基于MTBNet的唐山尾矿库提取  

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

作　　者：张昆仑 常玉光[2] 潘洁 卢凯旋 昝露洋 陈正超 ZHANG Kunlun;CHANG Yuguang;PAN Jie;LU Kaixuan;ZAN Luyang;CHEN Zhengchao(Airborne Remote Sensing Center of Aerospace Information Research Instiute,Chinese Academy of Sciences,Bejing 100094,China;School of Surveying and Land Information Engineering,Henan Polytechnic University,Jiaozuo 454000,Henan,China)

机构地区：[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

出　　处：《河南理工大学学报（自然科学版）》2022年第4期65-71,94,共8页Journal of Henan Polytechnic University（Natural Science）

基　　金：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)。

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。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.

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 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]