BGDNet: a boundary-guided deformable network for accurate farmland parcel segmentation in high-resolution remote sensing imagery

To address the challenges of complex farmland parcel shapes, blurred boundaries, and frequent adhesion between adjacent parcels in high-resolution remote sensing imagery, this study proposes BGDNet, a boundary-guided and deformable convolution collaboratively driven network for farmland segmentation. First, DCN is introduced into the encoder to enhance the model’s ability to adaptively represent irregular farmland geometries. Second, a MSAM and a FEM are constructed for multi-scale contextual aggregation and cross-layer feature calibration, respectively, thereby improving semantic representation capability in complex scenes. On this basis, a BGDF is proposed, which generates spatial gating weights through collaborative learning between boundary and semantic branches, enabling dynamic constraint of decoder features by boundary information. This alleviates parcel adhesion and improves boundary localisation accuracy. Experimental results on the FGFD demonstrate that the BGDNet achieves an IoU of 85.57% and an F1-score of 92.22%, outperforming several mainstream semantic segmentation models in terms of segmentation accuracy, boundary integrity, and visual consistency. Qualitative visualisation and ablation studies validate the effectiveness of each module and their synergistic contributions. The proposed method provides an effective solution for high-precision farmland extraction under complex terrain conditions and offers technical support for agricultural resource investigation, cultivated land monitoring, and smart agriculture management.