redundancy. In this way, hyperspectral image functions specifically designed as RGB primaries of a stan- denoising techniques were recently developed by, e.g., dard tristimulus display. A spatio-spectral approach allows employing spatial-spectral information [17] or employing to retain spatial details as well, and often generally generates tensor decompositions and multilinear algebra [18]. In [19], high-contrast images. Spatio-spectral methods that were restoration of hyperspectral images was proposed based on developed use e.g. wavelet transforms to fuse multiresolu- anisotropic diffusion filtering. Remark that all the above tion information of the image bands [21], Markov Random mentioned methods preserve the original spatial and spec- Fields that model the spatial relationship between neigh- tral sampling and thus do not improve the spatial resolution. boring pixels [22] or constrained optimization to enforce spatial smoothness [23]. In Fig. 5, four different color visualizations of an AVIRIS hyperspectral image of 224 spectral bands are shown, obtained by PCA and the methods of [20]–[22] respectively. Restoration Spatial-Spectral Data Fusion C. SPATIAL DATA FUSION (MULTI- (Single Frame Superresolution) FRAME SUPERRESOLUTION) The term (geometric) superresolution (SR) refers to the enhancement Spectral Data Fusion of the spatial resolution of imaging sensors by inferring information at the subpixel level. Subpixel image information is for instance available as subpixel shifts of mul- Multisource Data Fusion tiple low-resolution observations (multiframe SR). In practice, the images are subsampled by divid- ing each pixel into mm# subpix- Spatial Data Fusion (Multiframe Superresolution) els and interpolating the pixel gray levels. Then, corresponding areas FIGURE 4. A schematic overview of the five different hyperspectral data fusion methodologies. between the multiple observations

10 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

(a) (b)

FIGURE 5. Visualization of hyperspectral AVIRIS image, from (a) to (d) using PCA and the methods of [20]–[22], respectively. are detected, geometrically registered and combined to Since a low-resolution hyperspectral pixel contains a generate one image of high spatial resolution. spectral mixture of different materials, superresolution can When applying SR on hyperspectral images, each image be accomplished by a spatial localization of the materials band can be processed separately, but it is clear that a joint fractions at subpixel level. How to obtain the fractions of processing of all bands is superior [24]. Multiple low-res- the present materials will be explained in the section on VI olution hyperspectral observations from the same scene devoted to spectral unmixing. The fractions can be obtained are obtained for instance by overlapping flight lines, multi- as well by using probabilistic classifiers that assign classifica- angle data [25], or multiframe instances in time [26]. tion probabilities for each of the materials classes to the pixel. Then the pixel is subsampled. Subpixel mapping or super- D. SPATIAL-SPECTRAL DATA resolution mapping refers to techniques that try to spatially FUSION SUPERRESOLUTION organize the fractional spectra of the different materials Another way of performing superresolution is by fusion of within a pixel [28]. This mapping can then further be used to different parts of a single image (single frame SR), which in simulate a subsampled hyperspectral image [29], [30]. the case of hyperspectral data amounts to spatial-spectral fusion. In [27], the interband spatial subpixel shifts that E. MULTISOURCE DATA FUSION are intrinsically present in a hyperspectral data cube are A third way of performing superresolution of an HSI is applied for obtaining a SR image. by the use of other available image sources of high spatial

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 11

resolution, acquired by other sensors [e.g. mounted on Finally, some but certainly not all of the methods men- unmanned aerial vehicles (UAV)]. Several strategies are tioned explicitly make use of transfer function information possible. First of all, many methods originally designed for of the employed sensors. When available, it is expected pansharpening, i.e. fusion of multispectral images with a that this information improves the performance of data high spatial resolution panchromatic image [31] are eas- fusion methods. ily transferable to HSIs. A large majority of these methods rely in one way or another on the injection of high spatial III. HYPERSPECTRAL UNMIXING information of the panchromatic image into the hyper- The signal recorded by a hyperspectral sensor at a given spectral image bands. Another approach is to assume a band and from a given pixel, letting alone the effects of joint statistical model between the two image sources and the atmosphere, is a mixture of the “light” scattered by the apply Bayesian estimation techniques for enhancing the constituent substances located in the respective pixel cover- spatial resolution of the HSI [32], [33]. Alternatively, simi- age. Fig. 6 illustrates three types of mixtures owing to low lar subpixel mapping strategies as in the spatial-spectral spatial resolution of the sensor (a), presence of intimate fusion can be applied, in which the high-resolution image mixtures (b), and multiple light scattering in a two-layer can deliver the required materials spectral information media (c). As a result, when mixing occurs, it is not any- [34], [35]. Alternatively, the local correlation with the high more possible to determine what materials are present in spatial resolution image can be employed [36]. the pixels directly from the respective measured spectral vectors. This is to say that the key feature of the spectral F. CHALLENGES sensors, which is its ability to discriminate materials based All described methods for enhancement of the spatial res- on the their spectral responses, is compromised. This sec- olution will generate images at a higher spatial sampling tion addresses spectral mixing modeling, provide insights that show higher contrast and finer details, but this does on the spectral unmixing inverse problems, and point to not necessarily guarantee an improvement of the actual algorithms to solve them. spatial resolution [37]. In particular for HSIs, the described With the objective of recovering the ability to discrimi- methods will be very useful for exploratory analysis and nate materials, an impressive amount of research work has visualization purposes. However, a quantitative analysis been devoted to hyperspectral unmixing (HU) (see, e.g., [3], requires a high reliability of the obtained spectra. Further [39], and references therein). HU is, however, a hard inverse research needs to be conducted on validation methodolo- problem. The difficulties begin with its formulation. Put in gies of these fusion methods [38]. simple terms, given a measured spectral vector y ! Rnb, HU Moreover, with technological progress, the spatial aims at explaining y in terms of the spectral properties of resolution of sensors improves largely. Also the employ- the materials present the respective pixel an of its distribu- ment of UAV’s leads to very high spatial resolution data. tion. An useful treatment of this problem cannot be given While most spatial resolution enhancement methods aim without a formal model, y = f()i , where f()$ is the so-called at a resolution improvement of a factor of 2-3, in practice, forward operator, linking the measurements y to the scene resolution differences on the order of a factor of 10 need to parameters i. In conclusion, the precise meaning of HU be bridged. depends on the meaning of parameter vector i.

Linear Mixture Two-Layers: Canopies + Ground Intimate Mixture

y y y

m1 m2 m3

a a a = R a + R a 9 1 2 3 y imi ij mi mj y = f(i) Media Parameters ! = R a iij y i mi i Single Scattering Double Scattering

(a) (b) (c)

FIGURE 6. Schematic view of three types of spectral mixing. (a) Linear mixing in a checkerboard type surface. (b) Nonlinear (linear plus bilinear) mixing in a two-layer media. (c) Nonlinear mixing in an intimate (particulate) media.

12 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

RTT is a mathematical model for the transfer of energy HU as a semiblind approach in which the endmember as photons interacts with the materials in the scene, and identification is replaced with a SR over a library of spectral thus to the derive forward operators necessary to solve HU signatures, usually overcomplete, obtained in laboratory. problems. The core of the RTT is a differential equation Details of this approach are given in section III-G. describing radiance read by the sensor. It can be derived via the conservation of energy and the knowledge of the phase A. LINEAR UNMIXING function, which represents the probability of light with a Under the LMM (1), a given measured hyperspectral vec- given propagation direction be scattered into a specified tor can be written as yM=+a w, where Mm:[= 1 ,f , mp ] T angle solid around a given scattering direction. stands for the mixing matrix, a:[,= aa1 f ,p ] stands for In general, the forward operator f()i is not invertible, the fractional abundance vector, and w accounts for addi- unless we have partial knowledge of vector i, which usu- tive perturbations due to, for example, model mismatches ally depends on scene parameters often very hard to obtain. and additive noise. Because the Three notable exceptions to these scenario, schematized in components of a represent frac- Fig. 6, are the linear model, the bilinear model, and the tions, then they satisfy the con-

Hapke model [3], [40]. These are three approximations for straints ai $ 0, for ip= 1,,f and SPECTRAL UNMIXING IS / p a = the analytical solution to the RTT suitable to unsupervised i = 1 i 1, termed abundance THE INVERSE PROBLEM applications, i.e., when no prior knowledge exits about the nonnegativity constraint (ANC) and OF DETERMINING THE materials and its distributions. abundance sum constraint (ASC), SPECTRALLY PURE The linear mixing holds true when the mixing scale respectively. Owing to signature COMPONENTS is macroscopic and the incident light interacts with just variability, the ASC is seldom (ENDMEMBERS) PRESENT one material, as is the case in checkerboard type scenes observed in real applications. [40] schematized in Fig. 6. The light from the materials, Nevertheless, because the spec- IN MIXED PIXELS, AS although almost completely separated, is linearly mixed tral vectors are non-negative, is it WELL AS THEIR within the measuring instrument, owing to insufficient always possible to build rescaled ABUNDANCE FRACTIONS. spatial resolution. Formally, the measured spectral vector versions thereof, belonging to an = f T = f y : [,yy1 ,nb ], holding the radiances at bands in1,,,b affine set [41], and thus satisfying is expressed as the ASC (see [3] for details). We

p assume, therefore, that the ASC holds true. ym= / aii, (1) Before unmixing, the hyperspectral data set usu- = i 1 ally undergoes atmospheric calibration and dimension nb where mi ! R , for ip= 1,,,f is the spectral signature of reduction. The atmospheric calibration step converts the the ith material, termed endmember, and ai is the percent- measured radiance into reflectance, which is an intrinsic age that the ith material occupies inside the pixel, termed characteristic of the materials. However, the unmixing fractional abundance or simply abundance. Inspired in the inverse problem can also be formulated in the radiance linear mixing model (LMM), the HU problem is very often data, provided that the effects of atmosphere are pixel invar- defined as the unsupervised estimation of the endmembers iant. The dimension reduction step (see IV-A for additional and of the respective fractional abundances. details) identifies the subspace where the spectral vectors The LMM has been widely used in the past decade to live and projects them onto this subspace. Given that the address HU problems. The reason is threefold: a) despite identified subspace is generally of much lower dimension its simplicity, LMM is an acceptable approximation for than that of the spectral vectors, this projection yields con- the light scattering in many real scenarios; b) under suit- siderable gains in algorithm performance and complexity, able conditions of the data set, LMM yields well-posed data storage, and noise reduction. inverse problems; c) under the LMM, HU is interpret- Suppose we are given a hyperspectral data set contain- able as a blind source separation (BSS) problem or a non- ing n spectral vectors of size nb arranged in the matrix nnb # negative matrix factorization problem, which have been Yy:[,= 1 f ,] yn ! R . Defining the abundance fraction vastly researched in many signal processing areas. In sec- matrix A :[= a1 ,f ,an ], where ai represents the fractional tion III-A, we address in more details relevant aspects of abundance vector of the ith pixel, then the linear HU HU under the LMM. inverse problem can be stated as In spite of the LMM attractiveness, researchers are min <

Unmixing via sparse regression (SR) is still another the componentwise sense and 1p and 1p are column vectors direction recently introduced to circumvent part of the limi- with p and n ones, respectively. Note the inequality A0$ , is T tations of the blind linear HU. This line of attack formulates the ANC and the equality 1Ap = 1n is the ASC.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 13

green denote non-pure spectral vectors, whereas the points in red are pure spectral vectors, thus corresponding to the

m1 vertices of the simplex. Note that the inference of the mix- y = Ma ing matrix M amounts to identify the vertices of the simplex C. This geometrical point of view has been exploited by many unmixing algorithms, which can be mainly classi- fied either as pure pixel or non-pure pixel based. m2 B. PURE PIXEL BASED ALGORITHMS In the pure pixel based algorithms it is assumed the presence in the data of at least one pure pixel per endmember, meaning that there is at least one spectral vector on each vertex of C = conv{M} m3 = 2-simplex the data simplex. This geometric picture is illustrated in the left hand side of Fig. 8. The pure pixel assumption, though enabling the design of very efficient algorithms from the computational point of view, is a strong requisite that may Illustration of the simplex set C for p = 3 (C is the FIGURE 7. not hold in many datasets. These class of algorithms have convex hull of the columns of M, C = conv{}M ). Green circles been the most often used in linear HU applications, per- represent spectral vectors. Red circles represent vertices of the haps because of their light computational burden and clear simplex and correspond to the endmembers. conceptual meaning. Most of the pure pixel based algorithms exploit one of The optimization (2) is interpretable both as a linear the following properties of the endmember signatures: a) blind source separation problem and as a matrix factoriza- the extremes of the projection of the spectral vectors onto tion problem. In the former case the independent compo- any subspace correspond to endmembers; b) the volume nent analysis (ICA) come to mind to separate sources (i.e., defined by any set of p spectral vectors is maximum when the fractional abundances). ICA have in fact been consid- those are endmembers. Representative algorithms of class ered to solve spectral unmixing problems. Unfortunately, a) are pixel purity index (PPI) [43], vertex component analysis ICA is based on the assumption of mutually independent (VCA) [44], simplex growing algorithm (SGA) [45] successive sources, which is not the case of hyperspectral data, since volume maximization (SVMAX) [46], and the recursive algo- the sum of abundance fractions is constant, implying sta- rithm for separable NMF (RSSNMF) [47]; Representative algo- tistical dependence among them. This dependence com- rithms of class b) are N-FINDR [48], iterative error analysis promises ICA applicability to hyperspectral data as shown (IEA), [49], sequential maximum angle convex cone (SMACC), in [42]. and alternating volume maximization (AVMAX) [46]. 1) The convex geometry of linear unmixing: In order to shed light into the linear HU problem, we now give an interpre- C. NON-PURE PIXEL BASED ALGORITHMS tation of problem (2) based on convex geometry. The set Fig. 8, middle and right hand side, schematizes two data sets without pure pixels; the data set in the middle does not p contain pure pixels but contains at least p - 1 spectral vec- Cjp:{==yMa :/ aajj =101 ,$ , = ,,}f j = 1 tors on each facet. In this data set, the endmembers may be inferred by fitting a minimum volume (MV) simplex to the i.e., the convex hull of the columns of M, is a ()p - 1 -simplex data; this rather simple and yet powerful idea, introduced in Rnb. Fig. 7 illustrates a 2-simplex C for a hypothetical mix- by Craig in his seminal work [41], underlies several geo- ing matrix M containing three endmembers. The points in metrical based unmixing algorithms.

m1 m1 m1

m2 m3 m2 m3 m2 m3

FIGURE 8. Illustration of the concept of simplex of minimum volume containing the data for three data sets.

14 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

From an optimization point of view, the MV based description length (MDL) principle; 2) a generalized expec- unmixing algorithms are formulated as tation maximization (GEM) algorithm is derived to infer the model parameters. 2 min <

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 15

Soil 1 Shade Trees Soil 2 Soil 1 Shade Trees Soil 3 Grass

Soil 2

Soil 3

Grass

(a) (b)

Data Soil 1 N-FINDR VCA

Trees Grass

(e) (f)

FIGURE 9. Unmixing results: (a) TERRAIN HSI; (b) identified endmembers; (c) data projection onto the subspace defined by the first two eigen directions; (d) soil abundance map; (e) tree abundance map; and (f) grass abundance map.

that are very hard or impossible to obtain. For this reason, usually simpler strategies are applied using data-driven but the research on nonlinear HU is far more immature com- physics-inspired models, such as the bilinear and Hapke’s pared to linear HU. To avoid the complex physical models, models.

16 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

The bilinear model is valid when the scene can be par- the data without the need for a priori knowledge of the titioned in successive layers with similar scattering proper- mixture types. In addition, the explicit modeling of both ties. Fig. 6, in the middle, schematizes a two-layer scene and mixture types allows for direct estimation of the end- shows the expression for the measured light. The sum on members [65]. the left hand side accounts for the single scattering and is similar to the LMM; the sum on the right hand side accounts G. UNMIXING VIA SPARSE REGRESSION nb for the double scattering, where the vectors mmij9 ! R HU via SR has recently been introduced with the objective (symbol 9 stands for elementwise multiplication) account of coping with data sets not fulfilling the geometrical or for pairwise interactions. statistical assumptions on which the HU methodologies Fig. 6, right hand side, illustrates an intimate mixture, presented in the previous sections rely. In the SR formu- meaning that the materials are in close proximity and the lation, it is assumed that the measured spectral vectors mixture occurs at a microscopic level. The Hapke approxi- can be expressed as linear com- mation [40] for intimate mixtures models the reflec- binations of a small number of tance as a nonlinear function of a convex combination of pure spectral signatures known SPARSE REGRESSION the individual endmember albedos. The coefficients of in advance [6] (e.g., spectra col- TECHNIQUES ALLOW the linear combination are the relative geometric cross- lected on the ground by a field TACKLING THE UNMIXING sections of the components. When the endmember par- spectro-radiometer). Unmixing PROBLEM USING ticle sizes and densities are similar, the coefficients are then amounts to finding the good approximations for the mass fractions of the differ- optimal subset of signatures in (POTENTIALLY VERY ent endmembers. However, in general, one needs infor- a (potentially very large) spec- LARGE) GROUND mation concerning the particle sizes of the components to tral library (dictionary in the SPECTRAL LIBRARIES. relate the mass fractions and the relative geometric cross- SR jargon) that can best model sections [59]. each mixed pixel in the scene. Several strategies have successfully applied the bilinear In practice, this is a combina- model to treat the double scattering problem, such as Bayes- torial problem, which calls for efficient linear SR tech- ian algorithms, where prior models are chosen to satisfy the niques based on sparsity-inducing regularizers. Linear positivity and sum-to-one constraints [60]. On the other SR is an area of very active research with strong links to hand, kernel-based methods can design flexible kernels to compressed sensing [66]. handle the problem of intimate mixtures. Linear kernels, Let us assume that we are given a spectral library # radial-basis functions, polynomial, and physics-based A ! Rnmb containing m spectral samples. Usually, we have kernels were proposed [61]. To cope with both scattering mn2 b and, therefore, the linear problem in hand is under- and intimate mixture problems simultaneously, machine determined. Let x ! Rn denote the fractional abundance learning technologies have been proposed, where training vector with respect to the library A. With these definitions samples were used to train artificial neural networks for in place, we can now write our SR problem as nonlinearities (see [62] and references therein). Polynomial

<<02 <- <#d $ functions can be applied as well to model the nonlineari- minx xAxyx0 subject to ,,(4) ties [63].

A disadvantage of the above methods is that they where <

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 17

SUnSAL CLSUnSAL

(a) (b)

FIGURE 10. Sparse regression solutions X for 50 simulated spectral LMM generated with 4 endmembers and SNR= 30 dB and using a subset of the USGS library with 250 signatures. The horizontal axis represents pixels and the vertical axis represents endmembers. (a) Solution computed by the SUnSAL algorithm treating independently each pixel. (b) Solution computed by CLSUnSAL algorithm, enforcing collaborative sparsity.

b) the limitation imposed by the high correlation of the 50 pixels and 4 endmembers randomly extracted from spectral signatures is mitigated by the high level of spar- the library. The simulated measurement were contami- sity often observed in the hyperspectral mixtures. nated with additive noise and SNR = 30 dB. The image

The current research efforts to cope with the high cor- on the left hand side corresponds to the ,1 relaxation relation of the spectral signature are aligned with the of (4) computed with the SUnSAL algorithm [70], thus recent advances in the area of structured sparsity [71]. This treating each pixel independently. The image on the research direction exploits prior information about pat- right hand side is the solution of the group sparse prob-

terns of sparsity known to exist in specific applications. The lem using the mixed ,21, norm and was computed with types of structured sparsity exploited in HU are directly the CLSUnSAL algorithm introduced in [74]. The col- linked with two characteristics of hyperspectral data: i) laborative regularization yields a cleaner solution with the fractional abundance maps are piecewise smooth; and many rows set to zero. ii) the fractional abundance vectors for the different pixels in the HSI share the same support (i.e., the set of non- H. CHALLENGES zero elements). As a result of intense work in the last ten years, the The structured sparsity referred to in i) is linked to the research boundary in HU has advanced considerably. fact that, in piece-wise smooth maps, it is very likely that Many instances of HU are, however, hard inverse prob- neighboring pixels have very close values. These ideas are lems far from being solved in reasonable terms. The exploited in [72] by including the total variation (TV) reg- need for reliable unmixing results will continue to foster ularization term in the objective function (4), which pro- active research in HU, namely in areas of mixing mod- motes piecewise-smooth abundance maps. els, accounting for the measurement process, and data The structured sparsity referred to in ii), termed col- representation or prior knowledge. In the area of mixing

laborative sparsity, is promoted, for example, by the ,21, models, researchers are starting to derive and use nonlin- <<= / n <

18 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

IV. CLASSIFICATION so that an effective set of features can be identified prior Hyperspectral image classification has been a very active to classification. In this subsection we briefly outline some area of research in recent years [75]. Given a set of observa- of the available approaches for feature mining from hyper- tions (i.e., pixel vectors in a hyperspectral image), the goal spectral data sets. of classification is to assign a unique label to each pixel vec- 1) Feature extraction: Several strategies have been used tor so that it is well-defined by a given class. in the hyperspectral imaging literature to perform feature The availability of hyperspectral data with high spa- extraction prior to classification purposes. A distinguish- tial resolution has been quite important for classification ing characteristic of feature extraction methods is that techniques, as their main assumption is that the spatial they exploit all available spectral measurements in order resolution of the data is high enough to assume that the to extract relevant features. A widely used approach has data mostly contains pure pixels (i.e., pixels represented by been the generation of features in a new space, like those a single predominant spectral signature). In the opposite obtained from the PCA [3] or the minimum noise fraction scenario (i.e., the data mostly contains mixed pixels) it is (MNF) [77]. In these techniques, the hyperspectral data are preferable to use spectral unmixing techniques to perform projected onto a new space in which the first few compo- the analysis. nents account for most of the total information in the data, In this section, we outline some of the main techniques and therefore only the first few features could be retained. and challenges in hyperspectral image classification. We The segmented PCA [78] reduces the computational load focus mainly on supervised and semi-supervised classifi- significantly for feature extraction, compared with the cation, although techniques for unsupervised classification conventional PCA. Another spectral-based approach to and/or clustering have also been used in the literature [3]. generate new features has been the discrete wavelet trans- For instance, a relevant unsupervised method successfully form (DWT), which allows for the separation of high and applied to hyperspectral image data is Tilton’s recursive hier- low frequency components separately. This allows a form archical segmentation (RHSEG) algorithm11. Supervised of derivative analysis which has been also used to gener- classification has been more widely used [76], but it also ate features prior to hyperspectral image classification [79]. faces challenges related with the high dimensionality of the Another popular strategy has been canonical analysis [3], data and the limited availability of training samples [75]. which is focused on the extraction of features that maxi- In order to address these issues, feature mining [3], sub- mize the ratio between the variance among classes and the space-based approaches [58] and semi-supervised learning average variance within the classes. However, this approach techniques [1] have been developed. In feature mining and requires good estimates of the class covariance matrices, subspace approaches, the goal is to reduce the dimensional- and therefore a generally large number of training samples ity of the input space in order to better exploit the (limited) (which may not be available in practice) are often required. training samples available. In semi-supervised learning, the An alternative strategy to deal with this problem has been idea is to exploit the information conveyed by additional to use semi-supervised feature extraction [80], in which (unlabeled) samples, which can complement the available only a few labeled samples and additional unlabeled sam- labeled samples with a certain degree of confidence. In ples are used. Other widely used methods have been non- all cases, there is a clear need to integrate the spatial and parametric weighted feature extraction (NWFE) or decision spectral information to take advantage of the complemen- boundary feature extraction (DBFE) [75]. Another strat- tarities that both sources of information can provide [76]. egy for feature extraction has been grouping of neighbor- An overview of these different aspects, which are crucial ing bands, using techniques such as the weighted sum or to hyperspectral image classification, is provided in the average of each group [81]. A free Matlab toolbox for linear following subsections. and nonlinear feature extraction methods is simFEAT12. 2) Feature selection: In feature selection, the idea is to A. FEATURE MINING select a set of spectral bands from the initial pool of bands Hyperspectral imaging is characterized by the high spectral available prior to classification. A particular characteristic resolution available, which allows capturing fine details of of feature selection methods is that they tend to retain the the spectral characteristics of materials in a wide range of spectral meaning (while reducing the number of bands). applications. However, it has been demonstrated that the In unsupervised feature selection, the goal is to automati- original spectral features contain high redundancy [3]. Spe- cally find statistically important features. The advantage cifically, there is a high correlation between adjacent bands of unsupervised methods is that they do not need training and the number of the original spectral features may be too data. Quite opposite, supervised feature selection is based high for classification purposes [3], [75]. In addition, the on general/expert knowledge, and require labeled and original spectral features may not be the most effective ones (often) unlabeled training samples. Techniques in the lat- to separate the objects of interest from others. These obser- ter category comprise methods based on class separability vations have fostered the use of feature mining techniques measures using standard distance metrics (e.g., Euclidean,

11 12 http://opensource.gsfc.nasa.gov/projects/HSEG/______http://www.uv.es/gcamps/code/simfeat.htm

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE19

Mutual information, Bhattacharyya), or more sophisticated that maximizes the margin. If the data are not linearly sep- class separability measures in feature space [3]. In this regard, arable, soft margin classification with slack variables can be methods have been proposed that implement an exhaustive used to allow mis-classification of difficult or noisy cases. search of optimal features, such as the progressive two-class However, the most widely used approach in hyperspectral decision classifier [82]. Other advanced feature selection classification is to combine soft margin classification with a strategies (e.g., using kernels) are described in [83]. kernel trick that allows separation of the classes in a higher dimensional space by means of a nonlinear transformation. B. SUPERVISED CLASSIFICATION In other words, the SVM used with a kernel function is a Several techniques have been used to perform supervised nonlinear classifier, where the nonlinear ability is included classification of hyperspectral data. For instance, in dis- in the kernel and different kernels lead to different types criminant classifiers several types of discriminant func- of SVMs. The extension of SVM to the multi-class cases is tions can be applied: nearest neighbor, decision trees, linear usually done by combining several binary classifiers. Two functions, nonlinear functions, etc. In linear discriminant classical procedures are the one versus the rest and the one analysis (LDA) [84], a linear function is used in order to versus one [87]. maximize the discriminatory power and separate the avail- In the following, we illustrate the performance of SVMs able classes effectively. However, such a linear function (implemented using the Gaussian radial basis function may not be the best choice and nonlinear strategies such as kernel) by processing a widely used hyperspectral data set quadratic discriminant analysis (QDA) or logarithmic dis- collected by the Reflective Optics Imaging Spectrographic criminant analysis (LogDA) have also been used. The main System (ROSIS) optical sensor over the urban area of the problem of these classic supervised classifiers, however, is University of Pavia, Italy. The flight was operated by the their sensitivity to the Hughes effect. Deutschen Zentrum for Luftund Raumfahrt (DLR, the Ger- In this context, kernel methods such as the support vec- man Aerospace Agency) in the framework of the HySens pro- tor machine (SVM) have been widely used in order to deal ject, managed and sponsored by the European Commission. effectively with the Hughes phenomenon [85], [86]. The The image size in pixels is 610# 340, with very high spatial SVM was first investigated as a binary classifier [87]. Given resolution of 1.3 meters per pixel. The number of data chan- a training set mapped into an Hilbert space by some map- nels in the acquired image is 103 (with spectral range from ping, the SVM separates the data by an optimal hyperplane 0.43 to 0.86 nm). Fig. 11(a) shows a false color composite of

Asphalt Meadows Gravel Trees Metal Sheets Bare Soil Bitumen Bricks Shadows

(a) (b) (c)

FIGURE 11. (a) False color composition of the ROSIS University of Pavia scene. (b) Reference map containing 9 mutually exclusive landcover classes. (c) Training set used in experiments.

20 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

the image, while Fig. 11(b) shows nine reference classes of interest, which comprise urban features, as well as soil and TABLE 2. ACCURACIES OBTAINED BY DIFFERENT SUPERVISED CLASSIFIERS FOR THE ROSIS UNIVERSITY OF PAVIA SCENE. vegetation features. Finally, Fig. 11(c) shows a fixed training set available for the scene which comprises 3921 training METRIC LDA QDA LogDA SVM MLR MLRsub samples (42776 samples are available for testing). OA 77.95 77.95 78.41 80.99 80.11 67.08 AV 73.67 78.73 79.82 88.28 87.80 77.20 Table 2 illustrates the classification results obtained by kappa 0.606 0.770 0.720 0.761 0.750 0.703 different supervised classifiers for the ROSIS University of Pavia scene in Fig. 11(a), using the same training data in Fig. 11(c) to train the classifiers and a mutually exclusive extraction prior to classification of hyperspectral data by set of labeled samples in Fig. 11(b) to test the classifiers. extracting the first few principal components of the data As shown by Table 2, the SVM classifier obtained compara- using the PCA [3], and then building so-called extended tively superior performance in terms of the overall classifi- morphological profiles (EMPs) on the first few components cation accuracy (OA), average classification accuracy (AV) to extract relevant features for classification [94]. and kappa statistic [88] when compared with discriminant As shown by Table 3, the combination of EMP for feature classifiers such as LDA, QDA or LogDA. In this experiment, extraction followed by SVM for classification (EMP/SVM) the SVM was also slightly superior to the multinomial provides good classification results for the ROSIS Uni- logistic regression (MLR) classifier [89], which has been versity of Pavia scene. Recently, morphological attribute recently explored in hyperspectral imaging as a technique profiles (APs) [95] were introduced as an advanced mecha- able to model the posterior class distributions in a Bayesian nism to obtain a detailed multilevel characterization of a framework, thus supplying (in addition to the boundaries hyperspectral image created by the sequential application between the classes) a degree of plausibility for such classes of morphological attribute filters that can be used (prior [90]. A subspace-based version of this classifier, called MLR- to classification) to model different kinds of the structural sub [91], is also included in the comparison given in Table 2. information. According to the type of the attributes consid- The idea of applying subspace projection methods relies on ered in the morphological attribute transformation, differ- the basic assumption that the samples within each class can ent parametric features can be modeled. The use of different approximately lie in a lower dimensional subspace. How- attributes leads to the concept of extended multi-attribute ever, in the experiments reported in [91] for the MLRsub it profiles (EMAPs) which have been also used successfully was observed that spatial information needs to be included for hyperspectral image classification purposes [96]. in this (and other classifiers) in order to improve classifi- Another strategy in the literature has been to exploit cation performance. In the following subsection, we sum- simultaneously the spatial and the spectral information. marize some techniques for spatial-spectral classification. For instance, in order to incorporate the spatial context into kernel-based classifiers, a pixel entity can be redefined C. SPATIAL-SPECTRAL CLASSIFICATION simultaneously both in the spectral domain (using its spec- Several efforts have been performed in the literature in tral content) and also in the spatial domain, by applying order to integrate spatial-contextual information in spec- some feature extraction to its surrounding area which yields tral-based classifiers for hyperspectral data [76]. It is now spatial (contextual) features, e.g., the mean or standard commonly accepted that using the spatial and the spectral deviation per spectral band. These separated entities lead information simultaneously provides significant advan- to two different kernel matrices, which can be easily com- tages in terms of improving the performance of classifica- puted. At this point, one can sum spectral and textural dedi- tion techniques. Some of these approaches include spatial cated kernel matrices and introduce the cross-information information prior to the classification, during the feature between textural and spectral features in the formulation. extraction stage. Mathematical morphology [92] has been This simple methodology yields a full family of new ker- particularly successful for this purpose. Morphology is a nel methods for hyperspectral data classification, defined widely used approach for modeling the spatial character- in [97] and implemented using the SVM classifier thus pro- istics of the objects in remotely sensed images. Advanced viding a composite kernel-based SVM (SVM-CK) illustrated morphological techniques such as morphological pro- in Table 3 (using the summation kernel). Recently, compos- files (MPs) [93] have been successfully used for feature ite kernels have been generalized in [98], using the MLR

TABLE 3. OVERALL ACCURACIES OBTAINED BY DIFFERENT SUPERVISED SPATIAL-SPECTRAL CLASSIFIERS FOR THE ROSIS UNIVERSITY OF PAVIA SCENE.

METRIC ECHO LDA-MLL QDA-MLL LOGDA-MLL SVM-CK EMP/SVM SVM-W SVM-RHSEG MLR-MLL MLRsub MLL MLR-GCK OA 87.58 80.27 89.48 87.04 87.18 85.22 85.42 93.85 85.57 94.10 98.09 AV 92.16 78.05 91.91 83.32 90.47 90.76 91.31 97.07 92.54 93.45 97.76 kappa 0.839 0.739 0.864 0.872 0.871 0.808 0.813 0.918 0.818 0.922 0.974

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 21

classifier and EMAPs to define spatial context. The resulting D. SEMI-SUPERVISED CLASSIFICATION generalized composite kernel-based MLR (MLR-GCK) can A relevant challenge for supervised classification techniques linearly combine multiple kernels without any restriction of is the limited availability of labeled training samples, since convexity. This introduces a different approach with regards their collection generally involves expensive ground cam- to the SVM-CK and multiple kernel learning methods, in paigns [105]. While the collection of labeled samples is which composite kernels need to be convex combinations generally difficult, expensive and time-consuming, unla- of kernels. This approach provided the best classification beled samples can be generated in a much easier way. This result for the ROSIS University of Pavia scene in Table 3. observation has fostered the idea of adopting semi-super- Another approach to jointly exploit spatial and spectral vised learning techniques in hyperspectral image classi- information is to use Markov random fields (MRFs) for fication. The main assumption of such techniques is that the characterization of spatial information. MRFs exploit new (unlabeled) training samples can be obtained from a the continuity, in probability sense, of neighboring labels (limited) set of available labeled samples without signifi- [99]. In this regard, several techniques have introduced an cant effort/cost. MRF-based multinomial logistic level (MLL) prior which In contrast to supervised classification, semi-supervised encourages neighboring pixels to have the same label when algorithms generally assume that a limited number of performing probabilistic classification of hyperspectral data labeled samples are available a priori, and then enlarge the sets. As it can be seen in Table 2, such MLL prior can sig- training set using unlabeled samples, thus allowing these nificantly improve the results provided by spectral-based approaches to address ill-posed problems. However, in classifiers such as the LDA, QDA, LogDA, MLR and, most order for this strategy to work, several requirements need notably, the MLRsub described in the previous subsection. At to be met. First and foremost, the new (unlabeled) samples this point, it is worth noting that the combination of a sub- should be obtained without significant cost/effort. Second, space-based classifier such as the MLRsub with an MLL prior the number of unlabeled samples required for the semi- enforcing spatial homogeneity in the resulting segmentation supervised classifier to perform properly should not be too provides one of the best classification results in Table 3. high in order to avoid increasing computational complexity Several other approaches include spatial information as a in the classification stage. In other words, as the number of post-processing, i.e., after a spectral-based classification has unlabeled samples increases, it may be unbearable for the been conducted. One of the first classifiers with spatial post- classifier to properly exploit all the available training sam- processing developed in the hyperspectral imaging literature ples due to computational issues. Further, if the unlabeled was the well-known ECHO (extraction and classification samples are not properly selected, these may confuse the of homogeneous objects) [75]. Another one is the strategy classifier, thus introducing significant divergence or even adopted in [100], which combines the output of a pixel-wise reducing the classification accuracy obtained with the ini- SVM classifier with the morphological watershed (SVM- tial set of labeled samples. In order to address these issues, W) transformation [92] in order to provide a more spatially it is very important that the most highly informative unla- homogeneous classification. A similar strategy is adopted in beled samples are identified in computationally efficient [101], in which the output of the SVM classifier is combined fashion, so that significant improvements in classification with the segmentation result provided by the RHSEG seg- performance can be obtained without the need to use a mentation algorithm (SVM-RHSEG). These strategies lead to very high number of unlabeled samples. much improved classification results with regards to spec- The area of semi-supervised learning for remote sens- tral-based classification, as it can be observed in Table 3. A ing data analysis has experienced a significant evolution detailed overview of recent advances in spatial-spectral clas- in recent years. For instance, in [106] transductive SVMs sification of hyperspectral data is available at [102]. (TSVMs) are used to gradually search a reliable separating Last but not least, an important recent development hyperplane (in the kernel space) with a transductive process has been the use of sparse representation classifiers using that incorporates both labeled and unlabeled samples in the dictionary-based generative models [103]. In this case, an training phase. In [107], a semi-supervised method is pre- input signal is represented by a sparse linear combination of sented that exploits the wealth of unlabeled samples in the training samples (atoms) from a dictionary [103]. The clas- image, and naturally gives relative importance to the labeled sification can be improved by incorporating the contextual ones through a graph-based methodology. In [108], kernels information from the neighboring pixels into the classifier. combining spectral-spatial information are constructed by This can be done indirectly by exploiting the spatial correl- applying spatial smoothing over the original hyperspectral ation through an structured sparsity prior imposed in the data and then using composite kernels in graph-based clas- optimization process. As shown in [103], the performance sifiers. In [109], a semi-supervised SVM is presented that of different sparsity-based classifiers are comparable to the exploits the wealth of unlabeled samples for regularizing state-of-the-art SVM-CK classifier. Given sufficient training the training kernel representation locally by means of clus- data some researchers have also developed discriminative ter kernels. In [90], a new semi-supervised approach is pre- as well as compact class dictionaries to improve classifica- sented that exploits unlabeled training samples (selected by tion performance [104]. means of an active selection strategy based on the entropy of

