先验知识引导深度学习的耕地范围建筑物和大棚房监测方法

Prior Knowledge Guided Deep Learning for Monitoring Buildings and Greenhouses within Cultivated Land

精准、自动化的耕地非农化监测对严守耕地红线和保障社会经济的可持续发展具有重要意义。为精准监测建筑物和大棚房非法占用耕地现象并及时预警,本文提出了一种先验知识引导深度学习的耕地非农化监测方法。首先,根据“三调”数据库获取耕地的矢量范围和类别属性等先验知识;其次,利用前时相耕地矢量数据对后时相高分辨率遥感影像进行地块级分割,实现检测区域的快速定位;接着,使用本文设计的融入先验知识的潜在非农化图斑提取网络模型(SRAM-SegFormer)对潜在非农化图斑进行获取;最后,对提取结果进行镶嵌、重分类、叠加等后处理操作得到最终的耕地非农化监测结果。以徐州市沛县为研究区,对比分析常用的语义分割网络模型Deeplabv3+、PSPNet、U-Net、HRNet、SegFormer和本文设计的SRAM-SegFormer提取潜在非农化图斑的能力。实验结果表明:Deeplabv3+和PSPNet对复杂区域易出现漏检和误检现象;U-Net容易漏检大尺度的建筑物;HRNet提取的非农化图斑边界不规则;SegFormer对小尺度的建筑物和房提取能力较差,提取密集区域的建筑物和大棚房易出现粘连现象;SRAM-SegFormer提取的效果最佳,特别是针对小尺度的建筑物和大棚房的提取,在平均像素准确率(MPA)、平均交并比(MIoU)和总体精度(OA)上均取得了最高精度,分别为84.30%、73.76%和97.91%。因此,本文提出的方法能够更加高效、自动化的实现耕地非农化监测。

Accurate and automated monitoring of the non-agriculturalization of cultivated land has important implications for upholding the arable land red line and ensuring sustainable socio-economic development.This paper proposes a deep learning method guided by prior knowledge to achieve precise monitoring of the illegal occupation of cultivated land by buildings and greenhouses,with the ultimate goal of issuing timely warnings.Firstly,the vector range and category attributes are obtained from the third national land survey database,serving as prior knowledge.Then,parcel-level segmentation of high-resolution remote sensing images is performed using the front-phase vector data of the cultivated land to locate detection areas.Next,the SRAM-SegFormer model,integrated with prior knowledge,is employed to extract potential non-agricultural patches.Finally,post-processing operations such as mosaic,reclassification,and overlay are performed to obtain final monitoring results of arable land non-agriculturalization.Taking the Peixian County in Xu Zhou City as the study area,the performance of common sematic segmentation networks,including Deeplabv3+,PSPNet,U-Net,HRNet,SegFormer,and SRAM-SegFormer,in extracting potential non-agricultural patches are compared.The results show that Deeplabv3+and PSPNet are prone to omissions and false detections in complex areas;U-Net tends to miss large-scale buildings;HRNet exhibits irregular boundaries in extracted non-agricultural patches;SegFormer has poor extraction ability for small-scale buildings and greenhouses,and tends to merge buildings and greenhouses in densely populated areas;SRAM-SegFormer shows the best results,with the highest accuracy rate for Mean Pixel Accuracy(MPA)(84.30%),Mean Intersection-Over-Union(MIoU)(73.76%),and Overall Accuracy(OA)(97.91%),especially in extracting small-scale buildings and greenhouses.Therefore,the proposed method in this paper can achieve more efficient and automated monitoring of arable land non-agriculturalization.