FLAME 1.0: a novel approach for modelling burned area in the Brazilian biomes using the maximum entropy concept

Abstract. As fire seasons in Brazil lengthen and intensify, the need to enhance fire simulations and to comprehend fire drivers becomes crucial. Yet determining what drives burning in different Brazilian biomes is a major challenge, with a highly uncertain relationship between drivers and fire. Finding ways to acknowledge and quantify that uncertainty is critical in ascertaining the causes of Brazil's changing fire regimes. We propose FLAME (Fire Landscape Analysis using Maximum Entropy), a new fire model that integrates Bayesian inference with the maximum entropy concept, enabling probabilistic reasoning and uncertainty quantification. FLAME utilises bioclimatic, land cover, and human driving variables to model fires. We apply FLAME to Brazilian biomes, evaluating its performance against observed data for three categories of fires: all fires (ALL), fires reaching natural vegetation (NAT), and fires in non-natural vegetation (NON). We assessed burned-area responses to different explanatory variable groups. The model showed adequate performance for all biomes and fire categories. Together, maximum temperature and precipitation are important factors influencing burned area in all biomes. The number of roads and forest boundaries (edge densities), forests and pastures, and carbon in dead vegetation showed higher uncertainties among the responses. Overall, the uncertainties were larger for the NON category, particularly for the Pampas and Pantanal regions. Customising explanatory variable selection and fire categories based on biome characteristics could contribute to a more biome-focused and contextually relevant analysis. Moreover, prioritising regional-scale analysis is essential for decision-makers and fire management strategies. FLAME is easily adaptable and can be used in various locations and periods, serving as a valuable tool for more informed and effective fire prevention measures.