22 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

the samples). Here, unlabeled samples are used to improve with highest uncertainty from the obtained candidate set. the estimation of the class distributions, and the obtained This strategy relies on two main assumptions. The first classification is refined by using a spatial multi-level logis- assumption (global) is that training samples having the tic prior. In [110], a novel context-sensitive semi-supervised same spectral structure likely belonging to the same class. SVM is presented that exploits the contextual information The second assumption (local) is that spatially neighbor- of the pixels belonging to the neighborhood system of each ing pixels likely belong to the same class. As a result, this training sample in the learning phase to improve the robust- approach naturally integrates the spatial and the spectral ness to possible mislabeled training patterns. information in the semi-supervised classification process. In [111], two semi-supervised one-class (SVM-based) approaches are presented in which the information pro- E. CHALLENGES vided by unlabeled samples present in the scene is used to There are several important challenges when performing improve classification accuracy and alleviate the problem hyperspectral image classification. Supervised classifica- of free-parameter selection. The first approach models data tion faces challenges related with the unbalance between marginal distribution with the graph Laplacian built with high dimensionality and limited availability of training both labeled and unlabeled samples. The second approach samples, or the presence of mixed pixels in the data (which is a modification of the SVM cost function that penalizes may compromise classification results for coarse spatial more the errors made when classifying samples of the target resolutions). Another relevant challenge is the need to inte- class. In [112] a new method combines labeled and unla- grate the spatial and spectral information to take advantage beled pixels to increase classification reliability and accu- of the complementarities that both sources of informa- racy, thus addressing the sample selection bias problem. In tion can provide. These challenges are quite important for [113], an SVM is trained with the linear combination of two future developments and solutions to some of them have kernels: a base kernel working only with labeled examples been outlined in this section. Specifically, we have explored is deformed by a likelihood kernel encoding similarities techniques such as supervised and semi-supervised tech- between labeled and unlabeled examples, and then applied niques for hyperspectral image classification, strategies in the context of urban hyperspectral image classification. for integrating the spatial and the spectral information, In [114], similar concepts to those addressed before are or sparse classifiers that can bring solutions to the afore- adopted using a neural network as the baseline classifier. mentioned problems. However, some issues still remain. In [115], a semi-automatic procedure to generate land cover For instance, the geometry of hyperspectral data is quite maps from remote sensing images using active queries is complex and dominated by nonlinear structures. This issue presented and discussed. has undoubtedly an impact in the outcome of the classi- At this point, it should be noted that active learning fication techniques discussed in this section. In order to techniques have been mainly exploited in a supervised mitigate this, manifold learning has been proposed [120]. context, i.e. a given supervised classifier is trained with An important property of manifold learning is that it can the most representative training samples selected after a model and characterize the complex nonlinear structure (machine-human) interaction process in which the sam- of the data prior to classification [121]. Another remaining ples are actively selected according to some criteria based issue is the very high computational complexity of some of on the considered classifier, and then the labels of those these classifiers discussed in this section. In other words, samples are assigned by a trained expert in fully supervised there is a need to develop efficient classification techniques fashion [90], [116]–[118]. In this supervised context, sam- that can deal with the very large dimensionality and com- ples with high uncertainty are generally preferred as they plexity of hyperspectral data. In Section VII we discuss fast are usually more informative. At the same time, since the computing solutions for hyperspectral imaging algorithms. samples are labeled by a human expert, high confidence Last but not least, we emphasize that the techniques can be expected in the class label assignments. As a result, described in this section only represent a small sample classic (supervised) active learning generally focuses on (and somehow subjective selection) of the vast collection samples with high confidence at the human level and high of approaches presented in recent years for hyperspectral uncertainty at the machine level. image classification. For a more exhaustive summary of Recently, standard active learning methods have been available techniques and future challenges in this area, we adapted into a semi-supervised self-learning scenario [119]. point interested readers to [1]. The main idea is to obtain unlabeled samples (from a pool of samples) using machine-machine interaction instead of V. HYPERSPECTRAL TARGET DETECTION human supervision. The first (machine) level—similar to Hyperspectral imagery has been used in reconnaissance the human level in classic (supervised) active learning—is and surveillance applications where targets of interest are used to infer a set of candidate unlabeled samples with high detected and identified [7], [122]. The process of detect- confidence. In a second (machine) level—similar to the ing and identifying a target in hyperspectral imagery can machine level for supervised active learning—the machine be considered as consisting of two stages. The first stage is learning algorithm itself automatically selects the samples an anomaly detector [123], [124] which identifies spectral

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 23

t vectors that have significant spectral differences from their where nb is the estimated background clutter sample mean t surrounding background pixels. Man-made anomalies can and Cb is the estimated background clutter covariance. also be detected through change detectors [125], [126], The background mean and covariance matrix can be which are used to identify changes within a scene that estimated globally from the whole hyperspectral image or occur over time. The second stage is to identify whether or locally using a double concentric window approach [123]. t not the anomaly is a target or natural clutter. This stage can To estimate Cb globally the background pixels are usually be achieved if the spectral signature of the target is known modeled as a mixture of multivariate Gaussian distribu- which can be obtained from a spectral library [122] or from tions [137], linear subspace [138], [139], linear or stochastic a set of training data which could also be synthetically gen- mixture models [140] or by some clustering or segmentation erated [127]. Almost all the classical target detection tech- techniques [141]–[143] which are used to segment the back- niques in the literature [127]–[131] are based on a linear ground into several clusters. On the other hand the local process that only exploits the first and second order statis- background covariance matrix can be estimated by using a tics to identify anomalies or targets. Advanced nonlinear sliding double concentric window, centered at each test pixel, detection techniques based on statistical kernel learning which consists of a small inner window region (IWR) cen- theory [132] have also been developed in [133] that indi- tered within a larger outer window region (OWR), as shown rectly exploit the higher order statistics between the spec- in Fig. 12. The local background mean vector and covariance tral bands through a kernel function [132]. matrix are then computed from the spectral pixels falling within the OWR. The size of the inner window is assumed A. ANOMALY DETECTION to be the size of the typical target of interest in the image. A Anomaly detectors, outlier detectors, or novelty detec- guard band surrounding the IWR is also sometimes used tors are pattern recognition or statistical schemes that are to prevent the target pixels from corrupting the calculation used to detect objects that stand out from their cluttered of the background OWR statistics. The whole background background. In spectral anomaly detection algorithms probability density function has also been modeled by a sin- [123], [124], [134]–[136] pixels (materials) that have a sig- gle class support vector machine in [144] and spectral pixels nificantly different spectral signature from their neigh- that fall outside this model are considered as anomalies. boring background clutter pixels are identified as spectral Anomaly detection techniques formulated as eliminating anomalies. In such algorithms, no prior knowledge of the the whole or local background subspace from every pixel target spectral signature is utilized or assumed. In [134], a have also been investigated in [145]. spectral anomaly detection algorithm was developed for Several variations of the RX detector that attempt to detecting targets of unknown spectral distribution against alleviate the limitation of RX have been proposed in the lit- a background with unknown spectral covariance. This erature [135]–[137]. In [146], a modification to the R X algo- algorithm is now commonly referred to as the Reed-Xiaoli rithm called SubSpace RX (SSRX) was outlined that is based (RX) anomaly detector, has been successfully applied to on the PCA of the background covariance matrix. In the many hyperspectral target detection applications [7], [124], SSRX algorithm, several high-variance background dimen- [135], [136] and is considered as the benchmark anomaly sions are deleted before applying the RX algorithm as these detection algorithm for multispectral/hyperspectral data. are assumed to capture non-normal background clutter vari- The RX algorithm is a constant false alarm rate adaptive ance. Another consideration in RX implementation is poten- anomaly detector which is derived from the generalized tial ill-conditioning of the local covariance matrix due to the likelihood ratio test (GLRT). Assuming a single pixel target high correlation, high dimensionality of the hyperspectral y as the observation test vector, the results of RX-algorithm data and a limited background sample size. This ill-condi- is given by tioning is typically addressed by a regularization procedure such as PCA-based regularization or adding a scaled identity t T t -1 t RX()yy=- (nb ) Cybb ( -n ),(5) matrix to the background covariance matrix [147].

B. SIGNATURE-BASED TARGET DETECTION 19 In some applications, we have some prior knowledge about 9 the spectral characteristics of the desired targets. In these 7 Inner Window situations, the target spectral characteristics can be defined Region (IWR) 19 97 by a single target spectrum [148] or by a signal subspace Current Test [129]. The GLRT detector using a single targets spectrum Pixel is referred to as the spectral matched filter (SMF) and the maximum likelihood abundance estimate of the target in a Guard (Band) test pixel y is given in [130] as Outer Window Window Region (OWR) T t -1 = sC y DSMF ()y t - , (6) FIGURE 12. An example of a dual window. sCT 1 s

24 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

t where s is the spectral signature of the target and C rep- different detectors, such as the adaptive subspace detec- resents the estimated covariance matrix for the centered tor (ASD) [131] and orthogonal subspace projection (OSP) observation data. The SMF model is based on the assump- [128]. In ASD the target signature is represented by a lin- tion that the background clutter noise has a Gaussian ear subspace and the background statistics by a zero-mean t distribution N(,0C ) and the target distribution is also a Gaussian distribution whose covariance is estimated from t Gaussian N(,)asC with the same covariance statistics but the hyperspectral image data itself. In the case of OSP the with a mean of as, where a is an scalar abundance value target signature is represented by a single spectral vector s representing the target strength. and the background is represented by the spectral signa- The mean and covariance matrix appearing in (6) are tures of the undesired background endmembers B. The OSP estimated from the data either locally, or globally under the algorithm is based on maximizing the signal-to-noise ratio assumption that the number of target pixels contaminating (SNR) in the subspace orthogonal to the background sub- T = = # the estimation of the covariance matrix is insignificant. The space which is given by DOSP = sPyB where PIBBB =- spectral target signatures are usually obtained from the spec- is the background rejection operator and ()$ # denotes the tral library or from a target training data set. An essential pre- pseudo inverse matrix. processing stage to implement a signature-based detector is to estimate and compensate the atmospheric effects on the C. SPARSE REPRESENTATION TARGET DETECTION data [149] in order to transform the known target spectrum In the sparse representation classifiers (SRC) a test sample and measurement data into a common domain where an is approximately represented by very few training samples algorithm such as the SMF can be applied. Another consid- from both target and background dictionaries, and the eration in SMF implementation is potential ill-conditioning recovered sparse representation is directly used for classifi- of the covariance matrix due to the high correlation, high cation [152]. Given the concatenated training samples from dimensionality of the hyperspectral data and from a limited the target and background dictionaries as AAA= 6 bt@, the background sample size. Representing the inverse covariance sparse representation vector c satisfying Ayc = can be matrix in terms of its eigenvector-eigenvalue decomposition, obtained by solving the following optimization problem it becomes clear that the behavior of the inverse covariance cct == c matrix depends heavily on the small eigenvalues which arg min0 subject to Ay, (8) could render it unstable. In order to reduce SMF sensitivity to c , statistical and numerical errors, eigenvectors corresponding where 0, termed the 0-norm, is defined as the number to eigenvalues below an appropriate condition number or of non-zero entries in the vector c that also represents a estimated sensor noise level are discarded [150], or a scaled sparsity prior or penalty in (8). The above problem of mini- identity matrix is added to the background clutter covariance mizing the ,0-norm is NP-hard, but its approximate solu- before inverting, which is equivalent to including a regulari- tion can be obtained by greedy algorithms [68]. The ,0 zation term (penalty term) in the design of the SMF [151]. -norm can also be replaced by an ,1-norm regularization In a situation when the target and background pixel prior term in (8) [69], where standard convex optimiza- characteristics are modeled by two linear subspaces with tion algorithms [70] can be used. Once the sparse coeffi- additive noise the resulting detector is referred to as the sub- cient vector c is obtained, the class of the test pixel y can be =-A at space matched detector (MSD). The GLRT for such a two determined by comparing the residuals rbb()yy 2 =-bt at bt subspace model is given in [129] as and rtt()yyA2 , where and represent the recovered sparse coefficient vectors corresponding to the back- T ^ - h = yIPyB DMSD ()y T , (7) ground and target dictionaries, respectively. yI^ - PyTBh In the above process, the sparsity-based target detector where PB is the projection matrix associated with the back- is applied to each pixel in the test region independently ground subspace B , and PTB is the projection matrix without considering the correlation (contextual) between associated with the target-and-background subspace TB . neighboring pixels. To incorporate contextual information Usually, the eigenvectors corresponding to the significant within the SRC algorithm other sparsity priors or penalties, eigenvalues of the target and background covariance matri- such as joint sparsity (collaborative) or an l2-norm smooth- ces are used to generate the columns of T and B, respectively. ness constraint [153] can be considered. Fig. 13 shows the To generate appropriate target and background linear sub- receiver operating characteristic (ROC) curves for several spaces researchers have used scene samples from the hyper- different target detection techniques on a typical hyperspectral image itself or have used the MODTRAN software spectral image, the forest radiance I data collection (FR-I), package to generate a large number of synthetic target and obtained from a hyperspectral digital imagery collection background spectral pixels [127] for a given environment experiment (HYDICE) sensor which consists of 210 bands in order to estimate the two environmentally invariant sub- across the whole spectral range from 0.4 to 2.5 nm which spaces T and B. includes the visible and short-wave infrared bands. As seen Using some variations of the models used in SMF from the ROC curves in Fig. 13 the SVM-CK and sparsity- and MSD, researchers have also developed a number of based with smoothing classifiers outperform the classical

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 25

to their nonlinear versions [133], [154] by using the ideas in 1 kernel machine learning theory. Experimental results show 0.9 that typically the kernel-based algorithms outperform their 0.8 linear versions. 0.7 E. CHALLENGES 0.6 The major challenges in the classical anomaly and target 0.5 detection techniques (RX, SMF, MSD, ASD, OSP) are still the need for developing new approaches for estimating the 0.4 Sparsity−Based with Smooting background/target covariance matrices or their correspond- 0.3 Sparsity−Based Without Smoothing SVM−CK ing subspaces given a limited training data. Further research Probability of Detection 0.2 MSD is also needed in the classical techniques to incorporate the SMF 0.1 ASD spatial-contextual information about the targets that are 0 more than one pixel size. In the case of sparsity-based tech- 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 niques more research is needed to develop the appropriate False Alarm Rate class sub-dictionaries as well as compact discriminative dictionaries. More advanced structured sparsity priors are FIGURE 13. ROC curves for a typical hyperspectral image (FR-I) to be incorporated and their performance evaluated. Cur- with several military targets reproduced from [152]. rently most of the non-linear methods are based on kernel learning theory, other nonlinear approaches beside kernel- signature-based target detection techniques, the experi- based methods need to be introduced. mental detail and results for couple of other hyperspectral images can be found in [152]. VI. ESTIMATION OF LAND PHYSICAL PARAMETERS This section reviews the main problems and methods in D. NONLINEAR DETECTORS the field of model inversion and estimation of physical Almost all the anomaly and target detectors are based on parameters. Our main goal in remote sensing is to monitor only exploiting first and second order statistics in order the Earth and its interaction with the atmosphere. The to identify anomalies or targets. Kernel machine learning analysis can be done at local or global scales by looking at theory [132] has emerged as a new nonlinear-based learn- bio-geo-chemical cycles, atmospheric state and evolution, ing approach for extending the classical pattern recognition and vegetation dynamics [155], [156]. All these com- algorithms. The implicit exploitation of nonlinear features plex interactions can be studied through the definition of through kernels provides crucial information about a given physical parameters representing different properties for dataset which, in general, the learning methods based on land (e.g., surface temperature, biomass, leaf area coverage), linear models cannot achieve. The RX anomaly detection water (e.g., yellow substance, ocean color, suspended matter) algorithm, the statistical target detection algorithms and or the atmosphere (e.g., temperature and moisture pro- the sparsity-based target classifier have all been extended files at different altitudes). The field of physical parameter estimation is an intermediate modeling step necessary to transform the measurements13 into useful estimates [157]. The remote sensing modeling system is illustrated in Additional Variables Fig. 14. The forward (or direct) problem involves radiative transfer models (RTMs). These models summarize the physi- Forward cal processes involved in the energy transfer from canopies Radiative Problem Transfer and atmosphere to measured radiance. They simulate the reflected or emitted radiation transmitted through the Remote atmosphere for a given observation configuration (e.g., Variables Observation Sensing wavelength, view and illumination directions) and some of Interest Configuration Data auxiliary variables (e.g., vegetation and atmosphere characteristics). Solving the inversion problem implies the design Retrieval of algorithms that, starting from the radiation acquired by Inverse Algorithm Problem the sensor, can give accurate estimates of the variables of interest, thus “inverting” the RTM. In the inversion process, Prior a priori information of the variables of interest can also be Knowledge

13The acquired data may consist of multispectral or hyperspectral images FIGURE 14. Forward (solid lines) and inverse (dashed lines) provided by satellite or airborne sensors, but can also integrate spectra acquired by in situ (field) radiometers, GIS data that help to integrate geo- problems in remote sensing. Figure adapted from [157]. graphical information, or radiosonde measures for weather monitoring.

26 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

included to improve the performance, such as the type of used for mapping canopy nitrogen [162]. Nonlinear exten- surface, geolocation, or acquisition time. sion of PLS was introduced via kernels in [163] for chlo- Notationally, a discrete forward model can be expressed as rophyll concentration estimation. Recently, the support vector regression (SVR) has yielded good results in mod- Yx=+f(,i ) n , eling oceanic chlorophyll [164]. In the recent years, Gauss- ian Process (GP) regression has shown very good properties where Y is a set of measurements (such as the expected to tackle the physical parameter estimation: GPs are sim- radiance); x is the state vector that describes the system pler to train than other models, they show good numerical (e.g., parameters such as temperature or moisture); i contains performance and stability and provide sensible confidence a set of controllable measurement conditions (e.g., different intervals for the predictions [165]. combinations of wavelength, viewing direction, time, Sun 2) Physical inversion: Statistical approaches may lack position, and polarization); n is an additive noise vector; and transferability, generality, or robustness to new geographi- f()$ is a complex nonlinear continuous function that relates x cal areas. Physical inversion models can alleviate these with Y. The discrete inverse model is then defined as shortcomings by coupling models from lower to higher levels (e.g., canopy level models built upon leaf models). xYt = g(,~ ), Therefore, they provide a physically-based, interpretable and broad linkage between Earth observation data and bio- where g()$ is a possibly nonlinear function, parameter- chemical or structural state variables [166]. Running RTMs ized by weights ~, that approximates the measurement in forward mode enables creating a database covering a wide conditions, x, using a set of observations as inputs, Y. range of situations and configurations. Forward RTM simulations allow for sensitivity studies of canopy parameters A. TAXONOMY OF INVERSION METHODS relative to diverse observation specifications, and allow us Model inversion methods can be roughly divided into statis- to better understand the observed signal. The use of RTMs tical, physical or hybrid. In what follows, we review the main to generate datasets is a common practice, and especially contributions in each family for land parameter retrieval. convenient because acquisition campaigns are very costly The reader is referred to [1] for an updated review on atmo- (in terms of time, money, and human resources), and usu- spheric and ocean parameter retrieval methodologies. ally limited in terms of parameter combinations. RTMs are 1) Statistical inversion: Statistical inversion can be done also widely used in the preliminary phase of a new sensor either with parametric or non-parametric models. Para- design, which allows understanding both the limits and metric models rely on physical knowledge of the problem capabilities of the instrument for the retrieval tasks. There and build explicit parameterized expressions that relate a exist many RTMs implemented in software packages to few spectral channels with the bio-geo-physical parameter deal with the forward modeling. For example, PROSPECT of interest. Different narrowband vegetation indices (VIs) is an RTM accounting for leaf optical properties while SAIL have been proposed to study the vegetation status by esti- constitutes a canopy bidirectional reflectance model. Their mating chlorophyll content and other leaf pigments, as well combination led to the PROSAIL model. Other RTMs are as vegetation density parameters like the leaf area index more specific to ocean or atmosphere applications. (LAI) and the fractional vegetation cover (FVC) [158]. The Physical inversion models essentially try to reverse simple calculation of these indices has made possible deriv- (invert) data generated with an RTM. The basic assumption ing reasonable maps of vegetation properties in a quick and for inverting RTMs is that the forward model, f()$ , contains easy way. Furthermore, since the launch of imaging spec- all the necessary information about the problem, so its trometers into space crafts, these VIs have been applied at inversion may lead to accurate parameter estimates. When a canopy level on ecosystems across the globe [159]. Never- unique solution is not achieved, more a priori information theless, the majority of the indices only use up to five bands is required to overcome the ill-posed problem. After gener- thus under-exploiting the full potential information con- ating the dataset, {,},Yx the problem reduces to, given new tained in hyperspectral images [160]. spectra, assign the parameter of the ‘closest’ spectra. Several Alternatively, non-parametric models estimate the vari- approaches and metrics have been considered to solve the able of interest using a set of input-output training data problem: iterative optimization, look-up-tables (LUTs), sim- pairs, which come from concurrent measurements of the ulated annealing and genetic algorithms, and Markov chain parameter and the corresponding reflectance/radiance Monte Carlo (MCMC). See [156] and references therein. observation. A terrestrial campaign is thus necessary at the 3) Hybrid inversion: Hybrid inversion models combine same time the satellite overpasses the study area to measure the previous approaches: they exploit the input-output data the surface parameter. Several nonparametric approaches generated by RTMs simulations, {,},Yx and train statistical have been introduced for land parameter retrieval. In [161], methods (typically neural networks) to invert the model, biomass was estimated using common spectral band ratios, i.e., learn g()$ . Nonparametric statistical inversion is com- vegetation indices and linear/stepwise multiple regression putationally efficient and can replace more costly physical models. Partial least squares (PLS) regression has been inversion learning g()$ . Regression trees has been used for

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 27

example to estimate land surface variables like LAI, frac- tools and are not always easily accessible. A recent effort tion of photo-synthetically active radiation (FAPAR), and that is worth mentioning is the automated radiative trans- chlorophyll content [167]. However, the vast majority of fer models operator (ARTMO) Matlab toolbox. ARTMO hybrid inversion methods consider the use of neural net- provides all necessary tools for running and inverting a works [168]–[170] for canopy parameter retrieval, mainly suite of plant RT models, both at the leaf and at the canopy due to their capability to ingest huge databases. Neural net- level. The toolbox is freely available14, and will soon inte- work revealed less biased than standard LUT methods in grate advanced statistical regression for inversion. Currently, LAI retrieval [171]. In [172], neural networks were trained the simpleR toolbox15 provides easy Matlab code to develop on a reflectance database made of RTM simulations, and parametric and nonparametric retrieval algorithms. LAI, FAPAR and FCOVER were accurately retrieved as compared with ground measurements. Very recently, the B. EXPERIMENTS combination of clustering and neural networks inverted We here illustrate the performance of both empirical and simulated data with additive noise. Inclusion of multian- statistical approaches to retrieve chlorophyll concentration gle images improved the LAI estimations. Lately, in [173], from hyperspectral images. The data were obtained in the neural networks were successfully developed over RTMs to SPARC-2003 (SPectra bARrax Campaign) and SPARC-2004 estimate LAI, FCOVER and FAPAR. Only very recently, ker- campaigns in Barrax, La Mancha, Spain. The region consists nel methods [174] have been used: In [175], the SVR was of approximately 65% dry land and 35% irrigated land. The

used to retrieve LAI by inverting PROSAIL. methodology applied to obtain the in situ leaf-level Chlab RTMs have become important tools for the analysis of data consisted of measuring samples with a calibrated CCM- Earth observation data, providing meaningful links between 200 Chlorophyll Content Meter in the field. Concurrently, radiometry and environmental applications. However these we used CHRIS images Mode 1 (62 spectral bands, 34 m models are still often perceived as excessively complicated spatial resolution at nadir). The images were preprocessed, geometrically and atmospherically corrected. A total of n = 136 datapoints in a 62-dimensional space and the meas- TABLE 4. CORRELATION COEFFICIENT R RESULTS OF ured chlorophyll concentration constitute the database. STANDARD NARROWBAND AND BROADBAND INDICES ALONG Performances of a wide array of established vegeta- WITH RECENT NONPARAMETRIC NONLINEAR MODELS. tion indices, linear regression with all bands (LR), SVR, METHOD FORMULATION R and GP [165] were tested. Models were run for a total of

GI R672/R550 0.52 (0.09) 50 random realizations of the training and test data. Aver- - + GVI (R682 R553)/(R682 R553)0.66 (0.07)aged correlation coefficients are shown for the test set in - - - MCARI2 1.2[2.5(R800 R670) 1.3(R800 R550)] 0.71 (0.12) Table 4. Nonparametric methods show the best results both - + - mNDVI (R800 R680)/(R800 R680 2R445)0.77 (0.12)in correlation and stability, with GP performing best of the - + - mNDVI705 (R750 R705)/(R750 R705 2R445)0.80 (0.07)tested methods. - + mSR705 (R750 R445)/(R705 R445)0.72 (0.07)The best GP model was used for prediction on the - - - mTVI 1.2[1.2(R800 R550) 2.5(R670 R550)]) 0.73 (0.07) whole CHRIS image to generate a pixel-by-pixel map of - + NDVI (R800 R670)/(R800 R670)0.77 (0.08)Chl and its confidence map (see Fig. 15). The maps show - + NDVI2 (R750 R705)/(R750 R705)0.81 (0.06)clearly the irrigated crops (the circles in orange-red), the - + NPCI (R680 R430)/(R680 R430)0.72 (0.08)semi-natural areas (light blue) and the bare soil areas (dark - + + OSAVI 1.16(R800 R670)/(R800 R670 0.16) 0.79 (0.09) blue). Gaussian Processes also provide confidence inter- PRI (R531-R570)/(R531+R570)0.77 (0.07) vals for the predictions, Fig. 15(b), which may be helpful PRI2 (R570-R539)/(R570+R539)0.76 (0.07) to identify anomalies. For example, the high confidences PSRI (R680-R500)/R750 0.79 (0.08) (western part of the image) were the fields sampled the RDVI -+ 0.76 (0.08) ()/()RR800 670 RR 800 670 most, while low confidence predictions (center of the SIPI (R800-R445)/(R800-R680)0.78 (0.08) image) correspond to areas particularly underrepresented SPVI 0.4[3.7(R800-R670)-1.2(R530-R670)] 0.70 (0.08) in the training data, such as dry barley, harvested barley, SR1 R750/R700 0.74 (0.07) and bright bare soils. This product may be used to set sen- SR3 R750/R550 0.75 (0.07) sitivity margins of field instruments quite intuitively: areas SR4 R672/R550 0.76 (0.10) are thresholded with error levels above 10% of the total Chl SRPI R430/R680 0.76 (0.09) 2 -2 TVI 0.5[120R750-R550)-200(R670-R550)] 0.70 (0.10) range (e.g. 6.5 μg/cm ), see Fig. 15(c).

VOG R740/R720 0.76 (0.06) NAOC Area in [,]643 795 0.79 (0.09) C. CHALLENGES LR Least squares linear regression 0.88 (0.06) We reviewed the very active field of physical parameter SVR RBF kernel 0.98 (0.03) estimation from acquired images. We presented the main KRR RBF kernel 0.98 (0.04)

GP [165] Anisotropic RBF kernel 0.99 (0.02) 14 https://sites.google.com/site/jochemverrelst/ARTMO______15 ______http://www.uv.es/gcamps/code/simpleR.html

28 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

approaches in the literature and introduced the principles and standard terminology. The use and performance of the different existing approaches were illustrated in a real problem of vegetation monitoring which confirmed the excellent results obtained by nonparametric sta- (a) (b) (c) tistical approaches. The field encompasses both physics of FIGURE 15. Chlorophyll concentration estimated map (a), predictive standard deviation (b), and land and atmosphere, optimi- masked confidence map (c) generated with GP on the CHRIS 12-07-2003 nadir image. zation, and machine learning. The future poses some challenging problems: the community will be confronted to the sets. Specifically, we describe techniques based on different availability of huge amount of training data coming from types of hardware accelerators (see Fig. 16), such as clusters, RTMs, the design of more sophisticated and realistic RTMs, distributed platforms and specialized devices such as com- the combination of both statistical and physically-based modity graphics processing units (GPUs) or field program- models, and the specification of models that can adapt to mable gate arrays (FPGAs). The section concludes with a multitemporal domains. summary of the main challenges in the exploitation of HPC platforms in hyperspectral remote sensing missions. VII. FAST COMPUTING In this section, we outline several strategies to accelerate A. CLUSTERS AND DISTRIBUTED hyperspectral image computations using different kinds of PLATFORMS FOR HYPERSPECTRAL PROCESSING high-performance computing (HPC) architectures. As men- Clusters were originally developed with the purpose of creat- tioned before in this contribution, the improved spatial, ing a cost-effective parallel computing system able to satisfy spectral and temporal resolutions provided by hyperspec- specific computational requirements in different applica- tral instruments demand fast computing solutions that can tions. In remote sensing, the need for large amounts of com- accelerate the efficient exploitation of hyperspectral data putation was first identified for processing multispectral

Advanced Earth Increased Spatial, Observation Spectral and Capabilities Temporal Resolution

Spectral Signature Development Real-Time Hardware Hyperspectral Requirements Accelerators Pixel

Hyperspectral Image Spectral Signature

The Spectral Signature of a Pixel Is a Combination of the Reflected or Emitted Energy from All the Features That Fall Within That Reflectance Pixel Area. 0.4 2.5 Wavelength (μm)

FIGURE 16. Different types of hardware accelerators commonly used to improve computational performance of hyperspectral imaging applications.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 29

imagery with tens of bands. As sensor instruments incor- significantly increase the computational power of cluster- porated hyperspectral capabilities, it was soon recognized based and distributed systems (e.g., the fastest supercomput- that computer mainframes and mini-computers could not ers in the world are now clusters of GPUs18). provide sufficient power for effectively processing this kind Several efforts exploiting GPU technology can already of data. It is worth noting that NASA and ESA are currently be found in the hyperspectral imaging literature [15], [16], supporting massively parallel clusters for remote sens- [180]. For instance, only in the area of spectral unmixing ing applications including hyperspectral imaging, such as there have been many developments already. A GPU-based the Columbia supercomputer16 at NASA Ames Research implementation of an automated morphological end- Center. Another example of massively parallel computing member extraction (AMEE) algorithm for pure spectral facility which has been exploited for hyperspectral imaging signature (endmember) identification is described in [181]. applications is located at the High Performance Comput- In this case, speedups on the order of 15# were reported. ing Collaboratory (HPC2) at Mississippi State University17, A full spectral unmixing chain comprising the automatic which has several supercomputing facilities that have been estimation of the number of endmembers, the identifica- used in hyperspectral imaging studies. tion of the endmember signatures, and quantification of Homogeneous clusters have already offered access to endmember fractional abundances has been reported in greatly increased computational power at a low cost (com- [182] with speedups superior to 50#. Additional efforts mensurate with falling commercial PC costs) in a number towards real-time and on-board hyperspectral target detec- of hyperspectral imaging applications, such as classification and classification [183], [184] using GPUs have also tion or spectral unmixing [15]. However, a recent trend in been recently available. It should be noted that, despite the design of HPC systems for data-intensive problems, the increasing programmability of low-power GPUs such such as those involved in hyperspectral image analysis, is as those available in smartphones, radiation-tolerance and to utilize highly heterogeneous computing resources [176]. power consumption issues still prevent the full incorpor- In this regard, networks of heterogeneous workstations can ation of GPUs to spaceborne Earth observation missions. realize a very high level of aggregate performance in hyperspectral imaging applications, and the pervasive avail- C. FPGAS FOR HYPERSPECTRAL PROCESSING ability of these resources resulted in the current notions of An FPGA [178] can be roughly defined as an array of inter- grid and, later, cloud computing, which are yet to be fully connected logic blocks. One of the main advantages of exploited in hyperspectral imaging problems [177]. these devices is that both the logic blocks and their inter- Although hyperspectral processing algorithms generally connections can be (re)configured by their users as many map quite nicely to clusters or networks of CPUs, these times as needed in order to implement different combi- systems are generally expensive and difficult to adapt national or sequential logic functions. This characteristic to onboard remote sensing data processing scenarios, in provides FPGAs with the advantages of both software and which low-weight and low-power integrated components hardware systems in the sense that FPGAs exhibit more are essential to reduce mission payload and obtain analysis flexibility and shorter development times than application results in real-time, i.e., at the same time as the data is specific integrated circuits (ASICs) but, at the same time, collected by the sensor. In this regard, the emergence of are able to provide much more competent levels of perfor- specialized hardware devices such as FPGAs [178] or mance, closer to those offered by GPUs (but with much GPUs [179] exhibit the potential to bridge the gap towards lower power consumption). In fact, the power and energy onboard and real-time analysis of remote sensing data. efficiency of FPGAs has significantly improved during the last decade. FPGA vendors have achieved this goal improv- B. GPUS FOR HYPERSPECTRAL PROCESSING ing the FPGA architectures, including optimized hardware In recent years GPUs have evolved into highly parallel, multi- modules, and taking advantage of the most recent silicon threaded, many-core coprocessors with tremendous compu- technology. For instance, manufacturing companies such tational power, consumption and memory bandwidth [179]. as Xilinx19 or Altera20 have reported a 50% reduction in The combined features of general-purpose supercomputing, the power consumption when moving from their previous high parallelism, high memory bandwidth, low cost, com- generation of FPGAs. This feature, together with the avail- pact size, and excellent programmability are now making ability of more FPGAs with increased tolerance to ionizing GPU-based desktop computers an appealing alternative to radiation in space, have consolidated FPGAs as the current a massively parallel systems made up of commodity CPUs. standard choice for on-board hyperspectral remote sensing. The exploding GPU capability has attracted more and more In the following, we outline several hyperspectral analysis scientists and engineers to use it as a cost-effective high- techniques that have been recently implemented in FPGAs. performance computing platform, including scientists in If we consider the area of spectral unmixing, implemen- hyperspectral processing areas. In addition, GPUs can also tation of endmember extraction algorithms using a Xilinx

18 http://www.top500.org______16 19 ______http://www.nas.nasa.gov/Resources/Systems/columbia.html ______http://www.xilinx.com 17 20 ______http://www.hpc.msstate.edu ______http://www.altera.com

30 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Virtex-4 FPGA have been recently described in [185]. This such a grand challenge system could support a wider vari- FPGA model is similar to radiation-hardened FPGAs certi- ety of hyperspectral imaging applications. fied for space operation. The acceleration factor or speedup of this implementation, compared with a software descrip- VIII. CONCLUDING REMARKS tion developed in C language and executed on a PC with Among the remote sensing modalities, the role of the AMD Athlon 2.6 GHz processor and 512 Mb of RAM, is hyperspectral technology in the detection and identifica- 37x for the well-known AVIRIS Cuprite scene (16 endmem- tion of materials, determination of physical parameters, bers)21, 38x for a hyperspectral image collected also in the and change detection cannot be overstated. A few signs of Cuprite mining district by EO-1 Hyperion (21 endmem- the crescent importance of the bers), and 37x for an AVIRIS image collected over the Jasper hyperspectral remote sensing Ridge biological preserve in California (19 endmembers). technology are the increasing The speedup factor is quite constant across all the images, number of hyperspectral space- THE ADOPTION OF NEW even taking into account the differences in the number of borne and airborne sensors STATISTICAL METHODS endmembers. Similarly FPGA implementations of abun- and applications supported on IN HYPERSPECTRAL dance estimation algorithms have also been described in hyperspectral remote sensing IMAGE ANALYSIS IS A [186]. This implementation was tested in the same FPGA data. Another sign is the grow- FACT TODAY. NEW used in [185], and achieved a speedup factor of 10x when ing number of scientific pub- processing the AVIRIS Cuprite scene and over 12# when it lications, which has reached EXCITING ADVANCES IN comes to the AVIRIS Jasper Ridge scene. Authors also reach a figure comparable to that of THE THEORY AND the conclusion that, using FPGAs, the execution time scales radar remote sensing, as docu- APPLICATIONS ARE linearly with the size of the image. FPGA implementations mented in the Introduction. YET TO COME. of other classic unmixing algorithms have also been dis- Owing to several factors cussed in chapter 2 of [180]. among which we refer to the Other areas in which FPGA implementations have been high dimensionality and size particularly relevant is target detection and classification. In of the hyperspectral data, the spectral mixing (linear this context, [180] discusses several examples. Specifically, and nonlinear), and the degradation mechanisms asso- chapter 15 in [180] discusses the use of FPGAs in detection ciated to the measurement process such as noise and applications and provides specific application case studies. atmospheric effects, the extraction of information from Chapter 16 in [180] describes FPGA implementations of hyperspectral remote sense data relies on sophisticated techniques for hyperspectral target detection applications. and complex data analysis methods. In this paper, we pre- Chapter 17 in [180] describes an on-board real-time pro- sented a tour over a number of representative and attrac- cessing technique for fast and accurate target detection and tive hyperspectral data analysis methods and algorithms, discrimination in hyperspectral data. Real-time implemen- organized in six main topics: data fusion, unmixing, clas- tations of several popular detection and classification algo- sification, target detection, physical parameter retrieval, rithms for hyperspectral imagery have also been discussed and fast computing. In all topics, we described the state- in [187]. of-the-art, provided illustrative examples, and pointed to future challenges and research directions. D. CHALLENGES As the reader has noted, the remote sensing data analysis Despite the individual success of the different types of HPC chain is very broad. For this reason, we could not cover all architectures described in this section in different prob- the interesting and relevant aspects exhaustively. For exam- lems, a key aspect still missing is the integration of such sys- ple, we do not cover the important field of change detec- tems in complementary fashion. Although the role of each tion and multitemporal classification, which are very active type of accelerator depends heavily on the considered appli- areas of research. Also note that the relevant field of image cation domain, cluster-based systems seem particularly compression is missing, while the crucial steps of atmos- appropriate for efficient information extraction from very pheric and geometric corrections, co-registration, decon- large data archives comprising data sets already transmit- volution, or image restoration and quality assessment have ted to Earth, while the time-critical constraints introduced been treated only superficially. by many remote sensing applications call for on-board and, As it can be concluded from the material presented in the often, real-time processing developments which require the paper, hyperspectral remote sensing data analysis is a mul- use of specialized hardware architectures such as GPUs and tidisciplinary area using and adapting frontier concepts, FPGAs. What is still missing is an infrastructure in which frameworks, and algorithms from the field of signal and these computing resources are available on-demand, possi- image processing, statistical inference, and machine learn- bly from a distributed cloud resource that can support cou- ing. Note that the methods used in hyperspectral remote pled HPC codes with strict processing deadlines. Clearly, sensing data analysis are not always simple adaptations of well-known methods developed in the above fields. For 21 http://aviris.jpl.nasa.gov/data/free______data.html example, hyperspectral unmixing has provided a unique

