贾宏光1, 2, 3,,,
高放3,
李文涛3,
陆晴3
1.中国科学院长春光学精密机械与物理研究所 长春 130033
2.中国科学院大学 北京 100049
3.长光卫星技术有限公司 长春 130102
基金项目:吉林省重大科技攻关项目(20170201006GX)，长春市科技局重大科技攻关项目(SA13RP2018040101)，吉林省科技厅重点科技研发项目(20180201109GX)
详细信息
作者简介:卢俊言：男，1990年生，博士生，研究方向为基于深度学习的遥感影像数据挖掘
贾宏光：男，1971年生，研究员，博士生导师，研究方向为无人机总体技术，精确末制导技术，飞行器半物理仿真及小型快速机电伺服技术
高放：男，1987年生，工学博士，研究方向为遥感数据处理与应用
李文涛：男，1990年生，硕士，研究方向为遥感影像DSM, DOM, DEM生产
陆晴：女，1995年生，硕士，研究方向为基于深度学习的计算机视觉及数据挖掘
通讯作者:贾宏光 jiahg@ciomp.ac.cn
中图分类号:TP394.1计量
文章访问数:586
HTML全文浏览量:148
PDF下载量:49
被引次数:0
出版历程
收稿日期:2020-01-09
修回日期:2020-09-10
网络出版日期:2020-09-14
刊出日期:2021-04-20
Reconstruction of Digital Surface Model of Single-view Remote Sensing Image by Semantic Segmentation Network
Junyan LU1, 2, 3,Hongguang JIA1, 2, 3,,,
Fang GAO3,
Wentao LI3,
Qing LU3
1. Changchun Institute of Optics, Fine Mechanics, and Physics, Chinese Academy of Sciences, Changchun 130033, China
2. University of Chinese Academy of Sciences, Beijing 100049, China
3. Chang Guang Satellite Technology Co., Ltd, Changchun 130102, China
Funds:The Key Technologies of Jilin Province (20170201006GX), The Major Science and Technology Research Project of Changchun Science and Technology Bureau (SA13RP2018040101); The Key Science and Technology Research Project of Jilin Province Science and Technology Department (20180201109GX)
摘要
摘要:该文提出了一种仅依靠激光探测与测量数据，实现单视图遥感影像数字表面模型(DSM)重建的新方法。该方法基于深度学习技术设计了一种编码-解码结构的语义分割网络，该网络采用多尺度残差融合的编码块与解码(MRFED)块从输入图像中提取语义信息，进而逐像素预测高度值；采用特征图跳跃级联的策略保留输入图像的细节特征和结构信息。该文采用了一个包含DSM数据的遥感影像公开数据集训练与测试模型，实验结果表明：DSM重建结果与真值的平均绝对误差(MAE)为2.1e-02，均方根误差(RMSE)为3.8e-02，结构相似性(SSIM)为92.89%，均优于经典的深度学习语义分割网络。实验证实该方法能够有效实现单视图遥感影像的DSM重建，具有较高的精度，以及较强的地物分布结构重建能力。
关键词:语义分割网络/
编码-解码/
多尺度残差融合/
跳跃级联/
数字表面模型
Abstract:A novel method for Digital Surface Model (DSM) reconstruction of single-view remote sensing image is proposed which only relies on light detection and ranging data. Based on deep learning technology, a semantic segmentation network with an encode-decode structure is designed. The network uses Multi-scale Residual Fusion Encode and Decode (MRFED) blocks to extract semantic information from the input image, and then predicts the height value pixel by pixel, as well as adopts a strategy of skip connections with feature maps to preserves the detailed features and structural information of the input image. The model is trained and tested on a public dataset of remote sensing images containing DSM data. Experiments show that, the Mean Absolute Error (MAE) between DSM reconstruction results and true values is 2.1e-02, the Root Mean Square Error (RMSE) is 3.8e-02, and the Structural SIMilarity (SSIM) is 92.89%, which are all better than the classic deep learning semantic segmentation networks. Experiments confirm that the method can effectively reconstruct the DSM of single-view remote sensing images with high accuracy, as well as the structure of feature distribution.
Key words:Semantic segmentation network/
Encode-decode/
Multi-scale residual fusion/
Skip connections/
Digital Surface Model (DSM)
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=65f6cbf7-2399-4b1c-8dbb-407e7e564e7b
