Background distribution of multi-spectral detector output

In the detection of field fire or leakage, the target signature is very weak. Analyzing statistical models of actual spectral detector output distribution is crucial for setting detection thresholds, selecting statistical classifiers and designing constant false alarm rate detectors. Multispectral image data is processed with the CEM algorithm. Statistical model of detector output is discussed based on the stochastic mixing model. It is concluded that Gaussian mixture model is suitable. This is demonstrated with the ETM Satellite date. EM algorithm is adopted to estimate parameters of Gaussian mixture model. Computer simulation results show that the probability density function of statistics derived from Gaussian mixture model agrees with that of actual data. Curves of false alarm rate also show that Gaussian mixture model is effective at the condition of Pfa＞10-4.