Cropland segmentation leveraging a synergistic edge enhancement and temporal difference-aware network with Sentinel-2 time-series imagery

Semantic segmentation of cropland is critical for accurately extracting crop distribution from satellite remote sensing (RS) images. However, the dynamic temporal patterns caused by crop rotations and the heterogeneous spatial characteristics of cropland pose significant challenges for achieving high-precision segmentation. To tackle these issues, we propose a novel spatiotemporal feature-enhanced network (STFE) designed specifically for cropland segmentation in remote sensing time-series images (RSTI). The STFE network effectively integrates temporal and spatial features by introducing key innovations. First, we design an edge-guided spatial attention (EGSA) module to enhance spatial detail extraction, particularly for delineating ambiguous boundaries. Second, a progressive feature enhancement (PFE) strategy is developed to capture and fuse multi-scale features progressively across network layers. Third, for temporal feature extraction, we incorporate a differential awareness attention (DAA) module, built on ConvLSTM, to dynamically aggregate temporal information, enabling the model to better capture crop rotation patterns and temporal variations. Experimental results on three benchmark datasets—PASTIS, ZueriCrop, and DNETHOR—demonstrate the superior performance of STFE compared to state-of-the-art methods, achieving mean IoU improvements of 3.2% over the best-performing baseline. The model excels particularly in handling challenging scenarios such as irregular crop shapes and mixed cropping patterns. Its adaptability to complex and evolving agricultural landscapes provides a scalable and reliable solution for supporting sustainable farming practices and informed decision-making.