摘要
针对生态环境脆弱的寒旱区开展地物要素提取以及土地覆盖变化监测研究,对农业规划、城乡建设、生态环境监测与保护等具有重要意义。借助2015-2019年新疆莫索湾垦区Landsat-8影像构建数据集,对比3种传统方法:最大似然分类(Maximum likelihood classification,MLC)、支持向量机(Support vector machine,SVM)和随机森林(Random forest,RF)及5种语义分割模型:Deep⁃Labv3+(Xception)、DeepLabv3+(MobileNet)、SegNet(ResNet50)、U-Net(MobileNet)和PSPNet(Mo⁃bileNet),选取最优自动化地物提取模型对研究区1998-2020年农用地、建筑用地、水体和荒漠4种地物要素进行分类,并运用土地利用转移矩阵和动态度进行定量动态变化分析。结果表明:Deep⁃Labv3+(Xception)模型可以实现更准确、更高效的地物提取,总体精确度(OA)、Kappa系数和F1值分别为96.06%、0.96和0.86,其中所选模型的平均交并比(MIoU)较其他模型提升0.03~0.39。近23 a,莫索湾垦区的荒漠、农用地和建筑用地三者的土地结构转化较为明显,荒漠总面积减少15.00%,农用地总面积增加12.68%,建筑用地总面积增加2.53%,水体面积变化较为平稳。地物类型总体转变方向为荒漠向农用地转化、农用地向建筑用地转化。该研究可为深度学习技术应用于中分辨率遥感卫星影像领域中实现土地利用及变化动态监测提供参考。
In this study,we conducted researchon the extraction of ground elements and the dynamic monitoring of land cover change in cold and arid areas with fragile ecological environments.The Mosuowan reclamation area in Xinjiang,China was selected as the target area,and the Landsat series satellite images from 2015 to 2019 were cut into subimages with sizes of 416×416 px.Specifically,the training set,validation set and testset are 306 subimages,204 subimages and 120 subimages respectively.Three traditional methods were then evaluated:MLC,SVM,and RF.Five semantic segmentation models were also evaluated:DeepLabv3+(Xception),DeepLabv3+(MobileNet),SegNet(ResNet50),U-Net(MobileNet),and PSPNet(MobileNet).In the evaluation experiments,DeepLabv3+(Xception)was found to achieve the optimal segmentation effect and multiscale feature fusion using fewer parameters.The overall accuracy,Kappa coefficient,precision,recall,F1-score,and MIoU were 96.06%,0.96,87.69%,83.78%,0.86,and 0.77,respectively.The MIoU of the DeepLabv3+(Xception)model was significantly better than those of the other four models,improving from 0.03 to 0.39.On the basis of the land use classification results of the long-term timeseries remote sensing data from 1998 to 2020,we analyze spatial structure change in land use and the associated driving factors.Over the past 23 years,the total areas of desert,agricultural land,and construction land have been reduced by 15.00%,12.68%,2.53%,respectively.At the same time,the amount of water area has remained relatively stable.The overall transformation direction of land use is from desert to agricultural land and then from agricultural land to construction land.It can be seen from the results that the desertification control was effective,and urbanization rapidly developed.Consequently,this study can provide a reference for the application of deep learning in the field of medium-resolution remote sensing images,which can be used to realize the dynamic monitoring of land use and change.
作者
袁盼丽
汪传建
赵庆展
王学文
任媛媛
杨启原
YUAN Panli;WANG Chuanjian;ZHAO Qingzhan;WANG Xuewen;REN Yuanyuan;YANG Qiyuan(School of Information Science and Technology,Shihezi University,Shihezi 832003,Xinjiang,China;Research Center for Space Information Engineering Technology of XPCC,Shihezi 832003,Xinjiang,China;XPCC Industrial Technology Research Institute,Shihezi 832003,Xinjiang,China;Internet College of Anhui University,Hefei 230039,Anhui,China)
出处
《干旱区地理》
CSCD
北大核心
2021年第6期1717-1728,共12页
Arid Land Geography
基金
国家重点研发计划项目(2017YFB0504203)
新疆生产建设兵团科技计划项目(2017DB005)资助。
