Abstract: Using remote sensing imagery to detect changes is crucial for understanding land surface dynamics. In recent years, deep learning-based methods have become a focus area owing to their excellent feature extraction and representation ability. The fusion of multi-scale feature information is the key to improving change detection performance in fully convolutional network-based structural methods. Most of the previous methods use skip connection or dense connection structure, which improves the accuracy of change detection methods to a certain extent. However, such methods only fuse features at the same scale and lack sufficient information from multiple scales to achieve satisfactory results. In this paper, a full-scale feature aggregation network (FSANet) is proposed to solve the problem of remote sensing image change detection. Firstly, the features of the bi-temporal images are extracted using a siamese network, then the features are efficiently concatenated using a full-scale feature concatenation structure, and to prevent feature redundancy, the features are refined using a feature refinement module. Finally, to optimize model training, a multiscale supervision strategy is used. Multiple additional detections are output in the decoder, which calculate the final loss value together. To check the reliability of FSANet, we tested it on two public datasets, the LEVIR-CD and SVCD datasets. The experimental results show that the method outperforms other mainstream change detection methods, while having a good balance between accuracy and complexity.

Key words: high-resolution remote sensing image, change detection, full-scale skip connection, attention mechanism, multiscale supervision