摘要
针对现有安全风险防范工作中对尾矿库数量统计与"一库一策"的管理需求,该文提出了一种将深度学习与随机森林结合的典型尾矿库遥感目标识别方法。首先,运用Faster R-CNN模型进行基于遥感影像的尾矿库目标检测,实现大范围影像内尾矿库正确识别结果的保留;其次,为进一步提高识别精度,统计识别结果中图像的几何特征,采用U-Net分别设计尾矿库语义分割模型与尾矿库水、砂、坝场景分割模型,获取了尾矿库面积特征、场景元素几何特征与空间关系特征,通过改进的Res-Net图像分类模型得出尾矿库概率特征;最后,基于随机森林分类模型对单一特征与多特征组合进行模型训练与测试,得到最优特征参数组合,从而实现典型尾矿库的高精度识别。结果表明,该文提出的方法能够在稀少样本条件下,通过特征提取与最优特征组合实现不同区域尾矿库高精度识别与提取,可为尾矿库科学管理工作提供一定技术支持。
Aiming at the problems of low efficiency and slow update in the traditional tailings pond identification method based on manual recognition,this paper proposes a typical tailings pond recognition method that combines deep learning and random forest.Firstly,to detect tailings pond targets in remote sensing images based on the Faster R-CNN model,the correct recognition results of tailings ponds in the large scale images are retained;then,in order to further eliminate false alarm targets and improve the recognition accuracy,geometric features of the targets are calculated.Based on U-Net,the semantic segmentation model of tailings pond and the segmentation model of water,sand and dam scenes are designed separately to catch the area characteristics,geometric characteristics and spatial relationship characteristics.Res-Net image classification model is improved to get the probability characteristics of tailings pond.Finally,based on the random forest classification model,the single feature and multi-feature combination were trained and tested,working out the optimal feature parameter combination and the high precision detection results of the typical tailings pond.The results show that the method proposed in this paper can realize the identification of typical tailings ponds under rare samples,and realize the identification and extraction of tailings ponds in large scales through feature extraction and optimal feature combination,then provide technical support for the scientific management of tailings ponds.
作者
刘冰洁
邢旭东
吴浩
胡少华
昝军
LIU Bingjie;XING Xudong;WU Hao;HU Shaohua;ZAN Jun(College of Urban and Environmental Sciences,Central China Normal University,Wuhan 430079,China;Central Southern China Electric Power Design Institute,Wuhan 430070,China;School of Safety Science and Emergency Management,Wuhan University of Technology,Wuhan 430070,China;Hubei Production Safety Emergency Rescue Center,Wuhan 430070,China)
出处
《测绘科学》
CSCD
北大核心
2021年第12期129-139,共11页
Science of Surveying and Mapping
基金
湖北省技术创新专项(重大项目)(2019ACA143)。
关键词
遥感影像
目标检测
特征提取
随机森林
尾矿库识别
remote sensing image
target detection
image segmentation
random forest
tailings pond identification
