Detection of fires from satellite images using a nonparametric algorithm of pattern recognition in space of the informative parameters

摘要

The problem of early detection of small-sized fires is very urgent, especially for almost inaccessible and sparsely populated territories. In this connection, a necessity arises of using satellite information to solve problems of real-time detecting sources of thermal anomalies suspicious of a fire. Among the Earth's artificial satellites most frequently orbited above a given locality that allow problems of monitoring of the Earth's underlying surface (EUS) to be solved, it is natural to chose satellites of NOAA series and to use the data of the AVHRR instrument that records radiobrightness values in five spectral ranges, including thermal ones, in the form of images. Unfortunately, low resolution of the AVHRR instrument and comparatively narrow range of radiobrightness values registered with it do not allow the problem of early detection of small-sized fires to be solved efficiently by conventional methods. Let us consider an approach based on theoretical methods of statistical hypothesis testing (pattern recognition) [1-4] used for solving problems with high degree of statistical uncertainty and unknown conditional probability density functions. Probabilistic models of situations to be recognized, one of which belongs to class "Fires" and the remaining belong to "Fire-like" interference classes, are reconstructed in spaces of the parameters informative for the minimum risk criterion. The detection of fires based only on the temperature does not allow this problem to be solved efficiently. Let us assume that examined images have already been preprocessed: geometrical distortions have been eliminated from videodata, they have been fixed geographically, their fragment with the territory of the Tomsk Region (TR) and its environs has been cut out, and the radiobrightness correction and calibration of videodata recorded with the AVHRR instrument have been performed with determination of albedos in channels 1 and 2 and thermodynamic temperatures in channels 3, 4, and 5. Thus, we have a matrix of 1024 X 1024 five-dimensional vectors to be analyzed.