摘要
冰湖溃决洪水是一种严重的山地自然灾害,威胁着中国高寒区的居民及铁路公路等重要基础设施,自动高效的冰湖遥感制图方法是冰湖灾害评估、监测预警的基础,然而现有自动制图方法在实际冰湖提取应用上难以达到传统人工和半自动冰湖提取方法上的精度,仍需进一步提高。文章在原生UNet模型基础上,在各桥连接部分融合极化自注意力机制,将输入影像特征分别在空间和通道层保持高分辨率,并通过非线性合成输出细腻的特征,构建了一种改进的U-Net冰湖遥感深度学习制图方法,并将其成功应用在高原铁路关键区。研究结果表明:1)与PSPNet、DeepLabV3+、原生U-Net三种经典模型相比,改进模型在冰湖预测数据集上的各项指标上都有提升,精确率、召回率、交并比和F_1值分别达到了0.972 5、0.966 5、0.940 8和0.969 4,相较于原生U-Net网络,精确度、召回率、交并比和F_1值分别提高了5.01%、6.05%、10.73%和5.53%;2)基于Landsat-8卫星遥感数据,应用改进模型完成了2013—2022年帕隆藏布和易贡藏布案例区冰湖信息自动高效提取,如2020年冰湖总体精度为98.16%,与参照数据的重叠度达到96.66%,提取的精度满足冰湖灾害评估和监测预警研究需求,可用于铁路等重大工程沿线冰湖灾害防治的实践。
Glacial lake outburst floods(GLOFs)are a serious mountain natural disaster,threatening residents and important infrastructure such as railways and highways in China’s high-altitude regions.Automatic and efficient glacial lake remote sensing mapping methods are the basis for glacial lake disaster assessment,monitoring and early warning.However,the existing automatic mapping method is difficult to achieve the accuracy of traditional manual and semi-automatic ice lake extraction methods in actual ice lake extraction applications,and it still needs to be further improved.This study is based upon the original U-Net model and incorporates polar self-attention mechanisms at various bridge connections.The input image features with high resolution are maintained both spatially and channel-wise and refined through the synthesis of nonlinear output features.Then,an improved U-Net glacial lake remote sensing deep learning mapping method is constructed and successfully applied in key areas of the plateau railway.The results are as follows.1)Compared with three classical models,namely PSPNet,DeepLabV3+,and the original U-Net,the improved model has improved performance on various metrics in the glacial lake prediction dataset,with the precision,recall,IoU,and F1 values reaching 0.9725,0.9665,0.9408,and 0.9694,respectively.Relative to the original U-Net network,the precision,recall,IoU,and F1 values of the revised model have been increased by 5.01%,6.05%,10.73%,and 5.53%,respectively.2)Using Landsat-8 satellite remote sensing data,the improved model is applied to automatically and efficiently extract glacial lake information in the Palong Zangbo and Yigong Zangbo case study areas from 2013 to 2022.The mapping glacial lakes in 2020 have an overall accuracy of 98.16%and an overlap rate of 96.66%with the user-interactive mapped reference data,meeting the research requirements for GLOF assessment and monitoring.This method can be used in the practice of glacial lake disaster prevention and control in major engineering projects suc
作者
杨泞滔
聂勇
YANG Ningtao;NIE Yong(Institute of Mountain Hazards and Environment,Chinese Academy of Science,Chengdu 610299,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《航天返回与遥感》
CSCD
北大核心
2024年第1期41-52,共12页
Spacecraft Recovery & Remote Sensing
基金
国家自然科学基金(41971153,42171086)
西藏科技厅项目(XZ202301ZY0016G)。
