A CFAR algorithm for anomaly detection and discrimination in hyperspectral images

paper proposes an anomaly detection algorithm for hy- perspectral images. It is unsupervised (the researched spectra are not required a priori), discriminates the anomalies according to their spectra and has a constant false alarm rate (CFAR). The main specificity of this algorithm is to combine these three assets rather than make a tradeoff which is generally necessary with existing methods. It is based on a physically convenient probabilistic model of the fastICA gener ated independent components. We compare it with the adaptive cosine/coherence estimator (a reference supervised target detection algorithm) on a real HYDICE dataset.