Predicting forest fire kernel density at multiple scales with geographically weighted regression in Mexico.

Identifying the relative importance of human and environmental drivers on fire occurrence in different regions and scales is critical for a sound fire management. Nevertheless, studies analyzing fire occurrence spatial patterns at multiple scales, covering the regional to national levels at multiple spatial resolutions, both in the fire occurrence drivers and in fire density, are very scarce. Furthermore, there is a scarcity of studies that analyze the spatial stationarity in the relationships of fire occurrence and its drivers at multiple scales. The current study aimed at predicting the spatial patterns of fire occurrence at regional and national levels in Mexico, utilizing geographically weighted regression (GWR) to predict fire density, calculated with two different approaches -regular grid density and kernel density - at spatial resolutions from 5 to 50 km, both in the dependent and in the independent human and environmental candidate variables. A better performance of GWR, both in goodness of fit and residual correlation reduction, was observed for prediction of kernel density as opposed to regular grid density. Our study is, to our best knowledge, the first study utilizing GWR to predict fire kernel density, and the first study to utilize GWR considering multiple scales, both in the dependent and independent variables. GWR models goodness of fit increased with fire kernel density search radius (bandwidths), but saturation in predictive capacity was apparent at 15-20 km for most regions. This suggests that this scale has a good potential for operational use in fire prevention and suppression decision-making as a compromise between predictive capability and spatial detail in fire occurrence predictions. This result might be a consequence of the specific spatial patterns of fire occurrence in Mexico and should be analyzed in future studies replicating this methodology elsewhere.