我国南方丘陵山区耕地地块精准提取方法——以福建省浦城县为例  

Accurate Extraction of Farmland Parcels in Hilly Areas of Southern China:A Case Study in Pucheng County,Fujian Province

作　　者：吴瑞姣 陈光剑 WU Ruijiao;CHEN Guangjian(Fujian Geologic Surveying and Mapping Institute,Fuzhou 350011,China)

机构地区：[1]福建省地质测绘院,福州350011

出　　处：《遥感信息》2024年第3期55-60,共6页Remote Sensing Information

基　　金：福建省科技厅创新资金项目(2022C0024);福建省科技厅引导性项目(2022N0019)。

摘　　要：鉴于南方丘陵山区耕地地块具有多样的尺寸和种植结构,传统语义分割方法存在精度低、边界吻合度差等问题,提出了一种顾及地块边界和形状特征多任务学习和注意力机制的高分辨率遥感图像提取方法。该方法构建了多任务神经网络模型FPEM-Net,主要包括耕地分割主任务以及与地块提取密切相关的轮廓检测和距离估计两个辅助任务。引入了CBAM注意力模块以增强特征的表达能力并减少冗余特征,并通过共同优化多任务损失函数训练模型。该方法成功应用于福建省浦城县,实验结果显示在测试集上表现出色,Hausdorff距离最小,像素准确率和交并比分别达到93.12%和93.55%。与Psi-Net相比,像素准确率提升了1.42%,交并比提升了3.4%;相较于UNet,像素准确率提升了14.11%,交并比提升了15.34%。该方法在规则和不规则耕地地块以及复杂种植结构的边界方面表现出良好的泛化能力,与实际耕地地块分布格局更加吻合,具有较好的应用潜力。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.

关 键 词：耕地地块 高分辨率图像 深度学习 多任务学习 注意力机制 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置] P237[自动化与计算机技术—控制科学与工程]