The approach to the detection of fires suggested here is based on methods of pattern recognition in spaces of the informative parameters from information contained in indirect measurements, which in this case are five-channel videodata recorded with the AVHRR instrument placed onboard NOAA satellites. A problem of preliminary integrated normalization of satellite videodata, including a transition to constant dimensions of scanning spot projections on the Earth’s surface, an increase in the spatial resolution of images for a model of integration within the spot, and correction of the radiobrightness characteristics of the images, is considered. Normalized images are subsequently used to solve the problem of detecting small-sized fires with the help of a three-stage procedure by an algorithm of pattern recognition in space of the informative parameters. A natural criterion for estimating the information content for the class of detection and pattern recognition problems is the functional of the average risk. In this case, the informative set of parameters and the decision rule are found by minimization of this functional. Because conditional probability densities, being mathematical models of stochastic images, are unknown, the problem of reconstructing distributions based on teaching samples with the use of nonparametric estimates with modified Epanechnikov kernel is solved. Unknown parameters of distributions are determined by minimization of a functional of the empirical risk. A comparison between the results of operation by the algorithm and the operator work demonstrates high efficiency of the algorithm of detecting thermal anomalies of fire types.