基于BP神经网络模型的南方离子型稀土矿区开采遥感监测研究  

Remote Sensing Monitoring of Rare Earth Mining in Southern China Based on BP neural Network Model

作　　者：林建平 邓爱珍[3] 朱青 赵小敏[2] 刘洋洋 廖谌婳 王霆宵 Lin Jianping;Deng Aizhen;Zhu Qin;Zhao Xiaomin;Liu Yangyang;Liao Chenhua;Wang Tingxiao(School of Geography and Environmental Engineering,Gannan Normal University,Gan-zhou 341000,China;Key Laboratory of Poyang Lake Basin Agricultural Resources and Ecology,Jiangxi Province,Jiangxi Agricultural University,Nanchang 330045,China;Department of Sur-veying and Geo-Informatics,Jiangxi College of Applied Technology,Ganzhou 341000,China;Ganzhou Bureau of Natural Resources,Ganzhou 341000,China)

机构地区：[1]赣南师范大学地理与环境工程学院,江西赣州341000 [2]江西农业大学鄱阳湖流域农业资源与生态重点实验室,江西南昌330045 [3]江西应用技术职业学院测绘地理信息学院,江西赣州341000 [4]赣州市自然资源局,江西赣州341000

出　　处：《中国稀土学报》2022年第2期339-350,共12页Journal of the Chinese Society of Rare Earths

基　　金：国家自然科学基金项目(41361049);土壤与农业可持续发展国家重点实验室基金项目(0812201202)资助。

摘　　要：离子型稀土开采带来了一系列生态环境问题,日益引起人们的关注。以多时相Landsat遥感影像为数据源,通过波段提取、植被指数(Normalized Difference Vegetation Index,NDVI)、祼土指数(Bare Soil Index,BSI)和数字高程模型(Digital Elevation Model,DEM)等不同特征组合,构建“6-10-6”的三层式BP神经网络分类模型,对赣南近30年离子型稀土矿区开采损毁土壤时空变化特征进行了分析,结果表明:基于NDVI,BSI,DEM等不同特征组合的神经网络分类方法,能够有效地提高分类精度,解译制图精度、用户精度、Kappa系数比单纯基于光谱的神经网络分类分别提高了12.11%~17.75%,13.40%~16.39%,0.137~0.179。从稀土矿区开采损毁面积来看,1987~2017年,赣南稀土矿开采大致呈现为由少量开采到过度开采再到逐步恢复的过程,由1987年的1905.5 hm^(2)扩张至2007年的17165.5 hm^(2),2017年损毁面积又回落至6477.74 hm^(2)。从空间分布来看,赣南稀土开采主要集中在寻乌、定南、安远、信丰、宁都、赣县、龙南8个县(市、区)。该研究较好地揭示了赣南30年来稀土矿区开采土壤毁损与恢复过程,研究结果可为矿区的生态环境治理及可持续发展提供借鉴。Ionic rare earth ore mining has brought a series of ecological and environmental problems,which have increasingly attracted people's attention.Based on Ganzhou city of Jiangxi Province as the research area,Normalized Difference Vegetation Index(NDVI).Bare Soil Index(BSI)and Digital Elevation Model(DEM)as the auxiliary data source,this paper constructs a classification Model of"6-10-6"three-layer BP neural net⁃work.The spatial and temporal changes of soil damaged by ionic rare earth mining in southern Jiangxi Province in recent 30 years were analyzed.The research showed that:(1)The neural network classification method based on different feature combinations such as NDVI,BSI and DEM can effectively improve the classification accuracy.The interpretation mapping accuracy,user accuracy and Kappa coefficient of rare earth mining areas are improved by 12.11%~17.75%,13.40%~16.39%and 0.137~0.179,respectively,compared with the neural network classification based on spectrum alone.(2)From the perspective of changes in rare earth mining dam⁃age and recovery,from 1987 to 2017,the rare earth mining status in southern Jiangxi Province generally pres⁃ents a process from a small amount of mining to excessive mining and then to gradual recovery,expanding from 1905.5 hm^(2)in 1987 to 17165.5 hm^(2)in 2007.In 2017,the land damage area dropped to 6477.74 hm^(2).(3)Ac⁃cording to the spatial distribution of rare-earth mining,rare-earth mining is mainly concentrated in 8 areas(dis⁃tricts):Xunwu,Dingnan,Anyuan,Xinfeng,Ningdu,Ganxian and Longnan.This study can better reflect the land destruction and restoration process of rare earth mining area in southern Jiangxi Province in the past 30 years,and the results can provide the basis for ecological environment restoration in mining area.

关 键 词：BP神经网络模型 稀土矿区 遥感 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]