BGFFNet: a semantic boundary-guided feature fusion network for cultivated land extraction from high-resolution remote-sensing images

Accurate extraction of cultivated land information from high-resolution remote-sensing images is essential for ecological conservation, resource utilization, and sustainable development. To alleviate the problems of blurred boundaries and incomplete details in cultivated land extraction, this paper presents an innovative network termed BGFFNet. The model adopts a U-shaped architecture with ResNet50 as the encoder and a Transformer module as the decoder. To mitigate boundary blurring, a Boundary Branch and a Boundary Guidance Module (BGM) are designed to enhance the network’s ability to capture boundary information. A Local-Global Feature Fusion Block (LGFB) is introduced to improve the network’s capacity for modelling global contextual information. Additionally, a Multi-Scale Feature Fusion Block (MSFB) enhances the network’s ability to perceive cultivated land with diverse sizes and irregular shapes. Experiments are conducted on the three datasets (FGFD, iFLYTEK, and 2classes), and comparisons are made against state-of-the-art network architectures spanning three mainstream paradigms: CNNs, Transformers, and CNN–Transformer hybrids. Experimental results demonstrate that on the iFLYTEK dataset, the proposed BGFFNet achieves an F1-score of 97.09% and IoU of 94.34%, significantly outperforming mainstream comparative models. The proposed model effectively strengthens the network’s capability in learning boundary information and multi-scale cultivated land characteristics, thereby realizing more accurate cultivated land extraction.