Comparison and statistical analysis between models of probability of occurrence of forest fires for Honduras. Year 2019

Abstract

Honduras, due to its climatic and physiographic conditions in recent years, has suffered negative impacts due to wildfires that cause losses in the country’s economy, such as the closure of airports, in the health of urban and rural populations, incidence of pests and diseases in agriculture and forestry, greater soil degradation and erosion, alteration of water recharge systems, negative effects on biodiversity as well as the loss of scenic values. In 2019, 1,177 fires were reported with an affected area of 72,434.77 hectares. The departments with the highest incidence of wildfires were; Francisco Morazán, Olancho, and Copan. In this research, the performance of three models (ICF, Random Forest and Max-Ent) for the generation of susceptibility maps to wildfires from environmental variables was statistically analyzed. The models were calibrated with the fire report data generated by the ICF. Additionally, fire points were selected randomly obtained from the FIRMS System. The Random Forest model had the best performance with an AUC for test data of 0.973 and an AUC for MODIS-FIRMS data of 0.919. Using the average rankings of the environmental variable importance measures, proximity to human settlements was the best predictor of wildfire ignitions, closely followed by distance to unpaved roads and elevation, for models based on Machine Learning Algorithms.