COMPARISON OF PoISAR AND PolInSAR COHERENCE BASED LAND USE/LAND COVER CLASSIFICATION

摘要

Classification of SAR data is a challenging as well as an essential task of continuous monitoring of the earth's surface due to its all-weather capability. This paper investigates PolSAR and PolInSAR coherence concept for the extraction of the earth features. The methodology integrated with Polarimetric decomposition models, PolInSAR Coherence, Mahalanobis and Knowledge based classification for land use and land cover (LULC) classification using fully polarimetric RADARSAT- 2 data. Polarimetric decompositions helped to investigate and understand the scattering mechanism patterns to extract the information of the earth's features. Freeman, Yamaguchi and H/A/Alpha decomposition models were used to extract different scattering mechanisms. The results shown overestimation of the volume scattering in highly dense and oriented building, this leads to misinterpretation of these buildings as forest in PoISAR classification. This issue is addressed by incorporating PolInSAR coherence, which is sensitive to both volumetric structure and temporal change of the scatterer. Forest regions were strongly affected by the temporal and volume decorrelation with time, which show low coherence values compared to the permanent scatterers like built-up areas. The backscatter response and the coherent patterns of vegetation compared to urban area are different, this helped to extract built-up regions more accurate. Therefore, PolInSAR coherence is utilized for extraction of the features by combining polarimetric and coherence information. The PolInSAR coherence extracted from the repeat-pass RADARSAT-2 images of 24 days temporal baseline for distinguishing vegetation and built-up areas. The Mahalanobis classification algorithm was used for extracting different features. To further improve the results, Knowledge based classification was executed by forming rules based on the statistical analysis of the features. The overall accuracy and kappa statistics of the PoISAR and PolInSAR Coherence classification are 79.17% and 0.75, 86.67% and 0.84 respectively. This states that PolInSAR coherence classification helped to characterize discrete and volume scatterers more accurately than the PoISAR classification.