Comparisonof local anomaly detection algorithms based on statistical hypothesis tests

We present some results about the comparison of two families of anomaly detection algorithms, specific to hyperspectral images analysis, both based on local statistical Hypothesis Testing (HT). The study has involved the RX and the Gauss-Markov Random Fields (GMRF) anomaly detectors. These algorithms share a quite similar sliding window strategy and the approach of statistical HT but they differ notably in the way data are modeled and the parameters estimated. An underlying main question is actually the trade-off between describing the background clutter with many parameters (case of RX), harder to locally estimate, and considering a model with fewer parameters (case of GMRF), thus easier to estimate from the local clutter only. The study includes the issue of reducing the spectral dimension with linear second order and higher order statistics based methods. As an output of the study, RX will be included in the next release of the Orfeo Toolbox (OTB), an open source remote sensing image processing library developed by the CNES.