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 31

problem scenario to the development of new blind source [13] S. López, T. Vladimirova, C. Gónzalez, J. Resano, D. Mozos, separation solutions that are not interpretable as particular- and A. Plaza, “The promise of reconfigurable computing for izations of known solutions. The same is true, in different hyperspectral imaging on-board systems: Review and trends,” Proc. IEEE, to be published. degrees, in the remaining addressed topics. [14] R. Trautner. (2011). ESA’s roadmap for next generation payload As a conclusion, hyperspectral remote sensing data data processors. presented at DASIA Conf. [Online]. 1. Avail- analysis is a mature field in the intersection of signal and able: http://www.esa.int/TEC/OBDP/ image processing, statistical inference, and machine learn- [15] Special Issue on High Performance Computing for Hyper- ing, contributing actively with frontier cross-disciplinary spectral Imaging, Int. J. High Perform. Comput., vol. 4, no. 3, pp. 528–544, 2011. research activities. We hope that this paper be useful for [16] Special Issue on Architectures and Techniques for Real-Time researchers working in the field and foster curiosity in post- Processing of Remotely Sensed Images, J. Real-Time Image graduate students looking for a PhD theme and in research- Processing, vol. 4, no. 3, pp. 191–193, 2009. ers looking for a new research area. [17] Q. Yuan, L. Zhang, and H. Shen, “Hyperspectral image denoising employing a spectral-spatial adaptive total variation model,” IEEE Trans. Geosci. Remote Sensing,vol. 50,no. 10Pt 1, IX. ACKNOWLEGEMENT pp. 3660–3677, 2012. We gratefully acknowledge Prof. Paolo Gamba from the [18] X. Liu, S. Bourennane, and C. Fossati, “Denoising of hyper- University of Pavia, Italy, for providing the ROSIS Pavia spectral images using the PARAFAC model and statistical per- University data set. We also gratefully acknowledge Prof. formance analysis,” IEEE Trans. Geosci. Remote Sensing,vol. 50, Lorenzo Bruzzone from the University of Trento, Italy, for no. 10 Pt 1, pp. 3717–3724, 2012. [19] R. Mendez-Rial, M. Calvino-Cancela, and J. Martin-Herrero, his kind invitation to provide this contribution, as well as “Accurate implementation of anisotropic diffusion in the the anonymous reviewers for their constructive comments hypercube,” IEEE Geosci. Remote Sensing Lett., vol. 7, no. 4, and suggestions. pp. 870–874, Oct. 2010. [20] N. Jacobson and M. Gupta, “Design goals and solu- REFERENCES tions for display of hyperspectral images,” IEEE Trans. [1] G.Camps-Valls,D.Tuia,L.Gómez-Chova,S.Jiménez, and Geosci. Remote Sensing, vol. 43, no. 11, pp. 2684–2692, J. Malo, Remote Sensing Image Processing.San Rafael,CA:Mor- Nov. 2005. gan and Claypool, 2011. [21] Z. Mahmood and P. Scheunders, “Enhanced visualization of [2] J. A. Richards and X. Jia, Remote Sensing Digital Image Analy- hyperspectral images,” IEEE Geosci. Remote Sensing Lett.,vol. 8, sis: An Introduction. New York; Berlin, Germany; Heidelberg, no. 5, pp. 869–873, 2011. Germany: Springer-Verlag, 2006. [22] M. Mignotte, “A bicriteria optimization approach based di- [3] J.Bioucas-Dias,A.Plaza,N.Dobigeon,M.Parente,Q.Du, mensionality reduction model for the color display of hyper- P. Gader, and J. Chanussot, “Hyperspectral unmixing over- spectral images,” IEEE Trans. Geosci. Remote Sensing,vol. 50, view: Geometrical, statistical, and sparse regression-based ap- no. 2, pp. 501–513, 2012. proaches,” IEEE J. Select. Topics Appl. Earth Observ. Remote Sens- [23] K. Kotwal and S. Chaudhuri, “An optimization-based ing, vol. 5, no. 2, pp. 354–379, 2012. approach to fusion of hyperspectral images,” (special issue), [4] A. F. H. Goetz, G. Vane, J. E. Solomon, and B. N. Rock, “Imag- IEEE J. Select. Topics Appl. Earth Observ. Remote Sensing,vol. 5, ing spectrometry for Earth remote sensing,” Science,vol. 228, no. 2, pp. 501–509, 2012. no. 4704, pp. 1147–1153, June 1985. [24] T. Akgun, Y. Altunbasak, and R. Mersereau, “Super-resolution [5] R. Heylen, D. Burazerovic, and P. Scheunders, “Non-linear spec- reconstruction of hyperspectral images,” IEEE Trans. Image Pro- tral unmixing by geodesic simplex volume maximization,” IEEE cessing, vol. 14, no. 11, pp. 1860–1875, 2005. J. Select. Topics Signal Processing, vol. 5, no. 3, pp. 534–542, 2011. [25] J. C.-W. Chan, J. Ma, P. Kempeneers, and F. Canters, “Su- [6] M. D. Iordache, J. Bioucas-Dias, and A. Plaza, “Sparse unmix- perresolution enhancement of hyperspectral CHRIS/Proba ing of hyperspectral data,” IEEE Trans. Geosci. Remote Sensing, images with a thin-plate spline nonrigid transform model,” vol. 49, no. 6, pp. 2014–2039, 2011. IEEE Trans. Geosci. Remote Sensing, vol. 48, no. 6, pp. 2569– [7] M. T. Eismann, A. D. Stocker, and N. M. Nasrabadi, “Automat- 2579, 2010. ed hyperspectral cueing for civilian search and rescue,” Proc. [26] H. Zhang, L. Zhang, and H. Shen, “A super-resolution recon- IEEE, vol. 97, no. 6, pp. 1031–1055, June 2009. struction algorithm for hyperspectral images,” Signal Process- [8] J. Keef and T. U. of Arizona. Optical Sciences, Hyper-spectral ing, vol. 92, no. 9, pp. 2082–2096, 2012. Sensor Calibration Extrapolated from Multi-spectral Measure- [27] S. Qian and G. Chen, “Enhancing spatial resolution of hyper- ments. The University of Arizona, 2008. spectral imagery using sensor’s intrinsic keystone distortion,” [9] W. Moses, “Atmospheric correction of hyperspectral data: Exe- IEEE Trans. Geosci. Remote Sensing, vol. 50, no. 12, pp. 5033– cution of and comparison between FLAASH and TAFKAA_6S,” 5048, 2012. Cornell Univ., Ithaca,New York,Aug.2005. [28] K. C. Mertens, B. De Baets, L. P. C. Verbeke, and R. R. De Wulf, [10] W. Verhoef and H. Bach, “Simulation of hyperspectral and “A sub-pixel mapping algorithm based on sub-pixel/pixel directional radiance images using coupled biophysical and spatial attraction models,” Int. J. Remote Sens.,vol. 27,no. 15, atmospheric radiative transfer models,” Remote Sens. Environ., pp. 3293–3310, 2006. vol. 87, no. 1, pp. 23–41, 2003. [29] Y. Gu, Y. Zhang, and J. Zhang, “Integration of spatial-spectral [11] K.Staenz,A.Mueller,A.Held, and U.Heiden,“Technical com- information for resolution enhancement in hyperspectral mittees corner: International spaceborne imaging spectros- images,” IEEE Trans. Geosci. Remote Sensing,vol. 46,no. 5, copy (ISIS) technical committee,” IEEE Geosci. Remote Sensing pp. 1347–1358, 2008. Newsletter, no. 165, pp. 38–42, 2012. [30] F. A. Mianji, Y. Gu, Y. Zhang, and J. Zhang, “Enhanced self- [12] A. Plaza, J. Plaza, A. Paz, and S. Sánchez, “Parallel hyperspec- training superresolution mapping technique for hyperspectral image and signal processing,” IEEE Signal Processing Mag., tral imagery,” IEEE Geosci. Remote Sensing Letters, vol. 8, no. 4, vol. 28, no. 3, pp. 119–126, 2011. pp. 671–675, 2011.

32 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

[31] L.Alparone,L.Wald,J.Chanussot,C.Thomas,P.Gamba, [51] L. Miao and H. Qi, “Endmember extraction from highly mixed and L. M. Bruce, “Comparison of pansharpening algorithms: data using minimum volume constrained nonnegative matrix Outcome of the 2006 GRS-S data-fusion contest,” IEEE Trans. factorization,” IEEE Trans. Geosci. Remote Sensing,vol. 45,no. 3, Geosci. Remote Sensing, vol. 45, no. 10, pp. 3012–3021, 2007. pp. 765–777, 2007. [32] Y. Zhang, S. De Backer, and P. Scheunders, “Noise-resistant [52] A. Zare and P. Gader, “Sparsity promoting iterated constrained wavelet-based Bayesian fusion of multispectral and hyper- endmember detection for hyperspectral imagery,” IEEE Geosci. spectral images,” IEEE Trans. Geosci. Remote Sensing,vol. 47,no. Remote Sensing Lett., vol. 4, no. 3, pp. 446–450, 2007. 11, pp. 3834–3843, 2009. [53] J. Bioucas-Dias, “A variable splitting augmented lagragian [33] Y. Zhang, A. Duijster, and P. Scheunders, “A Bayesian resto- approach to linear spectral unmixing,” in Proc. IEEE GRSS ration approach for hyperspectral images,” IEEE Trans. Geosci. Workshop Hyperspectral Image SIgnal Processing: Evolution Remote Remote Sensing, vol. 50, no. 9, pp. 3453–3462, 2012. Sensing (WHISPERS), 2009, pp. 1–4. [34] R. Zurita-Milla, J. G. P. W. Clevers, and M. E. Schdepman, “Un- [54] T.Chan,C.Chi,Y.,Huang, and W.Ma,“Convex analysis based mixing-based Landsat TM and MERIS FR data fusion,” IEEE minimum-volume enclosing simplex algorithm for hyperspec- Geosci. Remote Sensing Lett., vol. 5, no. 3, pp. 453–457, 2008. tral unmixing,” IEEE Trans. Signal Processing,vol. 57,no. 11, [35] N. Yokoya, T. Yairi, and A. Iwasaki, “Coupled nonnegative ma- pp. 4418–4432, 2009. trix factorization unmixing for hyperspectral and multispec- [55] N. Dobigeon, S. Moussaoui, J.-Y. Tourneret, and C. Carteret, tral data fusion,” IEEE Trans. Geosci. Remote Sensing,vol. 50,no. “Bayesian separation of spectral sources under non-negativity 2, pp. 528–537, 2012. and full additivity constraints,” Signal Process., vol. 89, no. 12, [36] M. Elbakary and M. Alam, “Superresolution construction of pp. 2657–2669, Dec. 2009. multispectral imagery based on local enhancement,” IEEE [56] J. Nascimento and J. Bioucas-Dias, “Hyperspectral unmixing Geosci. Remote Sensing Lett., vol. 5, no. 2, pp. 276–279, 2008. based on mixtures of Dirichlet components,” IEEE Trans. Geos- [37]C.Thomas,T.Ranchin,L.Wald, and J.Chanussot,“Syn- ci. Remote Sensing,vol. 50,no. 3,2012. thesis of multispectral images to high spatial resolution: A [57] R. Basedow, D. Carmer, and M. Anderson, “Hydice system: critical review of fusion methods based on remote sensing Implementation and performance,” in Proc. SPIE’s 1995 Symp. physics,” IEEE Trans. Geosci. Remote Sensing,vol. 46,no. 5, OE/Aerospace Sensing and Dual Use Photonics, pp. 258–267. pp. 1301–1312, 2008. [58] J. Bioucas-Dias and J. Nascimento, “Hyperspectral subspace [38] K. Kotwal and S. Chaudhuri, “A novel approach to quantitative identification,” IEEE Trans. Geosci. Remote Sensing,vol. 46, evaluation of hyperspectral image fusion techniques,” Inform. no. 8, pp. 2435–2445, 2008. Fusion, vol. 14, no. 1, pp. 5–18, 2013. [59] J. Mustard and C. Pieters, “Quantitative abundance estimates [39] N. Keshava and J. Mustard, “Spectral unmixing,” IEEE Signal from bidirectional reflectance measurements,” J. Geophys. Res., Processing Mag., vol. 19, no. 1, pp. 44–57, 2002. vol. 92, pp. E617–E626, 1987. [40] B. Hapke, “Bidirection reflectance spectroscopy. I. Theory,” [60]A.Halimi,Y.Altmann,N.Dobigeon, and J.-Y.Tourneret, J. Geophys. Res., vol. 86, pp. 3039–3054, 1981. “Nonlinear unmixing of hyperspectral images using a general- [41] M. Craig, “Minimum-volume transforms for remotely sensed ized bilinear model,” IEEE Trans. Geosci. Remote Sensing,no. 11, data,”IEEE Trans. Geosci. Remote Sensing,vol. 32, pp. 542–552,1994. pp. 4153–4162, Nov. 2011. [42] J. Nascimento and J. Bioucas-Dias, “Does independent [61] J. Broadwater, A. Banerjee, and P. Burlina, “Kernel methods for component analysis play a role in unmixing hyperspec- unmixing hyperspectral imagery,” in Optical Remote Sensing tral data?” IEEE Trans. Geosci. Remote Sensing,vol. 43,no. 1, Advances in Signal Processing and Exploitation,L. B. S.Prasad and pp. 175–187, 2005. E. J. Chanussot, Eds. Springer, 2011, pp. 247–269. [43] J. Boardman, “Automating spectral unmixing of AVIRIS data [62] A. Plaza, G. Martin, J. Plaza, M. Zortea, and S. Sánchez, “Re- using convex geometry concepts,” in Proc. Ann. JPL Airborne cent developments in spectral unmixing and endmember Geoscience Workshop, 1993, vol. 1, pp. 11–14. extraction,” in Optical Remote Sensing, S. Prasad, L. M. Bruce, [44] J. Nascimento and J. Bioucas-Dias, “Vertex component analy- and J. Chanussot, Eds. Berlin, Germany: Springer-Verlag, 2011, sis: A fast algorithm to unmix hyperspectral data,” IEEE Trans. ch. 12, pp. 235–267. Geosci. Remote Sensing, vol. 43, no. 4, pp. 898–910, 2005. [63] Y. Altmann, A. Halimi, N. Dobigeon, and J.-Y. Tourneret, [45] C.-I. Chang, C.-C. Wu, W. Liu, and Y.-C. Ouyang, “A new “Supervised nonlinear spectral unmixing using a post-non- growing method for simplex-based endmember extraction al- linear mixing model for hyperspectral imagery,” IEEE Trans. gorithm,” IEEE Trans. Geosci. Remote Sensing,vol. 44,no. 10, Image Processing, vol. 21, no. 6, pp. 3017–3025, 2012. pp. 2804–2819, 2006. [64] R. Heylen and P. Scheunders, “Calculation of geodesic distanc- [46] T.-H. Chan, W.-K. Ma, A. Ambikapathi, and C.-Y. Chi, “A sim- es in nonlinear mixing models: Application to the generalized plex volume maximization framework for hyperspectral end- bilinear model,” IEEE Geosci. Remote Sensing Lett., vol. 9, no. 4, member extraction,” IEEE Trans. Geosci. Remote Sensing,vol. 49, pp. 644–648, 2012. no. 11, 2011. [65] R. Close, “Endmember and proportion estimation using [47] N. Gillis and S. Vavasis, “Fast and robust recursive algorithms physics-based macroscopic and microscopic mixture models,” for separable nonnegative matrix factorization,” arXiv preprint Ph.D. dissertation, Univ. Florida, CA, Dec. 2011. arXiv:1208.1237, 2012. [66] E. Candès, J. Romberg, and T. Tao, “Stable signal recovery from [48] M. E. Winter, “N-FINDR: An algorithm for fast autonomous incomplete and inaccurate measurements,” Commun. Pure spectral end member determination in hyperspectral data,” in Appl. Math., vol. 59, no. 8, pp. 1207–1223, 2006. Proc. SPIE Image Spectrometry V, 1999, vol. 3753, pp. 266–277. [67] B. Natarajan, “Sparse approximate solutions to linear sys- [49] R. A. Neville, K. Staenz, T. Szeredi, J. Lefebvre, and P. Hauff, tems,” SIAM J. Comput., vol. 24, no. 2, pp. 227–234, 1995. “Automatic end member extraction from hyperspectral data [68]Y.Pati,R.Rezaiifar, and P.Krishnaprasad,“Orthogonal for mineral exploration,” in Proc. Canadian Symp. Remote Sens- matching pursuit: Recursive function approximation with ap- ing, 1999, pp. 21–24. plications to wavelet decomposition,” in 1993 Conference Record [50] M.Berman,H.Kiiveri,R.Lagerstrom,A.Ernst,R.Dunne, and 27th Asilomar Conf. Signals, Systems and Computers, pp. 40–44. J. F. Huntington, “ICE: A statistical approach to identifying [69] E. Candès and J. J. Romberg, “Sparsity and incoherence in endmembers in hyperspectral images,” IEEE Trans. Geosci. Re- compressive sampling,” IEEE Trans. Image Processing,vol. 23, mote Sensing, vol. 42, no. 10, pp. 2085–2095, 2004. pp. 969–985, 2007.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 33

[70] J. Bioucas-Dias and M. Figueiredo, “Alternating direction [90] J. Li, J. Bioucas-Dias, and A. Plaza, “Semi-supervised hyper- algorithms for constrained sparse regression: Application to spectral image segmentation using multinomial logistic re- hyperspectral unmixing,” in Proc. IEEE GRSS Workshop Hyper- gression with active learning,” IEEE Trans. Geosci. Remote Sens- spectral Image SIgnal Processing: Evolution Remote Sensing (WHIS- ing, vol. 48, pp. 4085–4098, 2010. PERS), 2010, vol. 1, pp. 1–4. [91] J. Li, J. Bioucas-Dias, and A. Plaza, “Spectral-spatial hyperspec- [71] J.Mairal,R.Jenatton,G.Obozinski, and F.Bach,“Convex and tral image segmentation using subspace multinomial logistic network flow optimization for structured sparsity,” J. Mach. regression and markov random fields,” IEEE Trans. Geosci. Learn. Res., vol. 12, pp. 2681–2720, 2011. Remote Sensing, vol. 50, no. 3, pp. 809–823, 2012. [72] M.-D. Iordache, J. Bioucas-Dias, and A. Plaza, “Total variation [92] P. Soille, Morphological Image Analysis, Principles and Applica- spatial regularization for sparse hyperspectral unmixing,” IEEE tions,2nd ed.Berlin,Germany:Springer-Verlag,2003. Trans. Geosci. Remote Sensing, vol. 50, no. 11, pp. 4484–4502, [93] M. Pesaresi and J. Benediktsson, “A new approach for the Nov. 2012. morphological segmentation of high-resolution satellite [73] J. Tropp, A. Gilbert, and M. Strauss, “Algorithms for simulta- imagery,” IEEE Trans. Geosci. Remote Sensing,vol. 39,no. 2, neous sparse approximation. Part I. Greedy pursuit,” Signal pp. 309–320, 2001. Process., vol. 86, no. 3, pp. 572–588, 2006. [94] J. Benediktsson, J. Palmason, and J. Sveinsson, “Classification [74] M.-D. Iordache, J. Bioucas-Dias, and A. Plaza, “Collaborative of hyperspectral data from urban areas based on extended sparse regression for hyperspectral unmixing,” submitted for morphological profiles,” IEEE Trans. Geosci. Remote Sensing, publication. vol. 43, no. 3, pp. 480–491, 2005. [75] D. A. Landgrebe, Signal Theory Methods in Multispectral Remote [95] M. Dalla Mura, J. Atli Benediktsson, B. Waske, and L. Bruz- Sensing.Hoboken,NJ:Wiley,2003. zone, “Morphological attribute profiles for the analysis of very [76] A. Plaza, J. A. Benediktsson, J. Boardman, J. Brazile, L. Bru- high resolution images,” IEEE Trans. Geosci. Remote Sensing,vol. zzone, G. Camps-Valls, J. Chanussot, M. Fauvel, P. Gamba, 48, no. 10, pp. 3747–3762, 2010. J. Gualtieri, M. Marconcini, J. C. Tilton, and G. Trianni, “Re- [96] M. Dalla Mura, J. Atli Benediktsson, B. Waske, and L. Bruz- cent advances in techniques for hyperspectral image process- zone, “Extended profiles with morphological attribute filters ing,” Remote Sens. Environ., vol. 113, pp. 110–122, 2009. for the analysis of hyperspectral data,” Int. J. Remote Sensing, [77] A. Green, M. Berman, P. Switzer, and M. Craig, “A transforma- vol. 31, no. 22, pp. 5975–5991, 2010. tion for ordering multispectral data in terms of image quality [97] G. Camps-Valls, L. Goméz-Chova, J. Muñoz-Marí, J. Vila- with implications for noise removal,” IEEE Trans. Geosci. Re- Francés, and J. Calpe-Maravilla, “Composite kernels for hyper- mote Sensing, vol. 26, pp. 65–74, 1988. spectral image classification,” IEEE Geosci. Remote Sensing Lett., [78] X. Jia and J. A. Richards, “Segmented principal components vol. 3, pp. 93–97, 2006. transformation for efficient hyperspectral remote sensing [98] J.Li,P.Marpu,A.Plaza,J.Bioucas-Dias, and J. A.Benedikts- image display and classification,” IEEE Trans. Geosci. Remote son, “Generalized composite kernel framework for hyperspec- Sensing, vol. 37, pp. 538–542, 1999. tral image classification,” IEEE Trans. Geosci. Remote Sensing,to [79] L. Bruce and J. Li, “Wavelets for computationally efficient be published. hyperspectral derivative analysis,” IEEE Trans. Geosci. Remote [99] Y. Tarabalka, M. Fauvel, J. Chanussot, and J. Benediktsson, Sensing, vol. 39, pp. 1540–1546, 2001. “SVM- and MRF-based method for accurate classification of [80]W.Liao,A.Pizurica,P.Scheunders,W.Philips, and Y.Pi, hyperspectral images,” IEEE Geosci. Remote Sensing Lett.,vol. 7, “Semisupervised local discriminant analysis for feature ex- no. 4, pp. 736–740, 2010. traction in hyperspectral images,” IEEE Trans. Geosci. Remote [100] Y. Tarabalka, J. Chanussot, and J. Benediktsson, “Segmenta- Sensing, vol. 51, pp. 184–198, 2013. tion and classification of hyperspectral images using watershed [81] S. Kumar, J. Ghosh, and M. M. Crawford, “Best-bases feature transformation,” Pattern Recognit., vol. 43, pp. 2367–2379, 2010. extraction algorithms for classification of hyperspectral data,” [101] Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton, IEEE Trans. Geosci. Remote Sensing, vol. 39, pp. 1368–1379, 2001. “Multiple spectral-spatial classification approach for hyper- [82] X. Jia and J. A. Richards, “Progressive two-class decision classi- spectral data,” IEEE Trans. Geosci. Remote Sensing,vol. 48,no. fier for optimization class discriminations,” Remote Sens. Envi- 11, pp. 4122–4132, 2011. ron., vol. 63, pp. 289–297, 1998. [102] M. Fauvel, Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and [83] G. Camps-Valls, J. Mooij, and B. Schölkopf, “Remote sens- J. C. Tilton, “Advances in spectral-spatial classification of ing feature selection by kernel dependence estimation,” IEEE hyperspectral images,” Proc. IEEE, to be published. Trans. Geosci. Remote Sensing, vol. 7, pp. 587–591, 2010. [103] Y. Chen, N. M. Nasrabadi, and T. D. Tran, “Hyperspectral [84] T. V. Bandos, L. Bruzzone, and G. Camps-Valls, “Classification image classification using dictionary-based sparse represen- of hyperspectral images with regularized linear discriminant tation,” IEEE Trans. Geosci. Remote Sensing,vol. 49,no. 10, analysis,” IEEE Trans. Geosci. Remote Sensing,vol. 47,no. 3, pp. 3973–3985, Oct. 2011. pp. 862–873, 2009. [104]A.Castrodad,Z.Xing,J.Greer,E.Bosch,L.Carin, and [85] F. Melgani and L. Bruzzone, “Classification of hyperspectral re- G. Sapiro, “Learning discriminative sparse representations for mote-sensing images with support vector machines,” IEEE Trans. modeling, source separation, and mapping of hyperspectral Geosci. Remote Sensing, vol. 42, no. 8, pp. 1778–1790, 2004. imagery,” IEEE Trans. Geosci. Remote Sensing,vol. 49,no. 11, [86] G. Camps-Valls and L. Bruzzone, “Kernel-based methods for pp. 4263–4281, Dec. 2011. hyperspectral image classification,” IEEE Trans. Geosci. Remote [105] F. Bovolo, L. Bruzzone, and L. Carlin, “A novel technique for Sensing, vol. 43, pp. 1351–1362, 2005. subpixel image classification based on support vector machine,” [87] B. Schölkopf and A. Smola, Learning with Kernels: Support Vector IEEE Trans. Image Processing, vol. 19, pp. 2983–2999, 2010. Machines, Regularization, Optimization and Beyond. Cambridge, [106] L. Bruzzone, M. Chi, and M. Marconcini, “A novel transduc- MA: MIT Press, 2002. tive SVM for the semisupervised classification of remote- [88] G. Foody, “Thematic map comparison: Evaluating the statisti- sensing images,” IEEE Trans. Geosci. Remote Sensing,vol. 11, cal significance of differences in classification accuracy,” Pho- pp. 3363–3373, 2006. togramm. Eng. Remote Sens., vol. 70, no. 5, pp. 627–633, 2004. [107]G.Camps-Valls,T.Bandos, and D.Zhou,“Semi-supervised [89] D. Böhning, “Multinomial logistic regression algorithm,” Ann. graph-based hyperspectral image classification,” IEEE Trans. Inst. Statist. Math., vol. 44, pp. 197–200, 1992. Geosci. Remote Sensing, vol. 45, pp. 3044–3054, Oct. 2007.

34 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

[108] S. Velasco-Forero and V. Manian, “Improving hyperspectral [127] G. Healey and D. Slater, “Models and methods for automated image classification using spatial preprocessing,” IEEE Geosci. material identification in hyperspectral imagery acquired un- Remote Sensing Lett., vol. 6, pp. 297–301, 2009. der unknown illumination and atmospheric conditions,” IEEE [109] D. Tuia and G. Camps-Valls, “Semisupervised remote sensing Trans. Geosci. Remote Sensing,vol. 37,no. 6, pp. 2706–2717,1999. image classification with cluster kernels,” IEEE Geosci. Remote [128] J. Harsanyi and C.-I. Chang, “Hyperspectral image classifica- Sensing Lett., vol. 6, no. 2, pp. 224–228, Apr. 2009. tion and dimensionality reduction: An orthogonal subspace [110] L. Bruzzone and C. Persello, “A novel context-sensitive projection approach,” IEEE Trans. Geosci. Remote Sensing, semisupervised SVM classifier robust to mislabeled training vol. 32, no. 4, pp. 779–785, 1994. samples,” IEEE Trans. Geosci. Remote Sensing,vol. 47,no. 7, pp. [129] L. Scharf and B. Friedlander, “Matched subspace detectors,” 2142–2154, 2009. Signal Process., vol. 42, no. 8, pp. 2146–2157, 1994. [111]J.Muñoz Marí,F.Bovolo,L.Gómez-Chova,L.Bruzzone, and [130] F. Robey, D. Fuhrmann, and E. Kelly, “A CFAR adaptive G. Camp-Valls, “Semisupervised one-class support vector ma- matched filter detector,” IEEE Trans. Aerosp. Electron. Syst.,vol. chines for classification of remote sensing data,” IEEE Trans. 28, no. 1, pp. 208–216, 1992. Geosci. Remote Sensing, vol. 48, no. 8, pp. 3188–3197, 2010. [131]S.Kraut,L.Scharf, and L.McWhorter,“Adaptive subspace de- [112]L.Gómez-Chova,G.Camps-Valls,L.Bruzzone, and J.Calpe- tectors,” Signal Process., vol. 49, no. 1, pp. 208–216, 2001. Maravilla, “Mean MAP kernel methods for semisupervised [132] B. Schölkopf and A. J. Smola, Learning with Kernels. Cambridge, cloud classification,” IEEE Trans. Geosci. Remote Sensing,vol. MA: MIT Press, 2002. 48, no. 1, pp. 207–220, 2010. [133] H. Kwon and N. Nasrabadi, “A comparative analysis of kernel [113] D. Tuia and G. Camps-Valls, “Urban image classification with subspace target detectors for hyperspectral imagery,” EURASIP semisupervised multiscale cluster kernels,” IEEE J. Select. Top- J. Appl. Signal Process., vol. 2007, no. 1, pp. 193–193, 2007. ics Appl. Earth Observ. Remote Sensing, vol. 4, no. 1, pp. 65–74, [134] S. Reed and X. Yu, “Adaptive multiple-band CFAR detection Mar. 2011. of an optical pattern with unknown spectral distribution,” [114] F.Ratle,G.Camps-Valls, and J.Weston,“Semisupervised neu- IEEE Trans. Acoust., Speech, Signal Process.,vol. 38,no. 10, ral networks for efficient hyperspectral image classification,” pp. 1760–1770, 1990. IEEE Trans. Geosci. Remote Sensing,vol. 48,no. 5, pp. 2271– [135] C. Willis, “Comparison of anomaly detection methods for 2282, May 2010. hyperspectral imagery,” in Proc. SPIE,Orlando,FL,2005, [115] J. Muñoz-Marí, D. Tuia, and G. Camps-Valls, “Semisupervised vol. 5988, pp. B1–B12. classification of remote sensing images with active queries,” IEEE [136] C.-I. Chang, Hyperspectral Data Exploitation: Theory and Applica- Trans. Geosci. Remote Sensing, vol. 50, no. 10, pp. 3751–3763, 2012. tions.Hoboken,NJ:Wiley,2007. [116]D.Tuia,M.Volpi,L.Copa,M.Kanevski, and J.Muñoz-Marí, [137] A. Stocker and A. Schaum, “Application of stochastic mixing “A survey of active learning algorithms for supervised remote models to the hyperspectral detection problems,” in Proc. sensing image classification,” IEEE J. Select. Topics Signal Pro- SPIE,Orlando,FL,1997,vol. 3071, pp. 47–60. cessing, vol. 5, no. 3, pp. 606–617, June 2011. [138] H. Kwon, S. Der, and N. Nasrabadi, “Adaptive anomaly detec- [117] W. Di and M. M. Crawford, “Active learning via multi-view tion using subspace separation for hyperspectral images,” Opt. and local proximity co-regularization for hyperspectral image Eng., vol. 42, no. 11, pp. 3342–3351, 2003. classification,” IEEE J. Select. Topics Signal Process., vol. 5, no. 3, [139] K. Ranney and M. Soumekh, “Hyperspectral anomaly detec- pp. 618–628, 2011. tion within the signal subspace,” IEEE Trans. Geosci. Remote [118] S. Patra and L. Bruzzone, “A batch-mode active learning tech- Sensing Lett., vol. 3, no. 3, pp. 312–316, 2006. nique based on multiple uncertainty for SVM classifier,” IEEE [140] D. Stein, “Stochastic compositional models applied to sub- Geosci. Remote Sensing Lett., vol. 9, no. 3, pp. 497–501, May 2012. pixel analysis of hyperspectral imagery,” in Proc. SPIE,2001, [119] I.Dopido,J.Li,P.Marpu,A.Plaza,J.Bioucas-Dias, and J. A.Bene- vol. 4480, pp. 49–56. diktsson, “Semi-supervised self-learning for hyperspectral image [141] E. Ashton, “Detection of subpixel anomalies in multispectral classification,” IEEE Trans. Geosci. Remote Sensing, to be published. infrared imagery using an adaptive bayesian classifier,” IEEE [120] L. Ma, M. Crawford, and J. Tian, “Local manifold learning- Trans. Geosci. Remote Sensing, vol. 36, no. 2, pp. 506–517, 1998. based k-nearest-neighbor for hyperspectral image classifi- [142] S. Schweizer and J. Moura, “Hperspectral imagery: Clutter cation,” IEEE Trans. Geosci. Remote Sensing,vol. 48,no. 11, adaptation in anomaly detection,” IEEE Trans. Inform. Theory, pp. 4099–4109, 2010. vol. 46, pp. 1855–1871, 2000. [121] W. Kim and M. Crawford, “Adaptive classification for hyper- [143] M. Carlotto, “A cluster-based approach for detecting man- spectral image data using manifold regularization kernel made objects and changes in imagery,” IEEE Trans. Geosci. Re- machines,” IEEE Trans. Geosci. Remote Sensing,vol. 48,no. 11, mote Sensing, vol. 43, no. 2, pp. 374–387, 2005. pp. 4110–4121, 2010. [144] A. Banerjee, P. Burlina, and C. Diehl, “A support vector method [122] D. Manolakis and G. Shaw, “Detection algorithms for hyper- for anomaly detection in hyperspectral imagery,” IEEE Trans. spectral imaging applications,” IEEE Signal Processing Mag., Geosci. Remote Sensing, vol. 44, no. 8, pp. 2282–2291, 2006. vol. 19, no. 1, pp. 29–43, Jan. 2002. [145] H. Goldberg and N. Nasrabadi, “A comparative study of linear [123] S. Matteoli, M. Diani, and G. Corsini, “A tutorial overview of and nonlinear anomaly detectors for hyperspectral imagery,” anomaly detection in hyperspectral images,” IEEE Aerospace in Proc. SPIE,Orlando,FL,2007,vol. 6497. Elect. Systems Mag.,vol. 25,no. 7, pp. 5–27,July2010. [146] A. Schaum, “Joint subspace detection of hyperspectral tar- [124] D. W. J. Stein, S. G. Beaven, L. E. Hoff, E. M. Winter, A. P. Schaum, gets,” in Proc. IEEE Aerospace Conf.,2004,vol. 3. and A. D. Stocker, “Anomaly detection from hyperspectral im- [147] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statis- agery,” IEEE Signal Processing Mag., vol. 19, pp. 58–69, 2002. tical Learning. Springer, 2001. [125] A. P. Schaum and A. D. Stocker, “Subclutter target detection [148] L. Scharf, Statistical Signal Processing. Reading, MA: Addison- using sequences of thermal infrared multispectral imagery,” Wesley, 1991. in Proc. SPIE, 1997, vol. 3071, pp. 12–22. [149] D. Gu, A. Gillespie, A. Kahle, and F. Palluconi, “Autono- [126] M. Eismann, J. Meola, and R. Hardie, “Hyperspectral change mous atmospheric compensation (AAC) of high reso- detection in the presence of diurnal and seasonal variations,” lution hyperspectral thermal infrared remote-sensing IEEE Trans. Geosci. Remote Sensing,vol. 46,no. 1, pp. 237–249, imagery,” IEEE Trans. Geosci. Remote Sensing,vol. 38,no. 11, Jan. 2008. pp. 2557–2570, 2000.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 35

