基于稀疏降噪自编码神经网络的遥感影像变化检测  

Remote sensing image change detection based on stacked denoising auto-encoder neural network

作　　者：谢垂军 刘龙威 罗莉 XIE Chuijun;LIU Longwei;LUO Li(Land Survey Team of Lianping County,Heyuan Guangdong 517000,China;Guangdong Institute of Lands and Resources Surveying and Mapping,Guangzhou Guangdong 510670,China;Key Laboratory of Natural Resources Monitoring in Tropical and Subtropical Area of South China,Ministry of Natural Resources,Guangzhou Guangdong 510500,China;Guangdong Science and Technology Collaborative Innovation Center for Natural Resources,Guangzhou Guangdong 510553,China;Guangdong Surveying and Mapping Engineering Company Limited,Guangzhou Guangdong 510670,China)

机构地区：[1]连平县国土测量队,广东河源517000 [2]广东省国土资源测绘院,广东广州510670 [3]自然资源部华南热带亚热带自然资源监测重点实验室,广东广州510500 [4]广东省自然资源科技协同创新中心,广东广州510553 [5]广东省测绘工程有限公司,广东广州510670

出　　处：《北京测绘》2024年第1期19-23,共5页Beijing Surveying and Mapping

基　　金：广东省科技计划(2021B1212100003)。

摘　　要：为满足广东省自然资源调查监测地类变化自动发现迫切需求,利用一种针对高分辨率遥感影像语义信息提取的稀疏降噪自编码(SDAE)深度神经网络,在珠三角试验区选取BJ2、GF1、GF2、GJ1、ZY3、SP6、PL0卫星影像各一景,按照分辨率对样本数据归类,分别训练生成多传感器训练模型;在识别目标区域的同时,降低变化区域的虚警率,实现两期影像变化信息提取,并输出变化置信度、变化类型。试验表明:该模型在珠三角城乡接合部、建成区和乡村的查全率分别达到90%、83%、82%,正确率分别为61%、60%、71%,证明该方法是可行的。In order to meet the urgent need for automatic discovery of land type changes in natural resource survey and monitoring in Guangdong Province,a stacked denoising auto-encoder(SDAE)deep neural network was used to extract semantic information from high-resolution remote sensing images.One scene of BJ2,GF1,GF2,GJ1,ZY3,SP6,and PL0 satellite images in the Pearl River Delta experimental area was selected.The sample data was classified according to resolution,and the multi-sensor training models were trained and generated separately.While identifying the target area,the false alarm rate of the changing area was reduced,achieving the extraction of change information from images in two periods.In addition,change confidence and change type were output.The experiment shows that the recall rates of this model in the urban-rural fringe areas,built-up areas,and rural areas of the Pearl River Delta are 90%,83%,and 82%,respectively,and the accuracy rates are 61%,60%,and 71%,respectively,demonstrating that the method is feasible.

关 键 词：高分辨率影像 深度神经网络 训练模型 变化检测 

分 类 号：P237[天文地球—摄影测量与遥感]