基于卷积神经网络的高光谱图像分类  

Hyperspectral Image Classification Based on Convolutional Neural Networks

作　　者：李子轩 官云兰[1,2,3] 王楠 周世健 LI Zixuan;GUAN Yunlan;WANG Nan;ZHOU Shijian(School of Surveying and Geoinformation Engineering,East China University of Technology,330013,Nanchang,PRC;Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake of Ministry of Natural Resources,330013,Nanchang,PRC;Key Laboratory of River Basin Ecological Processes and Information in Jiangxi Province,330013,Nanchang,PRC;Nanchang Hangkong University,330063,Nanchang,PRC)

机构地区：[1]东华理工大学测绘与空间信息工程学院,南昌330013 [2]自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,南昌330013 [3]江西省流域生态过程与信息重点实验室,南昌330013 [4]南昌航空大学,南昌330063

出　　处：《江西科学》2025年第1期26-35,共10页Jiangxi Science

基　　金：国家自然科学基金项目(42064001);自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室项目(MEMI-2023-15)。

摘　　要：近年来,以卷积神经网络为代表的深度学习模型因具有提取高级抽象特征能力而受到研究人员的关注,并在高光谱图像分类中得到广泛应用。针对几种常用的卷积神经网络模型进行对比,分析各模型的优缺点,以Pavia University、Salinas和WHU-Hi-HongHu数据集为基础开展高光谱图像分类实验。结果表明,3D-CNN和ResNet34在分类精度上优于1D-CNN和2D-CNN,尤其在处理具有复杂空间特征的高光谱数据时表现更为突出。同时,不同数据集对不同模型的响应存在差异,3D-CNN在UP和更复杂的HongHu数据集上表现最好,而ResNet34在SA数据集上具有优势,实践中需要根据数据集特点选用适合的网络模型,进一步开发具有普适性的算法,以取得好的分类效果。In recent years,deep learning models,particularly convolutional neural networks(CNNs),have attracted the attention of researchers due to their ability to extract high-level abstract features,and have been widely applied in hyperspectral image classification.This study compared several commonly used convolutional neural network models,and analyzed their advantages and disadvantages.Hyperspectral image classification experiments were conducted based on the Pavia University,Salinas,and WHU-Hi-HongHu datasets.The results show that 3D-CNN and ResNet34 outperformed 1D-CNN and 2D-CNN in terms of classification accuracy,particularly when handling hyperspectral data with complex spatial features.Additionally,differences in the response of different models to different datasets were observed.3D-CNN performs best on the UP and more complex HongHu datasets,while ResNet34 demonstrates an advantage on the SA dataset.It is suggested that the appropriate network model should be selected based on the characteristics of the dataset in practical applications,and further development of more generalizable algorithms is needed to achieve better classification results.

关 键 词：深度学习 高光谱图像分类 卷积神经网络 残差结构 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]