DDMSA-U-Net: A Lightweight Deep Learning Framework for Multi-Spectral Change Detection for Agricultural Land Use Monitoring

The change detection approach using remote sensing is vital in agriculture monitoring; however, the existing approach lacks efficiency in terms of boundary accuracy and sensitivity to spatial scales, and suffers from background noise. Such constraints make it difficult to map transitions such as crop rotations, fallow, and urban encroachment, particularly in heterogeneous landscapes. This research proposes Depthwise Dilated Multi-Spatial Attention U-Net (DDMSA-U-Net), which is a light architecture that enhances the accuracy of change detection achieved with both Landsat- 8 and Sentinel- 2 satellite images. The model integrates depth-wise separable convolutions, dilated multiscale feature extraction, and multispatial attention mechanisms to enhance spatial discrimination and reduce computational costs. The methodology includes preprocessing of multi-temporal satellite imagery, supervised training, stratified sampling, and post-classification comparisons to assist in change detection. Validation was carried out based on the seasonal crop observations of 2021-2024. Compared to traditional methods, DDMSA-U-Net improved in terms of classification and change detection with an overall accuracy of 91.6-96.6 and Kappa values above 0.85 for all cases. These findings highlight the usefulness of the model for observations of agricultural transitions.