[150] P. Villeneuve, H. Fry, J. Theiler, and W. Clodius, “Improved [170] F. Baret and T. Fourty, “Estimation of leaf water content matched-filter detection techniques,” in Proc. SPIE, Orlando, and specific leaf weight from reflectance and transmittance FL,1999,vol. 3753. measurements,” Agronomie, vol. 17, nos. 9–10, pp. 455–464, [151] N. Nasrabadi, “Regularized spectral matched filter for target 1997. recognition in hyperspectral imagery,” IEEE Trans. Signal Pro- [171] H. Fang and S. Liang, “A hybrid inversion method for map- cessing Lett., vol. 15, pp. 317–320, 2008. ping leaf area index from MODIS data: Experiments and ap- [152] Y. Chen, N. Nasrabadi, and T. Tran, “Sparse representation for plication to broadleaf and needleleaf canopies,” Remote Sens. target detection in hyperspectral imagery,” IEEE J. Select. Topics Environ., vol. 94, no. 3, pp. 405–424, 2005. Signal Processing, vol. 5, no. 3, pp. 629–640, 2011. [172] C. Bacour, F. Baret, D. Béal, M. Weiss, and K. Pavageau, “Neu- [153] Y. Chen, N. Nasrabadi, and T. Tran, “Simultaneous joint sparsity ral network estimation of LAI, fAPAR, fCover and LAI#Cab, model for target detection in hyperspectral imagery,” IEEE Trans. from top of canopy MERIS reflectance data: Principles and Geosci. Remote Sensing Lett., vol. 8, no. 4, pp. 676–680, 2011. validation,” Remote Sens. Environ., vol. 105, no. 4, pp. 313– [154] C. Chen, N. Nasrabadi, and T. Tran, “Kernel sparse represen- 325, 2006. tation for hyperspectral target detection,” in Proc. SPIE,Balti- [173] A. Verger, F. Baret, and F. Camacho, “Optimal modalities for more,MD,2012,vol. 8390. radiative transfer-neural network estimation of canopy bio- [155] C. Rodgers, Inverse Methods for Atmospheric Sounding: Theory and physical characteristics: Evaluation over an agricultural area Practice.Singapore:World Scientific,2000. with CHRIS/PROBA observations,” Remote Sens. Environ.,vol. [156] S. Liang, Quantitative Remote Sensing of Land Surfaces.Hoboken, 115, no. 2, pp. 415–426, 2011. NJ: Wiley, 2004. [174] G. Camps-Valls and L. Bruzzone, Kernel Methods for Remote [157] F.Baret and S.Buis,“Estimating canopy characteristics from re- Sensing Data Analysis.Hoboken,NJ:Wiley,2009. mote sensing observations: Review of methods and associated [175] S. Durbha, R. King, and N. Younan, “Support vector machines problems,” in Advances in Land Remote Sensing: System, Modeling, regression for retrieval of leaf area index from multiangle im- Inversion and Applications.Germany:Springer-Verlag,2008. aging spectroradiometer,” Remote Sens. Environ.,vol. 107,nos. [158] J. Berni, P. Zarco-Tejada, L. Suárez, and E. Fereres, “Thermal 1–2, pp. 348–361, 2007. and narrowband multispectral remote sensing for vegetation [176]A.Plaza,D.Valencia, and J.Plaza,“An experimental com- monitoring from an unmanned aerial vehicle,” IEEE Trans. parison of parallel algorithms for hyperspectral analysis using Geosci. Remote Sensing, vol. 47, no. 3, pp. 722–738, 2009. homogeneous and heterogeneous networks of workstations,” [159] J. Verrelst, M. Schaepman, B. Koetz, and M. Kneubhler, “An- Parallel Compt., vol. 34, no. 2, pp. 92–114, 2008. gular sensitivity analysis of vegetation indices derived from [177]J.Brazile,R. A.Neville,K.Staenz, D. Schlaepfer, L. Sun, and CHRIS/PROBA data,” Remote Sens. Environ., vol. 112, no. 5, K. I. Itten, “Cluster versus grid for operation generation of pp. 2341–2353, 2008. ATCOR’s MODTRAN-based look up table,” Parallel Compt., [160]M.Schaepman,S.Ustin,A.Plaza,T.Painter,J.Verrelst, and vol. 34, pp. 32–46, 2008. S. Liang, “Earth system science related imaging spectros- [178] S. Hauck, “The roles of FPGAs in reprogrammable systems,” copy: An assessment,” Remote Sens. Environ., vol. 113, no. 1, Proc. IEEE, vol. 86, no. 4, pp. 615–639, 1998. pp. S123–S137, 2009. [179] J. Nickolls and W. J. Dally, “The GPU computing era,” IEEE [161] L. Sarker and J. Nichol, “Improved forest biomass estimates Micro, vol. 30, pp. 56–69, 2010. using ALOS AVNIR-2 texture indices,” Remote Sens. Environ., [180] A. Plaza and C.-I. Chang, High Performance Computing in Remote vol. 115, no. 4, pp. 968–977, 2011. Sensing. New York: Taylor & Francis, 2007. [162] P. Townsend, J. Foster, R. J. Chastain, and W. Currie, “Applica- [181] J.Setoain,M.Prieto,C.Tenllado,A.Plaza, and F.Tirado,“Par- tion of imaging spectroscopy to mapping canopy nitrogen in allel morphological endmember extraction using commodity the forests of the central Appalachian Mountains using Hyper- graphics hardware,” IEEE Geosci. Remote Sensing Lett.,vol. 43, ion and AVIRIS,” IEEE Trans. Geosci. Remote Sensing,vol. 41, no. 3, pp. 441–445, 2007. no. 6, pp. 1347–1354, June 2003. [182]S.Sánchez,A.Paz,G.Martin, and A.Plaza,“Parallel unmix- [163] J. Arenas-García and G. Camps-Valls, “Efficient kernel orthonor- ing of remotely sensed hyperspectral images on commodity malized PLS for remote sensing applications,” IEEE Trans. Geosci. graphics processing units,” Concurrency Comput. Pract. Exp.,vol. Remote Sensing, vol. 46, no. 10, pp. 2872–2881, Oct. 2008. 23, no. 13, pp. 1538–1557, 2011. [164]G.Camps-Valls,L.Bruzzone,J. L.Rojo-Álvarez, and F.Mel- [183] Y. Tarabalka, T. V. Haavardsholm, I. Kasen, and T. Skauli, gani, “Robust support vector regression for biophysical vari- “Real-time anomaly detection in hyperspectral images using able estimation from remotely sensed images,” IEEE Geosci. multivariate normal mixture models and GPU processing,” Remote Sensing Lett., vol. 3, no. 3, pp. 339–343, 2006. J. Real-Time Image Process., vol. 4, pp. 1–14, 2009. [165]J.Verrelst,L.Alonso,G.Camps-Valls,J.Delegido, and J.More- [184]S.Bernabe,S.Lopez,A.Plaza, and R.Sarmiento,“GPU im- no, “Retrieval of canopy parameters using Gaussian processes plementation of an automatic target detection and classifica- techniques,” IEEE Trans. Geosci. Remote Sensing,vol. 49,2011. tion algorithm for hyperspectral image analysis,” IEEE Geosci. [166] S. Jacquemoud, C. Bacour, H. Poilvé, and J.-P. Frangi, “Com- Remote Sensing Lett., vol. 10, no. 2, pp. 221–225, 2013. parison of four radiative transfer models to simulate plant [185] C. González, D. Mozos, J. Resano, and A. Plaza, “FPGA imple- canopies reflectance: Direct and inverse mode,” Remote Sens. mentation of the N-FINDR algorithm for remotely sensed hy- Environ., vol. 74, no. 3, pp. 471–481, 2000. perspectral image analysis,” IEEE Trans. Geosci. Remote Sensing, [167] M. Weiss and F. Baret, “Evaluation of canopy biophysical variable vol. 50, no. 2, pp. 374–374, 2012. retrieval performances from the accumulation of large swath sat- [186] C. Gonzalez, J. Resano, A. Plaza, and D. Mozos, “FPGA imple- ellite data,” Remote Sens. Environ., vol. 70, pp. 293–306, 19991. mentation of abundance estimation for spectral unmixing of [168] F. Baret, J. Clevers, and M. Steven, “The robustness of canopy hyperspectral data using the image space reconstruction algo- gap fraction estimates from red and near-infrared reflectance: rithm,” IEEE J. Select. Topics Appl. Earth Observ. Remote Sensing, A comparison of approaches,” Remote Sens. Environ.,vol. 54, vol. 5, no. 1, pp. 248–261, 2012. pp. 141–151, 1995. [187] Q. Du and R. Nekovei, “Fast real-time onboard processing of [169] S. Gopal and C. Woodcock, “Remote sensing of forest change hyperspectral imagery for detection and classification,” J. Real- using artificial neural networks,” IEEE Trans. Geosci. Remote Time Image Process., vol. 4, no. 3, pp. 273–286, 2009. Sensing, vol. 34, pp. 398–404, 1996. GRS

36 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

© PHOTODISC Spectroradiometric Field Surveys in Remote Sensing Practice: A Workflow Proposal, from Planning to Analysis

L. POMPILIO, Department of Psychological, Humanistic and Earth Sciences, University “G. d’Annunzio,” V. Dei Vestini, 31, Chieti, Italy, I-66013 P. VILLA, Institute of Information Science and Technologies “A. Faedo,” National Research Council (ISTI-CNR), Via G. Moruzzi 1, Pisa, Italy, 56124 M. BOSCHETTI, AND M. PEPE, Institute for Electromagnetic Sensing of the Environment, National Research Council (IREA-CNR), Via Bassini 15, Milan, Italy, 20133

Abstract—Spectroradiometric field surveys, especially and data management. The feasibility and advantages of when addressed at heterogeneous targets and within a such approach are demonstrated through a real case sce- complex environmental context, require careful organiza- nario, covering two spectroradiometric campaigns aimed tion and structuring. This work focuses on a proposal of an at geolithology studies and performed in September 2010 operational workflow that holds together different aspects and 2011 in the Dolomites (Northern Italy), where com- of in situ spectral data acquisition, from planning to data plex environmental and variable atmospheric conditions, analysis, and is ancillary to environmental analysis based as well as, large target variability and spectral heterogeneity on remote sensing. The main objective is the enhance- occur. The workflow here adopted allowed us to face all the ment of the information coming from acquired spectra, challenges in order to usefully accomplish our campaigns. which is accomplished through a set of established guide- Issues commonly arising in spectroradiometric field activi- lines for: campaign planning, measurement collection, ties are easily foreseen, tackled and mitigated, if necessary, while preserving a high degree of flexibility and generaliza-

Digital Object Identifier 10.1109/MGRS.2013.2261257 tion for exploiting the workflow over heterogeneous areas Date of publication: 26 June 2013 and application fields.

1. INTRODUCTION simultaneously to the measurements. In addition, the pectroradiometric field surveys generally aim at the in- retrieval of correct radiometric quantities—which is criti- Svestigation of the spectral variability on ground, at the cal in a wide variety of applications—requires high lev- local and close view scales, with high quality data, and the els of accuracy in data collection. As a consequence, the subsequent application of this information to the regional fulfillment of specific criteria in the accomplishment of scale observations through hyperspectral imagery [1], [2]. spectroradiometric field surveys is recognized as a funda- Such type of campaigns are specifically designed to account mental step toward the development of the measurement for the spectral variability on ground at the local and close standardization and data interoperability processes. The view scales, and they provide an invaluable support for formalization of a measurement and data protocol with calibration, data processing and validation of image analy- proper rules and guidelines for the selection of targets, sis [3]. In order to maximize the information coming from measurement collection (both in the field and in the lab- them, processing remotely sensed data with the support of oratory), and data encoding will enforce the successful ground spectroradiometric data requires that on one hand integration of different datasets acquired at different time the in situ observations are accurately collected and statisti- and by different operators. cally representative, and on the other hand, the information So far, much of the efforts to establish standard guide- is conveniently documented to be managed and retrieved. lines for the accomplishment of field measurements have The scientific literature fre- been focused on the formalization of a convenient set of quently addresses the topic of the metadata that consistently document the measurement practical application of spectrora- archive [14], [15]. Recent examples show efforts focusing OUR OPERATIONAL diometric campaigns in provid- on finding both solutions and advancements that enhance WORKFLOW FOR FIELD ing additional information for specific aspects of spectroradiometric field practice; nev- SPECTRORADIOMETRY pre-processing and processing of ertheless, a comprehensive approach that holds together remote sensing images, especially guidelines and best practices in field data acquisition is INCLUDES GUIDELINES when carried out simultaneously still not defined with sufficient maturity [16]. The diverse FOR CAMPAIGN to satellite or airborne hyper/ nature of such topic, which includes numerous as well as PLANNING, multi-spectral sensors overpass heterogeneous issues, requires a solution based on an holis- MEASUREMENT [3], [4]. In particular, they pro- tic approach, with as much standardization as possible, and COLLECTION AND DATA vide an invaluable support to the proper flexibility to ensure that the standard procedures EXPLOITATION. the calibration, processing and can be fruitfully applied in different domains and applica- validation of the remote data tion fields. covering various applications: Our objective here is to provide a proposal for a work- with examples going from atmo- flow that includes the requirements for planning, per- spheric effect correction [5]–[10], to land cover mapping and forming and exploiting spectroradiometric field surveys environmental parameters assessment [11]–[13]. in the context of remote sensing environmental analysis Although sometimes considered as a mere matter of tech- techniques. The formalization of a standard measurement nology, instruments availability and organization, the col- process is beyond the purpose of the present work, but the lection of in situ spectra is not a trivial task [4]. The rigorous structuring of an operational workflow approach is a step collection of in situ measurements has a four-fold objective: forward in this direction. (i) provide reference information for pre-processing and pro- The guidelines here proposed form a complement to the cessing operations (i.e., calibration, atmospheric correction, results of two field campaigns carried out in a very chal- pattern recognition, endmember retrieval) of remote sens- lenging environment and under complex conditions, in ing imagery; (ii) implement spectral libraries for archive and order to show their applicability, weakness, threats and simulation purposes (i.e., testing the capabilities of new sen- usefulness. The key ideas underlying our field data acqui- sors; setting the spectral resolution required for the observa- sition-storage work-flow are that: (a) acquisition planning: tion of particular features); (iii) study the spectral behavior the more focused is the field survey, the more accurate is of different materials and their intrinsic properties, via scal- the retrieval of information from image analysis; (b) storing from proximal up to regional scales, for applications ing data: storing well documented measurements in proper such as: the study of vegetation health and phenology, water data repositories will provide guidelines for planning future color and quality assessment, thematic mapping of lithot- surveys and eventually identifying spectral pseudo-invari- ypes, organic matter distribution, and soil moisture content; ant features at their locations [17]; (c) flagging data: provide iv) avoid the errors due to incorrect calibration and/or atmo- proper metadata, as the relational reference table to the spheric correction, and reduce the risks that the selection of repository of measurements will allow both data providers incorrect reference spectra could invalidate or bias the whole and users to interact with an interoperable database mak- processing of imagery data. ing easy to discover and access data, even through the inter- The radiometric variability on ground depends net; (d) exploiting spectral library: using such approach to on a number of parameters that should be collected carry out spectroradiometric field surveys with the special

38 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

purpose of remote sensors’ calibration/validation has the before, contemporarily and after the campaign, and orga- invaluable advantages of allowing the best target selection nized in the following main categories: 1) Campaign as possible and avoiding redundancy. planning; 2) Measurement collection; 3) Data manage- The following section provides an overview of the work- ment and exploitation. flow we propose, focusing on its three main components: In the demonstration of both the feasibility and the campaign planning, measurement collection and data advantages of our approach to the achievement of estab- exploitation. Section 3 focuses on the application of the lished purposes, we will show its applicability to a real sce- guidelines described in section 2 to a real scenario, outlin- nario (Section 3). In particular, two different field surveys ing the weakness and threats of the approach we followed addressed at the lithological mapping of a wide area (about and the solutions adopted. Concluding remarks are dis- 120 km2) in Northern Italy were carried out in 2010 and cussed in section 4. 2011. The study area is very challenging, because of accessibility, altitude, heterogeneity, vegetation coverage, and 2. FIELD ACTIVITY WORKFLOW PROPOSAL: slope related issues: the target region is within the Dolo- OVERALL APPROACH mites mountain district, and includes the Fassa Valley to Spectroradiometric applications in the field of water qual- the East, and the town of Nova Levante to the West. It is ity monitoring have benefited from the use of the de facto limited by the Kesselkogel mountains to the North and the standard protocol developed within the SeaWIFS project Latemar mount to the South (Figure 2). Here the reliefs are [18], [19], in terms of standardization and interoperability mainly carbonatic in composition (calcite and dolomite within the thematic community interested. Even when not are the most abundant mineral specimens). Basaltic lava fully formalized or supported by sectoral organizations— and related alteration products (e.g., hyaloclastite) occur such as happens for SeaWIFS protocol—a standard work- in places, both at the contact with and as dikes within the flow has to include, according to both the purpose of the carbonates [20], [21] . Even if the real case scenario treated survey and the subsequent data exploitation, some basic in this work is dealing with spectroradiometric activities criteria for: target definition, strategies for target selection, aimed at geolithological mapping, the workflow here measurement collection and management. depicted has to be intended as a practical approach to com- In our research activities, the definition of an opera- plement the remote sensing observations in general pur- tional workflow for spectroradiometric field surveys aims pose applications. Therefore, the principles ruled out here at addressing a four main issues: (i) allow to handle and can be successfully applied to surveys specifically designed reproduce the measurements; (ii) ensure that the deliv- for the accomplishment of, for example: calibration/ ered data are compatible to each other and respond to validation of remote data, forest monitoring and mapping specific requirements (such as, homogeneity, accuracy [22]–[24], agricultural and precision farming practice [25], of the parameters, measurement conditions and units, [26], water quality assessment [27], [28]. type of materials, target characteristics); (iii) enforce the interoperability of data collection and sharing; (iv) allow 2.1. CAMPAIGN PLANNING the best collection of data as possible in order to fruitfully In order to accomplish a suitable field survey for the radio- use them for subsequent remote sensing analysis. To this metric calibration/validation of remote data it is crucial to end, our operational workflow has to include require- have an understanding of the environmental context. The ments for the selection of targets and measurement col- first step of the workflow is the campaign planning, for lection; provide the dataset with proper metadata to fully which the activities can be summarized as to: a) provide characterize targets and querying the database; explore a general definition of the targets to be sampled, accord- the dataset and have a quick overview of the spectral ing to the different purposes of the survey; b) properly properties and variability within the area. characterize a selection of target sites suitable for spec- The general framework of the workflow adopted here troradiometric surveys, based on criteria of abundance, (Figure 1) includes a number of activities to be performed distribution, cardinality, heterogeneity and priority [29];

Proposed Workflow

Research Data Aims Campaign Measurement Data Analysis & Planning Collection Management & Requirements Exploitation

FIGURE 1. Spectroradiometric activities field survey operational workflow scheme.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 39

Legend Trentino-Alto Adige Region Field Measurements (September 2010)

Field Measurements Demo Study Area (September 2011) (Fassa Valley)

Kesselkogel (Catinaccio d’Antermoia) (Catinaccio) Rosengarten Group

Milan Fassa Valley Venice

(Val di Fassa) Italy

Latemar Group (Latemar) Rome

Naples

0 2 4 6 8 10 Km

FIGURE 2. Target area of demonstration case field campaign, located in Northern Italy, Trentino Alto Adige region, along the Fassa Valley Dolomites mountainous area, shown over ASTER imagery (RGB:321), with superimposed points sampled during the two field campaigns described in demonstration study case in section 3.

c) specify a program of the in-field activities for each site in time constrains and, in turn, requires that the selected order to guarantee economy, efficiency and effectiveness; targets are easily accessible, measurable, and distributed d) deal with all the logistics for the progress of the cam- within a relatively small area. paign, including the instruments to deploy in the field, The spectral targets to be measured, in order to prop- the required documents to travel around the area, arrange- erly characterize the dynamic range of radiance within the ments and the involvement of local authorities [30]. For study area, are selected according to their adequacy to the example, surveys addressed to data calibration have to be following criteria: accomplished as close as possible to the sensor overpass. ◗ high spatial homogeneity, compared with the spatial This requirement allows for highly accurate data calibra- resolution of the imagery dataset; in the ideal case, each tion, since each element in the in-situ and remote datasets target covers an area of approximately 33# squared pix- shall be acquired under similar conditions. However, the els in the reference imagery statistical representativeness of the in-situ measurements ◗ representativeness of the dynamic range of radiance in for data calibration purposes is strongly dependent on the the region

40 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

◗ low adjacency effects (targets located at adequate dis- (i.e., atmospheric and illumination conditions, standard tance from other volumetric scattering disturbances, reference used and measurement scheme adopted, calibra- such as in the case of trees, buildings or mountain walls) tion of the instrument, geometry of acquisition). ◗ low slope effects (targets with flat surfaces) According to the general purpose of collecting useful ◗ low temporal variability of the spectral response (targets data in the field for subsequent exploitation of data derived with stable spectral response, not displaying rapid from remote sensing acquisitions, the following criteria changes due to short term dynamics phenomena). have to be fulfilled: Although not easy to accomplish, in natural conditions ◗ The required measurements can be acquired using a in the field, all these requirements contribute to strongly portable spectroradiometer for punctual measurements. reduce the distortion effects that arouse from the use of In order to accomplish the in-situ measurements in remote sensing applications. requirements of studying In the planning stage, a preliminary recognition of the scale effects on spectral the region of interest is suggested. Depending on the data (i.e., from proximal, to MEASUREMENTS HAVE quality of the available data, the analyst has the oppor- regional scale observations), TO BE ACQUIRED WITH tunity of estimating the dynamic range of the radiance an imaging spectrometer can THE SAME MODALITY upwelling from the ground, and possibly identifying be used instead or in combi- AND CHARACTERISTICS desirable targets and their accessibility. In the real case nation with the more com- AND ALL THE ANCILLARY scenario at hand (Section 3), based on the distribution mon version. INFORMATION HAVE TO and areal extension of the outcropping rocks and vegeta- ◗ Standard reference mate- BE DOCUMENTED. tion cover, the dynamic of the spectral data is expected rial have to be measured at to vary between the high albedo carbonatic rocks and least once before each target the low albedo lavas. The wide vegetation cover, with measurement is acquired, in several different specimens, is considered as an inter- order to verify the correct functioning of the instru- mediate albedo target, as well as the artificial reflecting ment and measure the target radiance under similar materials with very flat surfaces and regular geometries. illumination and atmospheric conditions. The stan- In addition, as a common practice in spectroradiomet- dard reference has to be checked in the lab for its radi- ric field surveys, the use of plastic and textile canvases, ance properties before it is deployed in the field. As black and white in color, laid on an horizontal surface a support to the assessment of the atmospheric effect is recommended to sample stable spectral radiance and, in particular, the attenuation due to the optical response extremes. thickness of the atmosphere, the spectral radiance (Lm) A recognition field campaign also contribute to have an of the standard reference material have to be acquired understanding of the inherent characteristics of the ground both in the sun (in the direction of the principal in the sites of interest. A recognition survey is recommended plane), and in the shade. in order to: i) verify the instrument proper functioning; ii) ◗ Due to the variable degree of homogeneity of target identify the main target materials/surfaces thought to be samples, at least three measurements per target must be relevant to the purpose of the project; iii) establish a time acquired in order to retrieve a statistically representative schedule to move from one potential target to another in spectral signature. In case of evident heterogeneity, the order to maintain the synchronism with the sensor over- number of measurements per target will be increased, pass, as much as possible (Section 3 shows detailed activi- accordingly. ties over real case scenario). ◗ In case of wide surfaces are recognized as interesting targets, an acquisition following linear transects or grid 2.2. MEASUREMENT COLLECTION patterns can be eventually planned, for response upscal- Measurement collection refers to tasks that are accom- ing to ground pixel size (Figure 3). plished directly in the field, with the main purpose of ◗ the whole measurement set have to be acquired with acquiring a collection of accurate spectral data that encom- the same modality (i.e., measurement geometry, illu- pass the whole dynamic range of the upwelling radiance mination conditions). However, all the ancillary infor- within the area of interest. Here, we recognize three main mation to the measurements have to be documented aspects to be taken into account: 1) the spectral response (section 2.3). quantities to be measured, according to the specific pur- Figure 3 provide a graphic representation of the possible poses of the campaign and the environmental conditions at spatial sampling to be adopted in field spectral measure- the time of the survey, as well; 2) the statistics of sampling ments (point, grid—regular or random—and transect). (point measurements, regularly spaced measurements, The choice of different sampling strategies depends on the number of repetitions, total number of measurements); 3) characteristics of homogeneity, size and location of the the non-spectral variables to consider, due to their influ- target. When the acquisition of spectral data is difficult ence on the spectral signatures of targets, and their use- due to extreme environmental conditions, point measure- fulness in subsequent data management and exploitation ment is the preferred solution (Fig. 3a). Grid sampling is

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 41

Point Grid Transect

Regular Random

(a) (b1) (b2) (c)

FIGURE 3. Sampling scheme of field measurements: point (a), grid—regular (b1) or random (b2)—and transect (c).

an option for wide homogeneous areas. This approach is model is not fully validated by the scientific community commonly adopted for target such as artificial spectral tar- [14], we strongly believe that this model builds a solid base gets (Figure 3b1) or field plots in agronomic experiment for further standard definition. The evaluation of the meta- (Figure 3b2). Multiple measurements along linear transects data specifications is beyond the purpose of the present is accomplished to account for the spatial variability of the work, but in Section 3 we will demonstrate with a few prac- target. In case of long transects (Figure 3c) the acquisition tical examples the usefulness of its fulfillment. of automatic GPS measurements and digital photography Flagging the measurements with a proper set of meta- for each point can be helpful in the post processing and data has many advantages: analysis of the data. ◗ It allows to quickly display the spectral signature of a specific target sample, and its variability as a function of time, 2.3. DATA MANAGEMENT illumination conditions, homogeneity of the surface. The third thread of the operational workflow is the one ◗ It allows to reproduce the measurements. In principle, dealing with data management activities. With the aim the knowledge of the spatial information, geometry of of providing the in situ measurements with exhaustive acquisition, target definition and characteristics (e.g., documentation, and following a detailed evaluation of target type, name and pictures) will allow scientists to the existing resources, we decided to adopt and include recognize the specific target measured, especially in the in our workflow the metadata model provided by Bojin- case of multiple surveys in the same area (such as in the ski et al. [31]. This model has been already employed in case of multitemporal studies). the context of spectroradiometric field surveys, and imple- ◗ It ensures that the data acquired in the field are com- mented in the SPECCHIO database [14], [15]. Table 1 lists patible to each other and to datasets acquired by dif- the metadata variables borrowed from the SPECCHIO ferent operators at different target sites. In addition, database and here followed in order to fully document the they can be fairly compared with the measurements dataset and assist the data management and exploitation. belonging to different datasets, acquired under similar We added two entries to the original list—namely ‘Hand conditions. This enforces the interoperability of data Sample’ and ‘Reference Inter-calibration’—according to collection and sharing. our fieldwork experience and purpose. Hand samples are ◗ It allows to identify pseudo-invariant features and eval- generally collected in the field in order to archive a sample uate their radiometric spatial uniformity and temporal set for further analysis in the laboratory, especially in the stability. context of a geologic fieldwork. The entry ‘Reference Inter- ◗ It allows to statistically evaluate the effects due to illumi- calibration’ has been included in order to avoid errors in nation conditions and target homogeneity at the close the instrumental calibration due to the presence of dirt on view scale. the reference panel, which can affect the standard reflector ◗ It allows to organize and manage the measurement in the field. results. When suitably filled in, each table entry Although, as mentioned by the authors, the metadata represent a unique variable through which querying definition currently implemented into the SPECCHIO the database. We built the database of measurements

42 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

TABLE 1. LIST OF METADATA VARIABLES ADOPTED FOR THE ACCOMPLISHMENT OF SPECTRORADIOMETRIC FIELD SURVEYS, MODELED ON SPECCHIO DATABASE REQUIREMENTS [31]. GROUP VARIABLE DESCRIPTION General/campaign Campaign name Name of the sampling campaign Campaign description Textual information about the campaign Investigator Person responsible for the campaign File path File system path to the spectral campaign data Spatial and temporal Capturing date and time Date and time of the sampling in UTC information Latitude Spatial sampling position Longitude Altitude

Target information Target homogeneity Homogeneous or heterogeneous Landcover type Based on CORINE land cover (EC DG XI, 1993) Spectrum name Scientific and common names of the target Target type RSL internal designation of target type, e.g., snow, pasture Hand sample ID name of the sample directly collected in the field Pictures Images depicting the target Sampling geometry Sensor zenith angle Measured from nadir, i.e., nadir = 0 Sensor azimut angle Relative to the illumination angle, i.e., 180° for the principal plane opposite of illumination source Sensor distance Distance of the sensor to the target Illumination zenith angle illumination source zenith angle measured from nadir Illumination azimuth angle Measured from geographic North Illumination distance Distance between the illumination source and target

Measurement Number of averaged spectra Number of spectra averaged internally by the instrument details White reference White reference panel used White Inter-calibration Inter-calibration between Lab and Field Reference Sensor Sensor model Instrument Specific instrument identified by a serial number Instrument calibration number Number of the instrument calibration Foreoptic Additional optic than changes the field of view (FOV) in degrees Illumination source Type of illumination source, e.g., sun, Hg lamp Sampling environment Field or laboratory Measurement type Single, directional, temporal Measurement unit Reflectance, digital numbers, radiance, absorbance Goniometer model Name of the goniometer used

Environmental Cloud cover Amount of clouds covering the sky defined in octas conditions Ambient temperature Air temperature in degrees Air pressure Air pressure in hPa Relative humidity Relative humidity as percentage Wind speed Qualitative description of the wind speed: calm, breezy, windy, stormy Wind direction Direction classes in 45° steps, measured from geographic North

File information Auto number Automatic, consecutive number assigned by the spectroradiometer capturing software User comment Comment added by the user Spectral file name Name of the spectral file File format File format of the spectral file Data structuring information Hierarchical structure as gleaned from folder structure

as a file system structure. By using a simple client errors and redundancy. Interactivity allows the user application based on a Graphical User Interface the to perform some distortion removal (as for the water user enters a query into the database, thus starting the absorption bands), and some statistical calculation on information retrieval process. Data management and the dataset (i.e., mean/median calculation of targets information retrieval therefore become very efficient response, intra/inter-variability of spectra, spectral and can be used during the survey in order to con- resampling). Some examples of those possibilities are textually verify the proper data collection and avoid illustrated in Section 3.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 43

3. DEMONSTRATION CASE: FIELD maps), a recognition field campaign was accomplished SPECTRORADIOMETRY OVER GEOLOGICAL TARGETS during June 2010. These preliminary activities helped In the demonstration of the feasibility and advantages us defining a list of desired targets, map their position, of adopting the operational workflow so far described and sort them according to a priority level. In addition, to approach spectroradiometric field surveys, we show we established a minimum number of required samples the results of two surveys, accomplished within a region and the time schedule, in order to be as consistent as pos- between the Kesselkogel, Rosengarten Group, Latemar sible in terms of illumination/atmospheric conditions and Mount and Fassa Valley (Dolomites, Northern Italy, environmental conditions across the measurement set Figure 2), during September 2010 and September 2011 (Table 2). (in the following, campaign A and B, respectively). Both In order to properly acquire a collection of measure- surveys were addressed at sampling the spectral response ments with the aim of the lithological mapping of the area variability and signatures of several carbonatic, siliciclastic through hyperspectral data, we deployed in the field a set and volcanic terrains exposed within the whole region. The of instruments: ultimate goal was to evaluate the reliability of geolithologi- ◗ ASD FieldSpec Pro Full Resolution Spectroradiometer cal thematic mapping through hyperspectral aerial imag- (with optical fiber at 25 deg, contact probe and remote ery within a region where dramatic variations in elevation cosine receptor—integrating sphere), for the measure- occur, and outcrops, especially the volcanic rocks, are par- ment of reflectance and ground solar irradiance tially hidden by the wide vegetation cover. ◗ Sun Photometer EKO MS-120 (368, 500, 675 and An accurate plan of both the campaigns has been 778 nm) for measurements of Aerosol Optical Thick- carried out as a starting task of the workflow (section 2.1). ness (AOD) properties during the campaign, needed After a preliminary study of the area through available for modeling atmospheric conditions and effects to be instruments for remote investigation (e.g., Google included in radiative transfer models used for hyper- Earth, Landsat imagery, and geological and topographical spectral data preprocessing

TABLE 2. LIST OF A NUMBER OF POTENTIAL TARGET SAMPLES, TOGETHER WITH A BRIEF DESCRIPTION OF EACH, THE TIME SCHEDULE OF THE MEASUREMENTS, AND THE LEVEL OF PRIORITY (CAMPAIGN A PLANNING).

POINT PRIORITY SUPPOSED NO. LEVEL TIMETABLE LOCATION TARGET TO BE ACQUIRED POSITION (WGS84) NOTES

1 09:30 Dolomite detritus 46°29'22.78'' N, 11°39'43.57'' E (alluvial fan)

2 09:45 Dolomite detritus 46°29'21.81'' N, 11°39'39.90'' E (shallow landslide) North-western + Simple logistic access , easy HIGH 3 10:00Duron Valley Mixed alluvial detritus 46°29'28.54'' N, 11°39'53.42'' E and fast acquisition target (Duron river bed) 4 10:15 Grass (Prati della Vecchia) 46°29'27.76'' N, 11°39'53.38'' E 5 10:30 Vulcanites detritus 46°29'33.14'' N, 11°39'48.82'' E (shallow landslide) 6 10:40 Black canvas (artificial 46°29'33.74'' N, 11°40'16.36'' E Mid-lower target 1) + Key area of field activity, Duron Valley HIGH with artificial targets and sun 7 11:00(Sun photom- White canvas 46°29'33.45'' N, 11°40'16.94'' E photometer acquisition eter position) (artificial target 2) 8 11:20 Grass (wet, swamp grass) 46°29'31.90'' N, 11°40'17.11'' E 9 11:4 0 Hyaloclastite 46°29'54.47'' N, 11°39'25.49'' E + Large and horizontal (I Frati formation) surfaces, good exposition MEDIUM Duron Pass – 20 minutes’ walk to reach them 10 Depending on Snow 46°29'50.74'' N, 11°38'31.31'' E + Bedrock targets, high schedule altitudes – Complex targets, difficult to 11 Depending on Near to Tires Ladinic Vulcanites 46°29'45.98'' N, 11°38'51.64'' E LOW reach and time consuming; schedule Pass (pedogenic alteration) vertical and almost vertical 12 Depending on Dolomite rock (Denti di 46°29'52.84'' N, 11°38'3.52'' E surfaces for bedrock; snow schedule Terra Rossa) cover not predictable 13 14:00 Asphalt (chairlift car 46°31'6.89'' N, 11°38'18.59'' E + Easy to reach by car; MEDIUM- Corno d'Oro parking area) artificial, stable target HIGH Chairlift ski tow – 20 minutes ride to reach it

44 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

◗ 66m# 2 plastic and 88m# 2 textile canvas (black and to ground circular footprint of around 0.70 m radius. The white in color), as standard reflectors for sampling the measurements have been acquired, according to the charac- spectral radiance response extremes teristics of the target, in two different measurement modes: ◗ Portable Global Positioning System (GPS) Trimble single measure or transect/grid patterns. The standard refer- receiver, equipped with the Italian and European geoid ence has been acquired twice per target, once at the begin- model for altitude derivation, and Geographic Informa- ning and once at the end of each transect/grid pattern, and tion System (GIS) capabilities single measurements, as well. At each target site, the geo- ◗ Additional instrumentation: digital camera, compass, graphic position has been recorded, and the measured sur- barometer, anemometer, thermoradiometer. face has been photographed. In addition, the whole set of The spectroradiometer deployed in the field directly metadata listed in Table 1 has been recorded. measures the spectral radiance in the Visible to Shortwave Infrared range (350–2500 nm), with 2 nm spectral resolu- 3.1. CAMPAIGN A tion, and the reflectance after normalization to a perfectly The campaign A has led to the measurement of about 300 reflective Lambertian standard reference (Spectralon panel, single spectral signatures of targets, and reference panel, 25# 25c m2 in area). Measurements have been acquired wit h at each site, as well as distributed solar irradiance. A total nadiral geometry, at a constant distance of about 1.5 m, and of 43 targets were spectrally measured and documented, 25 deg instrumental FOV (Field Of View), corresponding and a subset of them is shown in Figure 4. Targets were

0.7 0.9 S10 S22 - Fabric Black Canvas Sample 4 - Grass 0.8 S11 S23 - Fabric White Canvas 0.6 Sample 15 - Wet Grass S12 S24 - Plastic Black Canvas Sample 33 - Grass 0.7 S13 S25 - Plastic White Canvas 0.5 Sample 38 - Grass 0.6 Sample 39 - Synthetic 0.4 Grass 0.5

0.3 0.4

Reflectance 0.3 0.2 0.2 0.1 0.1

0.0 0.0

400 600 800 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400 (a) (b) 0.6 0.8

0.7 0.5 0.6 0.4 0.5 Sample 30 - Sand Sample 41 - Asphalt Sample 40 - Sand Sample 36 - Asphalt 0.3 0.4 Sample 43 - Water Sample 31 - Asphalt S19 - Gravel S02 - Dark Soil

Reflectance 0.3 0.2 S32 - Gravel S06 - Dark Soil S42 - Gravel S21 - Hyaloclastite 0.2 0.1 0.1

0.0 0.0

400 600 800 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm) Wavelength (nm) (c) (d)

FIGURE 4. Subset of the total measurements acquired during campaign A, showing the spectral domain investigated and the dynamic range of brightness encompassed. The median spectrum of each represented targets is shown. Wavelength ranges dominated by the influence of the atmosphere has been removed from spectra. Plots are separated in different panels, for clarity. See text for description.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 45

properly distributed between natural and artificial sur- plastic canvas increases within the whole range of wave- faces with the requirements established in the section lengths from measurements acquired around 10:56 and 2.3, and within the range of brightness, as well. As shown 12:10 A.M., regardless of the color. Fabric canvases show an in Figure 4, several surfaces of the same type have been opposite behavior, as a consequence of different thermal measured during the survey, in order to maintain consis- inertia of the constituent materials. In the infrared region, tency and representativeness among data, and to encom- the brightest materials are represented by carbonate-rich pass the time range of the remote observations. Figure 4a gravel (Figure 4c) and sand (Figure 4d). shows some spectra of grass, with variations mainly due Apart from the canvases that are generally recognized as to the moisture content, together with an artificial grass useful targets for describing the brightness extremes within sampled in a volleyball court (dashed line). The artificial a remote acquired image, and therefore are widely used in grass is easily separable from natural grass, especially in spectroradiometric field surveys coupled to remote sensors the near-infrared domain, but it has remarkable similar- overpass, we identified additional interesting targets. An ity with some dark soils and asphalt shown in Figure 4c example is represented by the asphalt that has an almost and 4d, respectively. dark and featureless signature with very slight variations White canvases, both fabric and plastic-made (Figure 4b) within the area investigated (Figure 4d). are responsible of the highest reflectance in the visible Figure 5 shows some results of data exploitation range of wavelengths. Water sample coming from a swim- through the interactive use of the database structured on ming pool site (Figure 4d) also shows high reflectance in metadata collection (section 2.3). The average spectral visible wavelengths, due to the very bright blue tiled pool signatures of a few targets shown in figure 4 are plotted bottom. It is interesting to note that the reflectance of the together with the standard deviation of the mean and

0.8 0.8 S02 - Dark Soil 0.7 0.7 Median Mean 0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3 Reflectance

0.2 0.2

0.1 S42 - Gravel 0.1 Median Mean 0.0 0.0

400 600 800 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400 0.8 0.8 S31 - Asphalt 0.7 0.7 Median Mean 0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3 Reflectance

0.2 0.2

0.1 S11 - Fabric White Canvas 0.1 Median Mean 0.0 0.0

400 600 800 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm) Wavelength (nm)

FIGURE 5. Subset of the total measurements acquired during campaign A, showing the median, mean and standard deviation within the set of the target related measurements. Wavelength ranges dominated by the influence of the atmosphere has been removed from spectra.

