Abstract : The Thermical InfraRed (TIR) channel contains wave lengths sensitive to the emission of heat. The forest fires can be characterized by intensity peaks on TIR images. We present a fully automatic method of forest fi re detection from TIR satellite images based on the random fi eld theory. First, preprocessing is used to model the image as a realization of a Gaussian fi eld which presents interesting properties. The fire areas which are supposed to be a minority are considered as anomalies of that fi eld. We present a statistical analysis of Gaussian field to determine a degree of belonging of a cluster to the image background (i.e. a realization of a Gaussian fi eld). We then extend this application to the estimation of the fi re propagation direction.