Bi-Temporal Remote Sensing Image Change Detection Based on Joint Attention Mechanism and Residual Cross-Layer Block

As an important branch of remote sensing research, change detection (CD) is widely used in land use detection, disaster assessment, and other fields. However, the existing CD methods still have the problems of unclear detection of complex contours of changing objects and a high missed detection rate of small-sized changing objects. To solve the above problems, based on UNet++, a supervised encoder-decoder model, we design a novel CD structure that contains two important modules Residual Cross-Layer Block (RCLB) and Joint Attention Module (JAM), so we name it RJUNet++. RCLB is applied in the Siamese network of the encoder, which can fuse features from different receptive fields and different semantic levels to eliminate the semantic gap. Based on an attention mechanism, JAM is used to fuse multiple side outputs of UNet++ to facilitate the model to learn key information. We validate the effectiveness of this method on Change Detection Dataset (CDD). After quantitative comparison, our method improves on multiple evaluation criteria such as overall accuracy, F1 score, etc. compared to other state-of-the-art methods. Qualitative comparisons show that our method detects details more clearly and is less disturbed by noise.