SwinUCDNet: A UNet-like Network With Union Attention for Cropland Change Detection of Aerial Images

Cropland conversion disrupt local agricultural production systems and pose a serious threat to global food security. The use of remote sensing for change detection (CD) can detect and prevent such events in a timely manner. However, existing CD methods struggle to produce change detection results efficiently and accurately. Additionally, the limited receptive field of the convolution process prevents CNN-based approaches from catching long-range relationships. The Vision Transformer, on the other hand, excels in a variety of vision-related tasks, including picture classification, object detection, and semantic segmentation, and has significant promise for modeling long-range relationships. Because of this, in this research we suggest a UNet-like network with union attention for detecting changes in cropland using aerial remote sensing images. We utilize a Swin Transformer backbone as the encoder and an effective union-attention Transformer block to build the decoder in a Transformer-based encoder-decoder structure. The application of a multibranch prediction head with two CNN classifiers yields change maps and enhances deep layer supervision. The effectiveness and advantages of the SwinUCDNet have been demonstrated through comparative experiments with several CD methods.