Hyperspectral anomaly detection based on score-guided cooperative autoencoders

Abstract

ABSTRACT Hyperspectral anomaly detection (HAD) is an important application of hyperspectral technology and has received extensive research attention. The lack of available prior spectral information limits deep-learning-based HAD. During the model training process, background and anomaly data are input into the network indiscriminately, and the model tends to overfit, which results in a better representation of backgrounds and anomalies. Anomalies and backgrounds are highly mixed together; therefore, it is difficult to effectively separate them and measure the difference between them. To mitigate these challenges, this study proposes a HAD algorithm based on score-guided cooperative autoencoders (SGCAE). Through the cooperative training of two autoencoders, suspected ‘anomalies’ are continuously removed during the training process to purify the training data. The autoencoder’s ability to discriminate anomalies from backgrounds is enhanced using a score-guided network in the hidden layer of the autoencoder. The score-guided network encourages the model to learn from obvious-background and obvious-abnormal data, gradually distinguish highly mixed backgrounds and anomalies, and output end-to-end anomaly scores. We apply the proposed algorithm to four real hyperspectral datasets. The experimental results verified the effectiveness of the proposed SGCAE algorithm and demonstrated that it outperforms other state-of-the-art methods.