Computer Aided Recognition of Relief Patterns on Radar Images Using a Syntax Analysis

We present computerized recognition of relief patterns on an airborne SAR image, band X (3.2 cm), using a syntax analysis. The test-site is the Montagne Sainte-Victoire (Provence, France). The pixel size is 2m×2m. Severe filtering to eliminate the speckle has reduced ground resolution, but geologic and geomorphologic objects to extract—thalwegs, structural escarpments, lithologic boundaries, and landslides—have large dimensions (tens of meters). Far-range correction was necessary to compensate weakening of radar signal with aircraft distance. Significant retrodiffusion values are for horizontal surfaces, projected shadows, surfaces whose normal is subperpendicular to radar beam, steep radiometric gradients. Elements of the topographic profiles have been correlated with the elementary signals of radar image, permitting a codification for syntactic analysis. The radar image has been transformed into an image of elementary topographic symbols, from which crests and thalwegs can be directly extracted. Recognition of structural escarpments and landslides involves search for typical sequences of symbols in different profile orientations, and determination of a specific grammar (syntax). Assessment of the results was made by comparison with similar syntactic methods of pattern recognition applied to a digital elevation model of the same area, and field knowledge. The radar image classifications have produced results close to reality. Imperfections are mainly due to the filtering process, radar distance correction, and difficulty in measuring significant retrodiffusion values, for example, retrodiffusion of horizontal surfaces. Artifacts inherent to radar acquisition also introduced confusions.