Dr. Jie-Bang Stephen Yan Co-PI

Two Papers:  Direction-of-Arrival Analysis of Airborne Ice Depth Sounder Data  UAS-Based Radar Sounding of the Polar Ice Sheets IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 4, APRIL 2017 2239 Direction-of-Arrival Analysis of Airborne Ice Depth Sounder Data Ulrik Nielsen, Jie-Bang Yan, Member, IEEE, Sivaprasad Gogineni, Fellow, IEEE, and Jørgen Dall, Member, IEEE

Abstract— In this paper, we analyze the direction-of- determine the boundary conditions of the ice-sheet models. arrival (DOA) of the ice-sheet data collected over Jakobshavn Basal conditions largely impact on the ice flow velocity, and Glacier with the airborne Multichannel Radar Depth Sounder therefore, precise knowledge of them is especially important (MCRDS) during the 2006 field season. We extracted weak ice– bed echoes buried in signals scattered by the rough surface of the for the estimation of the mass balance [6]. fast-flowing Jakobshavn Glacier by analyzing the DOA of signals received with a five-element receive-antenna array. This allowed A. Multiphase-Center-Based Radar Ice Sounding us to obtain ice thickness information, which is a key parameter when generating bed topography of glaciers. We also estimated The weak nadir radar signals from the ice–bed interface ice–bed roughness and bed slope from the combined analysis of are often masked by off-nadir surface clutter, signals scattered the DOA and radar waveforms. The bed slope is about 8° and from extremely rough crevassed surfaces in ice-sheet margins. the roughness in terms of rms slope is about 16°. Synthetic aperture radar (SAR) processing can be used to Index Terms— Airborne radar, direction-of-arrival (DOA) suppress surface clutter in the along-track direction, but it is estimation, glacier, ice sounding, radar remote sensing, surface ineffective in reducing the across-track clutter. Large across- scattering. track antenna arrays can be used to obtain a narrow across- track antenna beam to suppress surface clutter in this direction. I. INTRODUCTION At the same time, to avoid excessive attenuation of the ATELLITE observations show that both the Greenland and signals reflected within the ice, radars are normally operated SAntarctic ice sheets are losing mass [1], [2]. Most of the in the very high frequency (VHF) part of the electromag- ice loss is occurring around ice-sheet margins and through fast- netic spectrum. The long wavelengths in this band require flowing glaciers [3]. Although satellites provide much-needed large antenna dimensions to obtain an antenna beam that information on ice-surface elevation, surface velocity, and total is sufficiently narrow to reduce across-track surface clutter. mass, there is currently no satellite-based sensor that is able to Such large antenna dimensions cannot be accommodated on measure ice thickness. Bed topography and basal conditions airborne platforms, and additional clutter suppression is, there- for areas losing ice are needed to improve ice-sheet models. fore, needed to compensate for these limitations. The current These models are essential to predicting the response of the research in this field is based on multichannel systems com- ice sheets to a warming climate. One of the key parameters bined with advanced coherent postprocessing of data. By using needed is ice sheet thickness, which can be extracted using multichannel-receivers to sample array elements individually, radar depth-sounding techniques [4], [5]. In addition, we are beamforming techniques can be utilized to synthesize adaptive interested in the basal conditions of the ice sheets as they antenna patterns that suppress the surface clutter from specific off-nadir angles, while a high gain is maintained in the nadir Manuscript received January 14, 2015; revised July 29, 2015, direction [7]. February 28, 2016 and September 10, 2016; accepted October 9, 2016. Date of publication January 16, 2017; date of current version February 24, B. DOA Estimation in Radar Ice Sounding 2017. This work was supported by the National Science Foundation under Grant ANT0424589. In addition to beamforming, the multiphase-center sys- U. Nielsen was with the Department of Microwaves and Remote Sensing, tems also provide the opportunity to perform direction-of- National Space Institute, Technical University of Denmark, DK-2800 Kongens Lyngby, Denmark. He is now with IHFood A/S, DK-1577 Copenhagen, arrival (DOA) estimation of the different signal components Denmark (e-mail: [email protected]). within the received returns. In relation to ice sounding, early J.-B. Yan was with the Center for Remote Sensing of Ice Sheets, The Univer- studies on airborne InSAR in [8] can be seen as a precursor sity of Kansas, Lawrence, KS 66045 USA. He is now with the Department of Electrical and Computer Engineering, The University of Alabama, Tuscaloosa, to DOA estimation. A ground-based radar configuration was AL 35487 USA (e-mail: [email protected]). used in [9] to perform actual DOA estimates of the bed return. S. Gogineni is with the Center for Remote Sensing of Ice Sheets, The In [10], DOA data are used as the primary data product to University of Kansas, Lawrence, KS 66045 USA (e-mail: [email protected]). J. Dall is with the Department of Microwaves and Remote Sensing, National produce swath measurements of both the ice surface and the Space Institute, Technical University of Denmark, DK-2800 Kongens Lyngby, bedrock topography. This paper is the first published work on Denmark (e-mail: [email protected]). DOA estimation applied to airborne ice sounding data. The Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. results reported in [10] are based on the data acquired by Digital Object Identifier 10.1109/TGRS.2016.2639510 the Multichannel Radar Depth Sounder (MCRDS) developed 0196-2892 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. 2240 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 4, APRIL 2017 by the Center for Remote Sensing of Ice Sheets at the University of Kansas. The radar system is, in this experiment, operated in ping-pong mode to provide 12 effective receive phase centers. Estimation of the DOA angles of the surface clutter and bed return is used to compute relative elevations in slant-range geometry, followed by a mapping to ground range to obtain the topographic map in Cartesian coordinates. DOA estimation based on the data acquired with an upgraded version of the system, Multichannel Coherent Radar Depth Sounder/Imager (MCoRDS/I) [11], has been used to support the investigation of the bed topography of more glaciers, including Jakobshavn [12]. In [13], DOA estimation has been applied to data acquired with the four-channel POLarimetric Airborne Radar Ice Fig. 1. Photograph showing the five-element subarray of folded dipole Sounder (POLARIS) [14] developed by the Technical Univer- elements mounted under the right wing of the Twin-Otter aircraft. sity of Denmark, to improve the performance of surface clutter suppression techniques. The DOA angles of the surface clutter are estimated and used to optimize the synthesis of the antenna patterns for improving clutter suppression. Recently, DOA estimation based on POLARIS data is used to show an along-track variation of the effective scattering center of the surface return caused by a varying penetration depth [15], which directly provides glaciological information. In this paper, we present further applications of the DOA estimation technique for radar ice sounding. We used MCRDS multiphase-center data collected over Jakobshavn Glacier during the 2006 Greenland field season to convert radar echograms into a DOA representation. With this representation of the radar data, we were able to detect some of the most challenging parts of the bed along the channel of the fastest Fig. 2. Flight track (red) over the Jakobshavn Glacier at the west coast of flowing glacier on the earth. A model-based approach was then Greenland in the 2006 field season. The blue line corresponds to the location of the glacier channel. The flight track corresponds to frame 5, segment 4 in used to interpret the DOA estimation of the bed return. Further the data set acquired May 30, 2006. analysis showed that the backscattering characteristics of the ice–bed could be estimated by combining the DOA data and folded dipoles mounted in the across-track direction. The the radar waveform data. Based on the data, the across-track array was divided into two five-element subarrays installed slope of the bed was estimated as a fitted model parameter. under each wing, as shown in Fig. 1. The left wing subarray Finally, information on the bed roughness in terms of the rms was used for transmission and the right for reception. All slope was obtained by forward modeling using the incoherent elements in the transmit array were excited with uniform μ Kirchhoff model (IKM). weights during transmission. The pulse length was 10 s with a total transmit power of 800 W. A multichannel receiver was C. Paper Outline used to sample signals from each receive-antenna element This paper is organized as follows. Section II provides individually. The spacing of the effective phase centers was . λ λ details on the MCRDS system and the associated data set. approximately 0 3 ,where is the wavelength in the free A signal model is presented in Section III along with algo- space of the center frequency. rithms for DOA estimation. In Section IV, the algorithms are Data acquired with the MCRDS system in 2006 at the applied to data and used to provide an alternative representa- Jakobshavn Glacier were used for the DOA analysis. The tion based on DOA. This representation is used for detection data were acquired according to the flight track shown in of the bed in Section V and for retrieval of its backscat- Fig. 2. Results for a segment perpendicular to the ice flow tering characteristics in Section VI. Finally, in Section VII, are presented in this paper. The segment was flown northward we summarize and conclude this paper. and is highlighted in red in Fig. 2. The segment represents a strong complex clutter scenario with high attenuation that is II. SYSTEM AND DATA DESCRIPTION difficult to sound. This scenario is well suited for illustrating MCRDS [16] is a high-sensitivity radar system developed the capabilities of the proposed methods. The altitude of the for the collection of ice-sheet data. During the 2006 Greenland flight track is approximately 270 m above the ice surface. field mission, MCRDS was installed on the DHC-6 Twin-Otter aircraft from de Havilland Canada Ltd. and was operated A. Signal Processing at 150 MHz with a bandwidth of 20 MHz. The system was A linear frequency-modulated chirp was used for transmit- effectively configured with a ten-element antenna array of ted pulses to employ pulse compression. The received data NIELSEN et al.: DOA ANALYSIS OF AIRBORNE ICE DEPTH SOUNDER DATA 2241 were compressed using a matched filter with a frequency- the Q signals, and s(t) is a vector collecting the Q signal domain Hanning window to suppress range sidelobes. components at time t,thatis SAR processing was used to improve the along-track s(t) =[s (t) ... s (t)]T. (5) resolution by synthesizing a long aperture. The frequency- 1 Q wavenumber (F-K ) focusing algorithm that exploits the fast The steering matrix A is a function of the DOA vector , Fourier transform for computational efficiency was used for which contains the Q DOA angles. processing. A single time instance of x is denoted a snapshot. By using pulse compression and SAR processing, a nominal A collection of M snapshots acquired at time instances . resolution in range and azimuth of 7 5m (50ns)and5m, t1,...,tm can be modeled as respectively, was obtained. X = A()S + E (6)

