Deeplab V3+高分遥感影像水体信息提取  

Extracting Water Body Information from High-resolution Remote Sensing Images by DeepLab V3+

作　　者：殷腾箐 周兴华 宋立松 YIN Tengjing;ZHOU Xinghua;SONG Lisong(Zhejiang Institute of Hydraulics&Estuary(Zhejiang Institute of Marine Planning and Design),Hangzhou 310017,Zhejiang,China)

机构地区：[1]浙江省水利河口研究院(浙江省海洋规划设计研究院),浙江杭州310017

出　　处：《浙江水利科技》2024年第6期88-93,共6页Zhejiang Hydrotechnics

基　　金：2021年浙江省水利科技计划项目(RC2150);2021年浙江省水利河口研究院院长科学基金重点项目(ZIHE21Z004);2024年浙江省自然科学基金联合基金资助项目(LZJWY24E090002)。

摘　　要：及时准确获取和掌握水体分布信息对于水域管理等具有重要的意义。以水网密布、分布复杂的浙北杭嘉湖水网平原为研究区,采用随机森林模型和基于扩张残差网络(DRN)的Deeplab V3+模型,对北京二号高分辨率遥感影像进行水体信息提取,比较了2种模型水体信息提取的准确性。结果表明,采用5个特征的随机森林模型的提取精度要好于只采用3个特征的随机森林模型;Deeplab V3+模型的提取精度要明显优于随机森林模型,它的提取总体精度为0.9766,kappa系数为0.8266,MIoU为0.8266,均明显高于随机森林模型;以目视解译结果为参考,Deeplab V3+的提取结果也明显好于随机森林模型,消除了随机森林提取结果中明显的“椒盐效应”,其原因可能与Deeplab V3+模型能充分利用高分遥感影像的光谱和空间纹理特征有关。因此,DeepLabV3+模型可以从高分遥感影像有效地提取水体的信息,即使在杭嘉湖平原这种复杂的环境下,为快速、准确获得水体分布信息提供了一个有效的途径和手段。Timely and accurate acquisition and mastery of water body distribution information are of great significance for water area management.This study takes the Hangjiahu water network plain in northern Zhejiang as the study area,and uses the random forest model and the Dilated Residual Networks(DRN)based DeepLab V3+model to extract water body information from Beijing No.2 high-resolution remote sensing images,and compares the results of the two models.The results show that the extraction accuracy of the random forest model using 5 features is better than the random forest model using only 3 features;the extraction accuracy of the DeepLab V3+model is significantly better than the random forest model that its overall extraction accuracy is 0.9766,and the kappa coefficient is 0.8266,MIoU is 0.8266,both are significantly higher than the forest model;taking the visual interpretation results as a reference,the extraction results of DeepLab V3+are also significantly better than the random forest model,eliminating the obvious“salt and pepper effect”in the random forest extraction results.The reason may be related to the fact that the DeepLab V3+model can make full use of the spectral and spatial texture characteristics of high-resolution remote sensing images.Therefore,the DeepLabV3+model can effectively extract water body information from high-resolution remote sensing images,even in complex environments such as the Hangjiahu Plain,providing an effective way and means to quickly and accurately obtain water body distribution information.

关 键 词：语义分割 遥感 深度学习 DeepLabV3+ 

分 类 号：P237[天文地球—摄影测量与遥感] TV882.9[天文地球—测绘科学与技术]