摘要
高精度监测土地利用对实现可持续发展有重要意义。然而,由于遥感传感器成像的限制和地物的复杂性,单一的高光谱和多光谱图像已经不能满足高精度土地利用分类的要求,充分利用高光谱和多光谱遥感图像的互补信息能克服仅采用单一遥感图像分类的不足。该研究设计双分支卷积神经网络协同高光谱和多光谱遥感图像进行土地利用分类。针对高光谱图像设计3维-1维卷积神经网络(3D-1D Convolutional Neural Networks,3D-1D CNN)分支自动提取高光谱图像的空间-光谱特征;针对多光谱图像,设计3维卷积神经网络(3D Convolutional Neural Networks,3D CNN)分支提取多光谱图像的空间-光谱特征;设计融合层将从高光谱和多光谱图像提取的特征进行融合,最后通过全连接层输出土地利用类别。研究表明,与决策树(Decision Tree,DT)、支持向量机(Support Vector Machine,SVM)以及1D、2D和3D CNN方法相比,该文提出的基于双分支卷积神经网络的方法在两个数据集上Kappa系数平均分别提升了15.9、8.1、5.4、5.4和2.7个百分点。
Landscape monitoring using high-resolution aerial images has become highly significant to achieve a sustainable development of land use in modern agriculture.However,only hyperspectral or multispectral images individually cannot meet the harsh requirements of high-precision classification for land use,due to the technical limitations of remote sensing sensor imaging,and the complex features of landscape images.The combined hyperspectral and multispectral remote sensing images can be expected to overcome the technical problems,when classifying land use using one single type of images.In this paper,a dual-branch convolutional neural network(DBCNN)was designed to integrate with hyperspectral and multispectral remote sensing images for land use classification.In hyperspectral images,3D-1D convolutional neural network branch were used to automatically extract spatial-spectral features,whereas,3D convolutional neural network branch was used in multispectral images.A fusion layer was designed to fuse the extracted features from hyperspectral and multispectral images.The land use category was finally output via the fully connected layer in the network.A dropout layer was added to the fully connected layer in the network,in order to avoid overfitting of imaging data.Image data sets were taken from two research areas in this study.One research area was in the rural and urban border area at Chikusei,Ibaraki,Japan,indicating mainly farm land in the whole regions.The hyperspectral and multispectral images were generated from the data set of Chikusei.The multispectral data was constituted three wavelength bands,corresponding to the center wavelength,(the 60th,40th,and 20th bands of Chikusei data set),where the spatial resolution of multispectral image was 2.5m.Another research area was selected in Pavia University,USA,where the 53th,31st,and 7th wavelength bands were set,while the spatial resolution was 1.3 min in multispectral data.A Gaussian down sampling method was used to obtain hyperspectral data.In the experiment,the obtained data set was divided randomly,where the ratio of training samples,validation samples and test samples were 25.5%,4.5%and 70%,respectively.Five comparative experiments were conducted,including decision tree(DT),support vector machine(SVM),1D CNN,2D CNN and 3D CNN.A confusion matrix of each classification method was obtained,together with the overall accuracy,average accuracy,and kappa coefficient,by calculating the difference between the classification of comparative experiments and the land composition from actual field.In Chikusei data set,the overall accuracy of 3D-1D CNN was improved by 7.06,5.72,and 2.90 percentage points,respectively,compared with DT,SVM and 1D CNN,indicating the highest accuracy of 3D-1D CNN in hyperspectral image classification.Compared with DT,SVM,and 2D-CNN,the overall accuracy of 3D CNN was improved by 6.16 percentage points,0.38%,and 7.40%respectively,indicating the highest in the classification of multispectral remote sensing images.The experimental results show that the overall accuracy of fusion classification was constantly enhanced from shallow to deep learning algorithms.Compared with DT,SVM,1D CNN,2D CNN and 3D CNN,the overall accuracy of DBCNN has increased by 7.80,5.05,3.43,2.60 and 0.90 percentage points,respectively.The classification of CNN in deep learning performed better than that of SVM and DT in shallow classification algorithms,especially on the identification of road traffic land,grassland,and water bodies.An optimum effect can be achieved,combined with the classification algorithm of land use and the dual-branch network-based hyper-multispectral imagery.The spatial and spectral features of hyperspectral and multispectral images can be used to enhance the model's convergence and generalization ability,and thereby to improve the accuracy of classification.The classification of water body,grassland,and habitation performed very well,while the misclassification of transportation land also significantly decreased,indicating the overall accuracy was 99.39%,and the Kappa coefficient was 0.9920.In Pavia University data set,the classification effect of 1D CNN,2D CNN and 3D CNN performed better than that of the shallow classification method,such as DT and SVM.In the DBCNN classification method,there was a clear distinction between the unused land and transportation land,while the misclassification of housing and construction land significantly decreased.Compared with other classification methods,DBCNN has the best classification effect.The accuracy of all classifications was above 90%in the data of Pavia University.Compared with DT,SVM,1D CNN,2D CNN and 3D CNN,the overall accuracy of DBCNN increased by 14.84,6.75,4.55,5.34,3.06 percentage points.
作者
刘帅
张旭含
李笑迎
田野
Liu Shuai;Zhang Xuhan;Li Xiaoying;Tian Ye(School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,China;Hebei Key Laboratory of Information Transmission and Signal Processing,Qinhuangdao 066004,China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2020年第14期252-262,共11页
Transactions of the Chinese Society of Agricultural Engineering
基金
国家自然科学基金项目(61601398)
河北省高等学校科学技术研究项目(QN2017146)。
关键词
土地利用
遥感
卷积神经网络
高光谱
多光谱
land use
remote sensing
convolutional neural network
hyperspectral
multispectral
