Ocean Feature Analysis Using Automated Detection and Classification of Sea-Surface Temperature Front Signatures in RADARSAT-2 Images by Chris T
Total Page:16
File Type:pdf, Size:1020Kb
Ocean Feature Analysis Using Automated Detection and Classification of Sea-Surface Temperature Front Signatures in RADARSAT-2 Images BY CHRIS T. JONES, TODD D. SIKORA, PARIS W. VACHON, AND JOSEPH R. BUCKLEY he Royal Canadian Navy’s Meteorology and fronts that are signatures of SST fronts. RADARSAT-2 Oceanography Centre (MetOc) Halifax currently images therefore represent a potentially significant T produces a semiweekly ocean-feature analysis additional data source for OFA production. (OFA) to provide the fleet with the location of major water mass boundaries in the Western North At- SST FRONT SIGNATURES IN SAR. SAR im- lantic Ocean, such as the Gulf Stream North Wall ages reveal spatial variations in small-scale ocean (GSNW). OFAs are generated using temperature surface roughness, forced primarily by near-surface measurements from a network of buoys, temperature winds and modulated by longer waves, currents, profile measurements from ships, satellite sea surface and surfactants. A positive buoyancy flux within temperature (SST) images from the Advanced Very the marine atmospheric boundary layer (MABL) on High Resolution Radiometer (AVHRR), and SST the warm side of an SST front can force convective climatology. An automated analysis integrates all of processes that transport momentum from the upper the data using geophysical interpolation techniques MABL toward the surface. Furthermore, the atmo- (kriging) to fill in gaps, such as those due to cloud spheric temperature gradient often present across an cover. Cloud cover is prevalent in the region, with SST front can produce a corresponding cross-front clear-sky percentage varying from about 50% in late atmospheric pressure gradient. These processes can summer to less than 5% in winter (according to the lead to an enhancement in near-surface wind speed International Satellite Cloud Climatology Project), on the warm side of the front, with concomitant hampering OFA production. intensification of surface roughness compared to The synthetic aperture radar (SAR) sensor on the the cold side. The processes can therefore lead to a RADARSAT-2 satellite measures variations in the brightness front in the SAR image that is collocated roughness of the ocean’s surface unrestrained by with the SST front. cloud cover. These variations manifest as dark (low microwave backscatter) regions and bright (high SPACEBORNE OCEAN INTELLIGENCE microwave backscatter) regions on the SAR image, NETWORK. The Spaceborne Ocean Intelligence and often include high-contrast edges or brightness Network (SOIN) is a six-year project, supported by the Canadian Space Agency via its Government Related Initiatives Program, mandated to develop AFFILIATIONS: JONES—Department of Mathematics and Statistics, Dalhousie University, Halifax, Nova Scotia, Canada; procedures that can automatically detect ocean SIKORA—Department of Earth Sciences, Millersville University, SST front signatures in RADARSAT-2 images. Millersville, Pennsylvania; VACHON—Defence R&D Canada— The centerpiece of the project is a detection-and- Ottawa, Ottawa, Ontario, Canada; BUCKLEY—Royal Military classification algorithm that uses an edge detector College of Canada, Kingston, Ontario, Canada to identify brightness fronts in SAR images that may CORRESPONDING AUTHOR: Chris T. Jones, 95 Central be SST front signatures. A statistical classification Avenue, Fairview, NS B3N 2H7, Canada algorithm is then used to discriminate SST front sig- E-mail: [email protected] natures from horizontal wind shear signatures not DOI:10.1175/BAMS-D-12-00174.1 associated with SST fronts. This automated process ©2014 American Meteorological Society achieves classification accuracy greater than 80% in the vicinity of the GSNW. PB | MAY 2014 AMERICAN METEOROLOGICAL SOCIETY MAY 2014 | 677 Unauthenticated | Downloaded 09/25/21 06:50 PM UTC TABLE 1. The contingency table shows the identity of the validated Canny edges with confidence due to lack (columns) partitioned by row according to the assigned class. Edges were clas- of validating SST data. sified as SST front signatures only when the mean angle between the edge The 495 Canny edges were and the wind direction was greater than 23°. This resulted in an accuracy of subjected to statistical analysis (276 + 133)/495 = 0.83, or about 80%. Fisher’s Exact Test showed a significant association between rows and columns with p-value < 0.000. with the objective of identify- ing textural or contextual Validated as SST Validated as not SST Total measures near the fronts that could be used for statistical Classified as SST 276 60 336 classification. The only in- formative measure identified Classified as not SST 26 133 159 was the mean angle between Total 302 193 495 a Canny edge and the wind direction. Using nearest-in- time near-surface QuikSCAT SUMMARY OF SOIN EDGE DETECTION Level 3 data (25 km in resolution) obtained from NASA’s AND CLASSIFICATION