Feature Fusion based Unsupervised Change Detection in Optical Satellite Images

Abstract

This paper proposes a feature fusion technique for unsupervised change detection. Features extracted from two different techniques are fused to get the final feature vectors. The first technique utilizes the Gabor wavelet at multiple orientations and scales, where maximum magnitude over all orientation in each scale is taken to create features of two multitemporal satellite images. The second technique applies canonical correlation analysis (CCA) on the combination of original multispectral bands and extracted local neighborhood information from all the bands. Next, the difference feature vectors obtained from individual techniques are fused to generate the final feature vectors. Furthermore, to get the binary change map, fuzzy c-means clustering is applied on final extracted features. In this feature fusion, the local neighborhood information from Gabor wavelet kernel is combined with joint change information from group of pixels extracted by CCA to produce more discriminant features.Experiments conducted on optical satellite images, which are collected by two sensors of Landsat satellite, and it shows the better performance of the proposed technique compared to earlier stated techniques.