Background Characterization of Anomalies Using Hyperspectral data

Anomaly detection algorithms identify pixels which appear different with respect to immediate neighborhood. However, the output may contain many false alarms and also does not provide any information about the background against which the pixels are flagged as anomalous. This paper describes a methodology which is two-step process, first identifying anomalies using RX detector and reducing the false alarms by using probabilistic distribution. In the second step background against which the filtered anomalous pixels appear is characterized as vegetation, water, soil or snow based on combination of spectral indices. For characterizing water, soil and vegetation narrow band vegetation index is used and likewise for snow also. The developed methodology is tested on multiple datasets acquired by recently launched HYSIS satellite.