Application of a hybrid fuzzy inference system to map the susceptibility to fires

This research mapped the susceptibility to outbreaks of fire using a hybrid fuzzy inference system (h-FIS) in the hydrographic basin of the Sorocabuçu River, in the municipality of Ibiúna, São Paulo, Brazil. We used 14 potentially influencing variables, four climatic factors, two anthropogenic, four topographic and four factors related to vegetation characteristics. The h-FIS method, together with Support Vector Machine (SVM) and Randon Forest (RF) was implemented for the year 2018, 2019 and 2020, based on training (70%) and test (30%) data. The most important variables selected by the Boruta algorithm were considered. The prediction success of each map was determined using the ROC curve, AUC and global accuracy. The results obtained showed that the most important factors for predicting fires are land cover and use, air humidity, soil moisture, precipitation and elevation. In the three years analyzed, the h-FIS model demonstrated slightly higher performance compared to the SVM and RF models. In 2018, 2019 and 2020, the h-FIS method identified 8.62%, 16.81% and 19.64% of the area as high and with very high susceptibility. The accuracy values based on independent data showed the h-FIS model to be of a good design (AUC = 92.5% and accuracy = 0.924 in 2018, AUC = 93.3% and accuracy = 0.9315 in 2019, and AUC = 90.4% and accuracy = 0.8991 in 2020) and, with the predicted susceptibility map, the relationship between the number of occurrences of fire observed in the high and very high fire susceptibility classes indicated a success rate of greater than 0.77%. These results confirm the efficiency of the proposed method, proving it to be suitable for the mapping of susceptibility to outbreaks of fire, and it can be used to assist public managers with fire prevention and mitigation.