Modelling fire probability in the Brazilian Amazon using the maximum entropy method

Abstract

Fires are both a cause and consequence of important changes in the Amazon region. The development and implementation of better fire management practices and firefighting strategies are important steps to reduce the Amazon ecosystems’ degradation and carbon emissions from land-use change in the region. We extended the application of the maximum entropy method (MaxEnt) to model fire occurrence probability in the Brazilian Amazon on a monthly basis during the 2008 and 2010 fire seasons using fire detection data derived from satellite images. Predictor variables included climatic variables, inhabited and uninhabited protected areas and land-use change maps. Model fit was assessed using the area under the curve (AUC) value (threshold-independent analysis), binomial tests and model sensitivity and specificity (threshold-dependent analysis). Both threshold-independent (AUC = 0.919 ± 0.004) and threshold-dependent evaluation indicate satisfactory model performance. Pasture, annual deforestation and secondary vegetation are the most effective variables for predicting the distribution of the occurrence data. Our results show that MaxEnt may become an important tool to guide on-the-ground decisions on fire prevention actions and firefighting planning more effectively and thus to minimise forest degradation and carbon loss from forest fires in Amazonian ecosystems.