46 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

the median spectra. It appears evident that the asphalt 3.2. CAMPAIGN B is a very homogeneous target compared with natural ter- Campaign B was carried out on September 21, 2011, in the rains. This property makes the asphalt a good marker Northern part of the Fassa Valley and along the freeway of low albedo features, although its signature is strictly SS241 between Vigo Di Fassa and Nova Levante, as shown dependent on materials composition and local geogra- in Figure 2. The planning activities allowed to identify a phy. The variability of the spectral response of the fabric- number of desirable targets on the reference map. The sur- made white canvas in the spectral domain investigated vey resulted in about 230 total measurements in the field. (Figure 5) is probably due to the higher sensitivity of high The number of useful targets was 17 and a subset of them is albedo materials to little variations in the illumination shown in Figure 6. conditions, as well as thermal inertia effects. The car- Following the findings of the previous campaign A, and bonatic sand, as well, can be exploited as relatively high taking into account temporal constraints and the number of reflectance target in the infrared region, higher than the people involved in field activities, we decided not to deploy white canvas used as a reference (Figure 4). Nevertheless, the canvases as reference targets in campaign B. Therefore, these targets are available within the area of interest and we acquired measurements on targets having known bright- therefore the chance to find useful reference surfaces is ness properties in the spectral domain considered here, in strictly dependent on the target availability on ground order to span the whole dynamic range of brightness in the and ultimately on the target area. area and investigate variations with time, as well. In the

0.7 0.40 Sample 2 - Grass 0.6 Sample 10 - Grass 0.35

0.30 Sample 7 - Asphalt 0.5 Sample 30 - Asphalt 0.25 Sample 19 - Asphalt 0.4 Sample 39 - Asphalt Sample 36 - Concrete 0.20 0.3 0.15 Reflectance 0.2 0.10

0.1 0.05

0.0 0.00

400 600 800 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400 (a) (b) 0.5 0.8

0.7 Sample 25 - Snow Sample 20 - Snow 0.4 0.6

0.3 0.5 0.4

0.2 0.3 Reflectance

0.2 0.1 Sample 28 - Cobblestone Sample 22 - Gravel Sample 35 - Gravel 0.1 Sample 14 - Gravel Sample 41 - Gravel Sample 38 - Gravel Sample 40 - Gravel 0.0 0.0

400 600 800 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 1000 1200 1400 1600 1800 2000 2200 2400 Wavelength (nm) Wavelength (nm) (c) (d)

FIGURE 6. Subset of the total measurements acquired during campaign B, showing the spectral domain investigated and the dynamic range of brightness encompassed. The median spectrum of each represented targets is shown. Wavelength ranges dominated by the influence of the atmosphere has been removed from spectra. Plots are separated in different panels, for clarity. See text for description.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 47

7 Asphalt 35 3638 42 Gravel Pasture 6 Snow 14 22 40 ST Reference Volley Court 5

4 2 67 912161819 23 26 28 29 30 37 39

Target Homogeneity (Harvey Balls) 3 01 2 3 4 Target Homogeneity Target

56 7 8 9 1

0 1 3 4 5 8 1011 13 15 17 2021 24 25 27 41 010203040 Samples

FIGURE 7. Plot of the Target type versus Homogeneity level (campaign B samples). The target samples are identified via numbering them according to the notation used in the database. Homogeneity levels are depicted with numbers between 0 and 7, with increasing heterogeneity level. Panel on the right shows the homogeneity scale used, represented by Hervey Balls metric: a font code for depicting the homogeneity levels according to their composition of bright and dark features, within the sample set (bottom right panel).

visible region, we identified snow (figure 6d) and asphalt the dataset in order to verify the representativeness of the (figure 6b), as ‘white’ and ‘black’ extremes, respectively. In spectral variability and figure out the total number of sam- the Near infrared region, extremes of brightness are repre- ples collected per type and homogeneity levels. Figure 7 is an sented by snow at the lower level and carbonatic rocks and example of the distribution of the target samples, according sediments (Figure 6c) at the upper level. Grass (figure 6a) to the homogeneity of the measured surfaces and general is always critical because strictly definition of the type of material. Standard reference mate- dependent of the moisture con- rials plot with homogeneity levels 0 and 4, corresponding tent and, in turn, on the weather to the highest homogeneity levels, for bright and dark sur- THROUGH THE PROPOSED conditions during the campaign faces, respectively. The other levels represented have inter- WORKFLOW APPROACH, activities and the days immedi- mediate homogeneity. ISSUES COMMONLY ately before its accomplishment. The data exploitation functionality of the database so ARISING DURING FIELD In addition, the wide spectral far implemented also allows for some advanced evalu- ACTIVITIES ARE MORE variability within the carbonatic ation of the dataset. Figure 8 shows an example of the EASY TO FORESEE, rock inventory and the occur- advanced statistics computed via querying the database TACKLE AND MITIGATE. rence of diagnostic absorption through a simple client application. The Spectral Angle features at wavelengths longer Mapper (SAM) [32] and the Jeffries-Matusita (JM) dis- than the atmospheric influence, tance [33] values provide us with a measure of similar- compromises the use of this ity and separability within the dataset, respectively. Both target as reference brightness, although they still remain the coefficients have been computed using the spectral important for investigating the spectral variability of the reflectance of the samples. Each target sample has been outcropping lithotypes. compared with the others in the dataset, as couples of The collection of metadata together with the spectral spectral classes are commonly compared in the statistical measurements (Section 2.3) allowed us to quickly explore analysis of spectral datasets. Since we acquired multiple

48 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Spectral Angle Mapper Jeffries-Matusita Sample10 Sample14 Sample19 Sample2 Sample20 Sample22 Sample25 Sample28 Sample30 Sample35 Sample36 Sample38 Sample39 Sample40 Sample41 Sample7 Sample10 Sample14 Sample19 Sample2 Sample20 Sample22 Sample25 Sample28 Sample30 Sample35 Sample36 Sample38 Sample39 Sample40 Sample41 Sample7 Sample10 Sample10 Sample14 Sample14 Sample19 Sample19 Sample2 Sample2 Sample20 0.25 Sample20 Sample22 Sample22 1.95 Sample25 0.20 Sample25 1.90 Sample28 0.15 Sample28 Sample30 Sample30 1.85 Sample35 0.10 Sample35 Sample36 Sample36 Sample38 0.05 Sample38 1.80 Sample39 Sample39 Sample40 0.00 Sample40 Sample41 Sample41 Sample7 Sample7 (a) (b)

FIGURE 8. Graphical representation of the SAM (a) and J-M distance (b) values computed for some statistical metric on the dataset. For clarity, only a subset of the dataset collected during campaign B is shown, using a threshold-based color coding highlighting best separation-separability scores. measurements of each single target sample, we can fairly and hyper-spectral remote sensing; first, workflow struc- consider them as a spectral class, identified through its ture and aspects were presented, with three main com- own mean and variance. Values of the SAM metric lower ponents illustrated in section 2: Campaign planning, than 0.10 rad (red pixels in Figure 8a) provide evidence of Measurement collection, and Data management; and good similarity between samples. Values of the JM metric second, a real case scenario, demonstration case of our closer to 2 (red squares in Figure 8b) provide an indica- approach application was described, covering two spec- tion of high separability among samples. troradiometric campaigns performed in mountainous, heterogeneous and environmentally complex areas, for 4. CONCLUSIONS geolithologic mapping purposes. This demonstration The work described in this paper origins form two decades case has shown issues, challenges and solutions adopted, of field spectroradiometry experience of our working group, using our workflow approach as general framework and holding together planning activities, laboratory technol- guideline, exploiting its specific capabilities shown in ogy, and environmental-based assessment of remote sens- Sections 3.2 and 3.3. ing ground practice. Spectroradiometric field surveys, Although the illustrated demo case and relevant les- especially when performed over heterogeneous targets and sons learned here are focusing on field campaign in the within complex environmental contexts, are not subjects context of geolithotypes hyperspectral mapping, the use- of improvisation and need careful organization and struc- fulness of the approach is broader than that. The three- turing. Our work here has focused on offering to scientific step structured workflow described in section 2 is in fact community a proposal for operational workflow that holds a general one, and the guidelines proposed can be fol- together best practices of different aspects of in situ remote lowed for a broad range of spectroradiometric field activi- sensing field data acquisition, into a structured and com- ties, not only geologic ones. Our demonstration case is prehensive approach. Through the proposed framework, taking advantage of most of the capabilities enabled by some issues commonly arising from spectroradiometric the proposed workflow data management, but for other field activities are more easy to foresee, tackle and miti- particular application fields, a number of measurement gate, if necessary. The main target is that of enhancing the features is additionally supported by the approach, as usefulness and sharing/reusability of information coming demonstrated by the operational usage of SPECCHIO- from spectroradiometric field campaigns, in particular as based metadata schema for a diverse set of in situ spectro- ancillary activity in the context of remote sensing environ- radiometric applications. For instance, vegetation studies mental analysis techniques. through in situ spectroradiometry may focus on angular The paper has introduced an operational workflow spectral response of targets, through Bidirectional Reflec- approach that covers field spectroradiometric activities, tance Distribution Function (BRDF) retrieval. This appli- from planning to data analysis, in the context of multi cations need the deployment of ad hoc instrumentation,

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 49

such as goniometers for measuring multiangular sun- to individuate and remove possible bias factors due to sensor acquisition geometries over the same target, and nonoptimal measurement conditions. Field measure- such instrumentation is already supported by our data ments acquired according to our proposed workflow are management schema (see Table 1), which is able to han- well suited to satisfy those needs for high quality stan- dle multiannual datasets and exploit information coming dards, as the demonstration case and lessons learnt pre- from them. Other specific features requested by different sented in the previous section show. in situ targets can be the need for multiscale spatial sam- As a future perspective in spectroscopy applied to pling to take into account pixel and sub-pixel variability of geology mapping, for example, we can foresee that, for land cover in highly heterogeneous areas, as well as con- the next decade, the remote exploration of Earth and tinuous sampling in situ of vegetation and atmospheric other planets (Mars, first of all) as well will be based on parameters for agricultural applications; again, such fea- highly advanced technologies to accomplish: reconnais- tures can efficiently tackled through measurement plan- sance, global mapping and detailed geological-environ- ning and collection comprising specific instrumentation mental analyses at regional to outcrop scales. Therefore, and sampling strategies, and can be stored easily through both predicting the capabilities and information poten- our data management approach, as the level of detail of tial of data from future instruments and establishing the Table 1 metadata fields shows. requirements to design instruments for detailed investi- Spectroradiometric field campaign are performed for a gations are critical issues. In situ studies and multi-scale wide range of domains related to hyperspectral and multi- approach to data collection play a crucial role to face spectral remote sensing techniques and applications, and this challenge. To this end, the definition of operational our operational workflow can be exploited for such targets standards, protocols and workflows, to build compliance and adapted to their particular features. Those remote sens- among data will soon be mandatory. Our approach is one ing domains are briefly summarized as: step forward this, and we hope other proposals will fol- ◗ sensor or instrument calibration, as well as intercalibra- low, thus to build a proper standard protocol for merging tion of multisensor dataset; together the different datasets coming from field spectro- ◗ spectral data radiometric calibration, as well as data radiometric surveys. This will help improving effective- fusion from different sensors/platforms (e.g. optical, ness of remote sensing of land processes, starting from in SAR, Lidar, DTM); situ data quality improvement. ◗ validation of processed spectral data (atmospherically corrected ground reflectance data coming from aerial 5. ACKNOWLEDGMENTS survey are compared to ground reflectance spectra, to The spectroradiometric activities described here are com- assess quality of atmospheric effect correction adopted); ing from the 20 years field activities experience of Institute ◗ spectral data resampling and simulation (sensor for Electromagnetic Sensing of the Environment (CNR- response simulation for new platforms testing and com- IREA), proximal remote sensing group, and in particular mitment, as well as performing remote sensing sensitiv- thanks to the efforts and knowledge of C. Giardino. In ity studies to environmental parameters change). particular, actual field campaign described as demonstra- As a main example of the importance of spectrora- tion case in section 3 have been carried out with the tech- diometric field quality data, we can look at vicarious nical support of M. Bresciani and M. Musanti (CNR-IREA. calibration of satellite or aerial multispectral sensors. Milan section). Vicarious calibration is an important tool for monitoring sensor performances throughout the operation time REFERENCES and allowing high quality data to be acquired. Accord- [1] E. J. Milton, M. E. Schaepman, K. Anderson, M. Kneubühler, and ing to reflectance-based methods, it is possible to esti- N. Fox, “Progress in field spectroscopy,” Remote Sens. Environ., mate the at-sensor radiance over selected test sites on vol. 113, suppl. 1, pp. S92–S109, Sept. 2009. the Earth’s surface starting from in situ measurements [2] M. E. Schaepman, S. L. Ustin, A. J. Plaza, T. H. Painter, J. Verrelst, of the spectral reflectance of ground targets and accurate and S. Liang, “Earth system science related imaging spectros- measurements of the optical depth and other meteoro- copy: An assessment,” Remote Sens. Environ.,vol. 113,suppl. 1, logical parameters, at the time of the sensor overpass. pp. S123–S137, Sept. 2009. The reflectance-based validation provides a common [3] M. Dinguirard and P. N. Slater, “Calibration of space-multispec- approach to monitor and verify the accuracy of both tral imaging sensors: A review,” Remote Sens. Environ., vol. 68, pre-flight calibration and atmospheric correction via no. 3, pp. 194–205, June 1999. radiative transfer model. The level of agreement between [4] A. F. H. Goetz, “Three decades of hyperspectral remote sensing ground-based and airborne hyperspectral data allows for of the Earth: A personal view,” Remote Sens. Environ.,vol. 113, the estimation of data accuracy. Of course, all this needs suppl. 1, pp. S5–S16, Sept. 2009. that the quality of spectroradiometric data acquired in [5] G. Ferrier and N. S. Trahair, “Evaluation of apparent surface situ is as high as possible, and that those data are well reflectance estimation methodologies,” Int. J. Remote Sens., documented in order understand calibration results and vol. 16, no. 12, 1995.

50 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

[6] R. Richter, “Correction of atmospheric and topographic effects [20] N. Preto, M. Franceschi, G. Gattolin, M. Massironi, A. Riva, for high spatial resolution satellite imagery,” Int. J. Remote Sens., P.Gramigna,L.Bertoldi, and S.Nardon,“The Latemar: A vol. 18, no. 5, 1997. Middle Triassic polygonal fault-block platform controlled [7] E. Ben-Dor and N. Levin, “Determination of surface reflectance by synsedimentary tectonics,” Sediment. Geol.,vol. 234, from raw hyperspectral data without simultaneous ground data pp. 1–18, 2011. measurements: A case study of the GER 63-channel sensor data [21] H. Boro, G. Bertotti, and N. J. Hardebol, “Distributed fractur- acquired over Naan, Israel,” Int. J. Remote Sens.,vol. 21,no. 10, ing affecting isolated carbonate platforms, the Latemar Platform 2000. Natural Laboratory (Dolomites, North Italy),” Marine Petroleum [8] K. J. Thome, “Absolute radiometric calibration of Landsat 7 ETM+ Geol., vol. 40, pp. 69–84, 2013. using the reflectance-based method,” Remote Sens. Environ., [22] D. L. Peterson, J. D. Aber, P. A. Matson, D. H. Card, N. Swanberg, vol. 78, nos. 1–2, pp. 27–38, Oct. 2001. C. Wessman, and M. Spanner, “Remote sensing of forest canopy [9] R. Richter and D. Schläpfer, “Geo-atmospheric processing of and leaf biochemical contents,” Remote Sens. Environ.,vol. 24, airborne imaging spectrometry data—Part 2: Atmospheric/ no. 1, pp. 85–108, Feb. 1988. topographic correction,” Int. J. Remote Sens.,vol. 23,no 13,2002. [23] R. A. Jago, M. E. J. Cutler, and P. J. Curran, “Estimating [10] L. Guanter, K. Segl, and H. Kaufmann, “Simulation of optical re- canopy chlorophyll concentration from field and airborne mote-sensing scenes with application to the EnMAP hyperspec- spectra,” Remote Sens. Environ.,vol. 68,no. 3, pp. 217–224,June tral mission,” IEEE Trans. Geosci. Remote Sensing,vol. 47,no. 7, 1999. pp. 2340–2351, July 2009. [24] M. Schlerf, C. Atzberger, and J. Hill, “Remote sensing of [11] P. M. Teillet, G. Fedosejevs, R. P. Gauthier, N. T. O’Neill, K. J. forest biophysical variables using HyMap imaging spectrom- Thome, S. F. Biggar, H. Ripley, and A. Meygret, “A generalized ap- eter data,” Remote Sens. Environ.,vol. 95,no. 2, pp. 177–194, proach to the vicarious calibration of multiple Earth observation Mar. 2005. sensors using hyperspectral data,” Remote Sens. Environ.,vol. 77, [25] M. S. Moran, Y. Inoue, and E. M. Barnes, “Opportunities and no. 3, pp. 304–327, Sept. 2001. limitations for image-based remote sensing in precision crop [12] W. A. Abdou, C. J. Bruegge, M. C. Helmlinger, J. E. Conel, S. H. management,” Remote Sens. Environ., vol. 61, no. 3, pp. 319–346, Pilorz, W. Ledeboer, B. J. Gaitley, and K. J. Thome, “Vicarious Sept. 1997. calibration experiment in support of the multi-angle imaging [26] F. M. Howari, “The use of remote sensing data to extract infor- spectroradiometer,” IEEE Trans. Geosci. Remote Sensing,vol. 40, mation from agricultural land with emphasis on soil salinity,” no. 7, pp. 1500–1511, July 2002. Aust. J. Soil Res., vol. 41, pp. 1243–1253, 2003. [13] A. Brook and E. Ben Dor, “Supervised vicarious calibration [27] V. E. Brando and A. G. Dekker, “Satellite hyperspectral remote (SVC) of hyperspectral remote-sensing data,” Remote Sens. Envi- sensing for estimating estuarine and coastal water quality,” ron., vol. 115, no. 6, pp. 1543–1555, June 2011. IEEE Trans. Geosci. Remote Sensing,vol. 41,no. 6, pp. 1378–1387, [14] A. Hueni and M. Tuohy, “Spectroradiometer data structuring, June 2003. pre-processing and analysis: An IT based approach,” J. Spatial [28] A. N. Tyler, E. Svab, T. Preston, M. Présing, and W. A. Kovács, Sci., vol. 51, no. 2, pp. 93–102, 2006. “Remote sensing of the water quality of shallow lakes: A mix- [15] A. Hueni, J. Nieke, J. Schopfer, M. Kneubühler, and K. I. Itten, ture modelling approach to quantifying phytoplankton in water “The spectral database SPECCHIO for improved long-term characterized by high-suspended sediment,” Int. J. Remote Sens., usability and data sharing,” Comput. Geosci.,vol. 35,no. 3, vol. 27, no. 8, 2006. pp. 557–565, Mar. 2009. [29] J. E. Bartlett, J. W. Kotrlik, and C. C. Higgins, “Organizational re- [16] K.Pfitzner,A.Bollhofer,A.Esparon,R.Bartolo, and search: Determining appropriate sample size in survey research,” G. Staben, “Standardised spectra (400–2500 nm) and associ- Inform. Technol. Learn. Perform. J.,vol. 19,no. 1,2001. ated metadata: An example from northern tropical Australia,” [30] J. Secker, K. Staenz, R. P Gauthier, and P. Budkewitsch, “Vicari- in Proc. 2010 IEEE Int. Geoscience Remote Sensing Symp. (IGARSS), ous calibration of airborne hyperspectral sensors in operational pp. 2311–2314. environments,” Remote Sens. Environ.,vol. 76,no. 1, pp. 81–92, [17] B. Clark, J. Suomalainen, and P. Pellikka, “The selection of Apr. 2001. appropriate spectrally bright pseudo-invariant ground targets [31] S.Bojinski,M.Schaepman,D.Schlaepfer, and K.Itten,“SPEC- for use in empirical line calibration of SPOT satellite imagery,” CHIO: A spectrum database for remote sensing applications,” ISPRS J. Photogramm. Remote Sens., vol. 66, no. 4, pp. 429–445, Comput. Geosci., vol. 29, pp. 27–38, 2003. July 2011. [32] F. A. Kruse, A. B. Lefkoff, J. B. Boardman, K. B. Heidebrecht, [18] C. D. Mobley, “Estimation of the remote-sensing reflectance A. T. Shapiro, P. J. Barloon, and A. F. H. Goetz, “The Spectral from above-surface measurements,” Appl. Opt.,vol. 38,no. 36, Image Processing System (SIPS): Interactive visualization and pp. 7442–7455, 1999. analysis of imaging spectrometer data,” Remote Sens. Environ., [19] R. W. Gould, Jr., R. A. Arnone, and M. Sydor, “Absorption, scat- vol. 44, pp. 145–163, 1993. tering, and, remote-sensing reflectance relationships in coastal [33] J. A. Richards, Remote Sensing Digital Image Analysis.Berlin: waters: Testing a new inversion algorithm,” J. Coast. Res.,vol. 17, Springer-Verlag, 1999, p. 240. no. 2, pp. 328–341, 2001. GRS

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 51

REMOTE SENSING SATELLITES

C. RUF, A. LYONS, University of Michigan, Ann Arbor, MI, USA M. UNWIN, SSTL, Surrey Research Park, Guilford GU2 7YE, UK J. DICKINSON, R. ROSE, Southwest Research Institute, San Antonio, TX, USA D. ROSE, AND M. VINCENT, Southwest Research Institute, Boulder, CO, USA

CYGNSS: Enabling the Future of Hurricane Prediction

1. INTRODUCTION precipitating regimes, as was the case for QuikScat. As urricane track forecasts have improved in accu- a result, in the absence of reconnaissance aircraft, the Hracy by about 50% since 1990, largely as a accuracy of wind speed estimates in the inner core of result of improved mesoscale and synoptic modeling the hurricane is often highly compromised. The added and data assimilation of the remotely sensed back- quality and quantity of surface wind data provided ground environment. In that same period, there has by CYGNSS in precipitating conditions significantly been essentially no improvement in the accuracy of improves estimates of intensity. intensity forecasts due to inadequate modeling and Mesoscale Convective Systems (MCSs) contrib- observing capabilities in the hurricane inner core. The ute more than half of the total rainfall in the tropics inadequacy in observations results from two causes: and serve as the precursors to TCs. Over the ocean, 1) much of the inner core ocean surface is obscured the organization of the fluxes depends on a complex from conventional remote sensing instruments by interaction between surface level winds and storm intense precipitation in the eye wall and inner rain dynamics. Their development and characteristics bands, 2) The rapidly evolving (genesis and intensifi- depend critically on the interaction between ocean cation) stages of the tropical cyclone (TC) life cycle are surface properties, moist atmospheric thermodynam- poorly sampled in time by conventional polar-orbit- ics, radiation, and convective dynamics. ing imagers. CYGNSS (Cyclone Global Navigation Satellite Sys- 2.2. THE VALUE OF FREQUENT tem) is specifically designed to address these two limi- WIND OBSERVATIONS tations by combining the all-weather performance of Most current spaceborne active and passive micro- GNSS bistatic ocean surface scatterometry with the sam- wave instruments are in polar low earth orbit (LEO). pling properties of a constellation of satellites [1], [2]. LEO maximizes global coverage but can result in This article will describe the motivation for using large gaps in the tropics. Schlax et al. (2001) [17] pres- a micro-satellite constellation, the mission design ent a comprehensive analysis of the sampling charac- and deployment module. To view a short video about teristics of conventional polar-orbiting, swath-based CYGNSS, click here: http://www.youtube.com/watch?v imaging systems, including consideration of so-

______=iruxa6F4OJ8&feature=player_embedded. called tandem missions. The study demonstrates that a single, wide-swath, high-resolution scatterometer 2. THE SCIENCE MOTIVATION system cannot resolve synoptic scale spatial detail FOR THE CYGNSS APPROACH everywhere on the globe, and in particular not in the tropics. The irregular and infrequent revisit times (ca. 2.1. THE VALUE OF WIND OBSERVATIONS 11–35 hrs) are likewise not sufficient to resolve syn- IN PRECIPITATING CONDITIONS optic scale temporal variability. As a striking exam- Previous spaceborne measurements of ocean surface ple, Figure 1 shows the percentage of time that the vector winds have suffered from degradation in highly core of every tropical depression, storm and cyclone from the 2007 Atlantic and Pacific seasons was successfully imaged by QuikScat or ASCAT. Missed core Digital Object Identifier 10.1109/MGRS.2013.2260911 Date of publication: 26 June 2013 imaging events can occur when an organized system

2007 Atlantic Hurricane Season 100 % of Storm Center Hit by ASCAT Swaths 90 % of Storm Center Hit by QuikScat Swaths 80

50 [%]

0 ET Andrea ET Chantal TS Erin TS Gabrielle TS Ingrid TS Jerry HU Lorenzo TD Fifteen Overall (5/8–5/14) (7/31–8/05) (8/13–8/18) (9/8–9/11) (9/12–9/18) (9/23–9/24) (9/25–9/29) (10/10–10/12) ET Barry HU Dean HU Felix HU Humberto TD Ten HU Karen TS Melissa HU Noel (5/31–8/5) (8/13–8/23) (8/28–9/5) (9/12–9/13) (9/21–9/22) (9/23–9/29) (9/28–10/5) (10/28–11/02) (a)

2007 Eastern Pacific Hurricane Season 100 % of Storm Center Hit by ASCAT Swaths 90 % of Storm Center Hit by QuikScat Swaths 80

50 [%]

0 TS Alvin TD Three-E TS Five-E TS Dalila HU Flossie HU Henriette TD Thirteen-E TS Kiko Overall (5/27–5/6) (5/11–5/15) (7/14–7/15) (7/22–7/30) (8/3–8/13) (8/30–9/5) (9/20–9/25) (10/15–10/27) TS Barbara TD Four-E HU Cosmo TS Erick TS Gil HU Ivo TS Juliette (5/29–6/2) (7/9–7/11) (7/14–7/24) (7/31–8/2) (8/29–9/3) (9/16–9/24) (9/29–10/5) (b)

FIGURE 1. Percentage of time the center of named storms was observed with either QuikScat (blue) or ASCAT (orange) polar-orbiting scat- terometers during the 2007 Atlantic (a) and Pacific (b) hurricane season. Poor performance results from the coverage gaps and infrequent revisit times that are characteristic of polar-orbiting wide-swath imagers. passes through an imager’s coverage gap or when its 2.3. MEASUREMENT METHODOLOGY motion is appropriately offset from the motion of the Figure 2 illustrates the propagation and scattering geom- imager’s swath. The figure highlights the many cases in etries associated with the GNSS approach to ocean surface which TCs are resolved much less than half the time. scatterometry. The direct GPS signal provides a coherent ref- One particularly egregious case is Hurricane Dean, erence for the coded GPS transmit signal. It is received by an which was sampled less than 5% of the time possible RHCP receive antenna on the zenith side of the spacecraft. by ASCAT. The quasi-specular forward scattered signal from the ocean

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 53

450 GPS Sat.

500 Direct Signal

RHCP 550

600 x = x LHCP n s c

Specular Point CA Code Chip 650 x = 0 Glistening Zone Scattered 700 Signal Annulus 750 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 # 4 Doppler Hz 10

(a) (b)

FIGURE 2. GPS signal propagation and scattering geometries for ocean surface bistatic quasi-specular scatterometry. (b) Spatial distribution of the ocean surface scattering measured by the UK-DMC-1 demonstration spaceborne mission—referred to as the Delay Doppler Map (DDM) [5].

surface is received by a downward looking, LHCP antenna imposes more relaxed requirements on instrument calibra- on the nadir side of the spacecraft. The scattered signal tion and stability than does the former. However, it derives contains detailed information about its roughness statis- its wind speed estimate from a wider region of the ocean tics, from which local wind speed can be derived [3]. The surface and so necessarily has poorer spatial resolution. scattering cross-section image produced by the UK-DMC-1 Development of wind speed retrieval algorithms from demonstration spaceborne mission is shown in Fig. 2. Vari- DDMs is an active area of research [5]. able lag correlation and Doppler shift, the two coordinates of the image, enable the spatial distribution of the scatter- 2.4. EXAMPLE OF SCIENCE COVERAGE ing cross section to be resolved [4], [5]. This type of scatter- A time-lapse simulation comparing CYGNSS and ASCAT ing image is referred to as a Delay Doppler Map (DDM). coverage of Hurricane Frances just before its U.S. landfall Estimation of the ocean surface roughness and near-surface is shown in Fig. 3. The simulation was created by project- wind speed is possible from two properties of the DDM. ing satellite coverage predictions for each mission onto The maximum scattering cross-section (the dark red region the archival storm track record for Frances. Each frame in Fig. 2) can be related to roughness and wind speed. This represents all samples taken within a 3-hour interval. The requires absolute calibration of the DDM. Wind speed can TC inner core is shown as a large blue dot in each frame. also be estimated from a relatively calibrated DDM by the ASCAT, with its relatively narrow swath width, only infre- shape of the scattering arc (the red and yellow regions in quently samples the inner core, whereas the much wider Fig. 2). The arc represents the departure of the actual bi- and more dispersed effective swath of the CYGNSS con- static scattering from the purely specular case that would stellation allows for much more frequent sampling. The correspond to a perfectly flat ocean surface, which appear average revisit time for TC sampling is predicted to be 4.0 in the DDM as a single point scatterer. The latter approach hour, and the median revisit time will be 1.5 hour.

CYGNSS Time Lapse Simulations Comparing CYGNSS and ASCAT Coverage of Hurricane Frances

ASCAT Time (3 hr Increments Beginning at 00:00 Z on 2 Sept. 2004)

FIGURE 3. Time lapse simulation comparing the spatial and temporal sampling properties of CYGNSS and ASCAT, if they had both been in orbit during the Hurricane Frances U.S. landfall on 2 Sept. 2004.

54 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

3. MISSION DESIGN 3.2. GPS REFLECTOMETRY AND UK-DMC For some years, GPS receivers have been used to provide 3.1. MICROSATELLITE OBSERVATORIES position, velocity and time knowledge to satellite platforms Each CYGNSS Observatory consists of a microsatellite in low Earth orbit in a similar way to ground-based satellite (microsat) platform hosting a GPS receiver modified to navigation receivers. measure surface reflected signals. Similar GPS-based In addition to navigation, GPS signals have also been instruments have been demonstrated on both airborne increasingly used for remote sensing. Signals at L-band with and spaceborne platforms to retrieve wind speeds as high a 2–20 MHz bandwidth are being broadcast globally from as 60 m/s (a Category 4 hurricane) through all levels of a 20,000 km altitude and can be used to measure, amongst precipitation, including the intense levels experienced in other things, tectonic plate motion and ionospheric and a TC eyewall [1]. tropospheric parameters. Furthermore, signals from other Each Observatory simultaneously tracks scattered GNSS are becoming available, and there will soon be more signals from up to four independent transmitters in the than 120 signal sources in space. operational GPS network. The number of Observatories Spaceborne GNSS Reflectometry uses GNSS signals that and orbit inclination are chosen to optimize the TC sam- have been scattered by the Earth surface to measure geo- pling properties. The result is a dense cross-hatch of sample physical parameters. The potential for GNSS Reflectometry points on the ground that cover the critical latitude band was demonstrated by the UK-DMC (British National Space between !35°. Centre Satellite) mission in 2003. The mission included a The Observatory is based on a single-string hardware GNSS-R sensor with a nadir-pointing antenna (gain just architecture (Fig. 4) with functional and selective redun- under 12 dBiC, 3 dB field of view approximately 20° # 70°) dancy included in critical areas. The microsatellite has permitting collection of as many as three reflected signals been designed for ease of manufacture, integration, and simultaneously. The primary mode of operation on the test to provide a low-risk, cost-effective solution across first experiment was the collection of sampled IF data into the constellation. a data-recorder, typically 20 seconds, and downloading for

OD Ant DDMI Centaur CDS

LNA Pri Pwr s3"#30ARC 8) 3 "AND8#62 LNA s.ANOSAT$ATA RS422 Storage DMR LV 4X"2& 2X"0& LNA s#4)& Pwr Nadir LV s,V)#4 sPaYLOAD)& Hybrid Ant (2) LNA Pwr Diplexer s3CI$ATA Coupler Compression RS422 s3CI$ATA3TORAGE sTEMP3ENSOR)& ADCS s!$#3)& SMT s3URFACE&INISHES s-,) Horizon RS422 Temperature s Sensors LV Pwr s2ADIATOR LV Pwr LVPS+ Sensors Dist RS422 s,Ow Voltage Power Magnetometer Generation Heaters 3C LV (3-Axis) sLV Pwr Dist )NCLUDES)NSTRument StrUCTURE Pwr SuRVIVal Htrs)

Momentum PPT Wheel LV Pwr s"ATTERy Chg Cnti "ATTERy 3 Ahr Li- EPS s0Ri Pwr Dist )ON s%033AFe Hold #3 Torquerods (3) s3!$EPLOYMENT!Ct Solar Array DrIVers Panels Pri Pwr sTEMP(EATER)& Deploy!CTUATORS Dist sTR DrIVers Modified No Mods

FIGURE 4. CYGNSS single-string architecture.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 55

Delay Doppler Map 150

200 376 378 250 380 382 300 384 CA Code Chip*10

CA Code Chips 386 350 388

6000 7000 8000 9000 10000 11000 9000 8000 7000 6000 5000 4000 3000 Doppler Hz 10000 ------Doppler Hz (a) (b)

FIGURE 5. Example delay-Doppler maps from UK-DMC GPS-R experiment. (a) Ocean reflection, (b) sea ice/water reflection.

post processing on the ground. The raw data were processed might be handled by three separate units on previous on the ground into DDMs using software receiver tech- spacecraft. niques to allow analysis of signal returns off ocean, land a) It performs all the core functions of a space GNSS and ice. Two example DDMs are shown in Fig. 5; they mea- receiver, with front-ends supporting up to 8 single or 4 sure the spread in energy away from the specular point, and dual frequency antenna ports. the spread grows as the surface becomes rougher. b) It is able to store a quantity of raw sampled data from A substantial effort into the modeling of signal returns multiple front-ends or processed data in its 1 GByte solid has been undertaken using data from the first UK-DMC state data recorder experiment with the intention to assess inversion of sea c) It has a dedicated reprogrammable FPGA co-processor state parameters [4], [5] and the retrieval of directional (Virtex 4). roughness information [7], [8]. Although severely band- The co-processor was specifically included for the limited, the collection of reflected Galileo signals (from real-time processing of the raw reflected GNSS data into GIOVE-A) was also demonstrated. Moreover, the col- DDMs. However, it has flexibility to be programmed in lection of signals over mixed sea and ice indicates the orbit as required for different purposes, for example to potential of GNSS reflectometry for ice edge mapping [9]. track new GNSS signals, or to apply spectral analysis to The UK-DMC experiment demonstrated the feasibility received signals. for many remote sensing applications but limited space- For the co-processor to generate DDMs of the sampled based data is available for robust assessment of the geo- reflected data, it needs to be primed with the PRN (pseudo- physical retrieval accuracy of GNSS-R. random noise) code of the transmitting GPS satellite, and the estimated time delay and Doppler of the reflection as 3.3. THE SPACE GNSS RECEIVER—REMOTE seen from the satellite. These are calculated by the proces- SENSING INSTRUMENT (SGR-ReSI) sor in conjunction with the main navigation solution—the AND DELAY DOPPLER MAPS data flow for this is shown in Fig. 7. Direct signals (received The UK-DMC experiment demonstrated that a microsat- by the zenith antenna) are used to acquire and track GNSS ellite-compatible passive instrument was able to make signals. From the broadcast ephemerides, the GNSS satel- scientifically relevant geophysical measurements using lite positions are known. Then, from the geometry of the GPS reflectometry. position of the transmit and receive satellites, the reflecto- Satellite Technology Ltd. (SSTL) teamed with the metion geometry can be calculated. National Oceanographic Centre in Southampton and other The processing of the Delay Doppler Map is per- partners to develop a new GNSS-R instrument for this pur- formed on the coprocessor using data directly sampled pose, the Space GNSS Receiver—Remote Sensing Instrument from the nadir antenna. In common with a standard (SGR-ReSI). GNSS receiver, the local PRN is generated on-board the A schematic of the SGR-ReSI [10] is shown in co-processor. As an alternative to synchronizing and Fig. 6. The SGR-ReSI in effect fulfils in one module what decoding the reflected signal in a standalone manner, the

56 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

direct signals can be used to feed the navigation data sense, and assist the Dual Freq Antennas synchronization. The sampled data LNAs is multiplied by a replica carrier and fed into a matrix that performs an RF F/E RF F/E RF F/E RF F/E FFT on a row-by-row basis to form L1 L1 L1/L2 L1/L2 the DDM, to achieve in effect a 7000 channel correlator, integrating over 1 Program GNSS Receiver Reflectometry millisecond. Each point is then accu- SRAM Core Processing Unit (1MB TMR) Interlink mulated incoherently over hundreds Flash FPGA Reprogrammable of milliseconds to bring the weak sig- Soft Core Processor Co-Processor nals out of the noise. and GNSS Correlators 16.367 SRAM DSP FPGA FLASH MHz This processing is performed in (16MB) real-time on-board the satellite, which greatly reduces the quantity of data Interfaces: required to be stored and for the satel- DDR2 RAM RS422/LVDS/ Power Supply Data Recorder lite’s downlink. CYGNSS plans to use CAN the SGR-ReSI primarily in an autono- 16-40V mous manner generating DDMs at a low data rate continuously, which will provide gap-free measurements of the FIGURE 6. GNSS reflectometry instrument configuration. ocean roughness throughout the tropical oceans. material stress levels with a first mode natural frequency of 3.5. MICROSATELLITE STRUCTURE 211 Hz in the launch configuration, avoiding harmonic cou- The microsat structure requirements are driven by physi- pling with the LV natural frequency of 75 Hz during launch. cal accommodation of the DDMI antennas, the S/As, and launch configuration constraints. Our design uses 3.6. MICROSATELLITE MECHANISMS the same principles as our heritage avionics chassis, Observatory mechanisms are limited to heritage S/A using milled Al piece parts bolted together to provide an deployment devices. The four “z-fold” S/A panels perform integrated, mass efficient solution for CYGNSS. Close a one-time deployment into a permanently locked position tolerance pins/holes ensure repeatability of structural planar with the fixed center panels. The S/As are held in alignment. The microsat’s shape is specifically config- place for launch using a cup/cone interface and deployed by ured to allow clear nadir and zenith FOV for the DDMI a combination of flight-proven TiNi Aerospace Frangibolt antennas, while its structure integrates the microsat and actuators and Sierra Nevada Corp. S/A single-axis, locking, instrument electronic boards directly by creating avion- spring-loaded hinges. ics and Delay Mapping Receiver (DMR) “bays.” The avionics and DMR bays form the core of the microsat; all 3.7. MICROSATELLITE THERMAL other components are mounted to this backbone with The CYGNSS Observatory thermal design meets require- structural extensions included to accommodate the Al ments to maintain all components within their tempera- honeycomb-based S/As and DDMI nadir antenna assem- ture limits during all operational modes by using Southwest blies. The structural configuration allows easy access to all Observatory components when the nadir DDMI antenna panel assemblies and micro- Zenith Antenna Direct Signal Extract sat endplates are removed for Obser- Raw Samples Acquisition Ephemerides vatory AI&T. Direct Signal Navigation The microsat primary structure’s Tracking Solution nadir baseplate is the DM mechani- Reflection cal interface for launch. Primary shear Reflection Geometry Tracking and axial loads are carried by the Calculation microsat primary structure, providing full compliance with the dynamic Nadir Antenna Delay Doppler Raw Samples Map launch vehicle envelope. Preliminary FEA of the Observatory results predict launch loads are well within allowable FIGURE 7. GNSS reflectometry dataflow.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 57

Research Institute’s flight-proven, well-characterized, ther- included in the battery assembly. Battery performance mal design techniques. The thermal control design pro- models were used to analyze the CYGNSS mission with vides thermal stability and minimizes thermal gradients predicted EOL nominal battery state-of-charge being through an integrated design of multilayer insulation blan- 87.6%. Battery charging uses a constant current, voltage- kets (MLI), surface treatments, and localized radiators. The temperature limited charge scheme based on four stored arrangement of internal equipment is used to aid thermal profiles matched to the CYGNSS battery. Charging is also control and eliminate the need for supplemental heaters Coulomb limited to 120% of discharge level. The primary except for Standby/Safe Hold operations. power bus voltage is modulated to maintain charge cur- Results from our thermal analysis were used to size the rent and termination voltage. The Coulombic charge limit thermal radiators (EOL). The primary radiator is located on is tracked with an A-min integrator and when the level

zenith surface in the S/A gap along the Observatory center exceeds 1.2 # IdisTeclipse (Amin), battery charging levels are line, with a second radiator on the nadir baseplate. These reduced to C/100. locations are chosen to provide a direct, cohesive thermal conductive path to the primary observatory dissipative PEAK POWER TRACKER loads. The radiators are coated with 5 mil ITO/Tef/Ag, while Battery charge regulation for the CYGNSS EPS is a peak MLI is used on non-radiating external surfaces. power tracking (PPT) type regulator. The PPT board, developed using SwRI internal funds, matches S/A con- 4. AVIONICS ductance to the Observatory load through pulse-width The CYGNSS Avionics consists of four boards, portions modulation (PWM) using an optimization control circuit of the EPS and CDS. The boards include the Peak Power that integrates S/A W-sec over a preset period of time. The Tracker, the Low Voltage Power Supply, the Centaur single PPT includes a ground support equipment (GSE) inter- board computer, and the Flexible Communication Plat- face that serves as the connection point for ground power form radio. A block diagram of the avionics unit is shown in and battery maintenance, conditioning, and pre-launch Figure 9. The avionics unit does not include a box; instead, trickle charging. the microsat structure itself provides mechanical mounting The PPT unit is based on a 40W DC-DC converter, and electrical interconnects over a backplane and cables. which produces 28!4 Vdc from a solar array voltage of 36 to 72 Vdc. The design was produced with multiple missions in 4.1. ELECTRICAL POWER SUBSYSTEM mind, from a long-duration, intense radiation environment The EPS design performance provides robust margins on to a short, LEO mission, CYGNSS being toward the latter of all requirements. The EPS is designed to perform battery these two extremes. The DC-DC converter output voltage is charging without interrupting science data acquisition. modulated by the PPT and battery charge regulator to meet load power and battery charging demands. Power from SOLAR ARRAY the solar array flows into the PPT through an over current The EPS is based on a 28!4 Vdc primary power bus with protection fuse, current sense resistor and EMI filter. S/A electrical power generated by an 8-panel rigid solar array current and voltage are sensed and conditioned before con- (S/A). The S/A design is composed of solar panels, hinges, nection to an analog multiplier within the PPT circuit. The and deployment actuators. Four of the eight panels are analog multiplier converts these signals into instantaneous “z-folded” for launch. Flight-qualified, triple-junction solar S/A power, which is processed by the PPT watt-second inte- cells are arranged with an 84% packing density on the solar grator to track the power peak. The PPT circuit generates an panel substrates, including cover glass to improve their ther- error signal (PPT Error), which is used to provide supervi- mal performance and ground handling robustness. The sory control of the DC-DC converter in conjunction with 0.71 m2 total area S/A provides a 30.3% margin during max the battery charge regulator. eclipse periods (35.8 min). Full mission duration simula- Housekeeping power is provided by a high input voltage tions were performed to analyze worse case solar Beta cases linear regulator, which provides +16 Vdc for control circuit (58). The design provides 43.4% margin during these peri- power and midpoint bias of +8 Vdc to operate single supply ods. When stowed, the z-fold design of the S/A allows the operational amplifiers. solar cells to face outward, combining with the two supple- Battery charge regulation consists of programmable mental ram/wake S/As to power the microsat indefinitely in charge current and end-of-charge voltage settings, which Standby mode before S/A deployment (22% margin). are each controlled via opto-isolated 4-bit interfaces. The opto-isolators are set up for 3.3 Vdc CMOS drive BATTERIES levels from the Centaur interface. No flight software is Electrical power storage for eclipse operations is provided required for the control electronics, except for configura- by two 1.5 A-hr Li-ion 8s1p batteries connected directly tion control. to the primary power bus. The batteries are configured The PPT is also used to switch +28 Vdc bus voltage to for 3 A-hr (EOL) at 28.8 Vdc nominal. Temperature sen- spacecraft components, including the S/A deployment actu- sors, and bypass diodes (to withstand a failed cell) are ators, the DMMI, heaters, and momentum wheels.

