计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (2): 244-251.DOI: 10.3778/j.issn.1002-8331.2008-0415

• 图形图像处理 • 上一篇    下一篇

基于双通道变分自编码器的高光谱图像分类

刘遵雄,石亚鹏,彭潇雨,王毅宏   

  1. 华东交通大学 信息工程学院,南昌 330013
  • 出版日期:2022-01-15 发布日期:2022-01-18

Hyperspectral Image Classification Based on Two-Channel Variational Autoencoder

LIU Zunxiong, SHI Yapeng, PENG Xiaoyu, WANG Yihong   

  1. College of Information Engineering, East China Jiaotong University, Nanchang 330013, China
  • Online:2022-01-15 Published:2022-01-18

摘要: 针对现有高光谱图像变分自编码器(variational autoencoder,VAE)分类算法存在空间和光谱特征利用效率低的问题,提出一种基于双通道变分自编码器的高光谱图像深度学习分类算法。通过构建一维条件变分自编码器(conditional variational autoencoder,CVAE)特征提取框架和二维循环通道条件变分自编码(channel-recurrent conditional variational autoencoders,CRCVAE)特征提取框架分别提取高光谱图像的光谱特征和空间特征,将光谱特征向量和空间特征向量叠加形成空谱联合特征向量,将联合特征送入Softmax分类器中进行分类。在Indian pines和Pavia University两种高光谱数据集上进行了分析验证,实验结果显示,与其他算法相比,提出的算法在总分类精度、平均分类精度和Kappa系数等评价指标上至少提高了3.40、2.75和3.57个百分点,结果显示提出的算法得到了最高的分类精度和更好的可视化效果。

关键词: 深度学习, 高光谱图像, 分类, 变分自编码器, 空谱联合特征

Abstract: Aiming at the problems of low utilization efficiency of spatial and spectral features in hyperspectral image variational autoencoder(VAE) classification algorithm, a deep learning classification algorithm for hyperspectral images based on two-channel VAE is proposed. One-dimensional conditional variational autoencoder(CVAE) and two-dimensional channel-recurrent conditional variational autoencoders(CRCVAE) feature extraction frameworks are used to extract spectral and spatial features of hyperspectral images, respectively. These eigenvectors are superposed to form a spatial-spectral feature and classified by using the Softmax classifier. The analysis and verification are conducted on two hyperspectral data sets of the Indian pines and the Pavia University. The experimental results show that compared with other algorithms in this paper, the proposed algorithm improves at least 3.40, 2.75 and 3.57 percentage points in terms of overall classification accuracy, average classification accuracy and Kappa coefficient. The results show that the proposed algorithm has the highest classification accuracy and better visualization effect.

Key words: deep learning, hyperspectral image, classification, variational autoencoder, spatial-spectral feature