RESEARCH. SOIN Physical Oceanography Distributed Active Archive fostered the research presented in 2012 and 2013 Center (PO.DAAC) website (http://podaac.jpl.nasa articles published by Jones et al. in the Journal of .gov), it was found that signatures of horizontal wind Atmospheric and Oceanic Technology, the latter of shear not associated with SST fronts tended to coincide which is outlined here. with near-surface winds oriented along the Canny edge, A total of 1,227 RADARSAT-2 ScanSAR Narrow A typically at angles less than 23° from the along-edge images (C-Band, VV polarization, 300 km by 300 km) direction. In contrast, near-surface winds tended to be were provided to the SOIN team by MetOc Halifax. oriented across SST front signatures at angles greater Those images were preprocessed prior to edge detec- than 23°. Classification of the 495 Canny edges using tion. This consisted of the application of a land mask, a mean wind angle with a decision boundary of 23° filter to remove signatures of strong target sources such resulted in a classification accuracy of 0.83 (Table 1). as ships, block averaging to reduce image size (and thus Using Fisher’s Exact Test, the frequency of agreement computational cost) from 12,000 by 12,000 pixels (cor- between validated identity and assigned class was found responding to pixel spacing of 25 m) to 1000 by 1000 to be significantly greater than that expected by chance pixels spacing (with pixel spacing of 300 m), and radio- alone (p value < 0.0000). metric flattening to account for systematic variations in backscatter due to changes in viewing geometry across OPERATIONAL APPROACH. In the SOIN the image. A Canny edge detector was then applied to context, a practical operational approach to SST each image to identify brightness fronts. front detection in RADARSAT-2 images will entail A set of 495 Canny edges consisting of 302 SST automated detection and classification of brightness front signatures and 193 signatures of horizontal wind fronts followed by manual input and evaluation. First, shear not associated with an SST front—all located in an image will be processed to produce candidate SST the vicinity of the GSNW—was collected from 252 front signatures using the Canny edge detector. To RADARSAT-2 images. Comparing each RADARSAT-2 reduce the number of edges to a manageable number, image with concurrent MODIS SST and surface only those on the order of 80 km or more in length weather analyses validated the identities assigned to will be retained for analysis. The mean angle between the Canny edges. Edges identified as SST front signa- the near-surface wind direction and each edge will tures were those matching the location, orientation, be calculated, and edges with a mean angle of more and shape of an SST front in a MODIS image. Edges than 23° will be classified as SST fronts. Winds from aligned with an atmospheric front in a surface analysis the Global Environmental Multiscale (GEM) Model, chart were identified as being signatures independent obtained from the Canadian Meteorological Centre, of SST fronts. Edges not included in the analysis were will be used to determine the average angle between those associated with readily identifiable signatures the wind and a detected brightness front. of processes such as atmospheric or oceanographic A manual analysis will then be carried out by gravity waves and large-scale convective squalls, as MetOc personnel trained to use contextual cues (e.g., well as edges that could not otherwise be identified the previous OFA, visual cues in the SAR image that are 678 | MAY 2014 AMERICAN METEOROLOGICAL SOCIETY MAY 2014 | 679 Unauthenticated | Downloaded 09/25/21 06:50 PM UTC commonly associated with SST fronts near the GSNW, surface weather analysis) to correct misclassifications and/or to check for potential SST front signatures not found by the edge detector. Water mass boundaries will then be drawn based on all available information (see Fig. 1). The efficacy of this semiautomated approach will be tested once the operational procedure is in place. FUTURE DIRECTIONS. Although SAR observa- tion of the ocean surface is less constrained by cloud cover than infrared radiometers, strong atmospheric processes can sometimes obliterate signatures pro- data and products ©MacDonald, Dettwiler, and Associates Ltd. (2009)—all rights reserved. Associates Ltd. and Dettwiler, ©MacDonald, data and products duced by SST fronts. The question remains as to whether an OFA that includes SST fronts derived from RADARSAT-2 SAR images is more informative than an OFA that does not. To answer this question, the probability that an SST front signature exists in a RADARSAT-2 image of a region near the GSNW, at a time when cloud cover FIG. 1. This figure shows a RADARSAT-2 ScanSAR Wide prevents front identification using satellite SST, must A image (C-Band, VV polarization, 500 km by 500 km) be determined. An estimate of this probability cannot of a region south of Nantucket acquired on 22 Sep yet be provided because in the present study, cloud-free 2011 at 2232 UTC.