RBSF-Net: a Refined Boundary-Semantic Fusion Network Based on Marker-Controlled Watershed for Delineating Agricultural Fields from High-Resolution Satellite Imagery

Accurate delineating individual agricultural fields (AFs) plays a key role in advancing precision agriculture and effective land management. Although deep learning (DL) techniques have shown great potential in automatically extracting AFs from satellite imagery, existing models still face significant challenges, including field adhesion, inaccurate boundary localization, and limited transferability. To address this, we developed a novel Refined Boundary-Semantic Fusion multi-task learning framework (RBSF-Net) that integrates precisely located shallow boundary features with deep semantic features to extract fine-scale AFs from high-resolution satellite imagery. Firstly, a dual-branch hybrid neural network was used to capture the boundary and semantic features of AFs, respectively. Secondly, a Boundary Patch Refinement Module (BPRM) was designed to postprocess fragmented boundaries by utilizing directional information. Finally, a Boundary-Semantic Interaction Module (BSIM), based on a marker-controlled watershed fusion strategy, was employed to effectively integrate boundary-semantic features, generating the final individual AF entities by using deep semantic features as markers and shallow boundary features as guide maps. The proposed RBSF-Net achieved significant improvements in AFs extraction compared to current state-of-the-art methods (i.e., HDNet, BSiNet, E2EVAP, Deeplabv3+, ResU-Net, and U-Net), with an accuracy of F 1 extent at 0.880, F 1 adge at 0.919, and F 1 total at 0.169. The RBSF-Net also achieved a satisfactory accuracy across three diverse regions with different imagery resolutions, diverse agricultural landscapes, and different terrain conditions, performing great generalization and robustness. These findings suggest that RBSF-Net, which is an accurate, robust, and generalized method, holds substantial potential for broad application in precise large-scale AFs identification.