A two-stage procedure for detection of thermal anomalies (such as fires) on the territory of some region from satellite data of the AVHRR device is considered. At the first stage, the field of thermodynamic temperature of the underlying surface is reconstructed using a nonlinear non-parametric regression adapted to particular observation conditions found from the coordinated data of meteorological services and data of the AVHRR device from onboard the NOAA satellite. At the second stage, the reconstructed temperature field is used for constructing the Bayes adaptive rule to detect thermal anomalies. The rule is based on the principle of the component identification in a mixed distribution and approximation of the conditional density functions by the Johnson curves. An example of detection of thermal anomalies from satellite video data at the territory of the Tomsk Region is presented.