58 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Centaur 3U

- Horizon Sensor 2 RS422 SPARC V8 - Timers BiL Momentum Wheels RS422 CPU WDOG Magnetometer RS422 - AX2000 FPGA EEPROM Mag Temp 12 ADC - Ana SRAM DMMI RS422 CCSDS Fmtr SDRAM LVs FCP LVDS Level-0 BiL

Low Voltage Power Supply (LVPS) 3U BiL Horizon Sensor +12V SW - Heaters 5 +5V SW - - Torque Rods 3 +5V SW - - Magnometer 2 +12V LVs - + Momentum Wheels 5V +3.3V DC-DC Converter +3.3V Soft +28V +5V DC-DC Converter +5V +12V DC-DC Converter +12V

+28V

Peak Power Tracker (PPT) 3U

S/C Bus DC-DC Converter DMMI +28V SW - BiL S/A Actuators 2 +28V SW - Charger Control - Battery 2 Battery Charger EMI Filter - LVs

PWM Interface Structural and Electrical Backplane Solar Array Bus PPT PPT Setpoint

Analog Ana Monitors

Flexible Communication Platform (FCP) 3U

Centaur LVDS Master Scrubber BiL FPGA FPGA Onboard 5V Antenna Programmable RF Flash Front-End SDRAM PROM

FIGURE 9. Block diagram of the Avionics Unit.

LOW VOLTAGE POWER SUPPLY CYGNSS design is based heavily on the Juno JADE LVPS, Low voltages required by the avionics boards as well as which was tailored specifically to lower power, embedded switched low voltages for several ADCS components are use, making it ideal for microsatellite missions. SwRI has generated by the Low Voltage Power Supply (LVPS). The produced LVPSs for Orbital Express, Deep Impact, Kepler,

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 59

WISE, and DoD flight missions. The board receives +28 on the WISE Mission Unique Board, which produced Vdc from the PPT and regulates low voltages, including CCSDS Telemetry TM Source Packets and Transfer +/-12 Vdc, +5 Vdc, and +3.3 Vdc, for use by the Cen- Frames with Reed-Solomon Codeblocks (E-16, I=5). taur, FCP, and PPT control circuitry. Further, the board The command algorithms are identical to those used includes low voltage switches (+5 Vdc) to power ADCS on Deep Impact, Orbital Express, Kepler, and WISE, components, including the magnetometer, momentum which produce CCSDS TC Transfer Frames with Vit- wheel, and horizon sensor. erbi (rate 1=2) encoding. The CCSDS File Delivery Proocol (CFDP) File Protocol for hardware accel- 4.2. COMMUNICATIONS AND DATA SUBSYSTEM eration of CFDP Protocol Data Units (PDUs) is used, Most of the hardware to implement the CDS resides within leveraging designs from MMS. Further, Level 0 telem- the CYGNSS avionics bay, with the exceptions of the S-band etry and commanded resets are generated by the CTC antennas, diplexer, and hybrid. without required intervention from the processor. In this manner, the ground station can reset the space- CENTAUR craft even with the processor in a non-responsive state. All on-board microsat processing is performed on SwRI’s ◗ CCSDS Command and Telemetry Circuitry: The Centaur board. The Centaur consists of our space-qualified CCSDS command and telemetry circuitry includes heritage Atmel SPARC8 processor with heritage CCSDS ADCs (Analog to Digital Converters), RS422 command compliant command and telemetry interface, instrument interfaces, and power switches, controlled by the data interface, and ADCS interface designs. The board Centaur but housed on the Peak Power Tracker. All architecture is based on the Juno JADE IPB (launched Aug components utilize the same circuitry as the Command 2011) and extensively reuses the command and telemetry and Telemetry Boards on Deep Impact, Orbital Express, circuitry from Deep Impact, Orbital Express, Kepler, and Kepler, and WISE. WISE only requiring a board relay out for CYGNSS. This ◗ General Purpose Interfaces: The Centaur design includes board was designed to be a very low power embedded LVDS, RS422, analog, and discrete (low-level) inter- microcontroller and was also designed with multiple mis- faces. The CYGNSS ADCS and DDMI are compatible sion requirements in mind. The CYGNSS version of the with these interfaces and do not require Centaur modi- board will be tailored to a LEO radiation environment, pro- fication to accommodate. viding a dense non-volatile memory, and ample interfaces to ADCS, CDS, DDMI, and thermal components through- 4.3. CDS FLIGHT SOFTWARE out the observatory. The simple operational nature of the DDMI and science The Centaur provides the following functionality: profile allows the CDS flight software to be designed for ◗ Processor: The LEON3 ASIC is the spacecraft computer, autonomous control during all normal science and commu- which provides all resources for on-board microsat nication operation using only on-board Level 0 command flight soft-ware processing. The LEON3 dual-core pro- capabilities of the Centaur, stored command sequences, cessor, successor to the LEON2 core, utilizes a 7-stage and CCSDS File Delivery Protocol (CFDP) processes. CDS pipeline, 8 register windows, a 4#4 kBytei-cache and flight software refers specifically to the portion of software d-cache, branch prediction, hardware multiply/divide, dealing with data upload and downlink, including com- and hardware watchdogs. It interfaces to EDAC-pro- mand upload, parsing, telemetry generation, and transmis- tected memories, including MRAM, SDRAM, and Flash. sion. Engineering operations require standard command External interfaces include multiple SpaceWire, 1553, services provided by our hardware-based heritage designs CAN, Ethernet, and UART ports. located on the Centaur. Command services include COP-0 ◗ Processor Support Circuitry: The processor requires uplink command processing with BCH error detect and additional parts, including memories, clock, reset, and correction. The Centaur also provides FSW-independent power management, and interface drivers. Memories execution of a Level-0 command set used for ground-based include MRAM for code storage, SDRAM for code execu- fault management. All other commands are passed to the tion, and Flash memory for data storage. The radiation FSW Command Manager for execution or to the Stored tested Flash parts are being used on MMS. Command Sequence Manager as onboard absolute and ◗ CCSDS Command and Telemetry Core (CTC) (Heri- relative time sequences. tage HDL in FPGA): Resident in the Centaur FPGA, The FSW Telemetry Manager provides collection and the CTC autonomously receives and routes ground high-level formatting of housekeeping data. These data commands from the transceiver, assembles and pack- are either downlinked in real-time or passed to the FSW etizes science data, and autonomously collects and Storage Manager to be stored for later downlink. The Stor- formats housekeeping telemetry for transmission to age Manager software controls data acquisition, record- the transceiver, significantly reducing flight software ing, and playback of housekeeping and science data processing loads. The telemetry algorithms to perform using the 4 GB on-board memory for data storage. The the CCSDS packetization are identical to those used heritage 4 GB Flash memory data store allows for >10

60 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

days of continuous science operations without downlink, and effectively tailor the hardware to the required appli- providing significant margin for contingency operations. cation, being sensitive to resource constraints such as on- A heritage hardware formatter from Orbital Express and board FSW processing, mass, and power. WISE forms CCSDS source packets into transfer frames and supports four separate Virtual Channel (VC) buffers ANTENNAS to enable optimized data routing and processing within The S-Band Microstrip Patch Antenna has a hemispheri- the CYGNSS Ground Data System. These channels have cal gain pattern, with a 0 dBiC gain drop out to 60 off the been designated as real-time housekeeping, stored sci- boresight. These characteristics make it ideally suited to the ence data, stored housekeeping data, and Level 0 house- design of the CYGNSS Observatories. The CYGNSS obser- keeping data. CFDP is used for reliable delivery of stored vatories will use 2 of these antennas, one on the nadir sur- data across the spacelink. face of the vehicle and one on the zenith surface to provide near 4r steradian coverage to allow communications from FLEXIBLE COMMUNICATION PLATFORM all attitudes. S-band communication links are provided to uplink command sets and downlink science and housekeeping data. 5. ATTITUDE DETERMINATION These links use two fixed omnidirectional micro-strip AND CONTROL SUBSYSTEM patch antennas, one on the nadir baseplate and one on the The CYGNSS ADCS enables a standard nadir-pointing, zenith panel, to provide near 4r steradian communications 3-axis, momentum-bias design derived from the Heat without interrupting science operations. Normal commu- Capacity Mapping Mission. CYGNSS is able to take advan- nications use the nadir antenna, while the zenith antenna tage of entirely off-the-shelf ADCS components, using is provided for anomalous pointing. pitch/roll horizon sensors and a 3-axis magnetometer for The S-band transceiver, or Flexible Communication attitude determination; a pitch momentum wheel and Platform (FCP), is a single card communication solution 3-axis torque rods provide attitude control (torque rods developed by SwRI to provide a low-cost, radiation-toler- also provide momentum wheel desaturation). The only atti- ant, software defined radio system. The FCP was designed tude “maneuver” required by CYGNSS is to recover from with flexibility in mind, compatible with either an on- deployment modulation separation tipoff rates and estab- board analog front end or a highly radiation tolerant lish a nadir-pointing configuration, allowing an extremely front end, and is configured to provide S-band (2 GHz) simple mode flow. communications. The FCP provides O-QPSK encoded All CYGNSS ADCS components are COTS units with transmit data at 1.25 Mbps (up to 5 Mbps) with an FSK high technology readiness level (TRL), helping to mini- uplink receiver supporting data rates to 64 kbps. The FCP mize non-recurring engineering (NRE) costs while pro- was developed in 2010 with internal research funds to viding reliability and functionality assurance. The 30 support small spacecraft platforms and forms the basis of mNm-sec nominal momentum wheel was flown on SwRIs recent System F6 wireless communication system CanX-2, launched in April 2008, and AISSAT-1, launched for DARPA. F6 utilizes a variant of the FCP as an intra- in July 2010. The momentum wheels are still fully opera- constellation satellite communication link. Functions of tion on both missions. The torque rods are 1 Am2 units, the FCP are listed below. which have successfully flown on the JAXA led FedSat and Micro-LabSat missions. The magnetometer is a three- SOFTWARE DEFINED RADIO CORE (FPGA) axis smart digital magnetometer to detect the strength The SDR FPGA on the FCP is responsible for the modulation, and direction of an incident magnetic field. The three demodulation, and functional control of the transceiver. It magneto-resistive sensors are oriented in orthogonal receives and transmits raw telemetry and command data directions to measure the X, Y and Z vector components (respectively) from the Centaur. Telemetry data is up-con- of a magnetic field. These sensor outputs are converted verted to an intermediate frequency and modulated using to 16-bit digital values using an internal delta-sigma FSK. This data is sent directly to the on-board RF front-end A/D converter. An onboard EEPROM stores the mag- which modulates to S-Band frequencies. Ground command netometers configuration for consistent operation. The data is received and down-sampled in the RF front-end and data output is serial full-duplex RS-232 or half-duplex demodulated by the FPGA. Commands are interpreted by RS-485 with 9600 or 19,200 data rates. It has flown on the Centaur. several missions, including CanX-1. CYGNSS uses two Earth Horizon Sensors to measure pitch and roll angles SUPPORT CIRCUITRY of the spacecraft. Each sensor has two thermopile detec- The FCP includes support circuitry, including FPGA con- tors which view the Earth limb and measure the dip angle figuration PROM, buffer memories, and housekeeping with respect to the horizon. components, with which SwRI has extensive experience. The ADCS has three primary states of operation: Keeping the entire observatory command and telemetry rate damping, nadir acquisition, and normal pointing. chain in house allows SwRI to respond quickly to issues The rate damping state is used initially after separation

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 61

from the launch vehicle and for anomaly recovery if network via TCP/IP to permit the use of external simulators rates exceed normal state capabilities. Rate damping uses to test the ADCS. a “B-dot” algorithm to command magnetic dipole RTEMS provides a small memory footprint and deter- moments opposed to the rate of change of the magnetic ministic timing. Software development tools include the vector, both measured in body coordinates. It only uses GCC compiler, the debug monitor, and the Software Veri- the sensed magnetic field, and does not rely on a correct fication Environment. The flight software team has signifi- orbital ephemeris or magnetic field model. Wheel speed cant flight development experience with this environment is off for launch and initial tip-off recovery, or set to its from the Fermi, Juno, and MMS missions. nominal value during anomaly recovery. After the body rates are damped, the system transitions 7. GROUND SEGMENT AND MISSION OPERATIONS into nadir acquisition, which monitors the pitch/roll horizon sensors to determine a rough Earth vector. The sen- 7.1. CONCEPT OF OPERATIONS sors are not assumed to be in their linear range; simple “on In developing the design concepts for the CYGNSS Obser- Earth” and “off Earth” measurements are used to establish vatories, the Systems Engineering team has kept in mind slow roll and pitch rates to bring the sensors into their lin- ensuring the safety of the Observatories without ground ear range (! 5°). The momentum wheel is also maintained intervention. Providing on-board systems which minimize relatively close to its commanded nominal speed, with a the need to develop time-tagged command sequences for desaturation gain much lower than normal. each Observatory for routine operations also supports When the ADCS brings the sensors within their linear a simplified operational cadence for maintaining the ranges, it transitions to normal operations. The normal state constellation. uses pitch and roll measurements from the horizon sensors to calculate pitch, roll, and filtered roll rate information. It LAUNCH THROUGH COMMISSIONING com-pares the measured magnetic field with a calculated Each Observatory is deployed with solar arrays stowed and model to determine yaw and filtered yaw rate information. the Observatories can remain in this ‘stowed’ configuration These measurements are used to control momentum wheel indefinitely. After deployment from the launch vehicle, torques for pitch and the electromagnets for roll and yaw each Observatory transitions automatically through the angle, and pitch wheel desaturation. initial three states to reach the Standby Mode where it can Normal control is capable of degraded operation safely remain indefinitely. (used in Standby mode) if the ephemeris and magnetic Deployment of the S/As will occur within a commu- field model are temporarily unavailable. Pitch and wheel nication pass allowing the CYGNSS operations and SC desaturation are controlled as before, but roll and B-dot teams to observe the deployment sequence and address (y axis) information (as in HCMM) are used to control any issues that may occur using real-time commanding. roll and yaw with slightly degraded accuracy. The torque Additional commissioning activities for the Observato- rod commanding is synchronized to permit accurate ries will begin once the S/As are deployed on every Obser- measurement of the local geomagnetic field. A Kalman vatory in the constellation and will continue for a period filter is used to estimate body rates and improve yaw atti- of 2 to 4 weeks. tude estimation. Orbit position is provided via GPS deter- Commissioning activities for a CYGNSS DDMI com- mination from the DDMI. mences once its microsat has completed its commissioning sequence. DDMI commissioning begins and lasts an addi- 6. MICROSATELLITE FLIGHT SOFTWARE tional 4 weeks. During this time, the DDMI is operated in The CYGNSS microsat flight software, which handles all two Engineering modes, which are used to verify on-orbit station keeping, is based on a cost-effective, component performance and tune the on-board Delayed Doppler Map architecture, enabling significant software reuse. It is devel- (DDM) generation and subsampling algorithms. At the end oped in the C Language, executing on the Centaur computer of the DDMI commissioning activities, the instrument will in the RTEMS real-time operating system environment. be transitioned into its Science mode where it will collect The modular architecture and components enable efficient data continuously. development and verification while directly supporting Commissioning activities for the microsats and then on-orbit modification. The flight software is table-driven the instruments may progress in an interleaved man- and includes provisions for memory, table, and program ner. Within a single communication pass activities will image uploads. Application components interface through be performed on a single Observatory, however it is not a software bus implementation (part of the Flight Core) to necessary to complete all commissioning tasks on one exchange CCSDS packets. Standard CCSDS protocols sim- Observatory before progressing to the next Observatory plify the integration of application components and pro- in the constellation. Since all Observatories are indepen- vide a reliable mechanism to install component stubs and dent, it is also unnecessary to ensure each Observatory simulations during software testing. During flight software is at the same ‘step’ in a commissioning sequence. This development, the software bus is bridged to an Ethernet independence allows a flexible scheduling approach to

62 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

be used in setting up commissioning passes and does not ◗ any incomplete transmissions from the previous pass, delay commissioning activities for all Observatories if a based on the protocol messages, will be downlinked by single Observatory requires extra time while an off-nom- the microsat CFDP engine inal issue is being addressed. ◗ science and engineering files placed on the downlink list in the microsat since the last pass will be transmitted NOMINAL OPERATIONS to the ground and collected at the antenna site Upon completion of commissioning activities, the Obser- ◗ at the end of the planned pass time, the MOC will send a vatories will be transitioned into the ‘Science’ mode of CFDP freeze command to stop the transmission of files operation. At this point the DDMI is set to Science mode for and a transmitter off command the duration of the mission, except for brief returns to Cal/ ◗ the AER system on board the microsat will have a backup Val mode performed bi-annually. In Science mode, sub- transmitter off command which will be triggered by a sampled DDMs are generated on-board and downlinked timer that is set when the transmitter is turned on to with 100% duty cycle. ensure the transmitter is not inadvertently left on for a The Observatories are designed to implement nomi- long period of time nal Observatory operations and science data collection ◗ post pass—the collected files will be transferred from without on-board time-tagged command sequences. the antenna site to the USN (Universal Space Network) With the DDMI in its continuous science mode, and the Network Management Center (NMC) where they can Observatory set to maintain all nominal operations with- then be transferred to the CYGNSS MOC for processing out additional commanding, the primary ‘routine’ activ- and distribution. ity performed on a regular basis is communication with The plan for CYGNSS operations is to flow the CFDP the ground network to downlink the accumulated science files from the remote USN antenna sites to the USN NMC and engineering data. after the completion of the pass. This flow decouples the Science and engineering data files are generated, file processing from the real-time flow of the pass which stored on-board, and automatically added into an on- simplifies the operations and does not levy any bandwidth board downlink file list. Retrieval of the science data requirements on the links form the remote antenna sites to occurs during communications passes which are planned the NMC. to occur at the rate of one pass per Observatory every Post pass, the files collected during the pass will be 1.5 to 2 days during the nominal operations period. On- flowed to the CYGNSS MOC where they will be processed board microsat data storage provides storage for greater through the CFDP engine to create the protocol messages than 10 days of science data allowing flexibility in pass that will be uplinked at the next contact with the Observa- scheduling and supporting recovery from loss of commu- tory. Complete science files will then be transferred to the nications during a pass. Science Operations Center (SOC). Incomplete files will be Downlink pass acquisition operations are automated saved at the MOC until they can be completed during the using an on-board Automated Event Recognition (AER) next pass with the Observatory. capability. The mission operations team will schedule passes for each Observatory and when the Observatory is ROUTINE MAINTENANCE AND CALIBRATION within range of the scheduled ground antenna asset—the The majority of post commissioning operations for CYG- antenna will illuminate the microsat with a Clear Channel NSS will occur using the automated features available in communication. On board, the AER will be set to switch the microsat and in the MOC. However, there will also be the microsat transmitter on when the receiver detects the routine microsat maintenance and DDMI calibration activ- ground network signal. Once the transmitter is enabled, ities that will occur throughout the operational period of housekeeping telemetry will be transmitted allowing the the constellation. ground antenna to synchronize with the microsat. Once Maintenance activities for the microsat do not need lock has been established, a notification of the acquisition to be scheduled on a specific cadence. Review of micro- status will be relayed to the CYGNSS Mission Operations sat systems and positioning information will be used to Center (MOC). assess the status of each subsystem as well as the loca- After establishing contact, the following steps are tion of each Observatory to determine when mainte- performed: nance activities may be needed. Based on the type of ◗ housekeeping data is continuously transmitted by activity, either real-time commanding—or a time-tagged the microsat, received on the ground and flowed to command sequences can be developed to perform the the MOC required activities. ◗ MOC sends the command to thaw the CCSDS File Cal/Val of the DDMI is planned to occur two times Delivery Protocol (CFDP) engine on board the microsat per year, nominally before and after hurricane season. ◗ MOC sends the CFDP protocol commands associated Cal/Val activities will be performed using on-board time- with the files downlinked during the last pass for this tagged command sequencing. Part of the Cal/Val process Observatory uses cooperative beacons on the ground and the time-tag

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 63

s &LIGHT$YNAMICS s EngineerING$ATA&ILES s -ISSION0LANNINGWITH s !$#3$ATA #ONSTRAINT#HECking s &375PLOADS!S.EEDED s #OMMAND'ENERATION ,OADS 2EAL 4IME #&$0 s #OMMAND&ILES!S.EEDED (AWAII

-ICRO3AT s #OMMAND3CHEDULING #HILe Eng -ISSION /PERATIONS s 3CIENCE$ATA&ILES #ENTER3W2) s EngineerING$ATA&ILES s #&$00ROTOCOL AUSTRALIA )NTERNET s #OMMAND&ILES!S.EEDED

)NSTRUMENT Eng 3CIENCE /PERATIONS s )NSTRUMENT$ATA #ENTER5-ICH s )NSTRUMENT5PLOADS!S.EEDED 53.'ROUND s #OMMAND&ILES!S.EEDED .ETWORk s 3CIENCE$ATA&ILES /PS#ENTER s EngineerING$ATA s )NSTRUMENT$ATA s #OMMAND&ILES!S.EEDED

FIGURE 10. CYGNSS ground system overview.

command sequencing allows the team to coordinate For all subsequent stages, the MOC schedules nomi- instrument activities with the time periods when the bea- nal passes for the USN stations for each Observatory in cons will be observable by the Observatory. the constellation per the USN scheduling process. Each Observatory can accommodate gaps in contacts with 7.2. GROUND SYSTEM OVERVIEW storage capacity for >10 days of data with no interrup- The CYGNSS ground system, as shown in Figure 10, con- tion of science. sists primarily of the MOC; existing USN Prioranet ground stations in Australia, Hawaii, and Santiago, Chile; and the 7.3. MISSION OPERATIONS CENTER (MOC) SOC facility. Additional interfaces between the MOC and During the mission, the CYGNSS MOC, located at the SwRI the microsat engineering team and the DDMI instrument Boulder location, is responsible for the mission planning, engineering teams are supported. The MOC coordinates flight dynamics, and command and control tasks for each operational requests from all facilities and develops long of the Observatories in the constellation. A summary of the term operations plans. primary MOC tasks includes: ◗ coordinating activity requests GROUND DATA NETWORK—USN ◗ scheduling ground network passes CYGNSS selected USN for the ground data network due ◗ maintaining the CFDP ground processing engine to their experience in autonomously acquiring S/C per ◗ collecting and distributing engineering and science data our baselined approach. Co-location of a back-up CYG- ◗ tracking and adjusting the orbit location of each Obser- NSS MOC server at the USN Network Management Center vatory in the constellation (NMC) can also be supported. ◗ trending microsat data The Observatories within the CYGNSS constellation ◗ creating real-time command procedures or command will be visible to three ground stations within the Univer- loads required to perform maintenance and calibration sal Space Network (USN)—located in Hawaii, Australia, activities and Santiago, Chile—for periods which average 470–500 ◗ maintaining configuration of on-board and ground seconds visibility per pass. Each Observatory will pass parameters for each Observatory. over each of the three ground stations 6–7 times each day, thus providing a large pool of scheduling opportunities SCIENCE OPERATIONS CENTER (SOC) for communications passes. The CYGNSS SOC, located at the University of Michigan, The MOC personnel will schedule passes as necessary will be responsible for the following items: to support commissioning and operational activities. High ◗ support DDMI testing and validation both pre-launch priority passes will be scheduled to support the Observa- and on-orbit tory S/A deployment for each of the constellation microsats. ◗ provide science operations planning tools

64 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

◗ generate instrument command requests for the MOC that the ground system must take into account, including ◗ process science data Levels 0–3 unique command constraints, telemetry conversions and ◗ archive Level 0–3 data products, DDMI commands, limit checking. The ITOS tools provide the database ele- code, algorithms and ancillary data at a NASA Distrib- ments necessary to support and maintain a constellation uted Active Archive Center (DAAC). configuration. The CYGNSS team will be using ITOS throughout the MISSION OPERATIONS CENTER spacecraft development including as the main control Another key aspect to providing cost effective support for system during system integration and environmental a constellation, is to have a set of tools supporting the mis- testing. This bench-to-flight approach allows for heavy sion operations team that allow the team to see issues with reuse of existing STOL (Spacecraft Test and Operations any single Observatory as well as supporting a view of the Language) procedures that will be baselined into the Mis- potential issues or interactions between Observatories. The sion Operations configuration management system as CYGNSS mission operations team has selected a set of tools the standard scripts and processes the team will use to with the feature sets available to address this issue as out- fly the mission. lined in the following paragraphs. The CYNGSS Mission Planning System takes inputs from flight dynamics, and science activities from the sci- COMMAND AND CONTROL SYSTEM ence operations center (SOC), as well as event files, such The requirements for the Mission Operations Center are as eclipse periods and ground tracks. In addition, it must to implement a command and control system that can resolve resource conflicts, such as power load, recorder handle all unique aspects of the CYNGSS mission. For usage, or over subscription of a ground antenna resource. uplink, it must support real-time commanding at 2000 The system must also check that planned events do not bps, including memory load-dump-compare operations. result in violation of flight constraints – either for a single On downlink, it must support ingesting CFDP data, Observatory or for the constellation. Resolving the con- Reed-Solomon decoding, derandomization and include flicts, the system can then generate a command load, real-time telemetry display, and long-term archival and when required, that is handed off to the command and analysis tools. For the ground segment, the tools need to control system for uplink to the spacecraft. be able to interface, configure and monitor the ground The CYGNSS mission chose FlexPlan as the basis for network. It is also important that the system is easily its mission planning system. FlexPlan, is specifically deployed, low cost and facilitates use by a team distrib- designed to manage multi-elements such as a spacecraft uted across the country. constellation and is a highly configurable tool, imple- The CYGNSS mission chose the Integrated Test and mented with customization in mind [16]. It contains five Operations System (ITOS) for its command and control major architectural components, Mission Environment system. ITOS is a suite of software developed by the Real- Preparation (MEP), Planning Input Customization (PIC), Time Software Engineering Branch at the Goddard Space Schedule Generator (SG), Conflict Resolution (CR) and Flight Center, and is supported by the Hammers Company. External Interfaces (EI). This Government Off-the-Shelf (GOTS) solution also has The MEP is an offline tool that is used to define the zero license costs for NASA missions and runs on inexpen- flight rules and mission rules, as well as event and resource sive Linux hardware [1]. availability for standard operations segments of the mis- ITOS itself is not uniquely customized from mission sion. It will be defined early in the mission cycle, and only to mission, instead mission customization is through redefined on an as-needed basis if there are large changes to database driven command and telemetry specifications the concept of operations of the mission. and a small set of configuration files. This obviates the The PIC module takes event triggers from external need for additional software development and training. inputs (for instance, Flight Dynamics, SOC or ground The database includes limit checking and engineering network events) and interfaces to the SG. The SG then unit configurations as well as highly customizable display takes the MEP and the PIC inputs to generate a first revi- pages for monitoring spacecraft data. The ITOS telemetry sion of a mission schedule. At this stage, the mission server can interface across a firewall to a public server, schedule still may not be conflict free, so the user must which can display telemetry and events remotely via execute the CR module. This module detects conflicts a web browser, which facilitates simple, real-time moni- due to timeline or resource constraints, and resolves toring of the spacecraft from a geographically diverse them with the user-in-the-loop. The required external mission team. data products are then created using the EI module, For the CYGNSS mission, it is critical for the command which uses an XML interface schema to easily adapt to and control system to be able to define eight unique and different external interface requirements. concurrent spacecraft, and be able to manage and display Satellite Tool Kit (STK) has been selected by the data unique to each. Though the spacecraft will be iden- CYGNSS team for the flight dynamics tool. During mis- tical by design, they will all likely have unique aspects sion development, STK will be used by the science and

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 65

systems teams to evaluate the science coverage of the minimized by averaging 4 push springs (Fig. 11) to reduce constellation as well as the dispersion of Observatories microsat cg location criticality and minimize the effects through various mission phases. The mission operations of spring tolerances. Screening the springs during DM team will pick up the scenarios developed and maintain assembly further reduces tip off errors. Each Observatory and use these scenarios to support the mission opera- is secured to the DM by torquing the Qwknut actuator tions Flight Dynamics tasks. into the microsat nadir baseplate, compressing the sepa- CYGNSS Flight Dynamics tasks are straight forward ration springs to achieve desired spring load for Obser- and include assessing satellite locations in support of vatory ejection. Tapered alignment pins, combined with ground station scheduling and working with the sys- the Qwknut actuator, rigidly constrain each Observatory tems team to assess, plan, and execute drag maneuvers to the DM for launch. Preliminary FEM quasi-static load as required to maintain constellation coverage and analysis of the fully integrated FS indicates that launch positioning. STK is an industry recognized tool with a loads have a 2.17 safety factor against ultimate loads. mature tool set fully capable of supporting satellite con- The FEA also indicates the first natural frequency of the stellation analysis. structure is a radial mode (Lobar) at 47.6 Hz in the launch configuration, avoiding harmonic coupling with the LV 8. DEPLOYMENT MODULE during launch.

8.1. DEPLOYMENT MODULE STRUCTURE 8.2. DEPLOYMENT MODULE AVIONICS The deployment module (DM) serves as the constel- The DM uses a heritage electronic sequencer to release lation carrier during launch and then deploys the the Observatories in a pre-determined sequence stored Observatories into their proper orbital configuration within the sequencer memory. The sequence is initiated once on orbit. via a standard LV discrete signal when the LV arrives The DM consists of 2 AL cylindrical sections or tiers, at the required orbit. The sequencer then performs the each with 4 mounting/separation assemblies (Fig. 10). deployment sequence by actuating the Qwknut actua- The tier design approach simplifies Observatory-DM tors. Sequence timing incorporates constellation separa- integration by enabling easy access of GSE while mini- tion requirements and deployment actuation tolerances. mizing potential for damage inherent in a single core Hardware safety is ensured through the use of a 2-stage structure. The mounting/separation assemblies are posi- command, single-fault tolerant actuator driver design tioned 90° apart to release the Observatories in pairs that includes a pre-flight Safe/Arm connector to fully dis- opposite each other, balancing deployment forces and arm the system. keeping disturbance torques well within LV capabilities. A 28 Vdc DC 140 W-hr Li-Ion battery is used to power Tier 2 is clocked 45° from Tier 1 to provide proper orbital the DM avionics and activates the deployment Qwk- dispersal vectoring. nut actuators. The battery is fully charged at launch with Deployment is initiated using flight-proven, high-reli- <5% of capacity required to complete the orbit insertion ability Qwknuts. Observatory separation tip-off errors are and deployment sequence. In support of pre-launch operations, the DM avionics route Obser- vatory battery trickle charge power from the GSE to the Observatory via separation connectors, with battery temperature signals acquired by the DM avionics and routed to the GSE for monitoring. Pre-launch Observa- tory command and telemetry handling is also provided by the DM avionics. The GSE command data stream is routed to each Observa- tory command hardline interface with only buffering provided by the DM. Specific command targeting is a function of S/C ID; the Observa- tory ignores the command if the S/C ID is not applicable. The DM enables Observatory pre-launch health and FIGURE 11. 2-tier deployment module provides balanced separation forces by using a status monitoring by multiplexing matched spring deployment mechanism. the Observatory hardline telemetry.

