Deep Siamese Multi-scale Convolutional Network for Change Detection in Multi-temporal VHR Images

Abstract

In this letter, we propose a powerful multi-scale feature convolution unit for change detection. The proposed unit is able to extract multi-scale features in the same layer. Based on the proposed unit, two novel deep Siamese convolution networks, deep Siamese multi-scale convolutional network (DSMS-CN) and deep Siamese multi-scale fully-convolutional network (DSMS-FCN), are designed for unsupervised and supervised change detection in multi-temporal very high resolution (VHR) images. For unsupervised change detection, we implement automatic pre-detection to obtain training patch samples, and the DSMS-CN fits the statistical distribution of changed and unchanged ground from patch samples for change detection through multi-scale feature extraction module and deep Siamese architecture. For supervised change detection, an end-to-end deep network DSMS-FCN is trained in any size of multitemporal VHR images, and directly output the binary change map. The experimental results with a GF data set and an open change detection data set confirm that the two proposed architectures perform better than the state-of-the-art methods.