An anomaly detection architecture based on a data-adaptive density estimation

In recent years, hyperspectral Anomaly Detection (AD) has become a challenging area due to the rich information content provided by hyperspectral sensors about the spectral characteristics of the observed materials. Within this framework, since no prior knowledge about the target is assumed, pixels that have different spectral content from typical background pixels are identified as spectral anomalies. The work presented here investigates this issue and develops a spectral-based algorithm for automatic global AD consisting in a two stage process. First, the background Probability Density Function (PDF) is approximated through a data-adaptive kernel density estimator. Then, anomalies are detected as those pixels that deviate from such a background model on the basis of the Likelihood Ratio Test (LRT) decision rule. Real hyperspectral data are employed to show the potential of data-adaptive background PDF estimation for detection of anomalies in a scene with respect to conventional non-adaptive PDF estimators