高分七号影像大场景DSM生成深度学习方法  

Deep learning method for large scene DSM generation of GF-7 imagery

作　　者：何升 张建凯 陈凤 李绅弘 江万寿[1] HE Sheng;ZHANG Jiankai;CHEN Feng;LI Shenhong;JIANG Wanshou(State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China;Beijing Geoway Spatial Information Co.,Ltd.,Beijing 100040,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079 [2]北京吉威空间信息股份有限公司,北京100040

出　　处：《测绘学报》2023年第12期2103-2114,共12页Acta Geodaetica et Cartographica Sinica

基　　金：高分城市精细化管理遥感应用示范系统(二期)(06-Y30F04-9001-20/22)。

摘　　要：立体匹配是利用卫星影像生成DSM的重要步骤,近年来基于深度学习的立体匹配方法有较好的性能,然而,由于模型预测的视差范围固定有限及缺少训练数据,深度学习很少直接用于大场景卫星影像的立体匹配。本文提出了一种分层动态匹配策略,根据上一层匹配结果来动态确定本层影像分块的区域,使左右核线影像块的视差相对较小,利于深度学习模型进行预测;本文提出了一套卫星影像立体匹配样本制作方案,利用人工编辑的DSM或LiDAR点云获取视差真值,构建了一个高分七号立体匹配数据集。使用该数据集和现有数据集训练Stereo-Net和DSM-Net并基于分层匹配策略,实现了结合深度学习技术的高分七号影像高质量DSM生成。3个城市的影像试验表明,本文方法匹配的视差图的平均视差绝对误差为1像素左右,错误视差像素比例不超过3.8%,生成的DSM质量优于传统方法。Stereo matching is an important step to generate DSM from satellite imageries.Recently,studies have shown that deep learning-based methods have better performance.However,due to the fixed and limited disparity range predicted by models and the lack of training data,deep learning is rarely directly applied to the stereo matching of satellite images in large scenes.In this paper,a hierarchical dynamic matching strategy is proposed to dynamically determine the region of image blocks of the current level according to the matching results of the previous level,so that the disparity between the left and right epipolar image blocks is relatively small,which is conducive to the prediction of deep learning models.Besides,a scheme for the production of samples is introduced,and a GF-7 dataset is constructed by using manually edited DSM or LiDAR point clouds to obtain ground truth disparity values.In the experiment,this dataset,together with an existing dataset,is used to train Stereo-Net and DSM-Net,and based on the hierarchical matching strategy,the generation of high-quality DSM from Gaofen-7 imagery combined with deep learning technology is achieved for the first time.Experiments in imageries from three cities show that the average endpoint error is about 1 pixel,and the fraction of erroneous pixels is less than 3.8%.The quality of the generated DSMs is better than that of the traditional method.

关 键 词：卫星影像 立体匹配 DSM 深度学习 高分七号 

分 类 号：P236[天文地球—摄影测量与遥感]