摘要
基于语义分割网络的遥感地物分类是近年来遥感解译所采用的新兴手段之一,其中Unet网络因结构简单、训练高效和易于构建等而成为最经典的语义分割模型。本文以最大限度扩增遥感数据信息量为前提,通过纹理、光谱、语义三方面特征提升数据信息丰度,主要工作包括:(1)引入Resnet残差模块对Unet网络模型加以改进;(2)采用窗口滚动输入输出方法解决数据喂入和相邻瓦块间“缝隙”问题;(3)采用数据串联耦合方法实现数据的三个方面特征融合。实验表明,改进Unet模型能在防止模型退化和保留纹理特征信息的前提下进一步加深网络层,提高了网络提取丰富语义特征的能力。数据串联耦合保留了数据的高层语义特征、纹理高频信息和数据辐射光谱信息,提升了SVM模型的分类精度,达到了良好的地物识别分割效果。
The classification of terrain objects in remote sensing data based on semantic segmentation model is becoming one of the newest methods adopted by RS interpretion task.Among those models,the Unet is likely to be the widely used one because of which of the succinct web structure,efficient training and easily construction.This paper aims to try hardest to strengthen the volume of RS data information through the way of promoting the volume of information in three aspects called texture,spectrum and semantics in data.The mainly jobs done among which contains as follows:(1)improve the classic Unet model by adopting the Resnet modules;(2)solve the problems of data-feed way and"chink"between adjacent data tiles;(3)adopt data seriescoupled method to realize the integration of the three features mentioned above.The consequence of experiment indicates that the advanced Unet model can hold deeper data-transform layers but at the same time prevent model degeneration and keep texture information transmission,and thus imporve the medel's capability of recognizing abundant semantic features.and that the method of data series-coupled integrates all of three of the high-level semantic feature,high-frequency texture feature and primary spectral information feature,and thus to help imporve the SVM model's precision of classification to achieve a better effect in terrain objects segmentation task.
作者
聂岩
蒋鹏飞
边防
贾方圆
NIE Yan;JIANG Peng-fei;BIAN Fang;JIA Fang-yuan(Dept of RS Information Center,ABDAS Space Information Technolog Co.Ltd,Zhengzhou,Henan 450000,China;Henan Branch Office,China United Network Communication Co.Ltd,Zhengzhou,Henan 450000,China)
出处
《新一代信息技术》
2023年第18期7-12,共6页
New Generation of Information Technology
关键词
Unet残差网络
语义特征
数据串联
信息特征
Res-Unet model
semantic feature
data series-coupled
informative characters
