Use of maximum entropy to improve validation and prediction of active fires in a Brazilian savanna region

Abstract

Monitoring fires using satellite images is a fundamental tool for managers of protected areas (PAs). However, false alerts hinder their ability and make their application unfeasible, since there are few fire brigades available for field checks. Thus, this research proposes the validation, identification, and prediction of active fires for the years 2001 to 2018 for a protected area of the Brazilian savanna. The validation was performed by counting the active fires that were effectively contained in a burning area by cross-referencing MODIS sensor images. Identification was performed by analyzing the spectral behavior of burning and non-burning in the active fire area. The maximum entropy was calculated from the spectral values and compared between an active focus with and without burning. In the prediction, the climatic conditions of the month before burning were analyzed, and the maximum entropy value was evaluated for model viability. The results indicated that only 30% of the active fires were actually in a burned area. The use of maximum entropy favors the improvement of the MODIS active fire product. The confirmed fires were in regions where the probability was above 75%, whereas the non-burning fires were approximately 60%. The variables evapotranspiration, wind speed, water deficit, and road proximity were the most relevant for the active fire prediction model. For the validation years 2019 and 2020, this method predicted the location of fires in 75% of the cases. This methodology, once employed by PA managers, can ensure accurate, efficient, and cost-effective monitoring of forest fire prevention and combat.