A Category-Aware Feature Disentanglement and Spatial-Aware Framework for Cultivated Land Change Detection in Diverse Scenarios

Cultivated land change detection using remote sensing images enables rapid and accurate identification of cultivated land destruction, facilitating land consolidation and agricultural structure adjustment. However, the current detection methods suffer from low monitoring accuracy due to insufficient geometric feature modeling, weak anti-counterfeiting change interference capability, and inefficient bi-temporal feature fusion. To overcome these limitations, we first propose a deformable high-order attention siamese network. This network adaptively captures the geometric shape of the changing area using deformable convolution and concurrently employs high-order attention mechanisms to uncover deep relationships among ground object features. We then introduce an adaptive center triplet loss function to achieve category-aware feature separation and aggregation strategy. This controls the compactness of similar cultivated land features and enhances the differentiation between cultivated and non-cultivated land, effectively reducing pseudo differences caused by seasonal variations. Moreover, we design a dual branch cross attention module based on transformer. This module refines spatiotemporal feature collaboration through cross-temporal feature interaction, improving the accuracy of difference feature computation. The experimental results show that the method proposed in this paper has significantly improved in detection accuracy and robustness. Moreover, in the boundary areas of cultivated land patches, the proposed method generates results that are spatially more continuous and have more refined geometric structures. This can provide more reliable technical support for applications such as cultivated land monitoring and ecological assessment.