An Improved Adaptive Weighted Deep Belief Network Autoencoder for Hyperspectral Images

Hyperspectral Anomaly Detection (HAD) in satellite missions allow us to autonomously detect regions of interest that notably differ from their surroundings. A weighting strategy together with an intensive pre-training for an autoencoder-based HAD method is proposed to increase detection accuracy. The proposed AE-based HAD method is evaluated in Matlab and extensively compared with baseline and state-of-the-art HAD algorithms in the paper. The results show that it achieves a high average accuracy on a set of standard hyperspectral images used to evaluate HAD algorithms.