Assessing the effect on fire risk modeling of the uncertainty in the location and cause of forest fires

Abstract

Wildfire risk assessments in Spain usually make little or no reference to the uncertainty of the results due to ignition data quality, or the implications that this potential uncertainty may have on wildfire management decisions. In Spain the autonomous regions have historically been the competent authorities in forest management and environmental protection as a result of the 1978 Constitution and, therefore, responsible on the operational application of the criteria defined for the country for wildfire classification and location. This competency framework has generated significant regional differences in the application of the criteria for wildfire classification among the different autonomous regions, arising potential uncertainty on wildfire assessments and fire risk models based on this historical series of data. This work explores six scenarios based on the classification of fire ignition causes and location data, reported in the General Statistics of Wildfires database (EGIF), to address the potential uncertainty from the point of view of the variability in predicted ignition probability and the changes in its spatial patterns. The analysis is focused on analyzing the effects on human-caused wildfires by using Random Forest algorithms to predict the ignition likelihood and cluster and outlier analysis (hot and cold spot) to detect changes in the spatial pattern of probability. Results suggest that there is significant uncertainty both in predicted human-caused ignition and spatial pattern related to the ignition source and location of fire events compiled in the EGIF database. The accuracy of the predictions ranges from AUC values of 0.90, when considering most of the records of the database, to around 0.76 in scenarios characterized by using only known-caused allocated fire events, probably due to differences in the proportions of unidentified and allocated fires within the mainland Spain.