III. DIRECTION-OF-ARRIVAL ESTIMATION where X and E are N × M matrices, A is N × Q,andS is × Several algorithms for DOA estimation exist. They Q M. Each column in X, S,andE corresponds to a spe- include the well-established MUltiple SIgnal Classification cific snapshot. For further details regarding the signal model, (MUSIC) [17] and maximum likelihood (ML) [18] algorithms. see [10] and [19]. Both of these algorithms have superresolution capabilities Before we move on to a review of MUSIC and ML, we first and other desirable properties, such as statistical consistency define the sample covariance matrix as and high accuracy in adverse situations, such as low SNR 1 M scenarios. Due to this as well as their applications in a R = x˜ (t )x˜ H(t ) (7) M m m number of fields, MUSIC and ML are the algorithms chosen m=1 for the study in this paper. Within the field of radar ice where (·)H is the Hermitian transpose and x˜ is a measured sounding, the algorithms have previously been applied a few array sample corresponding to the signal model from (3). times for different purposes. In [9], MUSIC has been applied In this way, the covariance matrix is estimated as an average to data acquired with a ground-based radar depth sounder over a given set of snapshots. In this paper, the snapshots are configuration, while ML has been applied to data from the extracted as a number of consecutive samples in azimuth— airborne experiments in [10], [13], and [15]. all at the same given range gate. We will now briefly describe the array signal model that is the basis for both algorithms. B. Multiple Signals Classification MUSIC exploits the eigendecomposition of R,thatis A. Signal Model =  H The signal received at time t by the N array sensors can be R U U (8) expressed in vectorial form as where  is a diagonal matrix containing the N eigenvalues x(t) = a(θ)s(t) + e(t) (1) of R,andU is an orthonormal basis consisting of the corresponding eigenvectors. where x(t) is an N × 1 vector, s(t) is the complex echo The DOA estimates are determined as the Q highest peaks signal at a reference sensor, e(t) is an additive Gaussian noise of the so-called MUSIC-spectrum [17] given by (θ) component, and a is the so-called array transfer vector 1 (or steering vector). This vector describes the phase shift at PMU(θ) = (9) aH(θ)U UHa(θ) each of the sensors corresponding to the interelement time n n delays determined by the array geometry and the given DOA, θ where Un is the subset of eigenvectors in U that corresponds   to the N − Q smallest eigenvalues. The subspace spanned − ω τ − ω τ T (θ) = (θ) j c 1 ... (θ) j c N a H1 e HN e (2) by Un is known as the noise subspace. T where (·) is the transpose operator, ωc is the center angular frequency, and τn is the time delay at the nth sensor relative C. Maximum Likelihood to an arbitrary reference sensor. Equation (2) also takes into The ML solution [18] of the DOA vector can be account the sensor transfer functions, Hn(θ). expressed as By applying the superposition principle to (1), Q simul- −  = min tr[A()(AH()A()) 1 AH()R] (10) taneously received echo signals with different DOAs can be ML  : described in the following way where tr[·] is the trace of the bracketed matrix. The ML esti- x(t) = A()s(t) + e(t) (3) mator with the assumption of Q signal components involves a computationally intensive Q-dimensional search. The com- where putation time can be reduced by applying the alternating projection algorithm [18] based on alternating maximization, A() =[a(θ1) ... a(θQ )] (4) which transforms the optimization problem into a sequence is the N × Q steering matrix formed by columnwise con- of much faster 1-D searches. The alternating projection algo- catenation of the steering vectors corresponding to each of rithm is a suboptimal approach due to nonexhaustive nature 2242 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 4, APRIL 2017

Fig. 3. Echogram based on coherently averaging of the receive channels. Fig. 5. ML-based DOA image. The black rectangles show regions of interest: glacier channel (Z1) and bedrock (Z2). chosen in order to ensure statistical stationarity in the rapidly changing scene. The array manifold, i.e., the set of steering vectors for the DOA interval of interest, is obtained from a full-wave electromagnetic simulation of the combined computer model of the antenna elements and the aircraft according to a similar procedure described in [20]. The outputs of the two algorithms are similar with respect to the large-scale content. The DOA of the near-range pixels is estimated with small (numerical) values, while the DOA of the far-range pixels is large. Dark blue and dark red represent far off-nadir signals, while green represents near- nadir returns. Parts of the ice–bed interface can be detected Fig. 4. MUSIC-based DOA image. as an abrupt transition from large to small estimated DOA angles, where the dominating signal component changes from of the search. However, except for the lowest signal-to- off-nadir surface clutter or noise, to the first (near nadir) return interference-plus-noise ratio cases, the global optimum is from the bedrock. With respect to the small scale content, almost always found. the MUSIC images are much noisier compared with the ML image. Furthermore, the ML image reveals large areas of off-nadir surface clutter (dark red) that appears due to a change IV. DOA REPRESENTATION OF RADAR ECHOGRAMS of sign in DOA angle compared with the background. The Now, we will utilize the DOA algorithms to obtain an alter- transition from ice to bedrock is much more significant in the native representation of the radar data. Consider the intensity ML image. In both images, a distinctive color sweep pattern in echogram in Fig. 3, which is generated by coherently averag- the estimated DOA angle is seen right after the first bed return. ing data from all receive channels. The DOA is estimated for Again, the phenomenon is more pronounced in the ML image. each pixel in the echogram. The number of signal components Based on this visual comparison of the MUSIC image and to be estimated can be difficult to determine for the individual the ML image, we conclude that the ML algorithm for this pixels. For this reason, and to simplify the processing and specific scene and clutter scenario is preferable for the further interpretation, the number of signal components is assumed to analysis. be one for all pixels, i.e., Q = 1, even though this is incorrect Sections V and VI address observations in the DOA repre- for some regions of the image. When this assumption does sentation in terms of the detectability of the ice–bed interface not hold, the DOA of the dominating signal component tends and the sweep-pattern in the estimated DOA angle at the bed. to be the one estimated, and in this way, the estimate is still meaningful. V. I CE–BED DETECTION By presenting the DOA estimates as an image with the pixel color representing the DOA angle, the procedure can By examining the echogram in Fig. 3, we can see that the be considered as a DOA representation of the echogram. The subsurface returns are highly contaminated by surface clutter. DOA representation of the echogram from Fig. 3 can be seen The bedrock is detectable at the beginning and end of the in Figs. 4 and 5 using MUSIC and ML, respectively. The frame (left/right of the glacier channel), but at the middle colormap is thresholded at ±40◦ as indicated by the colorbars. section (glacier channel), the weak bed return cannot be The covariance matrix is estimated based on five snapshots, discriminated from the clutter. Therefore, detection of the bed and the DOA images are filtered using a 5 × 5 median filter is not possible, which is unfortunate, since this data product to reduce noise and outliers. A low number of snapshots are and its derivatives are essential in glaciological modeling. NIELSEN et al.: DOA ANALYSIS OF AIRBORNE ICE DEPTH SOUNDER DATA 2243

Fig. 6. Echogram based on MVDR processing, where 11 snapshots are used for the estimation of the covariance matrix. A 5 × 5 mean filter is applied.

Fig. 7. MUSIC pseudospectral power derived image.

The MVDR beamformer can be used to reduce the surface clutter in the echogram. An echogram based on MVDR processing is shown in Fig. 6. When compared with Fig. 3, it is seen that the bed is more distinctive but that surface Fig. 8. Enlargement (Z1) of glacier channel, standard echogram (first), clutter is still limiting the detectability. As an alternative MVDR processed echogram (second), MUSIC pseudospectral power deriva- to MVDR, a similar visualization can be derived from the tive (third), and ML DOA image (fourth). MUSIC-spectrum in (9), as presented in [12]. For comparison, an MUSIC pseudospectral power derived image from [21] is showninFig.7. Now, we consider the ML DOA representation for bed detection. In Fig. 8, enlargements of the glacier channel in the echograms, MUSIC image, and DOA image are stacked for easy comparison. The colormaps of the enlarged images are scaled to enhance the local features. A 5 × 5 mean filter is applied to the intensity images. It is seen that the bed signal can be discriminated from the clutter in the DOA image, which is not possible in the standard radar-intensity echogram. The bed is more distinctive in the MVDR echogram compared with the standard echogram but detection is a challenge in the strong clutter region. A high Fig. 9. Bed detection (white dashed line) with interpolation (green dashed- amount of strong clutter is suppressed in the MUSIC image. dotted line), based on the ML DOA image. However, the weak parts of the remaining bed signal are hard to distinguish from the background noise. discontinuity from off nadir to near nadir is detected. This For the DOA image, on the other hand, even though the bed procedure corresponds to a tracing of the signal change from signal is flickering in the strong clutter region, the coverage is volume clutter to base return. In the case, where multiple basal sufficient to perform a reasonable trace of the interface with targets are present at a given along-track position, the one only a minor deviation at 8km, as shown in Fig. 9. The tracing closest to the radar is implicitly traced. In the strong clutter is done by scanning each line through range until a significant region, the detection might be based only on a few pixels 2244 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 4, APRIL 2017

Fig. 11. Geometry and notation associated with the illumination of sloped (across track) bed at different range gates.

Fig. 10. Enlargement (Z2) of the bed, echogram (top), and ML DOA image (bottom). are reflected corresponding to the left-hand side (LHS) and right-hand side (RHS) intersections of the wavefront with the ice–bed interface, as shown in the figure. It should be noted in range. The trace is interpolated at lines, where no bed signal that when referring to one of these two components, a specific is present at all. point on the bed can be described by either range, DOA, or In this way, the DOA image can be a powerful representa- (propagation) time. Therefore, the representations should be tion for discrimination and visualization of different types of read as being ambiguous or interchangeable if either the LHS targets, which can be used to interpret the echogram or for or RHS intersection is considered. A rough ice–bed interface direct applications, such as bed detection. is assumed such that energy is scattered back toward the radar. The across-track slope, φ(t0), and depth, si (t0), of the bed VI. ICE–BED BACKSCATTERING is estimated using radar and DOA data for the boxed region CHARACTERISTICS ESTIMATION in Fig. 10. Based on these parameters, a DOA simulation for Estimating surface roughness parameters from backscatter a flat sloped bed is conducted. The simulation is based on that is a well-known technique. However, the topography impacts the leading or trailing edge of the wave is characterized by a the local incidence angle, and when it comes to estimation of constant electrical distance surface roughness of glaciated bedrock, the ice complicates sa + nsi = ct (11) the problem by causing refraction and attenuation of the electromagnetic waves. In the following, we present a method where c is the speed of light in vacuum and n is the refractive for the estimation of bed roughness. The method is based on index of ice. This combined with Snell’s law of refraction the DOA representation of the data that allow us to compensate sin θ = n sin φ (12) for the bed topography, and the resulting change of incidence angle, refraction, and attenuation. is used to describe the wave within the ice. By specifying We start out by analyzing the DOA sweep-pattern observed the altitude h = sa sin θ, and the depth and across-track slope near the bed. An enlargement containing a part of the bed is of the bed, the DOA signal φ(t) for the bed return can be shown in Fig. 10. A subimage for further analysis is marked simulated. in the figure. The following analysis suggests that the DOA The DOA estimate of the boxed region in Fig. 10 is averaged pattern represents an off-nadir return from a rough sloped bed. in the along-track direction to a single line and plotted with The sounding geometry with notation associated with a the simulation as a function of time in Fig. 12. sloped (across track) bed is shown in Fig. 11. Since the The simulation consists of an approximately symmetric data are Doppler processed in the along-track direction, the two-legged curve, where each leg corresponds to the LHS and along-track extent of the resolution cell is small. In this RHS bed signal, respectively. It is seen that the estimation way, the extent of the resolution cell is (pulse) limited to and simulation fit very well, but clearly only one of the two the across-track direction at zero Doppler. At t0,thefirst components is estimated by the DOA algorithm. The reason bed return is reflected corresponding to the shortest electrical for this is that a one-signal (Q = 1) ML estimation was distance from the radar to the bed. The DOA of this first performed. In this case, the leg that is centered around nadir bed return, corrected through Snell’s law of refraction at the is the one estimated because of the transmit antenna pattern. air–ice interface, corresponds to the refraction angle φ of the Since the pattern is directed toward nadir, the given signal shortest ray path si , which corresponds to the across-track component is the one dominating the combined signal, hence slope of the bed. Later time, i.e., at t1, t2,..., two signals the one estimated by the DOA algorithm. At a less sloped NIELSEN et al.: DOA ANALYSIS OF AIRBORNE ICE DEPTH SOUNDER DATA 2245

