Unsupervised parameter estimation based on structure detection and clustering for polarimetric SAR speckle filtering

Abstract

A novel parameter estimation method for polarimetric synthetic aperture radar (PolSAR) speckle removal was introduced, which is a combination of adaptive structure detection and K-means clustering. First, an effective geometrical ratio detection approach was used to define training vectors. Based on the textural characteristic, the vectors were then classified into four types, i.e., homogeneous or slightly textured area, edge, line, and strong scatterer by clustering. These clustering results were then applied to estimate filter parameters. The effectiveness of the proposed method was demonstrated using NASA SIR-C/X-SAR, L-band, four-look PolSAR data of Tian Mountain, China. Finally, the multilook polarimetric whitening filtering algorithm was also used for comparing this method with three baseline ones, i.e., the Boxcar filter, the Prewitt edge detector, and the geometrical ratio detection method. Experimental results showed the superior performance of our proposed method on filter parameter estimation.