LaFormer: a Laplacian edge-guided transformer for farmland segmentation in remote sensing images

Accurate extraction of cultivated land from remote sensing imagery with complex backgrounds is essential for agricultural digitalization, land supervision, and farmland protection. However, existing methods face significant challenges when dealing with cultivated land images characterized by sparse foregrounds, blurred boundaries, and complex structures, often leading to inaccurate boundary localization, object adhesion, false positives, and missed detections. To address these issues, a lightweight edge-aware transformer segmentation network, named LaFormer, is proposed to enhance segmentation accuracy and boundary modeling in complex remote sensing scenarios. LaFormer is built upon a CNN–transformer hybrid architecture, integrating two key components: a Laplacian edge module (LEM) to enhance the spatial boundary perception of shallow encoders and the semantic detail representation of deep encoders, and an Edge-Aware Binary Loss (EABLoss) to enforce consistency between foreground responses and boundaries at the loss-function level. The LEM effectively alleviates insufficient boundary modeling in shallow features and limited small-scale perception in deep features, whereas EABLoss improves segmentation robustness under blurred boundary and sparse foreground conditions. Experiments were conducted on the a remote sensing land-cover dataset for domain adaptive semantic segmentation (LoveDA) dataset, which contains sparse urban farmland, and the fine-grained farmland dataset (FGFD), which comprises densely structured agricultural parcels with clear boundaries. Results show that LaFormer surpasses mainstream methods in key metrics, including intersection over union, F1-score, and recall, achieving IoU gains of 1.06% and 1.66% over the second-best model on LoveDA and FGFD, respectively. The method preserves foreground completeness while significantly enhancing boundary fidelity, demonstrating strong adaptability and generalization for high-precision farmland extraction in real-world remote sensing applications.