Unsupervised Hyperspectral Target Detection Using Spectral Residual of Deep Autoencoder Networks

Unsupervised hyperspectral target detection has attracted considerable attention in different applicable fields like environmental protection, agriculture monitoring, human rescue and etc. Due to the complexity of the background of hyperspectral images, accurate detection of target pixels using background estimation are faces lots of challenges which cause low accurate detection rates and high false alarm rates. In this paper we proposed an approach to estimate the background using a deep autoencoder neural networks architecture, then spectral residual error between the original hyperspectral image and the estimated one is considered for target detection using an exponential function. For evaluating the effectiveness and potency of the proposed method, several experiments are carried out on the well-known and challenging Hyperion dataset and the results are compared with some state-of-the-art methods. The visual and quantitative assessments show how spectral residual point of view using autoencoder networks can make high detection rates and low false alarms simultaneously.