A Novel Deep Framework for Change Detection of Multi-source Heterogeneous Images

Change detection of remote sensing images is to detect changes in multiple images (at least a pair of images) obtained at different time points. However, almost all existing change detection methods based on deep learning need to be trained repeatedly on different datasets (especially on heterogeneous datasets) to obtain satisfactory results. Besides, many of these methods have to be trained on large numbers of samples acquired by classical pre-classification methods. In this paper, a novel change detection framework based on meta-learning, called MLCD, is proposed. We model the change detection problem as a one-shot learning problem, although there are no labels for reference in reality. Inspired by active learning, we present two sample selection strategies to find a few representative samples. Based on the key idea of meta-learning, our framework mainly consists of a modified convolutional neural network and a graph neural network, and it is capable of learning to compare samples in the embedded feature space. Meanwhile, it only needs to be trained once with very few samples and can be directly transferred to test changes on other heterogeneous images. Experimental results on both Synthetic Aperture Radar (SAR) and multispectral images demonstrate the effectiveness and superiority of our framework over state-of-the-art methods.