Fig. 14. Illustration of the detrending procedure. Original echogram (left) and the corresponding detrended output (right). we will describe a procedure for estimating the backscattering pattern of the bed, which includes the corrections of the intensity waveform. Fig. 12. DOA estimation of the bed return along with a simulation based on the geometric model from Fig. 11. A. Detrending We are still considering the data region marked in Fig. 10. To get an accurate estimate of the DOA trace and the waveform of the bed return, both the DOA data and the intensity radar data are averaged in the along-track direction. However, the bed has an along-track slope, which distorts the shape of the DOA trace and the waveform, when the data are averaged. In order to avoid this distortion, the data are detrended with respect to the along-track slope before averaging. This is done by tracing the leading edge of the waveform and shifting each line in range accordingly. The procedure is equivalent to averaging in the surface parallel direction and is shown in Fig. 14. The resulting DOA trace and waveform after averaging are plotted as a function of time in Fig. 15.

B. Fitting of Bed Model To correct for attenuation and refraction at the air–ice interface and so on, the geometric model in Fig. 11 is adopted. Fig. 13. Two-signal ML DOA estimation and simulation of bed return. The model is fitted to the data shown in Fig. 15. As shown in the figure with the vertical dashed lines, the data are clipped in the range direction to capture the trailing edge part of the bed, it was possible with a two-signal estimation of the waveform and the valid part of the DOA trace. The to recover both of the signal components from the bed, as bed model is now fitted to the data by adjusting the slope seen in Fig. 13. In the case of a small slope, the geometry is parameter φ and the propagation time corresponding to the symmetric, which results in the bed signals of equal amplitude. closest approach. The error, which is minimized, is evaluated The retrieval of both bed signals in the low slope scenario in the DOA representation corresponding to the difference strengthens the hypothesis of the bed reflections being the between the data and the model in Fig. 12. The across-track mechanism behind the sweep pattern. slope of the bed is estimated by the fitted parameter to φ = 8◦, The case of a single dominating bed signal combined which for the specific flight segment corresponds to the slope with the DOA information makes it possible to estimate of the glacier channel. the backscattering characteristics of the bed for a range of incidence angles. This is done by combining the intensity C. Waveform Correction waveform with the corresponding DOA estimate. However, the backscattering information contained in the waveform is The data are now corrected for four mechanisms: affected by several factors, such as a varying propagation 1) receive gain; distance, antenna patterns, refraction at the air–ice interface, 2) transmit gain; and so on. These factors need to be taken into account to get an 3) attenuation loss; accurate estimate of the bed characteristics. In the following, 4) geometric spreading. 2246 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 4, APRIL 2017

Fig. 16. Transmit pattern in the across-track direction, based on simulation of the array manifold. The simulation was conducted using the HFSS software package by ANSYS. Fig. 15. Along-track averaged waveform of the bed return (top) and the correspondingly averaged DOA estimate (bottom). 3) Attenuation Loss: The electromagnetic propagation within the ice involves attenuation losses due to absorption 1) Receive Gain: To improve the signal-to-clutter ratio, and internal scattering. It is seen from the geometry in Fig. 11 suppress clutter and the secondary bed return, beamforming that the propagation distance in ice (s ) for the bed return is used to steer the receive beam toward the direction of the i varies with DOA. When the attenuation coefficient is assumed dominating bed return. constant, the attenuation loss is exponentially proportional to The output y of beamforming formulated as a spatial the propagated distance in ice, that is filtering process is given by asi L A ∝ 10 (15) y = hH x (13) where a is the attenuation constant. The attenuation loss varies where h is the N × 1 filter weight vector. In the case of with DOA through si , and can be taken into account. Under beamsteering, the filter weights are given by [19] the assumption of a constant ice temperature and by using the a(θ) model, s is calculated as a function of range and the waveform h = (14) i aH(θ)a(θ) is corrected accordingly. where θ represents the steering angle. The normalization 4) Geometric Spreading: The inverse-square law and the ensures unity gain in the θ-direction, and the correction for two-way propagation of the pulse result in the geometric the receive gain is in this way incorporated into the filtering spreading loss factor that is related to range in the following : process. way 4 DOA data are simulated based on the fitted model and are LGS ∝ R . (16) used as the steering angle in (14). A range varying beam is When s is the propagated distance in air, the range is defined in this way synthesized for, and applied to, each azimuth line. a as R = s + s , which takes the refraction at the air–ice The filtered data are then detected, detrended, and averaged a i interface into account. As for the attenuation loss, the geo- according to the procedure described earlier. metric model is used to calculate the range for each sample, 2) Transmit Gain: All transmit elements are used for trans- and the data are corrected accordingly. mission without any tapering. The resulting transmit pattern 5) Refraction Gain: Due to refraction at the air–ice inter- is shown in Fig. 16. By using the estimated DOA data in face, adjacent rays of a transmitted wave are focused into combination with the pattern, the waveform can be corrected a smaller area compared with a corresponding free-space for the antenna transmit gain. The antenna pattern is based scenario. This results in a gain factor known as the refraction on simulations [20] and does not take dynamic factors, such gain [22]. Simulations for the geometry of the given scenario as wing flexure and vibration into account. This affects the show that the variation of the refraction gain can be neglected true pattern particularly regarding the depth of the nulls. within the range of DOA angles under consideration. Based Furthermore, energy from the secondary bed return and from on this, no correction of the refraction gain is applied to the surface and volume clutter contributes to the received signal, data. which smoothens the waveform, when the transmit gain toward the bed is low. Therefore, if the waveform is corrected with D. Backscattering Pattern the unmodified simulated pattern with deep nulls, high ampli- fication of the clutter will occur at angles corresponding to The corrected waveform that represents backscatter from the the nulls. To avoid this clutter amplification, the nulls of the bed surface can be expressed as pattern are filled before the correction is applied. The modified P (θ)10asi R4 σ(θ) = K BS (17) transmit pattern is shown in blue in Fig. 16. Gt (θ) NIELSEN et al.: DOA ANALYSIS OF AIRBORNE ICE DEPTH SOUNDER DATA 2247

Fig. 17. Estimated and simulated backscattering pattern of the bed surface. Fig. 18. Backscattering pattern of the bed surface as in Fig. 17 but calculated with the assumption of a bed slope equal to zero. where PBS(θ) is the received power using beamsteering, and K is a product of factors independent of DOA, such as based on the fitted geometric model. Since the illuminated area the system gain and so on. The normalized backscatter is is rapidly changing for small incidence angles, backscatter is computed by dividing with the backscatter at zero incidence, only modeled for larger angles, where the estimate of the area that is is more accurate and robust. The IKM is fitted to the estimated σ(θ) data and is included in Fig. 17. A relative permittivity for σ(θ)ˆ = (18) ice equal to 3.2 is assumed. The bedrock permittivity enters σ(θ0) the model only through the Fresnel reflectivity that appears where θ is the DOA angle corresponding to zero incidence at 0 as a factor in the IKM [see (19)]. Since the IKM is fitted the bed, i.e., t = t in Fig. 11. Based on the model, the angle 0 to normalized data, see (18), the estimated rms slope does of incidence at the bed is calculated from the refracted angle φ not depend on the bedrock permittivity. Based on the fit of and the estimated bed slope. The normalized backscatter as a the IKM, the rms slope is estimated to 0.28 or 16◦,which function of incidence angle is plotted in Fig. 17. represents a measure of the bed roughness. For comparison, With the assumption of a random surface with a Gaussian a recent study [26] estimates bed rms slopes of Thwaites height distribution, the IKM [23], [24] is used to model the Glacier in West Antarctica based on radar ice sounding, but backscattering coefficient   with a different surface model and data acquired at a different tan2 α frequency, which is sensitive to another roughness scale. The σ 0 (α) = exp − (19) IKM 2 4 α 2 ◦ ◦ 2ms cos 2ms slopes are estimated to be between 6 and 8 . A solid validation of the estimated rms slope is difficult, where α is the angle of incidence, is the Fresnel reflec- since direct in situ measurements of the rms slope cannot be tivity [25] evaluated at normal incidence, and m is the root s obtained. Processing of different segments will not signifi- mean square (rms) slope of the surface given by [24] √ cantly improve the validation, since the roughness can vary 2σh with location. ms = . (20) λh To illustrate the importance of including DOA information for bed roughness estimation, a simulation has been conducted. The parameters λh and σh are the surface correlation length and rms height, respectively. The IKM only depends on the The procedure including all waveform corrections used to rms slope and is, therefore, invariant with respect to a common produce Fig. 17 is repeated except that the bed slope is set to λ σ zero. The estimated backscattering pattern and the fitted IKM scaling of h and h as long as the validity conditions [23] ◦ are fulfilled. When the surface height variation is Gaussian are shown in Fig. 18. The estimated rms slope is 0.42 or 24 , distributed, the validity conditions are given by [23] which differs significantly from the result estimated utilizing DOA information to take the bed slope into account. kλh > 6 (21) λ2 > . σ λ. h 2 76 h (22) VII. CONCLUSION Furthermore, application of geometric optics (stationary-phase Alternative applications of DOA estimation in relation approximation) requires that [23] to airborne radar ice sounding are presented in this paper. We use the MUSIC and ML estimators to convert the radar (2kσ cos α)2 > 10. (23) h data into a DOA representation, where the latter is seen to For ice with a relative permittivity of 3.2, λ = 1.12 m and provide superior performance. The DOA representation offers k = 5.62 m−1. a better visualization of the desired signals and clutter. Based The backscatter is obtained by multiplying the coefficient on this, we are able to discriminate the desired bed return with the time-varying illuminated area, which is calculated from strong surface clutter in the channel of the challenging 2248 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 4, APRIL 2017