66 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

9. CONCLUDING REMARKS [13] J. Dickinson, J. Alvarez, R. Rose, C. Ruf, and B. Walls, ”Avionics The CYGNSS mission introduces a new paradigm in low- of the CYGNSS Microsat constellation,” in Proc. IEEE Aerospace cost Earth science missions that employs a constellation of Conf.,Big Sky,MT,March2013. science-based microsats to fill a gap in capabilities of exist- [14] D. Rose, R. Rose, C. Ruf, and M. Vincent, ”The CYGNSS ground ing large systems at a fraction of the cost. segment: Innovative mission operations concepts to support a The CYGNSS observatories will make frequent wind micro-satellite constellation,” in Proc. IEEE Aerospace Conf.,Big observations, and wind observations in precipitating con- Sky, MT, March 2013. ditions, using GPS reflectometry to observe the TC inner [15]M. Unwin, M. Brummitt, and C. Ruf, ”The SGR-ReSI and its core ocean surface. These efforts will result in unprece- application for GNSS reflectometry on the NASA EV-2 CYG- dented coverage of winds within a TC throughout its life NSS mission,” in Proc. IEEE Aerospace Conf.,Big Sky,MT, cycle and thus provide critical data necessary for advancing March 2013. the forecast of TC intensification. [16] J. A. Tejo, M. Pereda, I. Veiga, J. P. Chamoun, G. Garcia, and T. Beech, ”FlexPlan: An operational mission planning and sched- 10. REFERENCES uling COTS used internationally,” presented at the IEEE Aero- [1] S. J. Katzberg, R. A. Walker, J. H. Roles, T. Lynch, and P. G. Black, space Conf, Mar. 2005. ”First GPS signals reflected from the interior of a tropical storm: [17] M. G. Schlax, D. B. Chelton, and M. H. Freilich, ”Sampling er- Preliminary results from hurricane Michael,” Geophys. Res. Lett., rors in wind fields constructed from single and tandem scat- vol. 28, pp. 1981–1984, 2001. terometer datasets,” J. Atmos. Oceanic Technol., vol. 18, pp. 1014– [2] S. J. Katzberg, O. Torres, and G. Ganoe, ”Calibration of reflected 1036, 2001. GPS for tropical storm wind speed retrievals,” Geophys. Res. Lett., vol. 33, p. L18602, 2006, doi: 10.1029/2006GL026825. [3] V. U. Zavorotny and A. G. Voronovich, ”Scattering of GPS sig- GRS nals from the ocean with wind remote sensing application,” IEEE Trans. Geosci. Remote Sensing, vol. 38, pp. 951–964, 2000. [4]S.Gleason,S.Hodgart,Y.Sun,C.Gommenginger,S.Mackin, M. Adjrad, and M. Unwin, ”Detection and processing of bi-stat- URSI Commission F ically reflected GPS signals from low earth orbit for the purpose of ocean remote sensing,” IEEE Trans. Geosci. Remote Sensing, Microwave Signatures 2013 vol. 43, no. 5, 2005. Specialist Symposium on Microwave Remote [5] S. Gleason, ”Remote sensing of ocean, ice and land surfaces us- Sensing of the Earth, Oceans, and Atmosphere ing bi-statically scattered GNSS signals from low earth orbit,” October 28-31, 2013 Ph.D. dissertation, Univ. Surrey, U.K., Jan. 2007. Espoo, Finland [6] M. Unwin, S. Gleason, and M. Brennan, ”The Space GPS reflectometry experiment on the UK disaster monitoring constellation URSI Commission F satellite,” in Proc. ION GPS 2003,Portland,OR. Aalto University [7] M. P. Clarizia, C. Gommenginger, S. Gleason, M. Srokosz, C. Sponsored by IEEE GRSS Galdi, and M. Di Bisceglie, ”Analysis of GNSS-R delay-Doppler maps from the UK-DMC satellite over the ocean,” Geophys. Res. Lett., 2009, doi: 10.1029/2008GL036292. [8] M. P. Clarizia, M. Di Bisceglie, C. Galdi, C. P. Gommenginger, and M. A. Srokosz, ”Simulations of GNSS-R returns for delay- Doppler analysis of the ocean surface,” in Proc. 2009 IEEE Int. General Chair: Geoscience and Remote Sensing Symposium (IGARSS) Earth Observa- Prof. Martti Hallikainen, Aalto University tion—Origins to Applications, Cape Town, South Africa. Abstract submission: [9] P. Jales, et al., ”First spaceborne demonstration of Galileo signals May 20, 2013 for GNSS reflectometry,” in Proc. ION GNSS 2008,Savannah,GA. One-page abstract [10] M. Unwin, et al., ”The SGR-ReSI: A new generation of space GNSS Email: [email protected] receiver for remote sensing,” in Proc. ION GNSS 2010,Portland,OR. Notification of abstract review outcome: [11] C. Ruf, et al., ”The NASA EV-2 Cyclone Global Navigation Satel- July 15, 2013 lite System (CYGNSS) mission,” in Proc. IEEE Aerospace Conf.,Big Advance registration: Sky, MT, March 2013. September 10, 2013 [12] R. Rose, C. Ruf, D. Rose, M. Brummitt, and A. Ridley, ”The CYG- Web Address: NSS flight segment: A major NASA science mission enabled by http://frs2013.ursi.fi micro-satellite technology,” in Proc. IEEE Aerospace Conf.,Big Sky, MT, March 2013. Digital Object Identifier 10.1109/MGRS.2013.2261365

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 67

TECHNICAL COMMITTEES

WILLIAM J. BLACKWELL, MIT Lincoln Laboratory, [email protected] BOON H. LIM, Jet Propulsion Laboratory, [email protected]

The IEEE GRSS Working Group on Remote Sensing Instruments and Technologies for Small Satellites

INTRODUCTION and include everything from “femto” to “mini” satellites he IEEE GRSS Instrumentation and Future Tech- in our purview: Tnologies (IFT) Technical Committee (henceforth, ◗ Large satellite: > 1000 kg IFT-TC) seeks to foster international cooperation in ◗ Medium satellite: 500-1000 kg advancing the state-of-the-art in geoscience remote ◗ Mini satellite: 100-500 kg sensing instrumentation and technologies that improve ◗ Micro satellite: 10-100 kg knowledge for the betterment of society and the global ◗ Nano satellite: 1-10 kg environment. The mission of the IFT-TC is to facilitate, ◗ Pico satellite: 0.1-1 kg engage and coordinate GRSS members and the commu- ◗ Femto satellite: <100 g nities-at-large to: assess the current Of particular interest recently is a class of pico/ state-of-the-art in remote sensing nano satellites called CubeSats, a type of miniaturized instruments and technology, iden- satellite for space research that usually has a volume IFT-TC MISSION: tify new instrument concepts and of exactly one liter (10 # 10 # 10 cm), has a mass of 1) ASSESS THE CURRENT relevant technology trends, and no more than 1.33 kg, and typically uses commercial- INSTRUMENTATION recognize enabling technologies for off-the-shelf (COTS) components. Larger CubeSats STATE-OF-THE-ART. future instruments. The IFT-TC is can be developed using multiple “1U cubes”, 3U and organized into six working groups 6U designs are now common. California Polytechnic 2) IDENTIFY NEW targeted at current and emerging State University and Stanford University developed INSTRUMENT CONCEPTS technology areas relevant to the the CubeSat specifications in 1999 to help universities AND TECHNOLOGY broad remote sensing community: worldwide perform space science and exploration. In TRENDS. 1) Active Microwave (RADAR and less than a decade, CubeSats have evolved from purely 3) RECOGNIZE ENABLING SAR), 2) Microwave Radiometry, educational tools to a standard platform for technol- TECHNOLOGIES FOR 3) Lidar, 4) Optical Instrumenta- ogy demonstration and scientific instrumentation. FUTURE INSTRUMENTS. tion, 5) Global Navigation Satellite System, and the newly formed, 6) THE EMERGENCE OF SMALL SATELLITES Remote Sensing Instruments and In the past, the preferred architecture for most space- Technologies for Small Satellites borne Earth remote sensing missions was a single (henceforth the “SmallSat Working Group”), the focus large spacecraft platform containing a sophisticated of this article. The SmallSat Working Group currently suite of instruments. Following the evolution of the has over 50 core members and continues to grow. computer from room-sized to pocket-sized, technology has paved the way for a similar shift in satellites. SMALL SATELLITE TAXONOMY Three distinct advantages arise from going ‘small’ There are several useful definitions of what it means to to compensate for the loss in mass, power and vol- be a “small satellite,” but for the purposes of the Small- ume. First, small satellites allow for cheap access to Sat Working Group, we offer the following guidelines space. By flying as secondary payloads and utiliz- ing excess capacity, launch costs can be reduced by

Digital Object Identifier 10.1109/MGRS.2013.2260912 an order of magnitude or more. Notably, the NASA Date of publication: 26 June 2013 CubeSat Launch Initiative (CSLI) has committed to

68 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

providing 89 CubeSat launches in the last 4 years at no cost to selected proposers. Second, small satellites allow Launch History of Nano and Pico Satellites for rapid development and lower costs through use of 20 common parts/frameworks. Economies of scale do exist 18 for small satellites, where parts and subsystems are rel- 16 evant for a wider variety of missions, instead of a single 14 application in traditional flagship missions. Third, small 12 satellites allow for a more relaxed risk posture due to 10 the significantly lower cost. In terms of distributed risk, 8 a single $100 million mission is inherently riskier than 6 one hundred $1 million missions. Radically new mis- Number of Satellites 4 sion architectures consisting of very large constellations or clusters of CubeSats promise to combine the temporal 2 0 resolution of GEO missions with the spatial resolution 1960 1970 1980 1990 2000 2010 of LEO missions, thus breaking a traditional trade-off in Launch Year Earth observation mission design [9]. Figure 1 shows the growing number of launches of nano and pico satellites FIGURE 1. Launch history of nano and pico satellites since 1955. since the 1990s [2]. and hurricanes. The CYGNSS mission comprises eight Low A SNAPSHOT OF CURRENT AND FUTURE MISSIONS Earth Orbiting 18-kg spacecraft (see Figure 2) that receive There have been many recent small satellite missions that both direct and reflected signals from GPS satellites to deter- have successfully launched or are currently in formulation mine position and ocean surface roughness. The Canadian that will demonstrate new technologies for Earth observa- CanX-4 and CanX-5 mission will demonstrate two nano- tion (see [5] for examples). The Cyclone Global Navigation satellites flying autonomously in precise formations with Satellite System (CYGNSS), with launch expected in 2016, relative position determination accurate to a few centime- will make frequent and accurate measurements of ocean ters using carrier-phase differential GPS techniques. Its suc- surface winds throughout the life cycle of tropical storms cess may someday enable a constellation of nanosatellite

59.0 cm (Stowed) 176.3 cm (Deployed) Stowed (Launch S/A Wing (×2) Configuration) Zenith S-Band Antenna Zenith DDMI Antenna

Thermal Radiator 42.5 cm Horizon Sensors Torque Rod (2 of 3) Momentum Wheel Avionics 28° Batteries (×2) 18.6 cm S/A (Ram) DMR S/A During Deployment

Nadir Sci Diplexer & S/A Deployed Antennas (2) Hybrid (Flt Config) Nadir S-Band Antenna

Z (Zenith) Nadir Baseplate & Radiator

X (Ram)

FIGURE 2. The CYGNSS Observatory. The exploded view shows individual subsystems, including the science payload’s Delay Doppler Mapping Imager (DDMI) antennas and receiver electronics (DMR).

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 69

receivers augmenting an existing SAR transmitter for UPCOMING SPECIAL SESSION AT IGARSS 2013 InSAR applications [6]. The European QB50 Project will There is a vibrant community within the GRSS that is tak- soon launch an international network of 50 2U CubeSats ing advantage of the small satellite platform. A full spe- to study the temporal and spatial variations of a number cial session at IGARSS 2013 in Melbourne, Australia will of key constituents and parameters in the lower thermo- be dedicated to past, present and future missions. Among sphere [4]. ESA’s 140-kg PROBA-V satellite will serve as a the topics to be discussed are: CubeSats for atmospheric miniature engineering lab in orbit. Less than a cubic meter monitoring using microwave radiometry, constellation in volume, PROBA-V is hosting approaches for improved mission performance, and new five technology experiments, sensing technologies offering extremely low size, weight, including innovative testing of and power. THE EUROPEAN QB50 fiber optics for space. The GEO- PROJECT WILL SOON CAPE ROIC In-Flight Perfor- REFERENCES LAUNCH AN INTERNA- mance Experiment (GRIFEX) is [1]W.Blackwell,G.Allen,C.Galbraith,T.Hancock,R.Leslie,I. TIONAL NETWORK OF a 3U CubeSat in development Osaretin, L. Retherford, M. Scarito, C. Semisch, M. Shields, M. that will perform engineering Silver,D.Toher,K.Wight,D.Miller,K.Cahoy, and N.Erickson, 50 2U CUBESATS TO assessment of an all-digital in- ”Nanosatellites for earth environmental monitoring: The Micro- STUDY THE TEMPORAL pixel high frame rate Read-Out MAS project,” in Proc. 2012 IEEE Int. Geoscience and Remote Sens- AND SPATIAL VARIATIONS Integrated Circuit (ROIC). This ing Symp. (IGARSS), pp. 206–209. OF A NUMBER OF KEY ROIC has an unprecedented [2] S. Janson, ”25 Years of small satellites,” in Proc. AIAA/USU Conf. CONSTITUENTS AND frame rate of up to 12 kHz Small Satellites,Aug.2011. PARAMETERS IN THE while consuming less than 2 W [3] B. Lim, D. Mauro, R. De Rosee, M. Sorgenfrei, and S. Vance, LOWER THERMOSPHERE. of power where the design of ”CHARM: A CubeSat water vapor radiometer for earth sci- analog-to-digital converters in ence,” in Proc. 2012 IEEE Int. Geoscience and Remote Sensing Symp. each pixel enables the all-digi- (IGARSS), pp. 1022–1025. tal design. MicroMAS [1], a 3U [4] J. Muylaert, et al., ”QB50: An international network of 50 Cube- CubeSat for 118 GHz sounding, utilizes LTCC SIW filters Sats for multi-point, in-situ measurements in the lower thermo- on the backend to provide channelization with a scanner sphere and for re-entry research,” in Proc. ESA Atmospheric Science motor assembly to achieve a cross-track swath. CHARM [3], Conf.,2009. a 3U CubeSat for 183 GHz sounding, utilizes a state-of-the- [5] C. D. Norton, M. P. Pasciuto, P. Pingree, S. Chien, and D. Rider, art Indium Phosphide low noise amplifier (<20 mW) and ”Spaceborne flight validation of NASA ESTO technologies,” in novel internal calibration. These two radiometer missions Proc. 2102 IEEE Int. Geoscience and Remote Sensing Symp. (IGARSS), in development have the capability of synergy in the future pp. 5650–5653. as a combined 118/183 GHz sounder. [6] E. H. Peterson, R. E. Zee, and G. Fotopoulos, ”InSAR microsatellite constellations enabled by formation flying and onboard pro- THE SMALLSAT WORKING GROUP cessing capabilities,” in Proc. 25th Annu. AIAA/USU Conf. Small Operational needs, such as weather forecasting, add a dis- Satellites,2011. tinct set of requirements for continual and highly reliable [7] P.Pingree,T.Werne,D.Bekker,T.Wilson,J.Cutler, and M.Hey- monitoring of global conditions [10]. A goal of the SmallSat wood, ”The prototype development phase of the CubeSat On- Working Group is therefore to address these diverse require- board processing Validation Experiment (COVE),” in Proc. 2011 ments and assesses how they might be met by small satel- IEEE Aerospace Conf.,Big Sky,MT. lites, identify the needed core technologies to enable and [8] C. S. Ruf, S. Gleason, Z. Jelenak, S. Katzberg, A. Ridley, R. Rose, J. facilitate small satellite mission development, and bridge Scherrer, and V. Zavorotny, ”The CYGNSS nanosatellite constel- the gap between small satellite and instrumentation tech- lation hurricane mission,” in Proc. 2012 IEEE Int. Geoscience and nologists and remote sensing mission planners. Universities Remote Sensing Symp. (IGARSS), pp. 214–216. have traditionally led the way in embracing the challenge [9] D. Selva and D. Krejci, ”A survey and assessment of the capabili- from the smaller scale. Space agencies worldwide can learn ties of CubeSats for Earth observation,” Acta Astronaut., vol. 74, to incorporate some of these practices that may be at odds pp. 50–68, 2011. with traditional space qualification grade missions (e.g., [10] Committee on Earth Studies, Space Studies Board, National Re- NASA’s Class-A/B missions). The broad membership of the search Council, The Role of Small Satellites in NASA and NOAA SmallSat Working Group can aid in this process. The co- Earth Observation Programs.Washington,DC:The National Acad- chairs invite readers who are interested in contributing to emy Press, 2000. GRS contact them for membership details.

70 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Building a Sustainable Earth Through Remote Sensing, IGARSS 2013

Remote sensing is an ever growing area of research and with sustainability a key issue in our society, this year’s IEEE International Geoscience and Remote Sensing Symposium will look at “Building a Sustainable Earth through Remote Sensing”. World renowned scientists and engineers will meet at the Melbourne Convention and Exhibition Centre between the 21–26 July to discuss the latest exciting developments in satellite imaging technologies, recent research breakthroughs, challenges and future directions of geosciences and remote sensing. Some of the key topics of discussion will include forest degradation, response of the Great Barrier Reef to climate change, disasters and hazard management, pollution and contamination, earthquake mitigation in New Zealand, sea level rise and much more.

The symposium will feature numerous keynote speakers including IEEE President Dr. Peter Staecker, an expert in microwave technologies which are today largely used in the field of remote sensing. Professor Melba Crawford, IEEE GRSS President, whose research in machine learning focuses on advanced methods for analysis of remotely sensed data, will highlight the many exciting latest developments in the field of remote sensing. The director of the Institute of Remote Sensing and Digital Earth, Professor Guo Huadong, will present the latest research specific to the Asia-Pacific region. Also addressing the audience will be Professor Mike Goodchild, one of the world’s pre-eminent thinkers in geographic information sciences and their role in supporting many other aspects of scientific development. Professor Goodchild will challenge us with his thoughts on the relationship between remote sensing and the broader GIS communities.

In addition there will be two keynote speakers from Australia, Dr. Chris Pigram and Dr. Rob Vertessy. Dr. Pigram is the Chief Executive Officer of Geoscience Australia, a world leader in providing first class geoscientific information and knowledge enabling the Australian government to make informed decisions about the use and management of resources, the environment, community well being and sustainable energy. His talk will crystallize the latest Australian developments in geoscience and remote sensing.

Dr. Rob Vertessy is Director of the Australian Bureau of Meteorology. The Bureau of Meteorology is Australia’s national weather, climate and water agency. Its expertise and services assist Australians in dealing with the harsh realities of their natural environment, including drought, floods, fires, storms, tsunami and tropical cyclones. Through regular forecasts, warnings, monitoring and advice spanning the Australian region and Antarctic territory, the Bureau provides one of the most fundamental and widely used services of the Australian government. Dr. Vertessy is currently leading a number of ground-breaking initiatives in the use of remote sensing and value-added spatial and information products systems and will speak about these at the Opening Ceremony.

For further details of this global conference, visit www.igarss2013.org.

Digital Object Identifier 10.1109/MGRS.2013.2261355

CHAPTERS

Distinguished Lecturer Program

he IEEE GRSS Distinguished Lecturer Program ◗ Lorenzo Bruzzone: Current Scenario and Chal- Tprovides GRSS Chapters with talks by experts on lenges in the Analysis of Multitemporal Remote topics of interest and importance to the Geoscience Sensing Images and Remote Sensing community. The purpose of the ◗ Melba Crawford: Advanced Methods for Classifica- program is to provide our members with an opportu- tion of Hypersectral Data nity to learn about the work being done in our disci- ◗ David Goodenough: Methods and Systems for pline and to meet some of the prominent members of Applications our Society. Information about the speakers and how ◗ Ya-Qiu Jin: Research on the Modeling and Simula- to take advantage of the program, including an appli- tion of Polarimetric Scattering and Information Re- cation from, are available at the GRSS website (look trieval for Microwave Remote Sensing for “Distinguished Lecturer” under the “Education” ◗ Yann Kerr: SMOS First Successes and Related Issues: tab). Further information including a manual with in- The First Global Soil Moisture and Sea Salinity Maps structions can be obtained by e-mailing the program are Coming chair, David Le Vine at [email protected]. ◗ Ricardo Lanari: Differential SAR Inteferometry: The program has been structured so that Chap- Basic Principles, Key Applications and New Advances ters will incur no cost in making use of this program. ◗ Keith Raney: Two Hybrid-Polarimetric Imaging Briefly a chapter contacts a lecturer from the list of Radars at the Moon: Their Design, Build and Results available speakers, and once the initial details of the ◗ Werner Weisbeck: Digital Beam-Forming in visit have been worked out (e.g. date and a rough Remote Sensing. budget) the chapter fills out the application form. Abstracts for the talks and background information for The Lecturer travels on his/her own expense with each Lecturer are available on the GRSS website: www.___ reimbursement made directly by IEEE to the Distin- grss-ieee.org. guished Lecturer. Suggestions on ways to improve the program or The list of speakers and topics for 2013 is: ideas for topics and/or speakers are always welcome. Please send your comments or questions to the pro-

gram chair at [email protected]. Your suggestions Digital Object Identifier 10.1109/MGRS.2013.2260914 Date of publication: 26 June 2013 would be greatly appreciated.

Bangalore Section Chapter of the GRSS

EEE Bangalore Section Chapter of the Geoscience The Chair of this chapter, that got formal approval on Iand Remote Sensing Society (IEEE-BSC-GRSS) was 24th Dec 2012, is Prof. B.S. Daya Sagar, Senior Mem- formally inaugurated on 26th March 2013 at Indian ber of IEEE. Inaugural talk on “GIS and Automation” Statistical Institute-Bangalore Centre, Bangalore, India. was delivered by Prof. N. Viswanadham, Life Fellow IEEE (Fig. 1). Followed by this inaugural talk, Dr. Raju Garudachar, Senior Member, has given a talk on “Cli- Digital Object Identifier 10.1109/MGRS.2013.2260918 Date of publication: 26 June 2013 mate Change: Microwave Radiometers and Sounding

72 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Prof. Daya Sagar and Prof. Saroj Meher have delivered a series of 14 lectures on the topics related to mathematical morphology and pattern recognition. The two tracks in this three-day workshop include “Mathematical Morphol- ogy: Theory and Applications” and “Pattern Recognition: Theory and Applications”. A host of algorithms developed based on the concepts of mathematical morphology were covered in a series of seven lectures by B. S. Daya Sagar. Applications of both classical and modern Pattern Recog- nition techniques were covered in a series of seven lectures by Saroj Meher. Conveners (Sagar and Saroj) have framed the sequence of lectures in such a way that there was an excellent coherence. These lectures were mainly intended for Postgraduate students, Ph.D. scholars, Post-Docs and young faculty members and scientists. This workshop has been attended by 60 participants (Fig. 2) drawn from academia, industry, and government organizations and labs. Both the conveners of this first technical event of IEEE- BSC-GRSS gratefully acknowledge the support and guid- FIGURE 1. Inauguration of the IEEE Bangalore Section Chapter ance provided by Prof. Wolfgang-Martin Boerner, Life Fel- of the Geoscience and Remote Sensing Society took place on 26 low IEEE. Further details about the inauguration of the March 2013 at the Indian Statistical Institute-Bangalore Centre. IEEE-BSC-GRSS, and the three-day workshop are avail- Prof. N. Viswanadham, FIEEE, delivering an inaugural talk on “GIS able at: http://www.isibang.ac.in/~mmprta. and Automation”. This note is prepared by B.S. Daya Sagar and Saroj Kumar Meher Radars for Antarctic Research”. This formal inauguration Systems Science and Informatics Unit (SSIU) was followed by a Three-Day workshop on “Mathemati- Indian Statistical Institute—Bangalore Centre cal Morphology and Pattern Recognition: Theory and Ap- 8th Mile, Mysore Road, R.V. College P.O plications”, which was technically sponsored by the IEEE Bangalore-560059, India. Bangalore Section. The two conveners of this workshop Homepage: http://www.isibang.ac.in/~ssiu

FIGURE 2. A photograph taken on the first day (26th March 2013) of the workshop, where the participants and conveners could be seen.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 73

GRSS CHAPTERS AND CONTACT INFORMATION

CHAPTER LOCATION JOINT WITH (SOCIETIES) CHAPTER CHAIR E-MAIL ADDRESS

Region 1: Northeastern USA

Boston Section, MA GRS William [email protected] Springfield Section, MA AP, MTT, ED, GRS, LEO Paul [email protected] Western New York GRS Jan van [email protected]

Region 2: Eastern USA

Washington, DC & Northern VA area GRS Miguel Roman [email protected]______

Region 3: Southeastern USA

Atlanta Section, GA AES, GRS Clayton Kerce [email protected]______Eastern North Carolina Section GRS Linda [email protected]

Region 4: Central USA Central Illinois Section LEO, GRS Weng Cho Chew [email protected] Chicago Section, IL AES, NPS, GRS, OE TBD TBD Southeastern Michigan Section GRS Adib Y. [email protected]

Region 5: Southwestern USA

Denver Section, CO AP, MTT, GRS Michael [email protected] Houston Section, TX AP, MTT, GRS, LEO Christi Madsen [email protected]______

Region 6: Western USA

Alaska Section, AK GRS Franz Meyer [email protected]______Los Angeles Section, CA GRS Paul A. [email protected]

Region 7: Canada

Ottawa Section, ON OE, GRS Yifeng Zhou [email protected]______Quebec Section, Quebec, QC AES, OE, GRS Xavier Maldague [email protected]______Toronto Section, ON SP, VT, AES, UFF, Sri [email protected] OE, GRS Vancouver Section, BC AES, GRS David G. Michelson [email protected] Steven McClain [email protected]______

Region 8: Europe, Middle East and Africa

Benelux Section AES, GRS Mark Bentum [email protected]______Central Italy Section GRS Simonetta [email protected] Croatia Section AES, GRS Juraj [email protected] France Section GRS Gregoire Mercier [email protected]______Germany Section GRS Irena [email protected] Islamabad Section, Pakistan GRS, AES M. Umar Khattak [email protected]______Russia Section GRS Anatolij [email protected] [email protected]

Saudi Arabia Section GRS Yakoub Bazi [email protected]______South Africa Section AES, GRS Meena Lysko [email protected]______South Italy Section GRS Maurizio [email protected] Spain Section GRS J.M. Lopez-Sanchez [email protected] Student Branch, Spain Section GRS Pablo Benedicto [email protected]______Ukraine Section AP, MTT, ED, AES, Kostyantyn V. [email protected] GRS, NPS

United Kingdom & Rep. of Ireland GRS, OE Yong [email protected] (UKRI) Section

Digital Object Identifier 10.1109/MGRS.2013.2260919 Date of publication: 26 June 2013

74 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

CHAPTER LOCATION JOINT WITH (SOCIETIES) CHAPTER CHAIR E-MAIL ADDRESS

Region 9: Latin America

Student Branch, Colombia Section GRS Leyini Parra [email protected] Student Branch, South Brazil Section GRS Marcus [email protected]

Region 10: Asia and Pacific

Australian Capital Territory and New GRS Xiuping [email protected] South Wales Sections, Australia

Beijing Section, China GRS Ji Wu [email protected]______Delhi Section, India GRS O.P.N. Calla [email protected]______Japan Section GRS Yoshihisa Hara [email protected]______

Nanjing Section, China GRS Feng [email protected] Seoul Section, Korea GRS Joong-Sun [email protected] Taipei Section, Taiwan GRS Yang-Lang Chang [email protected]______

Abbreviation Guide for IEEE Technical Societies AES Aerospace and Electronic Systems Society NPS Nuclear and Plasma Sciences Society AP Antennas and Propagation Society OE Oceanic Engineering Society ED Electron Devices Society SP Signal Processing Society GRS EMB Engineering in Medicine and Biology UFF Ultrasonics, Ferroelectrics, and Frequency Control Society LEO Lasers & Electro-Optics Society VT Vehicular Technology Society MTT Microwave Theory and Techniques Society

From Imagination to Market

“IEEE Standards Online IEEE Standards Online gives us access to Your Direct Connection to IEEE Standards up-to-date standards with minimal effort.” Continuous access to the comprehensive and growing collection of IEEE Standards. – Leading Technology Company Access tomorrow’s draft standards ﬁ rst and get a jump on the competition Find a standard quickly with fast and comprehensive search and retrieval features Immediate updates and automatic email alerts Substantial savings over purchasing individually Free Trial! Experience IEEE – request a trial for your company. www.ieee.org/standardsonline

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 75

ORGANIZATION PROFILES

SCOTT H. SCHAIRE, Small Satellite Projects Manager, [email protected]______

NASA/Goddard Space Flight Center (GSFC) Wallops Flight Facility (WFF)

1. GSFC WALLOPS MANAGED SUBORBITAL PROGRAMS or decades the NASA GSFC GSFC Wallops Managed Suborbital Programs Wallops managed suborbital s(ERitage Payload Support Systems (e.g., Attitude Control, Telemetry, F Groundstations, Power, Command) Reduce Mission Risk programs have been, and continue sTesting Process and Infrastructure Prevents Mission Failures to be, indispensable platforms for s%FFective Project Management Increases Mission Success Probability offering rapid access to space for s(IGH$EGREEOF#OMMONALITY,Owers Costs s2APID2ESPONSE)NCREASES3CIENCE2ETURn cutting-edge science experiments, developing and nurturing the ,Ow-Cost PATHWay to next generation of scientists and Mission Success engineers, and for testing and vali- dating new technologies and instru- Most “Bang fORTHE"UCk” for Science, TECHNOLOGY, and mentation. The figure below lists (ANDS /N%DUCATION some of the advantages associated with the NASA Sounding Rocket GSFC Wallops enables new and exciting science, technology, and educational and Balloon Programs resulting in missions, by providing low-cost value-added services and technologies at the request the most “bang for the buck” for of the principal investigator. science, technology and hands-on education missions. It was during this TacSat-3 project that WFF learned 2. SUPPORT FOR THE NATIONAL SCIENCE quite a bit about object debris analysis, safety, and the FOUNDATION (NSF) CUBESAT PROGRAM required CubeSat and P-POD testing. The very first CubeSat launched in the United States was a 3U Cubesat called GeneSat-1, manifested on the Tacsat-2 launch out of GSFC Wallops in Dec. 2006. WFF manifested CubeSats with the Tacsat-3 launch on a Minotaur-1 launch vehicle which launched in May of 2009. Planning for the launch began in 2007 working with the Hawk Institute for Space Sciences. CubeSats are typically launched and deployed from a mechanism called a Poly-Picosatellite Orbital Deployer (P-POD), developed and built by California Polytechnic State University (Cal Poly). WFF provided a spare P-POD for this launch.

Vibration testing of CubeSats manifested with TacSat-3 at Digital Object Identifier 10.1109/MGRS.2013.2260932 Date of publication: 26 June 2013 GSFC Wallops.

76 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Mainly because of WFF experience with the sounding ◗ Ground station support for high data rates over a rocket and balloon programs, and working with principal government frequency band investigators, WFF is ideally suited to support CubeSat ◗ Engineering assistance for reviews and resolving issues endeavors. ◗ Use of existing sounding rocket and balloon payload This was recognized by the NSF when they selected laboratories and test facilities for CubeSats. WFF to collaborate with their CubeSat activities in 2008. For RAX-1, WFF provided The NSF has a recurring CubeSat program for space the primary and spare P-POD’s weather research and WFF is primarily acting as the and cables for the launch. MAINLY BECAUSE OF WFF carrier integrator and launch vehicle interface so that the WFF coordinated safety experiment teams can concentrate on their CubeSats. WFF documentation and supported EXPERIENCE WITH THE continues to support this program. See http://www.nsf. a test of the cable and P-POD SOUNDING ROCKET AND gov/discoveries/disc_videos.jsp?cntn_id=124901&media_ with Orbital Sciences in Chan- BALLOON PROGRAMS,

______id=72804&org=NSF, “NSF’s Therese Moretto Jorgensen dler, Arizona. WFF engineering AND WORKING explains what CubeSats tell us about the atmosphere”. performed a re-entry casualty WITH PRINCIPAL analysis. WFF also was respon- INVESTIGATORS, WFF IS 3. NSF CUBESATS sible for assuring the integra- IDEALLY SUITED TO Of the first eight NSF CubeSat missions selected, tion and testing of the RAX SUPPORT CUBESAT four are in orbit and four are in work. The on-orbit mis- CubeSat into the P-POD, and ENDEAVORS. sions include: preparation for flight, support- ◗ Radio Aurora Explorer (RAX-2)—University of Michigan, ing vibration environmental SRI International, see http://rax.engin.umich.edu/ testing at the University of ◗ Dynamic Ionosphere CubeSat Experiment (DICE)— Michigan and at Cal Poly, pre- Utah State University, Embry-Riddle Aeronautical senting a pre-ship review with the Space Test Program University, Clemson University, see http://www.sdl.usu. (STP), and participating in Orbital Science’s integration

______edu/programs/dice of the P-POD onto the Minotaur IV into the launch vehi- ◗ Colorado Student Space Weather Experiment cle at the Kodiak Launch Complex. RAX-1 had a short (CSSWE)—University of Colorado, see http://lasp.

colorado.edu/home/csswe/overview/______◗ Cubesat for Ions, Neutrals, Electrons, MAgnetic fields (CINEMA)—University of California, Berkley, Kyung-Hee U., Imperial College, Applied Physics Lab,

Inter-American University of Puerto Rico, see ____http:// ______newscenter.berkeley.edu/2012/07/31/cinema-among-

tiny-cubesats-to-be-launched-aug-2/.______The four in-work missions include the following: ◗ Firefly—GSFC, Hawk Institute for Space Sciences, Siena

College, manifested to launch in Sep. 2013, see ____http:// www.nasa.gov/topics/universe/features/firefly.html______◗ Firebird—Montana State University, University of New Hampshire, Aerospace Corporation, manifested

to launch in Dec. 2013, see ______https://ssel.montana.edu/ category/cubesat/______◗ Composition Variations in the Exosphere, Thermo- sphere, and Topside Ionosphere (EXOCUBE)—Scientific Solutions Inc., Cal Poly, University of Wisconsin, GSFC, not yet manifested, see http://www.sci-sol.com/

Exocube-Oct2011.pdf______(See the picture of the Goddard Space Flight Center/Naval Research Labs Winds-Ion- Neutral Composition Suite (WINCS) instrument slated to fly on EXOCUBE and CADRE CubeSats.) ◗ CubeSat investigating Atmospheric Density Response to Extreme driving (CADRE)—University of Michigan,

Naval Research Labs, not yet manifested, see ____http://

exploration.engin.umich.edu/blog/?page_id=961.______Goddard Space Flight Center/Naval Research Labs Winds-Ion- Some of the services that WFF has been providing to the Neutral Composition Suite (WINCS) instrument slated to fly on NSF and their CubeSat teams include the following: EXOCUBE and CADRE CubeSats.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 77

presents a huge improvement over the 9.6 Kbit/s data rates otherwise available. The NSF sponsored DICE CubeSats launched on a Delta II vehicle from Vandenberg AFB in October 2011 with WFF acting as the sole communications node. The mission is mapping geomagnetic Storm Enhanced Den- sity (SED) plasma bulge and plume formations in Earth’s ionosphere. Two identical spinning spacecraft measure plasma density and electric fields to determine the how and why of variations in ionospheric plasma density affect the performance of communications, surveillance, and navigation systems on earth and in space. The DICE mission consists of two identical 1.5U Cube- Sats that were deployed simultaneously into the same orbit. Each carries two Langmuir probes which measure in-situ ionospheric plasma densities and electric field probes to measure direct current and alternate current electric fields. These measurements permit accurate identification of storm-time features, such as the SED bulge and plume, together with simultaneous, co-located electric previously been lacking. Following launch from Vandenberg AFB aboard a Delta II, Wallops Research Range Services (RRS) technicians at the UHF Radar immediately went into action locating and communicating with both the picosatellites as the sole uplink/downlink station supporting the project. WFF has since conducted daily tracking of the satellites and provided the information remotely to project principles at Utah State University. Problems with the satellites, interference, and GSFC Wallops UHF CubeSat groundstation. maintenance of the UHF Radar, have occasionally led to loss of tracking, but in all cases RRS technicians were able time in orbit. RAX-2 is now in orbit and is conducting to relocate and communicate with them ensuring ongoing experiments. success of the mission. The UHF CubeSat groundstation answers a growing 4. UHF CUBESAT need for high data rate from CubeSats over a govern- GROUNDSTATION ment licensed frequency. Government funded CubeSats THE UHF CUBESAT There is a Radar at WFF that is using amateur radio frequencies may violate the intent GROUNDSTATION being used as a satellite telem- of the amateur radio service and it is a violation of ANSWERS A GROWING etry ground station (See the National Telecommunications Information Administra- picture of the GSFC Wallops tion (NTIA) rules for a government funded groundsta- NEED FOR HIGH DATA UHF CubeSat groundstation). tion to use amateur radio frequencies to communicate RATE FROM CUBESATS This large dish is ideal for with CubeSats. OVER A GOVERNMENT UHF satellite communication LICENSED FREQUENCY. because of its 35 dBi gain. The 5. FUTURE DICE CubeSats are currently The success and wide community support for the NSF using the WFF groundstation CubeSat Program combined with the increasing num- and Firefly is planning on ber of NASA proposals that utilize CubeSats demon- using this UHF groundstation. strates the maturation of the CubeSat platform. The UHF In the past the Radar had been used for tracking and CubeSat groundstation is planned for communication study of reentry wakes in the upper troposphere. The support for the NSF/GSFC Firefly CubeSat, the GSFC Radar is currently being used as a CubeSat groundstation, Compact Radiation Belt Explorer (CeREs) CubeSat, the and will be used to support Global Precipitation Mea- Massachusetts Institute of Technology (MIT) Micro-sized surement (GPM) precipitation validation measurements. Microwave Atmospheric Satellite (MicroMAS) CubeSat In support of CubeSats, the UHF Radar with its high gain and has been proposed on a number of NASA and follow- antenna provides a government-licensed frequency allo- on NSF missions. cation that enables high data rates (1.5 Mbit/Sec). This GRS

78 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Geoinformatics 2013: The 21st International Conference on Geoinformatics June 20-22, 2013 Kaifeng (China)

Organized by The College of Environment and Planning at Henan University and The International Association of Chinese Professionals in Geographic Information Sciences (CPGIS)

Sponsored by IEEE

Geoinformatics 2013 Co-Chairs: Henan University and CPGIS Abstract submission: Before Feb. 28, 2013 Email: [email protected]______Register: Before June 10, 2013 ______Registration fees: Early bird on-line registration opens on Dec. 10, 2012 and Digital Object Identifier 10.1109/MGRS.2013.2261364 closes on May 10, 2013 at http://www.GeoInformatics2013.org Registration fee varies based on membership status and registration date. Web Address: http://www.GeoInformatics2013.org ICEO&SI 2013 Digital Object Identifier 10.1109/MGRS.2013.2261361 International Conference on Earth Observations and Societal Impacts June 23-25, 2013 NCKU – Tainan (Taiwan)

Taiwan Group on Earth Observations National Cheng Kung University

ICEO&SI2013Conference Chairman: Dr. Yuei-An Liou, TGEO President Abstract submission: By February 28, 2013 ______Email:[email protected];[email protected]______Register: By May 1, 2013 Registration fees: Indicated at http://2013.iceo-si.org.tw/important-dates Web Address: http://2013.iceo-si.org.tw

Digital Object Identifier 10.1109/MGRS.2013.2261358

Digital Object Identifier 10.1109/MGRS.2013.2261362

EDUCATION

MICHAEL INGGS, Senior Member, IEEE, University of Cape Town, South Africa

GRSS Educational Activities Planning for 2013

I. OVERVIEW devote space in this column. However, bear in mind the his article relates to the Educational activities of the deadlines that are forced on us by the publication cycle. TGRSS. There is a change in format to support the The Education Director can also apply to the AdCom quarterly GRS Magazine, which has just launched, emerg- for funding to assist educational activities. Budgets ing from the Newsletter. I have just taken over from our are tougher and tougher, as we all know, but please very able director, Antonio Plaza, who has moved on to approach me with education initiatives that might need become an Editor of the Society’s Transactions. Thus, some assistance. Remember also that funding is done a the first part of the article will discuss my plans for edu- year ahead, so good notice is required. A regular feature cational activities. This is a programme that has to fit of the budget is the assistance that we give to people to into the cycle of activities of the Society. You will notice attend the IGARSS every year. Clearly student applica- that I am very keen for members to use this forum to re- tions are strongly supported, as this is a very good way port on Educational drives, and that we should establish for students to see that advantages of being a member a regular set of topics related to Education. of the Society. A new innovation falling under the Director Educa- II. THE RESPONSIBILITIES tion is the GRSS Summer School. The first was held in OF THE EDUCATION DIRECTOR association with the IGARSS 2012 in Munich, and was The Administration Committee (AdCom) of our Soci- reported in the Newsletter. The next will occur in Mel- ety ensures that, based on input from the membership, bourne, Australia, in the week before IGARSS 2013 in members see benefit in being part of our community. An the same city. important part of why we are members of GRSS is that Associated with each IGARSS is a series of short (half we can interact technically, and, learn from colleagues or one day) tutorials, offered by volunteers who receive a through journal, conference and workshop contribu- small stipend. These are not necessarily introductory in tions. In exchange, we contribute to these. Although nature. They are popular, as supported by the IGARSS much of this interaction is via published works, there survey that we do every year. It is my intention to assist is an important need to develop the next generation of the IGARSS organizers with these, to make sure there is remote sensors: this is very much the business of the articulation with the new GRSS Summer School. Thus, Education Directorship. There is also a need for con- if you are interested, please contact me. tinuing professional development (CPD) for all levels of Student prizes at the IGARSS is an Educational activ- member, so that we can continue to contribute to our ity, and is run very successfully as part of the IGARSS employers and society. process by senior members of the Society, so at present, The intention is not to be prescriptive, but to act as a not changes seem necessary. Student achievements are coordinator, and through this column, to assist in publi- usually published as part of report on IGARSS that the cation of educational drives. There is an open invitation Magazine will carry, as did the Newsletter before. for you to let me know about your plans so that we can Under Education, we shall continue to collect newly published Ph.D. Theses. Contact information is given down below. Digital Object Identifier 10.1109/MGRS.2013.2260933 Date of publication: 26 June 2013

80 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

When Thursday and Friday, July 18 and 19. From 9:00 AM both days. Finishes at 3 PM on Friday. Cost $75 (Includes all materials, morning and afternoon tea both days plus a social event on Thurs- day, July 18). Where RMIT University, Building 80 (Lecture theatre 02.002 on Level 2), 445 Swanston Street, Mel- bourne. A 20-minute walk from the IGARSS conference site.

