A Novel Approach for Hyperspectral Change Detection Based on Uncertain Area Analysis and Improved Transfer Learning

Although a number of change detection (CD) methods have been proposed during the past years, most of them are developed based on the assumption that there are either training samples or no training samples for both the pretime and posttime images. Few studies have addressed the scenario of only small amounts of training samples are available only in a single-time image. In this article, we propose a novel approach that can detect multiple changes in bitemporal hyperspectral images when only a few training samples are available in one of the images (the source image). The proposed method consists of four main steps: first, unsupervised CD based on uncertain area analysis to generate the binary change map; second, classification of the source image (X1) according to active learning; third, classification of the target image (X2) by the use of improved transfer learning; and fourth, generation of the multiclass change map by postclassification comparison. The proposed method was tested on one simulated dataset and two pairs of real bitemporal hyperspectral images. Experimental results demonstrate that: first, uncertain area analysis can improve the binary CD accuracy; while active learning and improved transfer learning can enhance the classification accuracy of the source and target images, the multiple CD accuracy is increased by the use of the proposed method; and second, compared with the state-of-the-art methods, the proposed method produced best results.