Jakobshavn Glacier. We show how this can be used to detect [15] U. Nielsen and J. Dall, “Direction-of-arrival estimation for radar ice some of the most challenging parts of the bed along the sounding surface clutter suppression,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 9, pp. 5170–5179, Sep. 2015. channel. [16] A. Lohoefener, “Design and development of a multi-channel radar depth Furthermore, a geometric model is used to show how the sounder,” Ph.D. dissertation, Dept. Elect. Eng. and Comput. Sci., Univ. across-track slope of the bed is related to the DOA pattern of Kansas, Lawrence, KS, USA, Nov. 2006. [17] R. O. Schmidt, “Multiple emitter location and signal parameter esti- the bed return. In a low slope scenario where the associated mation,” IEEE Trans. Antennas Propag., vol. 34, no. 3, pp. 276–280, geometry gives rise to comparable amplitudes of the LHS and Mar. 1986. RHS bed signals, the DOA for both components is retrieved [18] I. Ziskind and M. Wax, “Maximum likelihood localization of multiple sources by alternating projection,” IEEE Trans. Acoust., Speech Signal and validated with the model. For larger slopes, it is shown Process., vol. 36, no. 10, pp. 1553–1560, Oct. 1988. that the bed component received closest to nadir is dominant [19] P. Stoica and R. L. Moses, Introduction to Spectral Analysis. due to amplification caused by the combination of the transmit Englewood Cliffs, NJ, USA: Prentice-Hall, 1997. [20] J.-B. Yan et al., “Measurements of in-flight cross-track antenna patterns pattern and asymmetric geometry. This is exploited to retrieve of radar depth sounder/imager,” IEEE Trans. Antennas Propag., vol. 60, bed characteristics by combining DOA data and waveforms no. 12, pp. 5669–5678, Dec. 2012. of the radar data. By fitting the geometric model to the data, [21] University of Kansas, Lawrence, KS, USA. (Dec. 2012). MUSIC Pseudo-Spectral Power Derived Images of the Jakobshavn Glacier the across-track slope is estimated. Based on the model, a Center for Remote Sensing of Ice Sheets. [Online]. Available: ftp://data. number of corrections are applied to the waveform to retrieve cresis.ku.edu/data/rds/2006_Greenland_TO/CSARP_music/2006053 the received backscatter of the bed surface as a function of [22] P. Gudmandsen, Electromagnetic Probing in Geophysics.NewYork,NY, USA: Golem Press, 1971, ch. 9, pp. 321–348. the local incidence angle. The backscattering pattern holds [23] F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote Sensing: information on the bed roughness. To further quantify the Active and Passive, vol. 2. Reading, MA, USA: Addison-Wesley, 1982. roughness, the IKM is fitted to the data and used to estimate [24] G. Picardi et al., “Performance and surface scattering mod- ◦ els for the mars advanced radar for subsurface and ionosphere a16 rms slope of the surface. sounding (MARSIS),” Planetary Space Sci., vol. 52, nos. 1–3, pp. 149–156, 2004. [25] F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote Sensing: REFERENCES Active and Passive, vol. 1. Reading, MA, USA: Addison-Wesley, 1981. [26] D. M. Schroeder, D. D. Blankenship, D. A. Young, A. E. Witus, and [1] E. Rignot, I. Velicogna, M. R. Van Den Broeke, A. Monaghan, and J. B. Anderson, “Airborne radar sounding evidence for deformable J. T. M. Lenaerts, “Acceleration of the contribution of the Greenland sediments and outcropping bedrock beneath Thwaites Glacier, Geophys. Res. Lett. and Antarctic ice sheets to sea level rise,” , vol. 38, West Antarctica,” Geophys. Res. Lett., vol. 41, no. 20, pp. 7200–7208, no. 5, p. L05503, Mar. 2011. Oct. 2014. [2] A. Shepherd et al., “A reconciled estimate of ice-sheet mass balance,” Science, vol. 338, no. 6111, pp. 1183–1189, Nov. 2012. Ulrik Nielsen received the B.Sc. and M.Sc. degrees [3] E. Rignot, J. Mouginot, M. Morlighem, H. Seroussi, and B. Scheuchl, in electrical engineering from the Technical Univer- “Widespread, rapid grounding line retreat of Pine Island, Thwaites, sity of Denmark, Copenhagen, Denmark, in 2008 Smith, and Kohler glaciers, West Antarctica, from 1992 to 2011,” and 2011, respectively, and the Ph.D. degree in array Geophys. Res. Lett., vol. 41, no. 10, pp. 3502–3509, May 2014. signal processing from the National Space Institute, [4] S. Gogineni et al., “Coherent radar ice thickness measurements over Technical University of Denmark, in 2015, with the Greenland ice sheet,” J. Geophys. Res., Atmos., vol. 106, no. D24, the focus on synthetic aperture radar tomography pp. 33761–33772, Dec. 2001. techniques for radar ice sounding. [5] J. Dall et al., “P-band radar ice sounding in Antarctica,” in Proc. Since 2015, he has been with IHFood A/S, Copen- IGARSS, Munich, Germany, Jul. 2012, pp. 1561–1564. hagen, where he has been involved in developing [6] M. Schäfer et al., “Sensitivity of basal conditions in an inverse model: computer vision technology. His research interests Vestfonna ice cap, Nordaustlandet/Svalbard,” Cryosphere, vol. 6, no. 4, include image analysis, machine learning, and statistical modeling. pp. 771–783, 2012. [7] J. Li et al., “High-altitude radar measurements of ice thickness over the Jie-Bang Yan (S’09–M’11) received the Antarctic and Greenland ice sheets as a part of operation icebridge,” B.Eng. degree (Hons.) in electronic and IEEE Trans. Geosci. Remote Sens., vol. 51, no. 2, pp. 742–754, communications engineering from The University of Feb. 2013. Hong Kong, Hong Kong, in 2006, the M.Phil. degree [8] J. J. Legarsky, “Synthetic-aperture radar (SAR) processing of glacial- in electronic and computer engineering from The ice depth-sounding data, Ka-band backscattering measurements and Hong Kong University of Science and Technology, applications,” Ph.D. dissertation, Dept. Elect. Eng. and Comput. Sci., Hong Kong, in 2008, and the Ph.D. degree in Univ. Kansas, Lawrence, KS, USA, 1999. electrical and computer engineering from the [9] J. Paden, T. Akins, D. Dunson, C. Allen, and P. Gogineni, “Ice-sheet University of Illinois at Urbana–Champaign, bed 3-D tomography,” J. Glaciol., vol. 56, no. 195, pp. 3–11, 2010. Champaign, IL, USA, in 2011. [10] X. Wu, K. C. Jezek, E. Rodriguez, S. Gogineni, F. Rodríguez-Morales, From 2009 to 2011, he was a Croucher Scholar and A. Freeman, “Ice sheet bed mapping with airborne SAR tomogra- with the University of Illinois at Urbana–Champaign, where he was involved phy,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 10, pp. 3791–3802, in MIMO and reconfigurable antennas. In 2011, he joined the Center for Oct. 2011. Remote Sensing of Ice Sheets, The University of Kansas, Lawrence, KS, [11] F. Rodríguez-Morales et al., “Advanced multifrequency radar instrumen- USA, as an Assistant Research Professor. He is currently an Assistant tation for polar research,” IEEE Trans. Geosci. Remote Sens., vol. 52, Professor of Electrical and Computer Engineering with The University of no. 5, pp. 2824–2842, May 2014. Alabama, Tuscaloosa, AL, USA. He holds two U.S. patents and a U.S. patent [12] S. Gogineni et al., “Bed topography of Jakobshavn Isbræ, Greenland, application related to novel antenna technologies. His research interests and Byrd Glacier, Antarctica,” J. Glaciol., vol. 60, no. 223, pp. 813–833, include the design and analysis of antennas and phased arrays, ultrawideband 2014. radar systems, radar signal processing, and remote sensing. [13] U. Nielsen, J. Dall, A. Kusk, and S. S. Kristensen, “Coherent sur- Dr. Yan was a recipient of the Best Paper Award in the 2007 IEEE (HK) face clutter suppression techniques with topography estimation for AP/MTT Postgraduate Conference, the Raj Mittra Outstanding Research multi-phase-center radar ice sounding,” in Proc. EUSAR, Nuremberg, Award, Urbana, IL, in 2011, and the NASA Group Achievement Award Germany, Apr. 2012, pp. 247–250. in 2013 for the NASA P3 Aircraft Antarctica Mission Team, and was the [14] J. Dall et al., “ESA’s polarimetric airborne radar ice Best Paper Finalist of the 2015 National Instruments Week. He serves as a sounder (POLARIS): Design and first results,” IET Radar, Sonar Technical Reviewer for several journals and conferences on antennas and Navigat., vol. 4, no. 3, pp. 488–496, 2010. remote sensing. NIELSEN et al.: DOA ANALYSIS OF AIRBORNE ICE DEPTH SOUNDER DATA 2249

Sivaprasad Gogineni (M’84–SM’92–F’99) Jørgen Dall (M’07) received the M.Sc. degree in received the Ph.D. degree in electrical engineering electrical engineering and the Ph.D. degree from from The University of Kansas (KU), Lawrence, the Technical University of Denmark, Copenhagen, KS, USA, in 1984. Denmark, in 1984 and 1989, respectively. He is currently the Deane E. Ackers Distinguished He has been an Associate Professor since 1993. Professor with the Department of Electrical He has been working with the Danish airborne SAR, Engineering and Computer Science, KU, where EMISAR, e.g., he led the development of onboard he is also the Director of the NSF Science and and offline SAR processors, was responsible for the Technology with the Center for Remote Sensing data processing and organized the EMISAR data of Ice Sheets. He will be joining the University acquisition campaigns in a five year period. Later he of Alabama, Tuscaloosa, AL, USA, in 2017. led the development of the POLARIS sounder and He developed several radar systems currently being used at KU for sounding SAR. His research interests include various aspects of ice sheet penetration, and imaging of polar ice sheets. He has also participated in field experiments e.g., InSAR elevation bias, PolInSAR extinction coefficients, tomographic ice in the Arctic and Antarctica. He has authored or co-authored over 90 archival structure mapping, and ice sounding. journal publications, 200 technical reports, and conference presentations. His research interests include the application of radars to the remote sensing of the polar ice sheets, sea ice, ocean, atmosphere, and land. Dr. Gogineni is a member of URSI, the American Geophysical Union, the International Glaciological Society, and the Remote Sensing and Photogrammetry Society. He was an Editor of the IEEE GEOSCIENCE AND REMOTE SENSING SOCIETY NEWSLETTER from 1994 to 1997. globe—© cartesia, plane—©1995 expert software, inc., satellite—wikimedia commons/richard-59 UAS-Based Radar Sounding of the Polar Ice Sheets