B. CONTACTS/ORGANIZERS

Xiuping Jia ([email protected]),______The University of New South Wales, Canberra Campus. Kim Lowell ([email protected]),______Cooperative Research Centre for Spatial Information, University of Melbourne.

C. PROGRAMME

______THURSDAY, JULY 18 09:00–09:10 AM Opening 09:10–10:30 Nathan Quadros (Lidar Processing) 10:30–10:50 Morning Tea (provided) III. GRSS SUMMER SCHOOL 10:50–12:10 Suzanne Furby (Image processing for The very last Newsletter has an in-depth article by carbon accounting) Antonio Plaza on the very successful, inaugural Summer 12:10–01:30 PM Lunch (on your own) School held in Munich. I will report back with the organiz- 1:30–02:50] Mihai Datcu (Image Information ers of the Summer School to be run before the Melbourne Mining) IGARSS 2014. Our society has a large pool of very talented 02:50–03:10 Afternoon Tea (provided) teachers, so it is well within our capabilities to provide 03:10–04:30 Lorenzo Bruzzone (Multi-temporal material on almost any topic. This is where you come in: imagery) please contact me an put forward your ideas for topics that 05:00–? Social event are not being covered, material you might like to teach, and so on. FRIDAY, JULY 19 The GRSS Summer School needs a name. Something 09:00–9:10 AM Housekeeping like ”GR4S”. We are keen to associate it with the GRSS. 9:10–10:30 Jocelyn Channusot (Hyperspectral Come up with a good idea, and we will reward you with imagery) a small prize. I will take suggestions until the 30th 10:30–10:50 Morning Tea (provided) June, 2013. 10:50–12:10 Phil Tickle (Developing Austra- lia’s high resolution coastal DEM IV. GRSS SUMMER SCHOOL 2013 MELBOURNE for sea level rise and coastal flood The Summer School Team in Melbourne have put together applications) an excellent programme, modelled closely on the success- 12:10–1:30 PM Lunch (on your own) ful Munich event. The key members of the team are Kim 1:30–2:50 John Richards (Radar) Lowell and Xiuping Jia, who have provided the information 2:50–3:10 Afternoon Tea (provided) for this section of the Education Corner. 3:10–4:10 Mihai Tanase and Rocco Panciera (Calibrating airborne radar for soil- A. DETAILS moisture extraction) What The IGARSS 2013 Summer School (actually a 4:10–4:20 Issue of certificate, photo, and Winter School since it is being held in the Aus- wrap-up. tralian winter) is a pre-conference event that provides an outstanding opportunity to obtain a V. FEATURED EDUCATION PROGRAMME broad knowledge of the fundamentals of a range I would like to establish a forum for members to present the of image data and analysis from leading Austra- Geoscience and Remote Sensing training programmes that lian and international authorities, and to meet they offer at their institution. There is a publication space lim- other attendees before the conference begins. itation, but I would like to make these quite comprehensive.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 81

You will see in the Editorial Plan further on, where I would Sometimes IGARSS is held earlier, so this might be the final like to place the material. This is a good opportunity to get update of the Summer School plans. some free publicity, but, more importantly, for the rest of us to learn about new techniques for training. B. JUNE 2013 ISSUE DEADLINE 09 APR. 2013 I would like to feature a university programme in Geosci- VI. EDUCATION CORNER EDITORIAL PLAN ence and Remote Sensing. I will approach members of Although we are already into 2013, here is the plan for the community to hear about their programmes in future 2013, in terms of content for the Education Corner. Please June issues. note the deadlines, and remember that I will need at least two weeks before that to merge the material into my col- C. SEPT. 2013 ISSUE DEADLINE 16 JULY 2013 umn. The Editor is necessarily strict on deadlines, so late Here I would like to discuss the format of training pro- breaking news may well have to move on to the next issue. grammes. Articles about existing programmes often I have included all the 2013 deadlines, since the pattern do not expose the underlying, pedagogical philosophy. will be similar for succeeding years. As soon as I have the Should we start thinking of establishing frameworks for dates for 2014 I will publish the new editorial plan. GRS courses? For example, our EE and ECE programmes often fall under the Washington Accord. Would it make A. MARCH 2013 ISSUE DEADLINE 17 JAN. 2013 sense to think of something similar? Contributions from This was covered by the previous Director, but it is members interested in discussing this will be featured. my intention to use the March issue of the Magazine to focus on funding opportunities associated with the D. DEC. 2013 ISSUE DEADLINE 15 OCT. 2013 upcoming IGARSS, especially in terms of travel sup- This will be an opportunity for all the educational activities port. Traditionally the results of the papers selections of the Society to be reviewed. The recent Summer School are available, and students will be applying, so this will be a flagship event, but I would like to hear about other information will be timely. successes. This issue will include a first call for sponsorship This issue will, in future, also give an update on plans requests for IGARSS travel funding. for the upcoming Summer School, and other training opportunities associated with our GRSS sponsored events. E. REGULAR REPORTS Every issue will carry information about: ◗ Recent Ph.D. Theses 2014 IEEE Radar Conference: ◗ Upcoming courses (please remember the deadlines) From Sensing to Information ◗ Funding opportunities (bursaries, stipends, postdoctoral). 19-23 May 2014 VII. RECENT PH.D. THESES Cincinnati, Ohio (USA) This is a good opportunity to showcase the work of your Cincinnati Marriott at RiverCenter research group. For publishing the PhD thesis information

you can contact Michael Inggs ([email protected])______or Dr. Lorenzo Bruzzone ([email protected]).______Ph.D. disserta- tions should be in the fields of activity of IEEE GRSS and should be recently completed. Please provide us with the General Chair: following: title of the dissertation, the students and advisors Prof. Brian Rigling – Wright State University names, the date of the thesis defense or publication, and a Technical Chair: link for downloading the electronic version of the thesis. Dr. Muralidhar Rangaswamy – US Air Force Research Lab VIII. CONCLUSION GRSS Liaison: Prof. Joel Johnson – The Ohio State University It will take a little while for the articles to settle down to a fixed format. I hope to receive feedback from the members, Abstract submission: 18 October 2013 especially in the form of input for the column. The regular (Up to 4 pages with figures) feature on Education innovation at different institutions is Author notification: 20 January 2014 important to all of us, and I will be approaching individuals Final papers: 21 February 2014 to make contributions. (Up to 6 pages with figures) We are also keen to hear opinions on education funding by GRSS: do we support students sufficiently to come Web Address: to our conference? What are the criteria we should use to http://www.radarcon2014.org grant funding? Digital Object Identifier 10.1109/MGRS.2013.2261366 GRS

82 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

ʹͲͳ͵ ͵ͲǦͶ ʹͲͳ͵ ǡ ǡ

Ǧ ʹͲͳ͵ Ǥ Ǧ Ǧ ǡ Ǥ Ȁǡ ǡ Ǥ ʹͷ Ǥ Ǥ ǡ

ǡ[email protected] ʹͺǡ ʹͲͳ͵Ǥ ǡ Ǥ ͵ͳǡ ʹͲͳ͵Ǥ ǡ

 ǡ ______ǣȀȀǤǦǤȀ Ǥ

Geoscience and Remote Sensing South Italy Chapter Geoscience and Remote Sensing Spain Chapter

Digital Object Identifier 10.1109/MGRS.2013.2261357

WOMEN IN GRS

GAIL SKOFRONICK JACKSON, GRSS Liaison to IEEE Women in Engineering

Unconscious Bias

nconscious bias can affect academic and work en- Other studies2 have shown that for postdoctoral Uvironments in ways that are still being discovered. fellowships women needed more publications than The Women in Geoscience and Remote Sensing Linke- men for the same competency ranking3. In terms of dIn Group has had several discussions on the topic of awards and according to the American Geophysical unconscious gender bias. What is unconscious bias? Union4, women tend to get nominated and awarded for It is a biased behavior, unknown and unrecognized early career awards (27%) and for service, teaching, by individuals, that likely results mentoring, and communication related awards (22%). from long-term experiences of In contrast, only about 11% of the senior scholarship cultural stereotypes. Studies1 have and research awards were given to women4. STUDIES HAVE SHOWN shown that both men and women What do we do about this unconscious bias? THAT BOTH MEN AND have a subtle unconscious nega- As was written by Ellen Druffel in 1994 in an WOMEN HAVE A SUBTLE tive bias of women. This bias can American Geophysical Union publication4, we need UNCONSCIOUS NEGATIVE manifest itself in hiring, mentor- to nominate worthy women for awards of all types, BIAS OF WOMEN. THIS ing, awards, and promotions. use genderless language, prioritize gender equity in BIAS CAN MANIFEST One relevant study involved awards, and increase the numbers and visibility of ITSELF IN HIRING, science faculty at United States of women in these fields. So the next time you review a America Universities1. The faculty candidate, proposal, or graduate student, think about MENTORING, AWARDS, were asked to evaluate potential how you might evaluate them if their name was Jennifer AND PROMOTIONS. “laboratory managers” essentially or John. If you are the applicant, consider using your an undergraduate student with a initials instead of a gender-identifying name. stated intention to go to graduate In closing, we look forward to providing school and offset costs by becoming a laboratory man- informative and interesting articles in future issues of ager. Resumes were exactly the same, except the names the GRSS Magazine. We welcome your suggestions were John or Jennifer. Faculty, both men and women, of material and topics for future columns. Please feel

at these universities were more likely to hire, mentor, free to contact us at [email protected].______and provide an increased starting salary for the “John” applicants1. GRS

1Moss-Racusin et al, Science faculty’s subtle gender biases favor male 2Handelsman et al., More Women in Science, 19 August 2005, Science, students, www.pnas.org/cgi/doi/10.1073/pnas.1211286109. www.sciencemag.org/cgi/content/full/309/5738/1190. 3Wenneras and Wold, Nature, 387,341, (1997). 4Holmes et al., Does Gender Bias Influence Awards Given by Societies? EOS Transactions, Vol 92, Number 47, November 2011. Digital Object Identifier 10.1109/MGRS.2013.2261260 4Druffel, Looking at gender distribution among AGU fellows, EOS Date of publication: 26 June 2013 Trans. AGU 75(39), 429, doi:10.1029/94EO01062.

84 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

5IJTQVCMJDBUJPOPčFSTPQFOBDDFTTPQUJPOTGPSBVUIPST *&&&0QFO"DDFTT1VCMJTIJOH

What does IEEE Open Access mean to an author? t5PQRVBMJUZQVCMJTIJOHXJUIFTUBCMJTIFEJNQBDUGBDUPST t*ODSFBTFEFYQPTVSFBOESFDPHOJUJPOBTBUIPVHIUMFBEFS t"DPOTJTUFOU*&&&QFFSSFWJFXTUBOEBSEPGFYDFMMFODF t6OSFTUSJDUFEBDDFTTGPSSFBEFSTUPEJTDPWFSZPVSQVCMJDBUJPOT t(SFBUXBZUPGVMmMMBSFRVJSFNFOUUPQVCMJTIPQFOBDDFTT

Learn more about IEEE Open Access Publishing: www.ieee.org/open-access

CONFERENCE REPORTS

JAMES L. GARRISON AND RASHMI SHAH, Purdue University, West Lafayette, IN

GNSS+R 2012 Workshop on Reflectometry Using GNSS and Other Signals of Opportunity http://www.gnssr2012.org

he Workshop on Reflectometry using GNSS and experiments, techniques, applications and mission TOther Signals of Opportunity (GNSS+R 2012) was concepts. 2) Meet as a community to define develop- held at Purdue University, West Lafayette, IN, USA from ment roadmaps and create advocacy for the future October 10 to 11, 2012. IEEE Geoscience and Remote support of reflectometry. 3) Make the broader Earth Sensing Society was a technical co-sponsor of this event, sciences community aware of the potential of reflec- along with NASA and the International Association of tometry measurements. Geodesy (IAG). Planning for GNSS+R 2012 began two The first goal was met through the organization of years ago at the conclusion of GNSS-R 2010, in Barce- seven sessions of oral papers, covering; ocean applica- lona, Spain. James Garrison, Associate Professor in the tions (2 sessions), land and cryosphere applications School of Aeronautics and Astronautics at Purdue, was (2 sessions), Instruments, Models and Signal Process- the GNSS+R 2012 Conference Chair. He was assisted by ing, and Missions. Generally, this organization put ses- Dr. Stephen Katzberg (NASA Distinguished Research sions having a focus on science and applications in the Associate) and Prof. Scott Gleason (Concordia Univer- first day and those focused on technology development sity) who served as Technical Program Chair and Publi- in the second day. A poster session was also held after cation Chair, respectively. lunch on the first day. In total, 28 oral presentations and Reflectometry refers to the re-utilization of exist- 6 posters were given. 46 people registered for the con- ing digital signals for remote sensing. Recognizing the ference, which attracted an international audience (18 recent progress that has been made in this emerging attendees (39%) arrived from outside North America). field, three goals were proposed for GNSS+R 2012: 1) There was also a diversity of affiliations, representing Provide a rigorous, peer-reviewed, forum for technical academia (48%), the government/public sector (41%) interchange of new findings in reflectometry theory, and commercial entities (11%).Presenters were offered the opportunity to make their presentation materi-

Digital Object Identifier 10.1109/MGRS.2013.2260935 als available on the conference web site (http://www. Date of publication: 26 June 2013 gnssr2012.org/)______and to submit optional proceedings

GNSS+R 2012 attendees in front of the Neil Armstrong statute at Purdue University.

86 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

papers for publication through IEEE Xplore (http://ieeexplore.ieee.org). Several student travel grants were awarded under NASA Grant NNX12AJ37G, which provided support for 11 students to attend the workshop and present papers. The field of reflectometry began just under 20 years ago, with the publication of a paper by Dr. Manuel Martín-Neira (ESA) on the potential of satellite altimetry using reflected Global Navigation Satellite Sys- tem (GNSS) signals [1]. In 1997, the first experiment demonstrating the sensitivity of reflected GNSS signals to ocean winds was conducted [2]. In the decade that followed, GNSS reflectometry (“GNSS-R”) has been 28 oral presentations and 6 posters were given. used to sense soil moisture, ocean altimetry, and ice properties. The UK-DMC satellite, launched in conference acronym of “GNSS+R” has been defined. The 2003, collected the first organized set of GNSS reflections “+” indicates an expansion to signals potentially available from orbit [3]. In 2011, NASA selected the CYGNSS mis- in nearly all microwave frequencies, from approximately sion under the Earth-Ventures 2 solicitation, representing 400 communication satellites presently orbiting the Earth. the first dedicated orbital demonstration of reflectometry A town-hall meeting was conducted after the conclu- applied to Earth observation. CYGNSS, led by the Univer- sion of the technical sessions. This meeting was held to sity of Michigan, will use GNSS-R measurements from a address the second and third goals of the workshop; define constellation of 8 nanosatellites, to observe tropical cyclone a roadmap and develop advocacy among the reflectom- development. etry community, and broaden awareness of reflectometry GNSS signals are particularly well adapted to reflec- with the wider Earth sciences community. Prof. Garrison tometry, a result of the pseudorandom noise used for delay chaired the town hall. Prof. Kristine Larson (University of estimation. GNSS-R has some limitations, however, due Colorado) and Dr. Derrek Burrage (Naval Research Labo- to low transmission power and L-band spectrum alloca- ratory, Stennis Space Center)reported on the discussions. tion for satellite navigation signals. Recently, reflectometry Two questions were posed to the participants for initial methods have been applied to digital signals from com- discussion; “What should be the priorities within NASA’s munication satellites, potentially expanding the utility of Earth Science program for GNSS+R?” and “What should this measurement to all microwave bands that penetrate be the priorities within NASA’s Technology program for the Earth’s atmosphere, at signal powers exceeding the GNSS+R?”. With this starting point, the resulting discus- power of GNSS by up to 30 dB. In order to recognize this sions were somewhat broad ranging, covering topics such burgeoning new class of measurements, while maintaining as; participation on NASA science teams, data products and a connection with the extensive heritage of GNSS-R, the accuracy thresholds for scientifically useful measurements,

A full-size model of the CYGNSS spacecraft was on display during the poster session.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 87

and trade studies of hosted secondary payloads vs. dedi- reflectometry experiments underway at the Agronomy cated constellations. A report of this town hall meeting has Center for Research and Education (ACRE). The two groups been submitted to the AGU Eos newspaper. then combined for a tour of the clean rooms at the Birck After the town hall, the process for selecting a venue for Nanotechnology Center before returning to Armstrong GNSS+R 2014 was discussed. It was decided to form a steer- Hall for a visit to Prof. Garrison’s Radionavigation Lab. ing committee to evaluate proposals for the next meeting, Rashmi Shah, another doctoral student of Prof. Garrison, reporting to the IEEE-GRSS. A meeting of Sub-commission provided an overview of GNSS+R and navigation research 4.6 of the IAG was held following the town hall. conducted in that laboratory. After the conclusion of GNSS+R 2012, many of the participants remained to join the hosts for one of two tours References of research facilities at Purdue. One tour group focused [1] M. Martín-Neira, “A passive reflectometry and interferometry on facilities at the Purdue University Airport. Graduate system (PARIS): Application to ocean altimetry,” ESA J.,vol. 17, student Oscar Garibaldi showed some of his UAV designs pp. 331–355, 1993. and the extensive development test facilities at the Aero- [2] J. L. Garrison, S. J. Katzberg, and M. I. Hill, “Effect of sea rough- space Sciences Laboratory. The tour then continued on to ness on bistatically scattered range-coded signals from the Glob- one of the hangars in the Department of Aviation Technol- al Positioning System,” Geophys. Res. Lett., vol. 25, no. 13, pp. ogy, where Professor Paul Shepson, director of the Purdue 2257–2260, July 1998. Climate Change Research Center, showed the Airborne [3]S.Gleason,S.Hodgart,Y.Sun,G.Gommenginger,S.Mackin, Laboratory for Atmospheric Research (ALAR). ALAR is a M. Adjrad, and M. Unwin, “Detection and processing of bistati- Beechcraft Duchess aircraft modified for in situ sensing of cally reflected GPS signals from low earth orbit for the purpose the atmosphere. The second tour, led by Zenki Lin, one of of ocean remote sensing,” IEEE Trans. Geosci. Remote Sensing,vol. Prof. Garrison’s graduate students, showcased soil-moisture 43, no. 6, pp. 1229–1241, June 2005.

Africa’s Largest Association, AARSE Had Its 9th Conference in El Jadida, Morocco

DR. TSEHAIE WOLDAI

nder the High Patronage of His Majesty King Moham- Honorable Minister of Higher Education and Scientific Umed VI, the African Association of Remote Sensing of Research of Morocco. The Honorable Guest highlighted the the Environment (AARSE), had its 9th International Con- benefits of geoinformation for good governance and the ference in El Jadida, Morocco from October 29 to Novem- ongoing developments in this field in Morocco. The honor- ber 2, 2012. The Conference was attended by 540 registered able guest accompanies by other Government officials then participants from over 45 countries. went to open the scientific exhibitions. With the word of Welcome by Prof. Oladjide Kufoniyi, The Conference featured: the President of AARSE, Dr. Kamal Lebassi, the Chairman ◗ Over 300 presentations and 110 posters, 10 keynote of Local Organizing Committee, congratulatory messages addresses in 4 plenary sessions, 31 technical sessions rep- followed from Presidents and representatives of partner resenting eight sub-themes on various aspects of Geoin- organizations: ISPRS, IEEE-GRSS, GEO, EARSel, EIS-Africa formation Sciences and Earth Observation, two poster and the UN-ECA. This was followed by Prof. Boumedian session, Various Workshops, four Special sessions, includ- Tannouti, President of Chouaib Doukkali University who ing the AARSE Executive Council and, AARSE General As- welcomed all participants and introduced the Governor sembly, Three Round tables and many business meetings. of El Jadida to deliver his good-will message to the partici- ◗ A total of 19 sponsors, comprising mainly of Gold pants. The opening speech was given by His Excellency the sponsors (GEO & South African GEO and ESRI; Silver sponsors: OCP, European Space Agency, Trimble, the Nigerian National Space Research and Development Digital Object Identifier 10.1109/MGRS.2013.2261655 Date of publication: 26 June 2013 Agency-NARSDA and ASTRIUM), Bronze sponsors

88 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

(SURREY, the Surveyors Council of Nigeria [SURCON], It is the highest award given by AARSE and in its 20 years European Space Imaging and The French Embassy in history has been given only to three distinguished sci- Morocco) and Moroccan Sponsors (Ministry of Foreign entists and one institution. In giving this award to Dr. Affairs of Morocco, Province of El Jadida, Conseil Ré- Woldai, the Prof. Kufoniyi gionale Abda-Doukkala, Conseil Provinciale El Jadida, recognized the dedication Conseil Municipal El Jadida, Centre National de la Re- and importance played by cherche Scientifique et Technique [CNRS] and Ministère Dr. Woldai in nurturing and AARSE, WITH MORE de L’Enseignement Supèrieur de la Recherche Scienti- shaping the association to THAN 1800 INDIVIDUAL fique et de la Formation des Cadres). be one of the most dynamic MEMBERS AND OVER ◗ A total of 20 exhibitors (ESRI, RCMRD-Kenya, Astri- and respected institution in 40 NATIONAL\ um-France, SA-GEO-Switzerland, Solution Mapping- Africa. Dr. Woldai was the INTERNATIONAL UK, Trimble-USA, GeoSystem-France, Vito-Belgium, founder of the defunct Afri- ORGANIZATIONAL CRTS-Morocco, SURREY-UK, AARSE-SA, SFPT-France, can Remote Sensing Society MEMBERS IS THE SACANEX-Russia, IEEE-GRSS-Australia, Satpalda- (ARSS) in 1982 and served LARGEST SCIENTIFIC India, Southern Mapping-SA, EIS-Africa-SA, RapidEye- as its most vocal representa- Germany and GEOEye-USA. tive in international forum ORGANIZATION IN ◗ 10 Keynote Speakers: T. Woldai (AARSE EC Member), until 1988. With its abolition THE FIELD OF REMOTE Ms. Aida Opoku Mensah—Director ISTD, UNECA, Dr. in 1990, he founded AARSE SENSING AND GIS Barbara J. Ryan—GEO Director, Prof. Tony Milne—Past in August 1992 in Colorado, APPLICATIONS IN AFRICA. Pres. IEEE-GRSS, Dr. Seidu Mohammed [Director Gen- USA and served the Associa- eral, NASRDA], Dr. Mecheline Tabache [Administratoe tion as its Secretary General, ESA], Prof. Laurent Polidori [President of SEPT], Dr. Jo- President (2004-2010) and han Stessens [Manager, RS Dept, Vito], Prof. Massimo immediate Past President (2010-current). Mementi [Delft Univ. of Technology] and Dr. Driss El ◗ Furthermore, after the opening keynote speech, Dr. Hadani [Director, CNRS]. Woldai was decorated with the ESRI “Making a Dif- ◗ The conference included three pre/post workshops and ference in Africa” Award. The latter is offered once three side events associated with the EU-GMES, GEO a year by the President of ESRI to honor personality and ESRI programmes, an ice-breaker and a gala dinner from around the world who have impacted others and offered by AARSE. advanced and promoted geoinformation technology ◗ The Conference involved several awards: One AARSE through their work. Dr. Salim Sawaya, in presenting highest achievement Award, five Presidential Citation the award on behalf of ESRI President Dr. Jack Dan- Awards, Seven AARSE-IEEE/GRSS Travel Fellowship germond, emphasized the commitment and dedica- Award, Three AARSE-European Space Agency Award for tion shown by Dr. Woldai, in promoting Geoinforma- best papers and posters, one ESRI “Making a Difference tion technology in Africa and for making a difference in Africa” award and one AARSE Certificate to the LOC. to many through his teaching. The nomination of Two Major Awards to Dr. Tsehaie Woldai: Dr. Woldai to the ESRI Award came from 79 distin- ◗ In the opening ceremony, AARSE President Prof. Kufonyi guished scientists and practitioners from 56 African presented the AARSE Achievement Award to Dr. Woldai. and Europe institutions.

ESRI—Federal/Global Affairs Account Manager, Dr. Salim Sawaya, Professor Kufoniyi presents the AARSE Achievement Award to TW presenting “making a difference in Africa” Award on behalf of ESRI on the opening ceremony. President Dr. Jack Dangermond.

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 89

HIGHLIGHTS OF THE CONFERENCE

Opening of the Exhibition by, His Excellency Mouaad Jamai, the Governor of El Jadida, Morocco.

A small portion of the exhibition booth during the conference.

90 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

IEEE-GRSS STAND THE 9TH AARSE CONFERENCE IN EL JADIDA, MOROCCO

IEEE-GRSS Stand with Michael Inggs, Tony Milne and Charles Luther and student registering to become IEEE-GRSS members.

President Citation Award to Dr. Charles Luther handed by the Association President Prof. Oladjide Kufoniyi for his outstanding work in fostering closer collaboration between IEEE-GRSS and AARSE.

Photo representing the winners of the IEEE-GRSS & AARSE Travel Fellowship Award with councilors of both Societies.

GRS

JUNE 2013 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 91

GRSS MEMBER HIGHLIGHTS

GRSS Members Elevated to the Grade of Senior Member in February 2013

◗ An attractive fine wood and bronze engraved Senior Nuria Duffo, Spain Section Member plaque to proudly display. Matthew Easley, Buenaventura Section ◗ Up to $25.00 gift certificate toward one new Society Attilio Gambardella, Benelux Section membership. Weimin Huang , Newfoundland-Labrador Section ◗ A letter of commendation to your employer on the Xiaoying Jin, Denver Section achievement of Senior Member grade (upon the re- Lim Peter, Boon-Lum, Singapore Section quest of the newly elected Senior Member). Jordi Mallorqui, Spain Section ◗ Announcement of elevation in Section/Society and/ Thomas Meissner, San Francisco Section or local newsletters, newspapers and notices. Shiv Mohan, Gujarat Section ◗ Eligibility to hold executive IEEE volunteer positions. Nazzareno Pierdicca, Italy Section ◗ Can serve as Reference for Senior Member applicants. Merce Vall-Llossera, Spain Section ◗ Invited to be on the panel to review Senior Member Stephen Volz, Washington Section applications. Ingo Walterscheid, Germany Section ◗ Eligible for election to be an IEEE Fellow. Applications for senior membership can be obtained from IEEE website: https://www.ieee.org/membership_

Senior membership has the following distinct benefits: ______services/membership/senior/application/index.

◗ The professional recognition of your peers for techni- ___html. You can also visit the GRSS website: ____http:// cal and professional excellence. www.grss-ieee.org.

Digital Object Identifier 10.1109/MGRS.2013.2260938 Date of publication: 26 June 2013 GRS

92 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

YOU KNOW YOUR STUDENTS NEED IEEE INFORMATION. NOW THEY CAN HAVE IT. AND YOU CAN AFFORD IT.

IEEE RECOGNIZES THE SPECIAL NEEDS OF SMALLER COLLEGES, and wants students to have access to the information that will put them on the path to career success. Now, smaller colleges can subscribe to the same IEEE collections that large universities receive, but at a lower price, based on your full-time enrollment and degree programs. Find out more–visit www.ieee.org/learning

CALENDAR

See also WWW.IEEE.ORG/CONFERENCES_EVENTS/INDEX.HTML______or WWW.TECHEXPO.COM/EVENTS______

2013 7TH INTERNATIONAL WORKSHOP AUGUST ON THE ANALYSIS OF MULTI- THE SECOND INTERNATIONAL JUNE TEMPORAL REMOTE SENSING CONFERENCE ON 33RD EARSEL IMAGES (MULTITEMP 2013) AGRO-GEOINFORMATICS SYMPOSIUM “TOWARDS June 25–27, 2013 (AGRO-GEOINFORMATICS 2013) HORIZON 2020: EARTH Banff, Canada August 12–16, 2013 OBSERVATION AND SOCIAL http://geog.ucalgary.ca/multitemp2013 Fairfax, Virginia, USA

PERSPECTIVES” E-mail: ______info@agro-geoinformatics2013. June 3–6, 2013 5TH WORKSHOP ON org__ Matera, Italy HYPERSPECTRAL IMAGE http://www.agro-geoinformatics2013.

E-mail: [email protected] AND SIGNAL PROCESSING org/__ http://www.earsel.org/ (WHISPERS 2013)

symposia/2013-symposium-Matera/______June 25–27, 2013 SEPTEMBER http://www.earsel.org/welcome.html Florida THE ASIA-PACIFIC CONFERENCE http://www.ieee-whispers.com/ ON SYNTHETIC APERTURE THE 21ST INTERNATIONAL RADAR (APSAR 2013) CONFERENCE ON 7TH INTERNATIONAL September 23–27, 2013 GEOINFORMATICS ASSOCIATION FOR CHINA Tsukuba, Japan (GEOINFORMATICS 2013) PLANNING CONFERENCE http://www.apsar2013.org/ June 20–22, 2013 (IACP 2013) Kaifeng, China June 29–July 1, 2013 OCTOBER E-mail: Geoinformatics2013@______Shanghai, China URSI COMMISSION F, gmail.com http://www.chinaplanning.org/conf/ MICROWAVE SIGNATURES

http://www.GeoInformatics2013.org index.php/iacp_7th/7thIACP_shanghai______2013 SPECIALIST SYMPOSIUM ON MICROWAVE REMOTE 6TH INTERNATIONAL JULY SENSING OF THE EARTH, CONFERENCE ON GI_FORUM 2013—CREATING OCEANS, AND ATMOSPHERE RECENT ADVANCES IN THE GISOCIETY (URSI-F 2013) SPACE TECHNOLOGIES July 2–5, 2013 October 28–31, 2013 (RAST 2013) Salzburg, Austria Espoo (Helsinki), Finland

June 12–14, 2013 E-mail: [email protected]______http://frs2013.ursi.fi/ Istanbul, Turkey http://www.gi-forum.org/ E-mail: [email protected] 2014 http://www.rast.org.tr/ INTERNATIONAL GEOSCIENCE AND REMOTE SENSING MAY INTERNATIONAL CONFERENCE SYMPOSIUM (IGARSS 2013) IEEE RADAR CONFERENCE: ON EARTH OBSERVATION July 21–26, 2013 FROM SENSING TO AND SOCIAL IMPACTS Melbourne, Australia INFORMATION

(ICEO&SI 2013) E-mail: [email protected] May 19–23, 2014 June 23–25, 2013 http://www.igarss2013.org/ Cincinnati, Ohio, USA Tainan, Taiwan http://www.radarcon2014.org http://2013.iceo-si.org.tw/ GRS

Digital Object Identifier 10.1109/MGRS.2013.2261261 Date of publication: 26 June 2013

94 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Proceedings of the IEEE: Pioneering technology from the inside out.

At Proceedings of the IEEE, we don’t want you to just read about technology. We want you to understand emerging breakthroughs—from beginning to end, from the inside out.

That’s why we give you multi-disciplinary technology coverage that explains how key innovations evolve and impact the world. Every month, you’ll ﬁ nd the comprehensive, in-depth research that only IEEE can provide.

Understand technology from every angle—subscribe today. www.ieee.org/proceedings

AD INDEX

The Advertisers Index contained in this issue is compiled as a service to our readers and advertisers: the publisher is not liable for errors or omissions although every effort is made to ensure its accuracy. Be sure to let our advertisers know you found them through IEEE Geoscience and Remote Sensing Magazine.

IEEE Marketing Department CVR 4 James A. Vick Midwest/Central Canada New England/ www.ieee.org/digitalsubscriptions Sr. Director, Dave Jones Eastern Canada Advertising Phone: +1 708 442 5633 Ilia Rodriguez +1 212 419 7767; Fax: +1 708 442 7620 Phone: +1 212 419 7768 Fax: +1 212 419 7589 [email protected]______Fax: +1 212 419 7589 [email protected] IL, IA, KS, MN, MO, NE, [email protected]______ND, SD, WI, OH ME, VT, NH, MA, RI Marion Delaney Canada: Manitoba, Canada: Quebec, Advertising Sales Saskatchewan, Alberta Nova Scotia, Director Newfoundland, +1 415 863 4717 Midwest/Ontario, Prince Edward Island, Fax + 1 415 863 4717 Canada New Brunswick [email protected]______Will Hamilton Phone: +1 269 381 2156 Southeast Susan E. Schneiderman Fax: +1 269 381 2556 Cathy Flynn Business Development [email protected] Phone: +1 770 645 2944 Manager IN, MI. Canada: Ontario Fax: +1 770 993 4423 +1 732 562 3946; [email protected]______Fax: +1 732 981 1855 West Coast/ VA, NC, SC, GA, FL, AL, [email protected] Mountain States/ MS, TN Western Canada PRODUCT Marshall Rubin Midwest/South Central/ ADVERTISING Phone: +1 818 888 2407 Central Canada MidAtlantic Fax: +1 818 888 4907 Darcy Giovingo Lisa Rinaldo [email protected]______Phone: +1 847 498 4520 Phone: +1 732 772 0160 AZ, CO, HI, NM, NV, Fax: +1 847 498 5911 Fax: +1 732 772 0164 UT, AK, ID, MT, WY, OR, [email protected][email protected] WA, CA AR, IL, IN, IA, KS, LA, NY, NJ, PA, DE, MD, DC, Canada: British Columbia MI, MN, MO, NE, ND, KY, WV SD, OH, OK, TX, WI Europe/Africa/ Canada: Ontario, New England/ Middle East/ Asia/Far Manitoba, Saskatchewan, South Central/ East/Pacific Rim Alberta Eastern Canada Heleen Vodegel Jody Estabrook Phone: West Coast/Southwest/ Phone: +1 774 283 4528 +44 1875 825 700 Mountain States/Asia Fax: +1 774 283 4527 Fax: +44 1875 825 701 Tim Matteson [email protected] [email protected] Phone: +1 310 836 4064 ME, VT, NH, MA, RI, CT, Europe, Africa, Middle Fax: +1 310 836 4067 AR, LA, OK, TX East, Asia, Far East, [email protected] Canada: Quebec, Pacific Rim, Australia, AZ, CO, HI, NV, NM, UT, Nova Scotia, New Zealand CA, AK, ID, MT, WY, Newfoundland, OR, WA Prince Edward Island, RECRUITMENT Canada: British Columbia New Brunswick ADVERTISING MidAtlantic Europe/Africa/ Southeast Lisa Rinaldo Middle East Thomas Flynn Phone: +1 732 772 0160 Heleen Vodegel Phone: +1 770 645 2944 Fax: +1 732 772 0164 Phone: Fax: +1 770 993 4423 [email protected] +44 1875 825 700 [email protected]______NY, NJ, CT, PA, DE, MD, Fax: +44 1875 825 701 VA, NC, SC, GA, FL, AL, DC, KY, WV [email protected] MS, TN Europe, Africa, Middle East

Digital Object Identifier 10.1109/MGRS.2013.2261354

96 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2013

Members share fascinating ﬁ rst-person stories of technological innovations. Come read and contribute your story.

IEEE Global History Network www.ieeeghn.org______

Instant Access to IEEE Publications

Enhance your IEEE print subscription with “IEEE is the umbrella that online access to the IEEE Xplore® digital library. allows us all to stay current with technology trends.”

Q Download papers the day they are published Dr. Mathukumalli Vidyasagar Head, Bioengineering Dept. Q Discover related content in IEEE Xplore University of Texas, Dallas

Q Signiﬁ cant savings over print with an online institutional subscription

Start today to maximize your research potential.

Contact: [email protected]______

______www.ieee.org/digitalsubscriptions