基于多尺度学习与深度卷积神经网络的遥感图像土地利用分类  

Land use classification of remote sensing images based on multi-scale learning and deep convolution neural network

作　　者：王协 章孝灿[1] 苏程[1] WANG Xie;ZHANG Xiaocan;SU Cheng(Institute of Spatial Information Technology,School of Earth Sciences,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]浙江大学地球科学学院空间信息技术研究所,浙江杭州310027

出　　处：《浙江大学学报（理学版）》2020年第6期715-723,共9页Journal of Zhejiang University（Science Edition）

摘　　要：土地利用信息是国土资源管理的基础和重要依据,随着高分辨率遥感图像数据的日益增多,迫切需要快速准确的土地利用分类方法。目前应用较广的面向对象的分类方法对空间特征的利用尚不够充分,在特征选择上存在一定的局限性。为此,提出一种基于多尺度学习与深度卷积神经网络(deep convolutional neural network,DCNN)的多尺度神经网络(multi-scale neural network,MSNet)模型,基于残差网络构建了100层编码网络,通过并行输入实现输入图像的多尺度学习,利用膨胀卷积实现特征图像的多尺度学习,设计了一种端到端的分类网络。以浙江省0.5 m分辨率的光学航空遥感图像为数据源进行了实验,总体分类精度达91.97%,并将其与传统全卷积网络(fully convolutional networks,FCN)方法和基于支持向量机(support vector machine,SVM)的面向对象方法进行了对比,结果表明,本文所提方法分类精度更高,分类结果整体性更强。Land use data is an important fundamental information for national land resources management.Following the availability of high resolution remote sensing image data,it is on urgent demand to have a fast and accurate land-use classification method.The object-oriented classification which has been widely applied at present has some problems such as low level utilization of spatial features and limited choice of features.In this paper,a multi-scale neural network(MSNet)model based on multi-scale learning and deep convolutional neural network(DCNN)is proposed.We built 100 layers encoding network based on residual neural network,and conducted several parallel input streams to accomplish multi-scale learning of input images,then utilized dilated convolution to accomplish multi-scale learning of feature images,finally designed an end-to-end classification network.Experiments were implemented on the optical aerial remote sensing images dataset of Zhejiang province with 0.5 m resolution,the overall accuracy of classification reached 91.97%.Compared with fully convolutional networks(FCN)network and the object-oriented method based on support vector machine(SVM),the MSNet method has a higher precision of classification and demonstrates more integrity of the scene.

关 键 词：高分辨率遥感图像 土地利用分类 多尺度学习 深度卷积神经网络(DCNN) 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]