C. Leuschen, R. Hale, S. Keshmiri, J.B. Yan, F. Rodriguez-Morales, A. Mahmood, and S. Gogineni Center for Remote Sensing of Ice Sheets, University of Kansas, Lawrence, Kansas, United States

Abstract—Both the Greenland and Antarctic ice sheets a dual-frequency radar that operates at approximately 14 are currently losing mass and contributing to global sea and 35 MHz. The radar transmits 100-W peak power at a level rise. To predict the response of these ice sheets to a pulse repetition frequency of 10 kHz, operates from 20 W warming climate, ice-sheet models must be improved by of DC power, and weighs approximately 2 kg. The UAS has incorporating information on the bed topography and a take-off weight of about 38.5 kg and a range of approxi- basal conditions of fast-flowing glaciers near their ground- mately 100 km per gallon of fuel. We recently completed ing lines. High-sensitivity, low-frequency radars with 2-D several successful test flights of the UAS equipped with the aperture synthesis capability are needed to sound and image dual-frequency radar at a field camp in Antarctica. The radar fast-flowing glaciers with very rough surfaces and ice that measurements performed as a part of these test flights rep- contains inclusions. In response to this need, CReSIS devel- resent the first-ever successful sounding of glacial ice with a oped an Unmanned Aircraft System (UAS) equipped with UAS-based radar. We also collected data for synthesizing a

Digital Object Identifier 10.1109/MGRS.2014.2306353 2-D aperture, which is required to prevent off-vertical scat- Date of publication: 8 April 2014 ter, caused by the rough surfaces of fast-flowing glaciers,

8 2168-6831/14/$31.00©2014IEEE ieee Geoscience and remote sensing magazine march 2014 from masking bed echoes. In this article, we provide a brief surface scatter and volume scatter from inclusions in the overview of the need for radar soundings of fast-flowing gla- ice can mask weak ice-bed echoes. The weak echoes are a ciers at low-frequencies and a brief description of the UAS result of much higher than normal attenuation of radar and radar. We also discuss our field operations and provide signals propagating through the ice due to the presence of sample results from data collected in Antarctica. Finally, we warm ice near the bed. For this reason, high-sensitivity, present our future plans, which include miniaturizing the low-frequency radars with 2-D aperture synthesis capa- radar and collecting measurements in Greenland. bility are required for ice-thickness measurements over fast-flowing glaciers. The low frequencies are required to 1. Introduction reduce volume scatter from ice inclusions, and the radars urrent satellite observations indicate that the Earth’s must be capable of 2-D aperture synthesis to obtain nar- Cpolar regions are undergoing significant changes, row antenna beams with low sidelobes, which prevents including the decrease of sea ice extent and thickness both along- and across-track surface scatter from mask- in the Arctic [1] and the thinning of the margins of the ing weak ice-bed returns. A narrow-antenna beam can Greenland and Antarctic ice sheets. Recently, Shepherd be generated in the along-track direction with traditional et al. [2] analyzed data from multiple satellite missions Synthetic Aperture Radar (SAR) techniques; however, it is and reported that mass loss from the Greenland ice sheet extremely difficult to accommodate a large, low-frequency increased from about 72 Gt/yr during 1993–1995—a 0.2 antenna array, even on aircraft like the NASA P-3 [8]. For mm/year contribution to sea level rise—to about 344 Gt/ this reason, we developed a small Unmanned Aircraft Sys- yr during 2005–2010—a 0.9 mm/year contribution to tem (UAS) that can be flown over closely-spaced lines in sea level rise. They also indicated that both the Antarc- the cross-track direction to synthesize a large aperture. We tic and Greenland ice sheets are losing mass. These large developed the UAS with a radar operating at 14 and 35 ice sheets contain enough fresh water to raise sea level by MHz for measurements over the ice sheets in Antarctica about 66 meters if they were to melt completely. While and Greenland. The fully instrumented UAS weighs about complete melting and disintegration are unlikely in the 38.5 kg with a range of 100 km for about one gallon of fuel. immediate future, their thinning and retreat are already The antennas for operating the radar at either 14 MHz with contributing to an accelerated rise in sea level. This sea about 1 MHz of bandwidth or 35 MHz with about 4 MHz level rise will have a strong impact on heavily populated of bandwidth are integrated into the wings and airframe of coastal regions [3]. the UAS. The radar transmits 100-W peak power signals at The Intergovernmental Panel on Climate Change a pulse repetition frequency of 10 kHz. The radar weighs (IPCC) reported that the global average sea level would approximately 2 kg and operates with about 20 W of DC rise between 26 and 98 cm by the end of this century [4]. power. We flight tested the radar-equipped UAS at the Sub- However, the upper bound reported by the IPCC is widely glacial Lake Whillans (SLW) field camp on the Whillans debated, as estimates using empirical models show that Ice Stream in Antarctica and accomplished the first suc- it could be as large as 2 m. Significant progress has been cessful sounding of ice with a radar on a UAS. made in the development of models that incorporate the In this paper, we provide a brief overview of the UAS, full stress tensor and use updated bed topography maps its radar system, experimental results from the field, and of both ice sheets [5–7], but substantial uncertainty still our future plans to miniaturize the radar and collect data surrounds the upper limit of projected sea-level rise. The over fast-flowing glaciers. large range and subsequent disagreement can be partly attributed to a lack of fine-resolution bed topography 2. Background for fast-flowing glaciers, particularly near their ground- Radars operating over a frequency range of about 1 MHz ing lines. These fast-flowing outlet glaciers, which are to 1000 MHz have been used to sound glacial ice for only few kilometers wide and yet carry vast reservoirs of decades [9]. The application of radars to ice sounding can ice into the ocean, are poorly represented in the current be considered to have started with the pioneering work of ice-bed maps and ice-sheet models. There is therefore an Amory Waite. He conducted experiments to measure ice urgent need to measure the ice thickness of fast-flowing thickness with a radar altimeter operating at 440 MHz on glaciers with fine resolution to determine bed topogra- the Ross Ice shelf in the late 1950s [10–11]. Waite is also phy and basal conditions. This information will, in turn, credited with conducting the first airborne ice measure- be used to improve ice-sheet models and generate accu- ments in the early 1960s. Following his pioneering work, rate estimates of sea level rise in a warming climate. With- the glaciological application of radars was expanded out proper representation of these fast-flowing glaciers, through significant contributions by Evans and his col- advancements in ice-sheet modeling will remain elusive. leagues at the Scott Polar Research Institute in the UK Unfortunately, radar sounding and imaging of outlet [12–14] and Gudmandsen and his colleagues at the Tech- glaciers near their grounding lines and calving fronts are nical University of Denmark [15]. Now, radar has become extremely challenging tasks, because off-vertical rough an invaluable tool in the study of ice sheets and glaciers. march 2014 ieee Geoscience and remote sensing magazine 9 other groups and are currently used for measurements over the Greenland and Antarctic ice sheets [22–23]. At the University of Kansas, we developed radars with 4–15 element cross-track arrays and multiple receivers for airborne and surface-based measurements [8, 21]. These radars include synthetic aperture radar (SAR) and array processing capabilities [8, 23] and have been successfully used to demonstrate 3-D imaging of ice sheets [24–26]. These systems have also been used to sound several chal- lenging areas of the ice sheets, including outlet glaciers and ice sheet margins [27–28], as well as for high-altitude measurements from long-range aircraft [29]. However, as mentioned previously, the performance of radars operat- Figure 1. G1X UAS on approach to land at SLW site in Antarctica. ing at frequencies of 50 MHz or higher severely degrades over fast-flowing glaciers due to rough surface scatter Radars have been used to sound ice, map both deep and and volume scatter. A few attempts have been made to near-surface internal layers, and image the ice-bed inter- develop coherent airborne HF radars operating at fre- face with fine resolution. quencies as low as 1 MHz and as high as 30 MHz [30–35] Most of the radars used for sounding ice during the for sounding glaciers with temperate ice; these have been 1960s and 1970s were incoherent. A few attempts were shown to be effective in sounding temperate ice under made to develop coherent radars for sounding ice in 1970s favorable conditions. Normally, low-frequency radars with very limited success because of the lack of suitable are operated with long, resistively-loaded wire antennas inexpensive technologies [16]. The first solid-state coher- trailing behind the aircraft. Although SAR processing can ent radar sounder was developed at the University of Kan- be used to reduce beamwidth in the along-track direc- sas in the late 1980s [17]. It was redesigned, upgraded, tion, the large beamwidth of trailing long-wire antenna and widely used for measurements over the Greenland in the cross-track direction results in significant reflec- ice sheet as a part of NASA’s Program of Arctic Assess- tions from the walls of the glaciers, which degrades radar ment (PARCA) initiative [18–21]. Following the success- performance. A UAS equipped with low-frequency radars ful demonstration of a fully coherent low-power radar for that can be flown over closely-spaced lines—as close as sounding ice sheets, coherent radars were developed by 5 m at 14 MHz—in the cross-track direction for synthe- sizing a 2-D aperture is needed to sound fast-flowing gla- ciers with fine resolution.

Table 1. G1X Technical Specifications. 3. Platform Overview Parameter Value Units The G1X unmanned aerial system (UAS) is a mid-wing, Dimensions/Geometry semi-autonomous, high-aspect ratio aircraft that has Length 2.84 m been modified by the University of Kansas specifically Height 1.11 m for scientific missions throughout the cryosphere. The Wing G1X UAS platform has a 5.3 m wingspan, 2.84 m fuse- Area 2.06 m2 lage length, and weighs approximately 38.5 kg when fully Span 5.29 m instrumented and fueled. When operating at a cruise Aspect ratio 11.75 n.d. speed of 28 m/s, the platform has a range of about 100 km Power Plant for approximately one gallon of fuel. Additional range Engine Desert Aircraft 100 cc and endurance can be enabled with a supplemental fuel Max power 7060 W tank. The aircraft can be configured with either wheels or Weights skis; recent field trials in Antarctica were all on skis. Take- Fuel 2.72 kg off and landing performance is verified at 90 m or less. Payload 9.07 kg Figure 1 shows a photograph of the G1X on approach to Empty 26.76 kg land at the SLW snow runway in West Antarctica. Table 1 Max takeoff 38.55 kg shows the G1X’s technical and performance information. Performance The integration of HF/VHF antennas for operating the Cruise speed 28 m/s radar on a small UAS drove aircraft requirements. The Range 100 km antenna’s physical length requirement demanded a rela- Endurance 60 min tively high-aspect ratio and wing span (Table 1). Winglets Takeoff/landing distance 90 m were used to improve the stability characteristics of the aircraft. The physics-based model, pilot rating, and system

