Multiband CFAR Detection of Thermal Anomalies Using Principal Component Analysis

This paper deals with the problem of CFAR detection of thermal anomalies in multispectral satellite data. The goal is to extend the algorithm proposed in [1], and successfully applied to MODIS data from band 21, to the case of multiband investigation. A multiple-channel model has been designed, where data from MODIS bands 21 and 31 are projected into a new coordinates system by adopting the principal component analysis (PCA). A preliminary statistical analysis has been performed on both the principal components of data to verify that the Weibull distribution can be adopted for background. Subsequently, a Kendall test has been used to check the level of dependency of the projected data and it has shown that channels independence can be assumed with high significance level. After PCA, a CFAR detection is applied to projected data and thanks to data independence the single detections are combined with an AND rule. The outcome of the AND operation gives the thermal anomalies detected in both channels with an assigned overall probability of false alarm (PFA). The Multiband CFAR algorithm has been applied to a 256 x 256 MODIS image from bands 21 and 31 and results have been compared with those from NASA-DAAC MOD14.