We propose a prediction/detection scheme for automatic forest fire surveillance by means of passive infrared sensors. Prediction takes advantages of the highly correlated environment in the infrared band to improve signal to noise ratio. We have observed that, in general, data are non-Gaussian distributed; thence nonlinear prediction allows improvements in the predictor performance. In particular, we consider the nonlinear Wiener system. In addition, the prediction step allows assuming Gaussianity for the detector design. A specific problem in the detection step is to distinguish uncontrolled fire from what we call occasional effects. This situation justifies basing the detection in a vector signature. We exploit the expected particular characteristics about fire signatures by means of two different detectors: a matched subspace detector and a detector that exploits the presence of increasing trends in the signature (increase detector). The problem with the fusion of the two decisions is also considered. Real data experiments validate the interest of the proposed scheme.