10 ieee Geoscience and remote sensing magazine march 2014 identification analysis all showed that the G1X UAS satis- However, the imaginary part of the dielectric constant fies every Level I handling quality requirement. associated with the conduction component is higher at The G1X is piloted by the Supplemental Pilot on takeoff, 14 MHz than at 35 MHz, as it is inversely proportional landing, and in the vicinity of the runway in the third per- to frequency. The trade-off is the difficulty in integrating son or external pilot mode using a line-of-sight S-band com- the 14 MHz antennas on a small platform. We used elec- munication link. Optionally, the G1X is commanded by the trically short antennas with matching networks to obtain Pilot Operator using the WePilot-2000 autopilot through a good impedance match and adequate efficiency. This the implementation of GPS-based state estimations. The limited the bandwidth to about 1 MHz at the lower band airborne WePilot communicates with the WePilot Ground and 4 MHz at the upper band. The frequencies of opera- Station via a UHF telemetry transceiver and an Iridium tion within each band are fully adjustable to match the satellite link. Communication with the science payload is optimal response of the antennas. For the radar electron- enabled through a supplemental UHF telemetry link to pro- ics, we have employed both off-the-shelf and custom- vide control, status, and quick-look results. The aircraft is designed components developed primarily from wireless equipped with a Guardian-3 Iridium Automated Flight Fol- communications and ultrasound imaging technologies. lowing (AFF) system, which transmits in the L band. This Range resolution determines the radar’s perfor- GPS-based system transmits position, altitude, bearing, and mance in resolving targets located at different ranges. It speed every two minutes in compliance with USDA AFF is inversely proportional to radar bandwidth, B, and can requirements. The information is transmitted to USDA-AFF expressed as tf= cB/,2 r where c is the speed of wave to enable continuous tracking of the aircraft. propagation in free space and fr is the relative permit- The G1X’s successful maiden flight took place during tivity of the propagation media, assumed to be equal to summer 2013 at Fort Riley Army Base, Kansas. This flight 3.15 for solid ice. For the operating bandwidths reported was fully manual with the pilot in the loop and was used to with the newly developed radar, the range resolution for assess and verify aircraft stability and handling qualities. the radar is estimated to be close to 85 m at the lower band (using 1 MHz of bandwidth) and close to 21 m at 4. System Overview the upper band (using 4 MHz of bandwidth). The range In this section, we present an overview of the radar elec- measurement uncertainty, on the other hand, is inversely tronics and the antenna structures used with the G1X proportional to both bandwidth and the square root of UAS. As mentioned in the introduction, the radar oper- the signal-to-noise ratio and is given by dtrk= //S Nl , _ i ates at center frequencies in the vicinity of 14 MHz (lower where ka= constant determined by the pulse shape and band, HF) and 35 MHz (upper band, VHF) with band- SN/ is the signal-to-noise ratio. An isolated bed return widths of 1 MHz and 4 MHz, respectively. We selected with an S/N ratio of 18 dB or better can be measured with these radar center frequencies to satisfy two require- an uncertainty of about 10 m for both bands. ments: (1) the antennas should be integrated into the The range resolution at the upper band is comparable UAS wings and airframe; and (2) the system should be to that of the airborne system flown on board the NASA capable of sounding ice in fast-flowing glaciers. We chose DC-8, which was estimated to be close to 18 m for a band- 14 MHz because of the success Arcone and his colleagues width of 9.5 MHz and accounted for the effect of a time- [36–37] had in sounding glaciers with temperate ice in domain window applied to the signal to reduce range sid- Alaska with a 12-MHz impulse radar. Similarly, we chose elobes [8, 29]. While the expected range resolution at the 35 MHz because of the success Blindow and his col- lower band (14 MHz) may seem coarse, it is important leagues [34–35] reported in sounding temperate ice in to consider, as mentioned earlier, that operation in this Patagonia with a 30-MHz impulse radar. The 35-MHz band is intended to fill the gaps where the VHF systems radar improves upon the performance of the 30-MHz fail due to high volume scattering. The performance of impulse radar by providing fine resolution and better the radar operating at the lower band is considered ade- range (ice thickness) measurement accuracy due to its quate to sound ice thicknesses ranging from a few hun- wider bandwidth. Likewise, the 14-MHz radar offers bet- dred meters to 2 km in such areas. ter performance than the 12-MHz impulse radar in areas with significant volume debris. The additional attenua- 4.1. Radar System tion (dB/m) due to volume debris can be modeled by a The radar consists of three main sections: (1) DC power complex effective permittivity [38] where the imaginary conditioning and distribution; (2) digital; and (3) radio part scales by frequency cubed, and the volume clutter frequency (RF). A simplified block diagram of the radar is due to Rayleigh backscatter scales by frequency to the shown in Figure 2. The DC power conditioning and distri- fourth power. For a simple scenario of 800 m-thick ice bution section is composed of a passive electromagnetic and meter scale spherical water inclusions at a volume interference (EMI) filter combined with a series of high- fraction of 2 percent, the additional attenuation of the reliability DC-DC converter and regulators modules. bed response due to these inclusions would be approxi- The digital section was implemented with a field pro- mately 30 dB at 35 MHz and negligible at 14 MHz. grammable gate array (FPGA) development kit, a high-speed march 2014 ieee Geoscience and remote sensing magazine 11 speed serial connection to the on-board autopilot, which, in turn, communi- Passive Control Matching RF-Section cates with the ground sta- EMI Filter (High-Power) Network tion using a 900 MHz link. Band-Select 28 VDC Power HPF Limiter Switch The digital section can be DC-DC Amplifier

T/R Antenna configured to function as a Converter 1 (28 V–28 V) Control vector network analyzer for in-flight impedance mea- Duplexer DC-DC surements of the antennas. Converter 2 The RF section is divided (28 V–7.5 V) Driver BPF into two sub-sections: low- Amplifier power and high-power. The Voltage Regulator RF Section BPF VGA LNA low-power RF subsection (7.5 V–5.0 V) (Low-Power) includes a driver amplifier and filters to condition the Power Waveform transmit signal from the Conditioning Generator and Distribution Clock digital waveform genera- Distribution ADC1 tor and two-receiver chan- Master Avionics Data Clock nels, each composed of a Acquisition ADC2 low-noise amplifier in cas- System 900-MHz Link: cade with a variable-gain UTC Time/1PPS amplifier and an anti-alias- Status/Control/ GPS Digital Section Quick-Look ing filter. The high-power RF subsection includes a Figure 2. Simplified block diagram of the radar. duplexer circuit for trans- mit/receive (T/R) operations Analog-to-Digital/Digital-to-Analog (AD/DA) card, and and a single, high-efficiency pulsed power amplifier built other peripherals. The section includes a digital waveform with GaN transistors. A high-power, high-pass filter and generator with full control over the frequency and envelope a low-power, single-pole double-throw (SPDT) switch are of the transmit signal. This allows the transmitter signal likewise included in this sub-section. The high-pass filter is frequency and bandwidth to be adjusted to match antenna placed in front of the power amplifier to eliminate low-fre- characteristics, as was done during the field experiment in quency transients from power amplifier switching, while Antarctica; this is discussed in the next section. the low-power SPDT switch is placed at the receive port of A Hanning window is typically used as the pulse enve- the duplexer to select between the 14 MHz and 35 MHz lope function. The digital section also includes two sepa- receiver channels in the low RF power subsection. A sum- rate digitizer channels capable of sampling at 50 MSPS. mary of relevant radar parameters is included in Table 2. Each channel is dedicated to one of the two bands of The radar is housed in a box with dimensions of operation. Multiple digital lines are available as control 20.3 cm # 15.2 cm # 13.2 cm. The radar electronnics are lines for the RF circuitry. The raw radar data are streamed stacked vertically with the DC power conditioning sec- to an on-board high-capacity SD card. Command, sta- tion at the bottom, the digital section and low-power RF tus, and data snapshots are transmitted through a low- subsection in the middle, and the high-power RF subsec- tion at the top. Figure 3 shows a photograph of the assem- bled radar with the top cover removed. Table 2. Radar Parameters. 4.2. Antenna Structures Parameter Value (typ.) Operating frequency (lower/upper) 14 MHz/35 MHz The G1X carries two separate antennas integrated onto its Bandwidth (lower/upper) 1 MHz/4 MHz wings and airframe, as shown in Figure 4. The 14 MHz Transmit power (peak) 100 W structure is a resistively-loaded dipole implemented with Estimated peak radiated power a45 W/77 W copper tape and removable wires that run around the Pulse duration 1 ns (adjustable) perimeter of the airframe. The 35-MHz antenna is a tapered Pulse repetition freq. 10–20 kHz (adjustable) planar dipole implemented with copper tape. Both anten- Sampling frequency 50 MHz nas are fed using ferrite baluns. Prior to system integration Weight /volume a2 kg/0.0041 m3 (as shown in Figure 4a), both antennas can operate simul- DC power consumption a20 W (using 28 VDC) taneously at their respective frequencies as they both have a natural resonance at their design frequencies. The simulated

12 ieee Geoscience and remote sensing magazine march 2014 20.3 cm 10 X (Resistive Load)

15.2 cm Feed

14 MHz Antenna (Red) 35 MHz 13.2 cm Antenna (Orange) Removable Wire (a)

Avionics/Radar

Carbon Fiber Spars Figure 3. Photograph of the radar electronics. Servo Wires return loss before integration was obtained using a High (b) Frequency Structure Simulator (HFSS) from ANSYS, shown in gray in Figure 5. After system integration, we configured Figure 4. HF/VHF radar antennas implemented on the G1X platform the radar to perform in-flight antenna impedance mea- (a) before system integration and (b) after system integration. surements in the field, and the results are given in red in Figure 5. Performance degradation was observed for both 0 antennas due to the close proximity of the servo wires and other conductive objects. To compensate for the frequency -5 shift and maximize the antenna bandwidth, two impedance -10 matching networks were designed, optimized, and built in the field using the in-flight measured impedance data. With -15

the added impedance matching networks, we have verified K S11 (dB) -20 14 MHz Antenna the antenna performance with another in-flight measure- -25 ment; the result is given in Figure 5 (green line). A 10-dB 35 MHz Antenna return loss bandwidth of 1 MHz is obtained at the lower -30 band and a bandwidth of 4 MHz is obtained at the upper 10 15 20 25 30 35 40 45 Frequency (MHz) band. These results are in agreement with a feasibility study performed before the field experiment. In Flight Radar Meas. (w/o MN) HFSS Sim. (After System Int.) HFSS Sim. (Before System Int.) 5. Field Operations In Flight Radar Meas. (w/ MN) Test flights and data collection with the radar-equipped UAS, including airborne and surface-based measure- Figure 5. In-flight measured and simulated return loss of the radar ments, were carried out as a part of ongoing research at antennas for different operating conditions. CReSIS during the 2013–2014 field season. The airborne program included two components: (1) multi-frequency radars on BT-67 aircraft over the grounding lines of In terms of the UAS mission, we have repeated the ini- Whillans and Bindschadler ice streams on the Siple Coast tial flight test in the field to ensure all flight-critical sys- of West Antarctica; and (2) UAS flight tests and data col- tems were functioning properly in local conditions. Phase lection with the UAS-based dual-frequency radar over I field testing began with line-of-sight flight tests and Sub-glacial Lake Whillans (SLW) and the WISSARD drill included additional flight tests to assess and verify com- site. The surface-based program included measurements munications, control the transition between manual and with the dual-frequency UAS radar from a sled to verify its autonomous flight, and assess and verify radar antenna functionality and measurements with a wideband radar performance and functionality. Phase II flight tests were operating over the frequency range of 600–900 MHz to autonomous, included over-the-horizon flight function- better estimate the bottom melt rates of ice shelves. The ality, and focused on improving ground track accuracy of graphic in Figure 6 illustrates the concept of our airborne the aircraft flight path and enhancing the performance of program, as well as the field camp. both the HF and VHF sounders. The G1X UAS was flown march 2014 ieee Geoscience and remote sensing magazine 13 0 0 -5 GPS Satellites 2 -10 -15 Iridium 4 Satellites -20 6 BT-67 -25 8 -30 Delay ( n s) - 10 35 G1X -40 12 Sled -45 Camp 14 -50 -11-0.5 0 0.5 Easting (km) (a) Figure 6. Field site location and operations in West Antarctica with airborne and surface measurements.

