迁移学习支持下的高分影像积云提取方法研究  

Research on Automatic Extraction of High Resolution Image Cumulus Based on Transfer Learning

作　　者：楚彬 郝建明 华亮春[1,2] 靳文凭[1,2] 李政 CHU Bin;HAO Jianming;HUA Liangchun;JIN Wenping;LI Zheng(Hunan Institute of Surveying and Mapping Technology,Changsha 410007,China;Chinese Academy of Surveying and Mapping,Hunan Branch,Changsha 410007,China;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]湖南省测绘科技研究所,湖南长沙410007 [2]中国测绘科学研究院湖南分院,湖南长沙410007 [3]西南交通大学地球科学与环境工程学院,四川成都611756

出　　处：《测绘与空间地理信息》2020年第3期93-96,共4页Geomatics & Spatial Information Technology

摘　　要：针对传统影像质量检查工作中积云提取存在人工作业量大、操作烦琐等问题,本文通过引入迁移学习机制,将已有数据集训练过程中得到的神经网络参数迁移到解译模型构建中,提出了一种适用于积云的自动提取方法。本文以湖南省不动产统一登记基础数据为实验对象进行了实验,结果表明,本文方法的浓积云提取总体精度可以达到90%以上,淡积云提取的总体精度可以达到87.3%,表明本文研究可用于高分影像积云自动提取。In this paper,a cumulus automatic extraction method based on transfer learning is proposed to solve the problem of large amount of manual operations and cumbersome operation in cumulus extraction in traditional image quality inspection,which neural network parameters obtained during the training of existing data sets are migrated to the interpretation model construction. Based on the unified data of real estate registration in Hunan Province,the results show that the overall accuracy of thick and thick cumulonimbus extraction can reach more than 90%,and the overall accuracy of thin cloud extraction can reach 87.3%. This paper can be widely used in the automatic extraction of cumulus in high resolution images.

关 键 词：积云提取 迁移学习 高分影像 面向对象分割 

分 类 号：P208[天文地球—地图制图学与地理信息工程]