Abstract: Change detection aims to observe the expression differences of ground objects in different time series. Deep learning has become the mainstream method to achieve this task. In the existing remote sensing change detection methods based on deep learning, they generally focus more on learning deep semantic features in images, while ignoring the semantic advantages and gaps between different levels of features resulting in insufficient detection performance. To this end, this paper proposes a change detection network that combines spatial constraints and difference feature aggregation. By controlling the flow of feature information in the network, the difference between the low-level feature and high-level semantic information of the detection object is eliminated, and the quality of prediction results is improved. Firstly, the siamese network is used in combination with the feature pyramid structure to generate multi-scale differential features; Then, the proposed coordinate self attention mechanism (CSAM) is used to constrain the low-level features, strengthen the edge structure of the change area and the accurate position information, and combine the classical convolutional attention module to fully capture the context change feature information; Finally, the gated fusion mechanism is used to extract the channel relationship and control the fusion of multi-scale features to generate a change image with clear boundary and complete interior. A large number of experiments are carried out on the change detection dataset CDD and LEVIR-CD, and compared with the existing change detection network models, the proposed method shows the best detection effect in different scenarios.

Key words: change detection, multi-scale difference features, spatial constraints, gated fusion mechanism, complex scene