Hyperspectral image change detection based on active convolutional neural network and spatial–spectral affinity graph learning

Abstract

The high spectral resolution of hyperspectral image (HSI) provides the possibility to capture the subtle changes associated with land-cover dynamic evolution process. Supervised deep leaning approaches have been extensively applied to HSI change detection task. However, their success is attributed to the large amount of annotated training samples. Moreover, the large receptive field in the convolutional layer and the presence of the pooling layer reduces the spatial resolution of the deepest FCN layer which will make the predicted change map tends to lack fine object boundary details. In order to boost the HSI change detection performance in a reduced labor of annotating data and enhance the object boundary details of the change map, in this paper, we propose a novel HSI change detection method that integrates both iterative active learning and affinity graph learning into a unified framework. The proposed method consists of two major branches, a unary HSI change detection network branch which learns the pixel-wise change probability, and a pairwise HSI affinity graph learning branch that learns the pairwise affinity of the hyperspectral difference image to refine the coarse probabilities through a random walk connection. To actively select the most informative unlabeled samples, we also propose an HSI change detection active learning strategy based on the spectral property of HSI difference image and the refined change probabilities. The procedures are conducted iteratively to obtain better change detection results progressively. Experimental results show that the proposed method can extract accurate subtle change information while properly preserving the edges and textures of the HSIs with significantly fewer labeled data.