Hyperspectral Anomaly Detection Based on Graph Regularized Variational Autoencoder

Nowadays, deep learning can play an important role in addressing the issue of hyperspectral anomaly detection (HAD). To further use the spatial information in hyperspectral images (HSIs), an anomaly detection (AD) method for HSIs is presented based on graph regularized variational autoencoder (GRVAE). First, the proposed method uses the superpixel segmentation algorithm to segment the HSI and constructs an adjacency matrix to evaluate the similarity between pixels. Second, a variational autoencoder is used to reconstruct the spectral vector of the HSI, and meanwhile, the spatial similarity of the image is shared in the feature space through the graph regularization term. Finally, the reconstructed background and the original input are used to obtain the spectral error map, and then attribute filtering is used to further refine the detection results. Performed on four datasets of abnormal target data with different shapes and different background complexity, the experiments show that the method has promising AD performance.

I. Introduction

Hyperspectral imaging is a remote sensing technology that collects electromagnetic spectral information for imaging, making it possible to identify different ground objects at a long distance through prior spatial–spectral features [1]. Hyperspectral image (HSI) data are always affected by lighting, environmental, atmospheric, and other conditions. The influence may reduce the accuracy of spectral-matching-based HSI analysis tasks such as classification, making such tasks challenging to adopt in natural environments’ application. As a blind signal detection technology, anomaly detection (AD) in HSIs can find abnormal pixels significantly different from the background without any prior target information [2]. This characteristic makes this technology widely used in military security, mineral exploration, and many other fields. Many hyperspectral AD (HAD) algorithms proposed in recent decades can be classified into distance-based, representation-based, and deep-learning-based algorithms.