Gabor and PCA Feature-Based Unsupervised Change Detection in Sar Images

The detection rate of the unsupervised change detection algorithm is affected by the incorrect identification of the unchanged and changed pixels. The reduction of such false pixels in the binary change map is still a challenging issue. This work aims to reduce false pixels by combining Gabor features and Principal Component Analysis (PCA) features. First, log transform is applied to input Synthetic Aperture Radar (SAR) images. Next, a Difference Image (DI) is generated using subtraction operator. Then, 2-D Gabor wavelet transform and PCA are employed to extract the features from the DI. Further, both features are combined to produce the final feature vector. Finally, a two-stage clustering technique is used to obtain change map. Experiments have been carried out on San Francisco and Bern datasets. The proposed method produces less overall error (768 and 271) as well as higher kappa values (0.9153 and 0.8775) for the two datasets.