MSGNet: Multiscale Gabor Priori Guided Network for Farmland Texture Feature Extraction

Accurate delineation of farmland boundaries is vital for agricultural resource management and environmental sustainability. However, existing deep learning approaches often face three intertwined challenges: large variations in farmland scales and orientations, ambiguous transitional regions between adjacent parcels, and complex background interference from vegetation and built-up areas. To address these difficulties, this study proposes MSGNet, a Multi-Scale Gabor Priori Guided Network that integrates explicit texture priors into deep feature learning. The core idea is to formulate farmland texture extraction as a Gabor parameter learning problem, allowing the network to capture directional and scale-sensitive texture cues that conventional isotropic convolutions often miss. Specifically, MSGNet introduces a multi-scale Gabor encoder to model anisotropic texture structures, while a Transformer-based module is employed to enhance global contextual reasoning and maintain boundary consistency across scales. Furthermore, a Gabor-based Boundary Enhancement Module (GBEM) dynamically refines spatial attention, strengthening the model’s sensitivity to local edge variations amid complex backgrounds. To preserve both texture continuity and geometric integrity, a composite loss function combining cross-entropy, Dice, and Hausdorff distance is designed, ensuring stable and topology-aware learning. Experiments on the Guangdong Farmland Dataset demonstrate that MSGNet achieves superior accuracy and robustness, with 83.80% IoU, 90.88% precision, 91.49% recall, and 91.18% F1-score, outperforming state-of-the-art methods. Tests on the France Farmland Dataset further confirm the generalization ability of the proposed network across diverse agricultural landscapes. The code is available at the following link: [https://github.com/YuchunHuang/FarmLand_Gabor].