Robust multiband detection of thermal anomalies using the Minimum Covariance Determinant estimator

This paper deals with the problem of Constant False Alarm Rate (CFAR) detection of thermal anomalies in multispectral satellite data. The goal is to provide robustness to the algorithm proposed in [1], with respect to the presence of outliers in the analysis window. In [1], data from 4 ¿m and 11 ¿m MODIS bands, that are statistically correlated, are re-projected through a Principal Component Analysis (PCA) to obtain uncorrelated data, a necessary condition for the final stage of CFAR detection. Unfortunately, the sample covariance matrix used in the PCA can be strongly affected by the presence of thermal anomalies, therefore a robust estimator is needed. To this aim, the Minimum Covariance Determinant estimator (MCD) is introduced in the PCA yielding an analysis that is little influenced by the presence of anomalies while provides results similar to the usual PCA for uncontaminated data. Experimental results have shown that many detections can be missed if the MCD estimator is not used in the presence of anomalies, even if their number is not so high but their values are able to modify significantly the sample covariance matrix. The robust multiband CFAR algorithm has been applied to a MODIS image and results have been compared with those from NASA-DAAC MOD14.