【作者】 张俊；

【导师】 张军；

【作者基本信息】 云南大学 ， 地图学与地理信息系统， 2024， 硕士

【摘要】 利用遥感影像进行梯田检测和田埂提取在控制坡地水土流失、蓄水、保土和粮食增产等方面具有重要意义。然而,现有研究大多基于传统的浅层特征进行梯田提取,这类方法忽略了梯田的细节特征,且自动化程度较低。此外,梯田和田埂在一定程度上是相互依赖的,梯田检测约束了田埂的外观,田埂提取增强了梯田的几何特征。但现有基于高分辨率影像的梯田信息提取均将两个任务分别进行处理,并没有考虑它们之间的密切关系,难以获得梯田地区精细化制图。本文提出了一种多尺度、多任务的梯田信息提取网络(DTRE-Net),建立两者之间的关系,同时完成梯田检测和田埂提取任务。论文的研究结果如下:(1)本文所提出的方法在视觉效果和精度指标方面相比四种方法更加优越。针对T1和T2地区的两种不同形态的梯田,在梯田检测任务中,其交并比值分别达到了85.18%和86.09%,精确率、召回率和F1评分均超过了90%的得分。对田埂的提取任务,其交并比值分别为59.79%和73.65%,精确率、召回率和F1评分均超过了70%的得分。(2)相对于单任务学习,多任务学习在T1区域在检测结果中精确率、召回率、F1评分和交并比方面分别增长了7.47%、5.5%、6.65%和8.06%,T2区域在检测结果中精确率、召回率、F1评分和交并比方面分别增长了4.36%、4.77%、4.56%和6.61%。此外,近红外波段对于梯田信息提取的重要性。四波段数据相较于三波段数据在梯田检测中精确率、召回率、F1评分和交并比方面分别提高了0.46%、1.5%、0.98%和2.12%。在田埂提取中精确率、召回率、F1评分和交并比方面分别提高了2.08%、0.6%、1.42%和 1.95%。(3)消融实验显示,引入多尺度特征融合模块和多尺度残差修正模块能够帮助模型训练更加稳定。相比于基础网络,精确率、召回率、F1评分和交并比相比分别上升了 1.17%、0.85%、1.01%、1.68%,且结果通过t-student假设检验样本IoU值证明普遍性提升。在高分辨率遥感影像语义分割领域,本文提出的多尺度、多任务梯田信息提取网络(DTRE-Net)在梯田检测和田埂提取方面表现出色,显著优于传统方法。DTRE-Net的提出为实现梯田地区精细化制图提供了一种高效、准确的解决方案,有望在梯田保护和农业生产中发挥重要作用。更多还原

【Abstract】 The use of remote sensing imagery for terrace detection and ridge extraction is of significant importance in controlling soil erosion on sloped land,water retention,soil conservation,and increasing crop yield.However,existing studies mostly rely on traditional shallow features for terrace extraction,which overlook detailed features of terraces and exhibit a low level of automation.Additionally,terraces and ridges are somewhat interdependent;terrace detection constrains the appearance of ridges,while ridge extraction enhances the geometric features of terraces.Nonetheless,current high-resolution image-based terrace information extraction approaches handle these two tasks separately,failing to consider their close relationship,making it difficult to achieve refined mapping of terrace areas.This paper proposes a multi-scale,multi-task terrace information extraction network(DTRE-Net)that establishes the relationship between the two tasks,simultaneously performing terrace detection and ridge extraction.The research findings of the paper are as follows:(1)The proposed method in this paper is superior in visual effect and accuracy metrics compared to four other methods.For the two different types of terraces in the T1 and T2 areas,the Intersection over Union(IoU)values in the terrace detection task reached 85.18%and 86.09%,respectively,with Precision,Recall,and F1-Score all exceeding 90%.In the ridge extraction task,the IoU values were 59.79%and 73.65%,with Precision,Recall,and F1-Score all exceeding 70%.(2)Compared to single-task learning,multi-task learning increased the Precision,Recall,F1-Score,and IoU in the T1 area by 7.47%,5.5%,6.65%,and 8.06%,respectively,and in the T2 area by 4.36%,4.77%,4.56%,and 6.61%,respectively.Additionally,the importance of the near-infrared band in terrace information extraction is highlighted.Compared to threeband data,four-band data improved the Precision,Recall,F1-Score,and IoU in terrace detection by 0.46%,1.5%,0.98%,and 2.12%,respectively,and in ridge extraction by 2.08%,0.6%1.42%,and 1.95%.(3)Ablation experiments show that the introduction of multi-scale feature fusion modules and multi-scale residual correction modules can help stabilize model training.Compared to the base network,Precision,Recall,F1-Score,and IoU increased by 1.17%,0.85%,1.01%,and 1.68%,respectively,with results validated by the t-student hypothesis test demonstrating a general improvement in IoU.In the field of high-resolution remote sensing image semantic segmentation,the proposed multi-scale,multi-task terrace information extraction network(DTRE-Net)excels in both terrace detection and ridge extraction,significantly outperforming traditional methods.The introduction of DTRE-Net provides an efficient and accurate solution for achieving refined mapping of terrace areas,with the potential to play a crucial role in terrace conservation and agricultural production.更多还原