摘要
针对卷积神经网络(convolutional neural networks,CNNs)需求的训练样本量多,而高光谱图像中存在大量的未标签样本未得到充分利用的问题,文章充分挖掘标签样本及其近邻的未标签样本的空谱信息,提出了一种基于灰度共生矩阵(gray-level co-occurrence matrix,GLCM)和三维卷积神经网络的空谱特征联合训练的高光谱图像分类方法。首先,通过灰度共生矩阵提取高光谱图像的纹理特征;然后,利用相关性分析剔除近邻未标签样本中的冗余信息,将标签样本与未标签样本的信息融合;最后,利用三维卷积神经网络提取深空谱特征进行分类。该方法不但充分挖掘了高光谱图像的深度空谱联合特征,而且利用近邻未标签样本的信息实现对样本信息的增强,降低了对训练样本数量的要求,具有较好的分类性能。在3个公共数据集上的实验结果表明,相比其他方法,该方法可以利用较少的训练样本获得较高的分类精度。
Convolutional neural networks(CNNs)require more training samples,and a large number of unlabeled samples in hyperspectral images are not fully utilized.Therefore,this paper fully exploits the spectral-spatial information of the labeled sample and the neighboring unlabeled sample,and proposes a hyperspectral image classification method with joint training of spectral-spatial features based on the gray-level co-occurrence matrix(GLCM)and the three-dimensional convolutional neural network.Firstly,the texture features are extracted by the gray-level co-occurrence matrix.Then,the redundancy information in the unlabeled samples is eliminated by correlation analysis,and the information of the labeled sample and the unlabeled sample is fused.Finally,the deep spectral-spatial features are extracted by using the three-dimensional convolutional neural network.This method not only fully exploits the deep spectral-spatial-joint feature of hyperspectral images,but also uses the information of neighboring unlabeled samples to enhance the sample information,reduces the requirement of the number of training samples,and has better classification performance.Experimental results on three common datasets show that the proposed method can achieve higher classification accuracy with fewer training samples than other methods.
作者
韩彦岭
高仪
王静
张云
洪中华
HAN Yanling;GAO Yi;WANG Jing;ZHANG Yun;HONG Zhonghua(College of Information Technology, Shanghai Ocean University,Shanghai 201306, China)
出处
《遥感信息》
CSCD
北大核心
2020年第5期19-30,共12页
Remote Sensing Information
基金
国家自然科学基金项目(61806123、41871325)。
关键词
高光谱图像分类
空谱特征
灰度共生矩阵
三维卷积神经网络
未标签样本
hyperspectral image classification
spectral-spatial feature
gray-level co-occurrence matrix
three-dimensional convolutional neural network
unlabeled sample
