Three-Branch Hybrid Network for Farmland Segmentation in Remote Sensing Images

Abstract

The accurate segmentation of remote sensing imagery is critical for precision agriculture but challenging due to spectral complexity and ambiguous interclass boundaries. The convolutional neural networks are limited in modeling global context, while transformer-based methods incur high computational overhead. This letter proposes a three-branch hybrid network (TBHNet). A collaborative global branch introduces an attention-enhanced visual state-space module (AVSSM) by integrating Mamba-based visual state-space blocks with an attention mechanism to enhance global context modeling. Parallel spatial and boundary branches preserve local structures and refine edges for complementary multibranch feature fusion. A lightweight multireceptive group convolution head (MRHead) is further designed to improve prediction efficiency. Experiments on FGFD and LoveDA datasets show that TBHNet achieves mean intersection over union (mIoU) of 85.90% and 68.58% and overall accuracy (Acc) of 92.69% and 84.46%, respectively, striking a superior balance between segmentation performance and computational cost.