MASNet: A Multitask Parallel Dual-Encoding Deep Learning Network for Farmland Parcel Segmentation

With the widespread application of remote sensing technology in the agricultural sector, the precise extraction of farmland parcel information has become one of the key technologies supporting modern agricultural management and resource planning. However, the complex and variable nature of farmland scenes imposes higher demands on the adaptability and robustness of segmentation models. To address this, this paper proposes a multi-task deep learning method named MASNet, to efficiently extract field information from remote sensing imagery under diverse farmland scenarios. The MASNet method adopts a parallel dual-encoder structure based on Mamba-Inspired Linear Attention (MILA) and Attentive Dilated-Separable CNN (ADS-CNN), significantly enhancing the model's feature extraction capabilities; simultaneously, the decoding stage incorporates the Spatial Group-wise Enhancement (SGE) attention mechanism, effectively improving the fusion efficiency of multi-scale features, thereby enhancing overall segmentation accuracy and model robustness. We thoroughly validated MASNet on the Solafune Competition Farmland Parcel Dataset and the JiLin-1 Farmland Parcel Public Dataset. Experimental results show that the Global Over-classification Error (GOC), the Global Under-classification Error (GUC), and the Global Total Error (GTC) of MASNet on the two datasets are 2.33%, 2.28%, and 2.37%, and 2.20%, 1.60%, and 1.94%, respectively; In the boundary extraction task, the Intersection over Union (IOU) and the mean Intersection over Union (mIOU) reach 39.51% and 68.20%, and 46.27% and 72.24%, respectively. Compared with existing mainstream methods, MASNet demonstrates superior performance across multiple metrics, showcasing its effectiveness and broad application potential in complex agricultural land parcel segmentation tasks.