14 times at the SLW field camp. In all flights, the air- (b)(b) craft was instrumented with the dual-frequency HF/VHF radar. Seven of those flights were autonomous, including Bed ReflectiviReflectivityty ((dB)dB) an over-the-horizon flight. 0.01 0 -10 -0.01 -20 6. Field Results -0.02 Multiple UAS/radar measurements were made during the -0.03 -30

Northing (km) -0.04 field deployment to the SLW camp. Prior to radar testing, 0.27 0.29 0.31 0.33 0.35 0.37 we performed return loss measurements of both anten- Easting (km) nas using a hand-held network analyzer to determine the (c) center frequency and pulse width settings for the radar and arbitrary waveform generator. Based on these results, Bed Phase (radians) 3 the radar was operated at 13.5 MHz and 35 MHz, slightly 0.01 0 different than the originally planned 14 and 35 MHz. -0.01 0 Next, we conducted initial tests of the HF radar on the -0.02 UAS by mounting the UAS on a Nansen-sled and towing -0.03 Northing (km) -0.04 -3 the sled behind a snowmobile. We performed a multiple- 0.27 0.29 0.31 0.33 0.35 0.37 pass survey of the ice runway at 13.5 MHz along 11 tracks, Easting (km) each spaced by approximately 5 meters. The primary (d) objectives of this survey were to: (1) verify the operation of the system; and (2) determine the amplitude and phase Figure 7. Results from an 11-pass sled survey of the ice runway at coherence across multiple tracks to enable synthetic aper- SLW camp. (a) Echogram of a single pass. (b) Phase of the bed reflec- ture radar (SAR) processing to improve overall system tion on all passes. Expanded (c) amplitude and (d) phase of the bed sensitivity. Figure 7a shows an echogram of a single pass reflection to show coherence across multiple ground tracks. along the runway. The flat reflector occurring at approxi- mately 10 ns is the ice-bedrock interface and corresponds to an ice thickness of about 800 m, which is confirmed We started flight operations after verifying system by previous measurements in the area. Figure 7b shows performance on the ground. Airborne activities included the phase response of the bed-reflection mapped along in-flight return loss antenna characterization; multiple- all survey lines. The 360 degree phase change across the pass runway surveys operated in a remote controlled (RC) surveyed portion of the runway indicates a change in ice configuration; multiple-pass, fully autonomous, line-of- thickness of approximately 6.5 meters. It is difficult to sight surveys of the runway; and finally a fully autono- interpret the phase and amplitude coherence in Figure 7b, mous, over-the-horizon survey of the SLW WISSARD drill so an expanded section is provided in Figure 7c and 7d to site. We performed in-flight return loss measurements by show fading of both the amplitude and phase, respectively, configuring the radar with a directional coupler to per-

along the survey lines. These results show excellent coher- form measurements of the scattering parameter S11 . This ence in the fading response across multiple survey lines. was done by first calibrating the set-up using standard

14 ieee Geoscience and remote sensing magazine march 2014 0 0 0 0 - 2 10 2 -10 4 -20 4 -20 6 -30 6 One Loop 8 -30 Delay ( n s) -40 8

10 Delay ( n s) -40 10 12 -50 -50 14 12 -60 0 0.511.5 2 14 -60 Distance (km) 0 5101520253035404550 (a) Distance (km) Bed Reflectivity (dB) (a) -84.209 -4 -84.210 Bed Reflectivity (dB) -8 -84.211 -4 -12 -84.21 -84.212 -6 Latitude -16 -84.213 -20 -8 -154.58 -154.54 -154.5 -154.46 -84.22 -10 Longitude (b) -12 Latitude -84.23 Surface Clutter (dB) -14 -84.209 -49 -16 -84.210 -84.24 -51 -84.211 -18 -53 -154.5 -154.2 -153.9 -153.6 -84.212 Latitude -55 Longitude -84.213 -57 (b) -154.58 -154.54 -154.5 -154.46

Longitude Figure 9. Results from the autonomous over-the-horizon survey of (c) the WISSARD drill site. (a) Echogram showing takeoff, survey, and landing. (b) Bed reflectivity with expanded section of the survey loops. Figure 8. Results from the manual remote controlled survey of the ice runway. (a) Echogram showing take-off and the first few loops. lated between multiple passes. Although it may not initially (b) Bed reflectivity over all survey loops. (c) Integrated power be apparent in the data, there is a significant amount of sur- between 4 and 8 ns to indicate variations in surface clutter. face clutter due to the wide radiation pattern of the dipole- antenna illuminating the surface. To show how the antenna terminations (open, short, and 50-Ohm load) and finally pattern can modulate this clutter, Figure 8c shows the inco- measuring the antenna impedance during flight. Addi- herently averaged power between 4 and 8 ns mapped into tionally, the 900 MHz communication link provided a latitude and longitude, similar to the reflectivity map in real-time, quick-look display of radar a-scopes to verify Figure 8b. The image clearly shows that the clutter power operation during flight. This feature has been tested dur- changes when the broad side of the dipole antenna rotates ing over-the-horizon operations at ranges up to 11.2 km. off nadir as the platform rolls during sharper turns. We Figure 8a shows an echogram of the first few runway will process these multi-pass data to synthesize a cross- passes at 35 MHz when the system was operating in remote track antenna beam with low sidelobes to demonstrate the control mode. The radar started collecting data on the reduction of surface clutter in the next few months. ground prior to take-off, and as a result the location of the Finally, Figure 9a shows an echogram of data collected bed reflection transitions from 10 ns to 11 ns as the aircraft during the fully autonomous over-the-horizon flight to the gains altitude. Figure 8b shows the bed reflection ampli- WISSARD drill site, and Figure 9b shows the bed reflection tude mapped onto the flight track during the 15 passes. amplitude during the survey. The measured ice thickness of The amplitude decreases due to the antenna pattern effect, about 800 m agrees with previously reported results. Again, which is caused by the aircraft rolling during turns. The we observe a strong correlation of echoes between multiple results show that amplitude variations are strongly corre- passes, indicating 2-D aperture synthesis is realizable. march 2014 ieee Geoscience and remote sensing magazine 15 The tests performed and data collected during the Acknowledgments field deployment, including results generated in the This work was completed at the University of Kansas with field, demonstrate a strong potential for the integrated funding from the National Science Foundation (NSF) system to collect data over fast-flowing glaciers. We suc- Center for Remote Sensing of Ice Sheet (CReSIS) grant cessfully performed an over-the-horizon survey consist- ANT-0424589 and with matching grants from the State of ing of four closely-spaced tracks, which demonstrated Kansas. The authors would like to acknowledge ANSYS for the platform’s capability to collect data in a manner providing a dedicated HFSS license used for the antenna that allows for the application of array processing tech- modeling and optimization in the field. We would also niques across several flight lines. The ground-based sur- like to acknowledge the support of many on the CReSIS vey showed consistent phase and amplitude coherence team, including but not limited to J. Fuller, L. Metz, and across multiple tracks, further indicating the capability P. Place for design and assembly of electronic parts; B. for the application of array processing. Finally, the radar Camps-Raga for component design and electronic testing demonstrated the capability to collect sounding data and support; A. Paden and S. Vincent for designing and fab- detect bedrock with high SNR from a small UAV at both ricating the housing for the electronics; A. Bowman, T.J. 14 and 35 MHz. Stastny, M. Ewing for avionics integration and UAS flight test support; and J. Hunter for UAS structure support. We 7. Conclusion would also like to thank Ms. J. Collins for editing and We developed a small UAS integrated with a compact, formatting the article. Finally, we acknowledge the work light-weight low-frequency radar for measurements over of Ms. E. Post in creating several graphics. rapidly changing areas of the Greenland and Antarctic ice sheets. The integrated system was successfully flight References tested in West Antarctica. We sounded ice with a UAS- [1] O. M. Jeffries, J. E. Overland, and D. K. Perovich, “The Arctic shifts based radar for the first time and collected data over a to a new normal,” Phys. Today, vol. 66, pp. 35–40, Oct. 2013. grid with closely spaced lines to demonstrate the 2-D [2] A. Shepherd, et al., “A reconciled estimate of ice sheet mass bal- SAR processing needed to reduce surface scatter from the ance,” Science, vol. 338, no. 6111, pp. 1183–1189, 2012. extremely rough surfaces typical of fast-flowing glaciers. [3] Wild Climate Research Programme, “Understanding sea-level rise It extremely difficult (nearly impossible) to collect and variability,” in Proc. Workshop Report. IOC/UNESCO, Paris, data with manned aircraft over a closely-spaced grid for France, June 6–9, 2006. determining bed topography with a resolution of about [4] Intergovernmental Panel on Climate Change, “Climate Change 100–200 m for 5–7 km wide glaciers. Fine-resolution bed 2014: Synthesis report, contribution of working groups I, II and topography is required to model ice dynamics near the III to the 5th assessment report of the intergovernmental panel grounding lines of fast-flowing glaciers [39–40]. A small on climate change,” Geneva, Switzerland, Tech. Rep., 2014. UAS equipped with radar and GPS receivers is extremely [5] F. M. Nick, A. Vieli, M. L. Andersen, I. Joughin, A. Payne, T. L. well-suited for this application. Additionally, a small UAS Edwards, F. Pattyn, and R. S. W. van deWal, “Future sea-level rise can be used to collect data over closely-spaced lines, as from Greenland’s main outlet glaciers in a warming climate,” close as approximately 5 m, in the cross-track direction Nature, vol. 497, pp. 235–238, May 2013. for synthesizing a narrow antenna beam for reducing sur- [6] F. Pattyn, “A new three-dimensional higher-order thermome- face clutter. A small UAS also uses several orders of mag- chanical ice sheet model: Basic sensitivity, ice stream develop- nitude less fuel per hour than the traditional manned ment, and ice flow across subglacial lakes,” J. Geophys. Res., vol. aircraft used today for ice sounding. In remote locations, 108, no. B8, p. 2382, 2003. such as Antarctica, the cost associated with transporting [7] F. Pattyn, L. Perichon, A. Aschwanden, B. Breuer, B. de Smedt, and caching fuel is very high. O. Gagliardini, G. H. Gudmundson, R. C. A. Hindmarsh, A. Our future plans include processing and analyzing Hubbard, J. V. Johnson, T. Kleiner, Y. Konovalov, C. Martin, A. data collected during this field season, miniaturizing J. Payne, D. Pollard, S. Price, M. Ruckamp, F. Saito, O. Soucek, S. the radar further and reducing its weight to 1.5 kg or Sugiyama, and T. Zwinger, “Benchmark experiments for higher- less, and increasing the peak transmit power to about order and full Stokes ice sheet models (ISMIP- HOM),” Cryo- 300 W. Over the next few months, we plan to perform sphere Discuss., vol. 2, pp. 111–151, 2008. additional test flights in Kansas to further evaluate the [8] F. Rodriguez-Morales, S. Gogineni, C. J. Leuschen, J. D. Paden, J. Li, avionics and flight control systems, as well as to measure C. C. Lewis, B. Panzer, D. G.-G. Alvestegui, R. D. Hale, E. J. Arnold, in-flight impedance of the antennas. We will use the mea- L. Smith, C. M. Gifford, D. Braaten, and C. Panton. Advanced mul- sured antenna impedances in-flight to design optimized tifrequency radar instrumentation for polar research. IEEE Trans. matching networks to extend the radar bandwidth. Dur- Geosci. Remote Sensing. to be published [Online]. Available: http:// ing the 2014 or 2015 field seasons, we are planning to ieeexplore.ieee.org/stamp/stamp.jsp?ar number=06557071 deploy the UAS to Greenland to collect data over areas [9] C. Allen. (2008, Sept.). A brief history of radio echo sounding with extremely rough surfaces and fast-flowing glaciers, of ice. Earthzine [Online]. Available: http://www.earthzine. such as Jakobshavn. org/2008/09/26/

