Support vector data description for anomaly detection in hyperspectral imagery combined with neighboring clustering segmentation

Support vector data description (SVDD) is good at detecting little anomalies in hyperspectral imagery, but it gives the high miss rate resulted from preconcerted shape of anomaly and blindfold background selection, and has to process vast amounts of data during traversing all over the imagery. This paper presented a new detection method for anomaly detection in hyperspectral imagery. Firstly the method segmented imagery by the method of neighboring clustering segmentation based on spectral information and regarded those small imagery blocks as the potential anomalies, and then selected adaptively the background windows to collect the background pixel samples according to the shape and size of the potential anomalies, lastly confirmed the anomalies quickly and exactly based on SVDD. The experiments on the HYMAP data show higher detection rate is obtained than SVDD. Moreover, operation redundancy is avoided during traversing the whole imagery in SVDD.