In response to the varied sizes and planting structures of farmland parcels in the southern hilly regions,traditional semantic segmentation methods face challenges such as low accuracy and poor boundary alignment,a high-resolution remote sensing image extraction method is proposed,which considers plot boundaries and shape features through multi-task learning and attention mechanisms.The method establishes a multi-task neural network model,FPEM-Net,consisting of the primary task of farmland segmentation and two auxiliary tasks closely related to plot extraction,which are contour detection and distance estimation.The CBAM attention module is introduced to enhance feature expression and reduce redundant features.The model is trained by jointly optimizing the multi-task loss function and is successfully applied to Pucheng county in Fujian province.Experimental results demonstrate its outstanding performance on the test set,with the minimum Hausdorff distance,pixel accuracy,and intersection over union rates reaching 93.12％and 93.55％,respectively.Compared to Psi-Net,there is a 1.42％increase in pixel accuracy and a 3.4％improvement in intersection over union.In comparison to UNet,pixel accuracy is improved by 14.11％,and intersection over union is increased by 15.34％.The method exhibits strong generalization capabilities in delineating boundaries of both regular and irregular farmland parcels,as well as complex planting structures.It aligns well with the actual distribution pattern of farmland parcels,demonstrating promising application potential.
万方数据