CR-CLCD: A cross-regional cropland change detection framework with multi-view domain adaptation for high-resolution satellite imagery

• A cross-regional cropland change detection framework, CR-CLCD, is proposed. • Enables high-precision change detection in target regions without samples. • Validated across different datasets, encompassing multiple study areas. • Effective in agricultural scenarios with varying radiance and change types. • It holds significant importance for cross-regional and large-scale monitoring. Cropland non-agriculturalization (CNA) monitoring is a typical change detection (CD) problem based on remote sensing imagery, aimed at tracking cropland outflow changes, which holds significant importance for cropland protection and food security. Recently, numerous advanced CD methods have been proposed to address the CNA problem. However, applying these methods to cross-regional or large-scale CNA detection presents several challenges: (1) Radiance-feature differences of croplands across regions i.e., crop type and phenology differences arising from variations in planting structures and seasonality; (2) Change-pattern differences of croplands across regions, i.e., differences in predominant change types resulting from distinct regional economic development characteristics. These cross-regional differences, when coupled together, result in insufficient adaptability of CD methods across regions. To address these issues, a Cross-Region Cropland Change Detection (CR-CLCD) framework with Multi-View Domain Adaptation (MVDA) is proposed. Specifically, Pattern Distribution Contrastive (PDC) sub-module achieves feature alignment from the semantic view by imposing contrastive constraints across inter-domain categories. Radiative Discrepancy Adversarial (RDA) sub-module, performs inter-domain global and local feature confusion by identifying regions of local uncertainty and applying enhanced adversarial training. MVDA is a flexible, plug-and-play domain adaptation module that can be seamlessly integrated with any existing change detection backbone network (e.g., CNN, Transformer), enabling rapid generalization to new data under unsupervised conditions. The experimental results demonstrate that the proposed CR-CLCD method achieves the best or second-best accuracy compared to other domain adaptation methods across different baselines.