基于深度学习的遥感影像目标自动提取技术研究  

Automatic Extraction of Target Information for Remote Sensing Images Using Deep Learning Methods

作　　者：王诗洋 李淳 于兴超 Wang Shiyang;Li Chun;Yu Xingchao(32023 Troops,Dalian 116023,China)

机构地区：[1]32023部队,辽宁大连116023

出　　处：《地理信息世界》2021年第2期120-124,共5页Geomatics World

摘　　要：为研究重要目标信息的自动化快速提取,本文基于PyTorch、CUDA等环境并在GPU运行下,利用CycleGAN算法进行深度学习的模型训练。通过GeoServer对谷歌影像进行地图切片等处理,制作上海市浦东新区的船舶、机场、岸线的训练数据集。模型训练后其模型测试的结果实现了基于高分辨率遥感影像的目标信息自动化提取的科学目标。比较目标地物如船舶、机场、岸线等的信息提取效果,结果表明3种瓦片等级尤其17级瓦片下进行的多种目标信息自动化提取效果均较好,且在地物目标监测及军事应用方面具有重要意义。In order to realize the automated extraction of important target rapidly,the CycleGAN algorithm for deep learning training based on PyTorch,CUDA and GPU is used in this study.Google images were made into map tiles through GeoServer,to make the training data sets of ships,airports and shoreline in pudong new area of Shanghai.The results of model test realized the automatic extraction of target information based on remote sensing images.These results show that the automatic extraction of target information in this study was of great significance in military application,according to the extraction effects of target information such as ship,airports and shoreline in different tile level.

关 键 词：深度学习 CycleGAN GEOSERVER 自动化提取 

分 类 号：P236[天文地球—摄影测量与遥感]