Abstract
The aim of this article is to investigate the capabilities of multitemporal RADARSAT-2 fine-beam polarimetric synthetic aperture radar (SAR) data and RADARSAT-2 ultra-fine-beam C-band single-polarization HH SAR (C-HH SAR) data for detailed urban land-cover mapping using a contextual approach. With an adaptive Markov random field and a spatially variant finite mixture model, contextual information was effectively explored to improve the mapping accuracy. A texture enhancement in FMM was further proposed to improve the classification accuracy. Moreover, a rule-based approach exploring object features and spatial relationships was employed to extract road, street, and park. Three-date RADARSAT-2 fine-beam polarimetric SAR (PolSAR) and three-date RADARSAT-2 ultra-fine-beam C-HH SAR data over the Greater Toronto area were used for the evaluation. For 10 major classes, the overall accuracy (OA) is 51% for C-HH SAR data and 79% for PolSAR data. Compared with C-HH SAR, PolSAR data produced better results for identifying various urban patterns. Although with multi-date, the C-HH SAR data showed low capability to distinguish high-density residential area and industry commercial area (Ind.). Considerable low-density residential area (LD) was misclassified as forest. Identification of the construction site (Cons.) and golf course were poor. Nevertheless, the efficiency of the multitemporal C-HH SAR textures for distinguishing the built-up areas was observed. By texture enhancement with the synergy of the PolSAR and C-HH SAR data, the mapping results could be significantly improved, especially for LD, forest, and crops. The OA is improved by 2.7% for PolSAR data, and 11.1% for C-HH SAR data. Road, street, and park could be extracted by the rule-based approach with OA about 77% for 13 classes.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have