Multitemporal RADARSAT-2 Polarimetric SAR Data for Urban Land Cover Classification Using Support Vector Machine

This research investigates the various RADARSAT-2 polarimetric SAR features for urban land cover classification using object-based method combining with support vector machine (SVM) and ruled-based approach. Six-dates of RADARSAT-2 fine-beam polarimetric SAR data were acquired in the rural-urban fringe of Greater Toronto Area during June to September, 2008. The major landuse/land-cover classes were high-density built-up areas, low-density built-up areas, roads, forests, parks, golf courses, water and several types of agricultural crops. The polarimetric SAR features examined are the parameters from Pauli, Freeman and Cloude-Pottier decompositions as well as the elements from coherence matrix and the intensities and their logarithm form of each channel. For urban land cover classification, SVM is combined with rule-based method for the object-based classification. The image objects containing the multitemporal polarimetric features were classified using the SVM classifier first. The SVM classification results were further refined using a rule-based approach. Rules were built to recognize specific classes defined by the shape features and the spatial relationships within the context. In terms of the effectiveness of different SAR ploarimtric parameters, the results indicated that the processed Pauli feature set could produce best classification result while the use of all the polarimetric features did not produce the best classification result. The raw Pauli parameters could generate similar result as all T elements. The logarithm parameters such as log intensity and processed Pauli parameters perform better than the intensity and raw Pauli respectively. The proposed object-based classification using SVM and rule-based approach yielded higher classification accuracies than the object-based classification using nearest neighbor classifier.