面向高光谱图像分类的局部Gabor卷积神经网络  

Local Gabor Convolutional Neural Network for Hyperspectral Image Classification

作　　者：王燕[1] 王丽[1] WANG Yan;WANG Li(College of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学计算机与通信学院,兰州730050

出　　处：《计算机科学》2020年第6期151-156,共6页Computer Science

基　　金：甘肃省重点研发计划(18YF1GA060)。

摘　　要：针对高光谱图像特征利用不足的问题,提出了一种新的基于空谱联合特征的高光谱图像分类方法。该方法首先利用主成分分析(Principal Component Analysis,PCA)和线性判别分析(Linear Discriminant Analysis,LDA)对高光谱图像进行组合降维;其次引入Gabor核,设计了一种基于Gabor核的卷积(Local Gabor Convolutional,LGC)层;最后基于LGC层设计了一个新的卷积神经网络(Local Gabor Convolutional Neural Network,LGCNN)进行分类。在Indian Pines和Salinas Scene数据集上对所提方法进行验证,并将其与其他经典分类方法进行比较。实验结果表明,该方法不仅能大幅度减少可学习的参数,降低模型复杂度,而且具备较好的分类性能,其总体精度达到99%,平均分类精度达到98%以上,Kappa系数达到98%以上。In order to solve the problem of insufficient utilization of hyperspectral image features,a new classification method based on spatial-spectral features was proposed.Firstly,principal component analysis(PCA)and linear discriminant analysis(LDA)are used to reduce the dimension of hyperspectral images.Secondly,the Gabor kernel is introduced to design a Local Gabor Convolution(LGC)layer based on the local Gabor kernel.Finally,a new convolutional neural network(LGCNN)is designed based on the LGC layer for classification.The proposed method is validated on Indian Pines and Salinas scene datasets and compared with other classical classification methods.Experiment results show that this method not only saves the learning parameters greatly,reduces the complexity of the model,but also shows good classification performance.Its overall accuracy can reach 99%,the average classification accuracy can reach more than 98%,and the Kappa coefficient can reach more than 98%.

关 键 词：高光谱图像分类 GABOR滤波 空间-光谱信息 卷积神经网络 深度学习 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]