A difference enhancement and class-aware rebalancing semi-supervised network for cropland semantic change detection

Changes in cropland are among the most widespread transitions on the Earth surface, significantly impacting food security, ecological conservation, and social stability. Compared to conventional change events, cropland changes involve complex dynamic transformations of semantic representations within the land system, requiring the identification of both the locations and categories of changes. Despite numerous remote sensing change detection methods have been proposed in previous studies, two challenges in cropland semantic change detection (SCD) still deserve further discussion: 1) transition confusions between similar categories and 2) under-labeling and class imbalance related to semantic labels. To address these challenges, we propose a difference enhancement and class-aware rebalancing semi-supervised network (Semi-DECRNet) for cropland SCD. The proposed Semi-DECRNet is implemented in a multi-task three-branch architecture, incorporating a multi-scale semantic aggregation difference enhancement module to couple the semantic and initial differential features at both global and local levels to model the temporal and causal relationships among the binary change detection and semantic segmentation branches. Additionally, a class-aware rebalancing self-training strategy is developed to adaptively calibrate the pseudo-label thresholds and further mine the semantic knowledge in unchanged areas. Experiments and analysis on three benchmark datasets demonstrate the effectiveness and superiority of the proposed Semi-DECRNet method for the cropland SCD task. Code is available at https://github.com/DaiAnjin/Semi-DECRNet.