Anomaly detection in hyperspectral imagery by fuzzy integral fusion of band-subsets

Anomaly detection in hyperspectral imagery is gaining increasing interest. However, most covariance matrix-based detectors are applied directly to all the hyperspectral data bands without considering the spectral variation and their possible different contribution to detection. Besides, the limited sample number as compared with the high dimensionality of the data can lead to the imprecise estimation of the covariance matrix and even the singularity problem. In this paper, a band-subset fuzzy integral fusion (BS-FI) detection method is presented to solve these problems. The complete set of hyperspectral data bands is first partitioned into several lower dimensional band-subsets, whose detection results are obtained separately and finally merged by a fuzzy integral fusion method. We adopt a non-parametric fuzzy support function which can utilize more statistical information and avoid the model discrepancy that might be brought in by a fixed distribution model. In addition, the fuzzy density is assigned by the ratio between the target signal and noise, which is the key to the target detection problem through an adaptive eigenvalue-based approach. In the experiments on real OMIS-I hyperspectral imagery, the proposed method outperforms the RX detector both on the complete set of bands and on each band-subset, and other band-subset fusion detectors.