混合卷积神经网络的高光谱图像分类方法

Hyperspectral image classification based on hybrid convolutional neural network

为了解决高光谱图像领域中,传统卷积神经网络因部分特征信息损失而影响最终地物分类精度的问题,采用一种基于2维和3维的混合卷积神经网络的高光谱图像分类方法,从空间增强、光谱-空间两方面分别进行了特征提取。首先从空间增强角度提出一种3维-2维卷积神经网络混合结构,得到增强后的空间信息;其次从光谱-空间角度利用3维卷积网络结构,得到光谱-空间的综合可分性信息;最后将所得信息进行特征融合并分类。用该方法在两个数据集上进行了实验并与其它方法进行了对比。结果表明,该方法在Indian Pines与Pavia University数据集上分别取得了99.36%和99.95%的分类精度,其分类精度和kappa系数都优于其它方法。该方法对高光谱图像的分类表现出竞争优势。

The traditional convolutional neural network method can loss some feature information,which may lead to unsatisfied terrain classification accuracy in the field of hyperspectral.In order to solve the problem,a new hyperspectral images classification method based on the 2-D and the 3-D,named hybrid convolutional neural network,was proposed.This method mainly extracted features from the spatial enhancement aspect and the spectral-spatial aspect.Firstly,a 3-D-2-D convolutional neural network hybrid structure was proposed for enhance spatial information.Secondly,the 3-D convolutional neural network structure was used for joint feature extraction from the aspect of spectral-spatial,and then the spectral-spatial comprehensive separability information was obtained.Finally,the separately obtained information was feature fused and classified.This method was used for classification experiments on hyperspectral data sets and compared with other methods.The results show that the classification accuracy of this method is 99.36%and 99.95%respectively in Indian Pines and Pavia University data set,and its classification accuracy and kappa coefficient are also better than other methods.This method has a competitive advantage in the classification of hyperspectral images.