地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (2): 259-268.doi: 10.12082/dqxxkx.2019.180519

• 遥感科学与应用技术 • 上一篇    下一篇

基于卷积神经网络与条件随机场方法提取乡镇非正规固体废弃物

刘懿兰1,2(), 黄晓霞1,*(), 李红旮1, 柳泽3, 陈崇3, 王新歌3   

  1. 1. 中国科学院遥感与数字地球研究所,北京 100101
    2. 中国科学院大学,北京100049
    3. 住房和城乡建设部城乡规划管理中心,北京 100835
  • 收稿日期:2018-10-17 修回日期:2018-12-03 出版日期:2019-02-20 发布日期:2019-01-30
  • 通讯作者: 黄晓霞 E-mail:18649052480@163.com;hxx@irsa.ac.cn
  • 作者简介:

    作者简介:刘懿兰(1994-),女,天津人,硕士,研究方向为深度学习,地理信息系统等。E-mail: 18649052480@163.com

  • 基金资助:
    国家重点研发计划(2017YFB0503905);国土资源部城市土地资源监测与仿真重点实验室开放课题(KF-2016-02-012、KF-2018-03-032)

Extraction of Irregular Solid Waste in Rural based on Convolutional Neural Network and Conditional Random Field Method

Yilan LIU1,2(), Xiaoxia HUANG1,*(), Hongga LI1, Ze LIU3, Chong CHENG3, Xin'ge WANG3   

  1. 1. Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100101, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
    3. Ministry of Housing and Urban-Rural Development of the People's Republic of China, Beijing 100835, China
  • Received:2018-10-17 Revised:2018-12-03 Online:2019-02-20 Published:2019-01-30
  • Contact: Xiaoxia HUANG E-mail:18649052480@163.com;hxx@irsa.ac.cn
  • Supported by:
    National Key Research and Development Program of China, No.2017YFB0503905;The Project Supported by the Open of Key Laboratory of Urban Land Resources Monitoring and Simulation,Ministry of Land and Resources, No.KF-2016-02-012, KF-2018-03-032

摘要:

随着村镇经济建设发展,生活垃圾和工业固体废弃物造成的污染问题日益突出,已经成为制约新农村建设发展和生态文明建设的关键问题,而目前针对乡镇非正规固体废弃物的调查与统计主要依赖全国各乡镇相关部门逐级调查上报,工作量较大。本文基于高分辨率遥感影像,将深度学习模型和条件随机场模型相结合引入到乡镇固体废弃物的提取研究中,探索一种基于深度卷积神经网络的乡镇固体废弃物提取模型。由于固体废弃物在影像上表现为面积小,分布破碎等特点,为了提高工作效率,将模型特分为识别和提取2个部分:① 通过全连接卷积网络(CNN)对固体废弃物进行快速识别判断,筛选感兴趣区域影像块;② 在传统的全卷积神经网络(FCN)的基础上加入条件随机场模型(CRF)提取固体废弃物边界,提高整体分割精度。根据安徽、山西等地区相关部门上报固体废弃物堆放点以及住房与城乡建设部城乡规划管理中心进行野外检查的结果,实验最终识别精度达到86.87%以上;形状提取精度为89.84%,Kappa系数为0.7851,识别与提取精度均优于传统分类方法。同时,该方法已经逐步应用于住房和城乡建设部有关成都、兰州、河北等部分乡镇非正规固体废弃物的核查工作,取得了较为满意的结果。

关键词: 高分辨率遥感影像, 深度学习, 固体废弃物, 卷积神经网络, 条件随机场

Abstract:

With the development of rural economic construction, the pollution problem caused by domestic waste and industrial solid waste has become increasingly prominent, which has become a key problem to restrict the construction of the new rural developing and ecological civilization. At present, the investigation and statistics of informal solid waste in rural areas mainly depend on the reports of departments of each township step by step, and the workload is large. So based on high-resolution remote sensing images, this paper combines Deep Learning model with Conditional Random Field model to the study of rural solid waste extracting, and explores a recognition and extraction model of rural solid waste based on Deep Convolution Neural Network. Due to the solid waste in images is characterized by small size, distribution ,fragmentation and so on, in order to improve the efficiency, the model is divided into two parts: Recognition and Extraction. In the first part, a Full-connected Convolution Network (CNN) is used to identify and judge solid wastes quickly, and the image blocks include the interesting regions are screened. In the second part, Conditional Random Field model (CRF) is added to the traditional Full Convolution Neural Network (FCN) to extract boundary of solid waste and improve the overall segmentation accuracy.According to the relevant reports about solid waste of some rural areas in Anhui and Shanxi province and the field inspection by the urban and rural planning and management center of the Ministry of Housing and Urban-Rural Construction, Compared with the test results of the model in this paper,the results show the recognition accuracy is 86.87%,the shape extraction accuracy is 89.84%,and the Kappa coefficient is 0.7851. So the recognition and extraction accuracy of the paper's method is proved to be superior to the traditional methods. At the same time, this method has been gradually applied to the investigation of informal solid waste in countryside in Chengdu, Lanzhou, Hebei and other provinces, and achieved satisfactory results.

Key words: High-resolution remote sensing image, deep Learning, solid waste, convolution neural network, conditional random field