Hyperspectral Image Classification Approach Based on Wasserstein Generative Adversarial Networks

Hyperspectral image classification is an important research direction in the application of remote sensing technology. In the process of labeling different types of objects based on spectral information and geometric spatial characteristics, noise interference often exists in continuous multi-band spectral information, which brings great troubles to spectral feature extraction. Besides, far from enough spectral samples will restrict the classification performance of the algorithm to some extent. In order to solve the problem of small amount of original spectral sample data and noisy signal, Wasserstein generative adversarial networks (WGAN) is used to generate samples similar to the original spectrum, and spectral features are extracted from the samples. In the case of small samples, the original materials are provided for the classification of hyperspectral images and a semi-supervised classification model WGAN-CNN for hyperspectral images based on Wasserstein generation antagonistic network is proposed in this paper. This model combines with CNN classifier and completes the classification of terrain objects according to the label for the synthesized samples. The proposed method is compared with several classical hyperspectral image classification methods in classification accuracy. WGAN-CNN can achieve higher classification accuracy in the case of small sample size, which proves the effectiveness of the proposed method.

1. Introduction

Hyperspectral imagery has the characteristics of large amount of information and noisy information. Even if the spectral resolution can reach the level of 10^-2λ, the features of ground objects will be concealed due to the interaction of electromagnetic radiation and atmosphere[l]. These interferences have always been a difficult problem in the research of hyperspectral image classification. Although the traditional classification algorithm can deal with high-dimensional data and noise signals, it can not solve the problem of too few hyperspectral samples, which greatly restricts the performance of the classification algorithm. For hyperspectral image classification, if the excellent classification network can be selected at the same time of increasing the training sample size, a high accuracy classification model will be obtained. Therefore, this kind of model has far-reaching research significance.