Enhanced crop classification using gated spatiotemporal features and parcel boundary enclosure from Satellite time series

Crop mapping from Satellite Image Time Series (SITS) is a challenging task. Traditional methods often segment land parcels followed by crop classification, neglecting the interaction between crop boundaries and high-level semantic features, which limits the accurate delineation of boundaries. To address this, we propose a boundary-aware model that integrates parcel semantics into spatiotemporal feature extraction, focusing on classifying crop types with clear boundaries and achieving parcel-based crop classification in an end-to-end manner. Our approach incorporates spatial and temporal compression modules to extract multi-level spatial and temporal features. These features are then aggregated using a spatiotemporal gated feature extractor (SGFE) that captures representative features across both dimensions and generates multi-scale feature maps. To enhance boundary awareness, we introduce the boundary-semantic interaction module (BSIM), which enables deep interaction between semantic features and boundary information. Additionally, a dense multi-branch aggregation module (DMAM) improves contextual integration, enabling the model to extract fields of varying sizes and shapes. Experiments on the public PASTIS and MTLCC datasets and the newly constructed HTB dataset show that the proposed method achieves the highest mIoU scores of 67.64%, 82.18%, and 88.90%, respectively, with OA scores of 85.08%, 93.20%, and 95.28%. Compared with the strongest baseline methods, it improves mIoU by 2.51, 6.84, and 0.30% points on PASTIS, MTLCC, and HTB, respectively. The method also improves the SBC metric across the three datasets, indicating better consistency between crop labels and parcel boundaries. In the five-fold boundary evaluation on PASTIS, it further achieves higher BF1 and BIoU scores, confirming improved pixel-level boundary matching and spatial overlap. These results demonstrate that explicit boundary–semantic interaction improves both crop-type discrimination and parcel-level spatial consistency in SITS-based crop mapping.