A performance comparison of machine learning models for wildfire occurrence risk prediction in the Brazilian Federal District region

Wildfire frequency and the total area burned in Brazil have steadily increased over time. However, in the Federal District, wildfire hazards have risen due to wildfires reaching several of its protected areas. Wildfire occurrence/incidence risk refers to the likelihood or probability of a wildfire event occurring in a particular geographic area. The risk is typically determined by factors such as weather conditions, topography, vegetation, and human activities and often measured through analyses of historical wildfire data and predictive models that estimate the likelihood of future wildfire events. High wildfire occurrence/incidence risk areas are those more likely to experience a wildfire incident, whereas low-risk ones are less susceptible to it. This study compares 8 machine learning models that predict wildfire occurrence risk worldwide so that they can be adopted in Brazil. They considered correlations among climate conditions, spatial location, topographic features, anthropogenic explanatory features, and fire occurrence and a dataset enriched with Brazilian governmental open data was comprised of observations on 16 climate features of 5 monitoring stations and satellite data on fires occurred over the past two decades and topographic, hydrographic, and anthropogenic features, such as Normalized Difference Vegetation Index (NDVI), urbanization index, and distance to rivers/roads. According to the results, fire risk can be predicted with 0.99 accuracy and the models showed more sensitive to NDVI, atmospheric pressure, and relative humidity, as demonstrated by a study on the impact of features.