16 ieee Geoscience and remote sensing magazine march 2014 [10] A. H. Waite and S. J. Schmidt, “Gross errors in height indication [27] S. Gogineni, J. Li, J. Paden, L. Smith, R. Crowe, A. Hoch, C. Lew- from pulsed radar altimeters operating over thick ice or snow,” is, E. Arnold, F. Rodriguez-Morales, C. Leuschen, R. Hale, A. R. in Proc. Institute Radio Engineers, Int. Convention Rec., 1961, pt. 5, Harish, and D. Braaten, “Sounding and imaging of fast flowing pp. 38–53. glaciers and ice-sheet margins,” in Proc. 9th European Conf. Syn- [11] V. Bogorodsky, C. Bentley, and P. Gudmandsen, Radioglaciology. thetic Aperture Radar, Apr. 23–26, 2012, pp. 239–242. Dordrecht, Holland: D. Reidel Publishing Company, 1985. [28] S. Gogineni, J. Paden, F. Rodriguez-Morales, J. Li, A. Hoch, L. [12] S. Evans, “Progress report on radio echo sounding,” Polar Rec., Smith, J. Meisel, C. Leuschen, and D. Braaten, “Bed topography vol. 13, no. 85, pp. 413–420, 1967. for Jakobshavn, Helhiem, and Kangerdlussuaq glaciers,” in Proc. [13] S. Evans, Ed., “Review of radio echo system performance in Gud- Int. Union Geodesy Geophysics General Assembly, Melbourne, Aus- mandsen, P. E.,” in Proc. Int. Meeting Radioglaciology, Lyngby, Den- tralia, June 27–July 7, 2011. mark, 1970, pp. 100–102. [29] J. Li, J. Paden, C. Leuschen, F. Rodriguez-Morales, R. Hale, E. [14] S. Evans and B. M. E. Smith, “A radio echo equipment for depth Arnold, R. Crowe, D. Gmez-Garcia, and P. Gogineni, “High-alti- sounding in polar ice sheets,” J. Sci. Instrum. (J. Phys. E), ser. 2, tude radar measurements of ice thickness over the Antarctic and vol. 2, no. 2, pp. 131–136, 1969. Greenland ice sheets as a part of operation ice bridge,” IEEE Trans. [15] P. Gudmandsen, “Glacier sounding in the polar regions: A sym- Geosci. Remote Sensing, vol. 51, no. 2, pp. 742–754, Feb. 2013. posium—Part II. Airborne radio echo sounding of the Green- [30] R. D. Watts and D. L. Wright, “Systems for measuring thickness land ice sheet,” Geograph. J., vol. 135, no. 4, pp. 548–551, 1969. of temperate and polar ice from the ground or from the air,” J. [16] S. Gogineni, Private communication. Glaciol., vol. 27, no. 97, pp. 459–469, 1981. [17] G. Raju, W. Xin, and R. Moore, “Design, development, field ob- [31] Z. Rodrigo, D. Ulloa, G. Garcia, R. Mella, J. Uribe, J. Wendt, A. S. servations, and preliminary results of the coherent Antarctic ra- Rivera, G. Gacitua, and G. Casassa, “Airborne radar sounder for dar depth sounder of the University of Kansas, USA,” J. Glaciol., temperate ice: Initial results from Patagonia,” J. Glaciol.,” vol. 55, vol. 36, no. 123, pp. 247–254, 1990. no. 191, pp. 507–512, 2009. [18] T. S. Chuah, “Design and development of a coherent radar depth [32] H. Conway, B. Smith, P. Vaswani, K. Matsuoka, E. Rignot, and P. sounder for measurement of Greenland ice sheet thickness,” RSL Claus, “A low-frequency ice-penetrating radar system adapted for Tech. Rep. 10470-5, Jan. 1997. use from an airplane: Test results from Bering and Malaspina Gla- [19] S. Gogineni, T. S. Chuah, C. Allen, K. Jezek, and R. K. Moore, “An ciers, Alaska, USA,” Ann. Glaciol., vol. 50, no. 51, pp. 93–97, 2009. improved coherent radar depth sounder,” J. Glaciol., vol. 44, no. [33] E. Rignot, J. Mouginot, C. F. Larsen, Y. Gim, and D. Kirchner, 148, pp. 659–669, 1998. “Low-frequency radar sounding of temperate ice masses in [20] S. Gogineni, D. Tammana, D. Braaten, C. Leuschen, T. Akins, J. Southern Alaska,” Geophys. Res. Lett., vol. 40, no. 20, pp. 5399– Legarsky, P. Kanagaratnam, J. Stiles, C. Allen, and K. Jezek, “Co- 5405, 2013. herent radar ice thickness measurements over the Greenland ice [34] N. Blindow, “The University of Munster Airborne Ice Radar sheet,” J. Geophys. Res. (Climate and Phys. Atmos.), vol. 106, no. (UMAIR): Instrumentation and first results of temperate and D24, pp. 33,761–33,772, 2001. polythermal glaciers,” in Proc. 5th Int. Workshop Advanced Ground [21] J. J. Legarsky, P. Gogineni, and T. L. Akins, “Focused synthetic- Penetrating Radar, Granada, Spain, 2009, p. 13619. aperture radar processing of ice-sounder data collected over the [35] N. Blindow, C. Salat, and G. Casassa, “Airborne GPR sounding Greenland ice sheet,” IEEE Trans. Geosci. Remote Sensing, vol. 39, of deep temperate glaciers—Examples from the Northern Pata- no. 10, pp. 2109–2117, 2001. gonian Icefield,” in Proc. 14th Int. Conf. Ground Penetrating Radar, [22] F. Hélière, C.-C. Lin, H. Corr, and D. Vaughan, “Radio echo sound- pp. 664–669, 2012. ing of Pine Island Glacier, West Antarctica: Aperture synthesis pro- [36] S. A. Arcone, D. E. Lawson, A. J. Delaney, and M. Moran, “12-100 cessing and analysis of feasibility from space,” IEEE Trans. Geosci. MHz depth and stratigraphic profiles of temperate glaciers,” in Remote Sensing, vol. 45, no. 8, pp. 2573–2582, Aug. 2007. Proc. 8th Int. Conf. GPR, Gold Coast, Australia, May 23–26, 2000, [23] E. M. Peters, D. D. Blankenship, S. P. Carter, S. D. Kempf, D. A. pp. 377–382. Young, and J. W. Holt, “Along-track focusing of airborne radar [37] S. A. Arcone, “Airborne-radar stratigraphy and electrical struc- sounding data from West Antarctica for improving basal reflec- ture of temperate firn: Bagley Ice Field,” J. Glaciol., vol. 48, no. tion analysis and layer detection,” IEEE Trans. Geosci. Remote 161, pp. 317–334, 2002. Sensing, vol. 45, no. 9, pp. 2725–2736, Sept. 2007. [38] A. Kong, “Effective permittivity for a volume scattering medi- [24] J. Paden, T. Akins, D. Dunson, C. Allen, and P. Gogineni, “Ice- um,” in Electromagnetic Wave Theory. New York: Wiley, 1986, ch. sheet bed 3-D tomography,” J. Glaciol., vol. 56, no. 195, pp. 3–11, 6, sec. 7, pp. 550–563. Jan. 2010. [39] G. Durand, O. Gagliardini, L. Favier, T. Zwinger, and E. Le Meur, [25] K. Jezek, X. Wu, P. Gogineni, E. Rodriguez, A. Freeman, F. Rodri- “Impact of bedrock description on modeling ice sheet dynam- guez, and C. Clark, “Radar images of the bed of the Greenland ics,” Geophys. Res. Lett., vol. 38, no. 2, 2011. ice sheet,” Geophys. Res. Lett., vol. 38, Jan. 2011. [40] I. Joughin, B. E. Smith, D. E. Shean, and D. Floricioiu, “Brief [26] X. Wu, K. C. Jezek, E. Rodriguez, S. Gogineni, F. Rodriguez-Mo- communication: Further summer speedup of Jakobshavn Is- rales, and A. Freeman, “Ice sheet bed mapping with airborne SAR bræ,” Cryosphere, vol. 8, pp. 209–214, 2014. tomography,” IEEE Trans. Geosci. Remote Sensing, vol. 49, no. 10, pp. 3791–3802, Oct. 2011. grs march 2014 ieee Geoscience and remote sensing magazine 17