Radiometric Normalization on SAR Images for Oil Spill Detection
Singha, Suman1; Topouzelis, Konstantinos2; Vespe, Michele3; Trieschmann, Olaf4
1Deprtment of Geography, University of Hull, UNITED KINGDOM; 2Department of Marine Sciences, University of the Aegean, Greece, GREECE; 3Centre for Maritime Research and Experimentation (CMRE), NATO STO, Italy, ITALY; 4Satellite Based Monitoring Services, European Maritime Safety Agency (EMSA), Portugal, PORTUGAL

Oil spills on the European waters are observed relatively often. Pollution due accidental and deliberate oil spill from ships and rigs represents a serious threat to the marine and coastal environment. Operational oil spill monitoring such as 'CleanSeaNet' hosted by European Maritime Safety Agency (EMSA) is currently using SAR based maritime surveillance for oil spill and ship detection. A total of 2069 potential spills were reported by EMSA in 2012 reduced from 3311 reported spill in 2008. Primarily, conventional wide swath mode X-band (2.4-3.75 cm) and C-band (3.75-7.5 cm) SAR products have been used for operational purpose due to its large coverage and cost effectiveness
. Wide swath images are generally affected by trends of reduced radar backscatter of the sea induced by the incidence angle increase at far range. This particular phenomenon, which is intrinsic to SAR images, needs to be taken into account when automatic oil spill detection is approached using statistical methodologies. This is commonly done either by implementing adaptive threshold estimation throughout the image, or by introducing image segmentation stage that subdivides the image in regions characterised by homogeneous backscatter statistics. Alternative normalisation approaches are here presented, motivated by a reduced complexity and computational burden to downstream dark feature detection algorithms. Three different techniques are introduced for compensating backscatter trend with particular focus on the oil spill detection problem: i) Backscattering shape function derivation. This approach consists in applying a theoretical backscattering profile in elevation direction, scaled by a calibration factor. This methodology leads to a short execution time but is less effective when local differences in sea state are present in the area of interest (particularly likely for wide swath SAR images). ii) 1-D profile extraction and removal. This approach consists in low-pass filtering a profile representing the backscattering along the across-track (elevation) direction, averaged over the along-track (azimuth) direction. This methodology is more time-consuming than the shape function derivation, and needs prior land-masking implementation. Nonetheless, the resulting corrected image is less influenced by local sea state variations; iii) 2-D background estimation and removal. The backscatter background is calculated by averaging and smoothing the image after land-masking operation. This image background removal process depends on the backscatter data and is therefore only indirectly linked to the incidence angle. Land-masked images are divided into sub-images, subsequently normalized using local backscatter statistics. This approach is more accurate than the previous ones in terms of normalisation performance since it locally estimates the backscattering floor, significantly removing the background variations due to the sea state conditions.