MSTransCDNet: a remote sensing image change detection method with multi-scale transformer

Due to issues of limited receptive fields, convolutional neural network-based change detection (CD) methods often fail to exploit the long-range spatial context information, resulting in sub-optimal CD performance. To overcome the limitations, a remote sensing image CD network with a multi-scale transformer (MSTransCDNet) is proposed, which follows a classic encoder-decoder architecture based on a multi-scale transformer (MSTransformer) unit. In particular, in MSTransformer, by introducing multi-scale key-value vectors, a multi-scale multi-head attention unit is constructed, allowing for the extraction of multi-scale features with rich spatial context information, thereby effectively addressing multi-scale targets and irregular geometric structures. In addition, a depthwise separable convolution (DWConv) layer is employed in the feed-forward network of MSTransformer, which significantly enhances the local information. The effectiveness of the proposed method is verified on change detection dataset (CDD) and learning, vision and remote sensing change detection (LEVIR-CD) datasets. Experimental results demonstrated the superiority of the proposed MSTransCDNet, which outperforms several state-of-the-art CD methods by a large margin, achieving an F1-score of 96.49% and 90.91% for CDD and LEVIR-CD, respectively.