深度学习方法在光伏用地遥感检测中的应用  

Application of deep learning method in remote sensing detection of photovoltaic land

作　　者：宋业冲 李英成[1,2,3] 耿中元 丁晓波 裴亚健[2,3] SONG Yechong;LI Yingcheng;GENG Zhongyuan;DING Xiaobo;PEI Yajian(Chinese Academy of Surveying and Mapping,Beijing 100036,China;Key Laboratory for Aerial Remote Sensing Technology of Ministry of Natural Resources of P.R.C,Beijing 100039,China;Beijing Engineering Research Center of Low Altitude Remote Sensing Data Processing,Beijing 100039,China)

机构地区：[1]中国测绘科学研究院,北京100036 [2]自然资源部航空遥感技术重点实验室,北京100039 [3]北京市低空遥感数据处理工程技术研究中心,北京100039

出　　处：《测绘科学》2020年第11期84-92,共9页Science of Surveying and Mapping

基　　金：国家重点研发计划子课题项目(2016YFC0803109,2016YFC0803104)。

摘　　要：针对传统方法提取新增光伏用地精度低的问题,该文提出了一种基于集成学习的U-Net双网络变化信息融合的深度学习方法用于新增光伏用地的提取。首先对U-Net网络进行改进得到性能较好的两个变化检测网络模型,然后分别训练两个网络模型用于在高分辨率卫星影像上检测新增光伏用地,将训练好的两个网络模型的分类图融合再经过后处理得到最终的变化检测结果。通过实验表明:该方法明显优于传统变化检测方法,也提高了单网络模型变化检测结果的精度。According to the problem of low precision of extraction of newly-increasing photovoltaic land by using traditional methods,this paper presenteda new deep learning method that based on ensemble learning theory by fusing the detection results of two different modified U-Net network models to extract newly-increasing photovoltaic lands.At first,we modified the U-Net network to get two stable performance network models for change detection.Next,we trained them separately to detect newly-increasing photovoltaic land on high-resolution satellite imagery,and then the classification results of the two trained network models were fused and post-processed to obtain the final change detection result.The results of experiments indicated that this method was obviously superior to the traditional method for change detection and improved the accuracy of detection results based on single network model.

关 键 词：深度学习 U-Net网络 新增光伏用地提取 集成学习 

分 类 号：P237[天文地球—摄影测量与遥感]