SBCL-Net: Discrepancy-Guided Semantic–Boundary Interaction for Robust Agricultural Parcel Delineation From Remote Sensing Imagery

Abstract

Accurate agricultural parcel delineation from remote sensing imagery is essential for farmland registration, land-use monitoring, and field-level agricultural management. However, reliable parcel delineation remains challenging due to irregular field geometries, weak or ambiguous boundary cues, and substantial geographic heterogeneity across agricultural landscapes. To address these challenges, we propose SBCL-Net, a discrepancy-guided semantic–boundary interaction framework for agricultural parcel delineation. The framework consists of a semantic stream for region-level parcel representation and an independent boundary stream for structure-sensitive contour perception. To enhance semantic–boundary coupling, a Discrepancy-Guided Selective Modulation Module explicitly models shared and discrepant contexts between the two streams and selectively refines target features at discrepancy-sensitive locations. In addition, a consistency-constrained multi-task supervision strategy jointly optimizes mask prediction, boundary prediction, and their structural agreement. Experiments on FHAPD, FTW-France, and AI4Boundaries demonstrate strong and competitive performance across region-, boundary-, and object-level evaluations, with IoUs of 96.08% and 96.41% on FHAPD-JS and FHAPD-XJ, respectively, 71.49% on FTW-France, and 74.06% on AI4Boundaries. Cross-region evaluations further demonstrate the robustness of SBCL-Net under both within-country and cross-continental domain shifts. The code is available at https://github.com/Sabo-D/SBCL-Net