DEEP LEARNING ENSEMBLES FOR HYPERSPECTRAL ANOMALY DETECTION

Hyperspectral imaging brings lots of opportunities in Earth observation, as such images acquired in orbit offer enormous spatial scalability, and the detailed spectral information allows to identify targets of interest. In hyperspectral anomaly detection, such targets are determined without any prior knowledge. In this paper, we tackle this issue and propose heterogeneous deep learning ensembles, benefiting from a variety of detection algorithms to precisely determine targets within hyperspectral imagery. Our experimental validation, performed over five benchmark hyperspectral scenes, shows that the ensemble approach outperforms all base learners, resulting in the average area under the receiver operating curve (AUC) of 0.968, with the AUC values of 0.889, 0.943, 0.942, 0.899, and 0.881 obtained using the base techniques included in the ensemble.