Predição da susceptibilidade a incêndios na Apa Triunfo do Xingu co ferramentas de machine learning

Abstract

The behavior and characteristics of fire are determined by a variety of factors, including topography, vegetation, climate, and socioeconomic aspects, which makes it a complex task to predict in which areas resources should be allocated for combatting and preventing forest fires. Machine learning tools have demonstrated promising results for fire prediction, which have included the generation of models that have been developed across a large range of contexts and locations, with applications to the Brazilian context. This research aims to map areas susceptible to forest fires within the Triunfo do Xingu Environmental Protection Area, employing machine learning algorithms to ascertain the influence of environmental, topographic and socioeconomic factors on fire occurrence. For this purpose, the Random Forest and Extreme Gradient Boosting regression models were used to predict kernel density values calculated over 15,291 confirmed burn points between 2010 and 2020, using 11 predictor factors, including Elevation, Slope, Aspect, Topographic Wetness Index, Precipitation, Temperature, Proximity to Inhabited areas, Proximity to Roads, Land Use and Cover, Vegetation Continuous Fields, and the Normalized Difference Vegetation Index. To evaluate the performance of the algorithms, the metrics used were Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R2 ). The test results showed that the models had similar performance, and both the Random Forest (RMSE = 36.26, MAE = 17.45, and R2 = 0.99) and the Extreme Gradient Boosting (RMSE = 35.73, MAE = 18.74, and R2 = 0.99) demonstrated good predictive capacity, with a statistically significant difference only in the MAE values, according to the Mann-Whitney statistical test. The elaborated map presented areas of high and very high susceptibility occupying 39% of the total area of the conservation unit, mainly located in the central-east and central-west regions. The variables with the greatest importance and contribution to the final model were environmental and socioeconomic variables, notably precipitation, distance from inhabited areas, and land use type.