EDSFINet: efficient double stages feature interaction network for remote sensing image change detection

Accurate remote sensing change detection contributes to more rational protection of the environment. How to perform more efficient and sufficient feature interaction and feature fusion between two corresponding high-resolution time series images has always been a problem in this area. Most existing methods use a one-stage network to perform feature interaction between two time series images, which often leads to insufficient and inefficient feature interaction. Therefore, this article proposes an efficient double-stage feature interaction network (EDSFINet) aimed at achieving more comprehensive and efficient feature interaction. This is the first exploration of using a two-stage feature interaction framework to accomplish this task, marking an innovative attempt in this field. For the interaction of shallow features, we designed a shallow feature interaction module (SFIM), which is mainly based on spatial and channel attention for image interaction, and can more effectively extract low-level information of the image, including edges, textures, and color distribution. For the interaction of deep features, we designed a deep feature interaction module (DFIM), which is mainly based on Transformer attention for image interaction, which can more effectively capture global features and long-distance dependencies. To achieve more efficient feature fusion, we designed a fusion feature enhancement module, which is implemented using the attention of the Transformer with linear computational complexity and can perform feature fusion enhancement more efficiently. To alleviate the class imbalance challenge, we adopted multi-scale supervised training during the training process to better utilize the multi-scale information of the images. According to experimental results, the F1 scores of EDSFINet on the LEVIR-CD is 91.36%, and the IoU indicators is 84.09%, which are better than the existing methods.