张昆仑, 常玉光, 潘洁,等.基于MTBNet的唐山尾矿库提取[J].河南理工大学学报（自然科学版）,2022,41(4):65-71.

ZHANG K L, CHANG Y G, PAN J, et al.Tailing pond extraction of Tangshan City based on Multi-Task-Branch Network[J].Journal of Henan Polytechnic University(Natural Science) ,2022,41(4):65-71.

基于MTBNet的唐山尾矿库提取

张昆仑1, 常玉光2, 潘洁1, 卢凯旋1, 昝露洋1, 陈正超1

1.中国科学院 空天信息创新研究院航空遥感中心，北京100094;2.河南理工大学 测绘与国土信息工程学院，河南 焦作 454000

摘要:针对尾矿库位置及边界难准确提取问题，在传统实例分割网络基础上，提出一个具有多任务分支结构的实例分割网络（Multi-Task-Branch Network ， MTBNet），并利用国产GF-1数据进行唐山地区的尾矿库提取试验。结果表明，召回率为95.8％时，尾矿库的检测准确率可达 78.8％。新方法进一步优化了尾矿库目标框和轮廓质量，增强了模型的特征学习能力，有效提升了尾矿库的实例分割精度，可为唐山地区尾矿库动态监测作支撑，辅助尾矿库开采或生态保护。

关键词:深度学习;GF-1遥感影像;尾矿库实例分割;唐山地区;MTBNet

doi:10.16186/j.cnki.1673-9787.2020070005

基金项目:中国科学院A类战略性先导科技专项（XDA23100304）；国家重点研发计划项目（2016YFB0500304 ）

收稿日期:2020/07/02

修回日期:2020/09/22

出版日期:2022/07/15

Tailing pond extraction of Tangshan City based on Multi-Task-Branch Network

ZHANG Kunlun1, CHANG Yuguang2, PAN Jie1, LU Kaixuan1, ZAN Luyang1, CHEN Zhengchao1

1.Airborne Remote Sensing Center of Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;2.School of Surveying and Land Information Engineering,Henan Polytechnic University,Jiaozuo  454000,Henan,China

Abstract: Aiming at the difficulty of accurate extraction of tailing pond location and boundary, an instance segmentation network with Multi-Task Branch structure based on the traditional instance segmentation network was proposed. Meanwhile, well-trained MTBNet was used to detect tailing ponds from GF-1 data for the Tangshan area. This novel method could achieve the detection accuracy of 78.8% while keeping the recall rate in a high level (95.8%). The method further optimized the qualities of the bounding box and mask, and enhanced the feature learning ability and effectively improved the instance segmentation accuracy of tailings pond, and it could be used as a data basis for dynamic monitoring of tailing ponds in Tangshan area， assisting tailing pond mining and reducing environmental pollution.

Key words:deep learning;GF-1 data;tailing pond instance segmentation;Tangshan area;MTBNet

基于MTBNet的唐山尾矿库提取_张昆仑.pdf