Classification of hyperspectral images based on a convolutional neural network and spectral sensitivity

作     者：Cheng-ming YE Xin LIU Hong XU Shi-cong REN Yao LI Jonathan LI Cheng-ming YE;Xin LIU;Hong XU;Shi-cong REN;Yao LI;Jonathan LI

作者机构：Chongqing Engineering Research Center of Automatic Monitoring for Geological HazardsChongqing 401120China Key Laboratory of Earth Exploration and Information Technology of Ministry of EducationChengdu University of TechnologyChengdu 610059China National Breeding Base of Technology and Innovation Platform for Automatic-monitoring of Geologic HazardsChongqing Institute of Geology and Mineral ResourcesChongqing 401120China Key Laboratory of Mountain Hazards and Earth Surface ProcessChinese Academy of SciencesChengdu 610041China Department of Geography and Environmental ManagementUniversity of WaterlooWaterlooN2L 3G1Canada 

出 版 物：《Journal of Zhejiang University-Science A(Applied Physics & Engineering)》 (浙江大学学报（英文版）A辑（应用物理与工程）)

年 卷 期：2020年第21卷第3期

页      面：240-248页

核心收录：

学科分类：0810[工学-信息与通信工程] 08[工学] 0805[工学-材料科学与工程（可授工学、理学学位）] 081002[工学-信号与信息处理] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Project supported by the Strategic Priority Research Program of the Chinese Academy of Sciences(No.XDA23090203) the National Key Technologies Research and Development Program of China(No.2016YFB0502600) the Key Program of Sichuan Bureau of Science and Technology(No.2018SZ0350),China 

主　　题：Hyperspectral imaging Deep learning Convolutional neural network(CNN) Spectral sensitivity 

摘      要：In recent years,deep learning methods have gradually come to be used in hyperspectral imaging *** of the peculiarity of hyperspectral imaging,a mass of information is contained in the spectral dimensions of hyperspectral ***,different ob jects on a land surface are sensitive to different ranges of *** achieve higher accuracy in classification,we propose a structure that combines spectral sensitivity with a convolutional neural network by adding spectral weights derived from predicted outcomes before the final classification ***,samples are divided into visible light and infrared,with a portion of the samples fed into networks during ***,two key parameters,unrecognized rate(δ)and wrongly recognized rate(γ),are calculated from the predicted outcome of the whole ***,the spectral weight,derived from these two parameters,is ***,the spectral weight is added and an improved structure is *** improved structure not only combines the features in spatial and spectral dimensions,but also gives spectral sensitivity a primary *** with inputs from the whole spectrum,the improved structure attains a nearly 2%higher prediction *** applied to public data sets,compared with the whole spectrum,on the average we achieve approximately 1%higher accuracy.