Deep Canonical Correlation Analysis Network for Scene Change Detection of Multi-Temporal VHR Imagery

Change detection at semantic scene level has now been an important topic of high spatial resolution remote sensing imagery analysis. In this paper, combining with Deep Canonical Correlation Analysis (DCCA), we proposed an end-to-end network (DCCA-Net) for scene change detection. DCCA-Net firstly utilizes a pretrained Convolutional Neural Network (CNN) to extract high-dimensional features of the input scene pairs. Then, the DCCA module is deployed to project the extracted features into a new feature space and maximize the correlation of the projected feature pairs. Finally, based on the transformed feature vectors, the semantic label of each scene image could be obtained by a softmax classifier. The binary scene changes could be obtained using a binary classifier based on the transformed features. The objective function of DCCA is integrated into the loss function of classification and binary change detection, so that they could be optimized simultaneously. We implemented the proposed network and performed experiments on a VHR remote sensing imagery dataset (Multi-temporal Scene - Wuhan, Mts-WH) for scene change detection. The experimental results have shown the effect of the DCCA-Net, and our method could outperform other conventional scene change detection methods and the baseline methods based on CNN.