Multitemporal RADARSAT-2 ultra-fine beam SAR data for urban land cover classification

Abstract

High-resolution optical satellite images have been widely used to update land cover information and monitor changes in urban areas. Several spaceborne synthetic aperture radar (SAR) systems are now providing SAR imagery with a spatial resolution comparable to high-resolution optical systems. Although SAR data is more reliably available than optical data, it takes more effort to employ high-resolution SAR imagery for urban applications owing to the difficulty in interpreting the complex content in SAR imagery over urban areas. The objective of this research was to develop effective object-based and rule-based methods for classification of high-resolution SAR imagery over urban areas. Multitemporal RADARSAT-2 ultra-fine beam C-HH SAR images with a pixel spacing of 1.56 m were acquired over the north part of the Greater Toronto Area during June to September in 2008. The SAR images were preprocessed and then segmented by means of a multiresolution segmentation algorithm. A range of spectral, geometrical, and textural features were selected and calculated for image objects. The image objects were classified based on these features using support vector machines (SVM). Compared with the nearest neighbor classifier, the object-based SVM produced much higher urban land cover classification accuracy (Kappa 0.43 vs. 0.63). The SVM classification result was then improved by developing specific rules to resolve the confusion among some classes. The final result indicated that the proposed methods could achieve a satisfactory overall accuracy (81.8%) for urban land cover classification using very high-resolution RADARSAT-2 SAR imagery.