Empirical fire propagation potential from a balanced dataset

Abstract

In order to assess fire and wildfire risk one must address various features and analyse the danger and vulnerability aspects. Besides fire ignition probability, one of the most important variables for addressing fire danger is fire propagation potential. Fire propagation potential (FPP) can be described as a quantitative description of the circumstances under which, if fire ignites, it leads towards propagation of fire. This means that not all ignitions cause propagation of significant fires. Some ignitions are easily extinguished and pose no danger to vulnerable assets. On the other hand, some ignitions result in large and mega fires, causing large, burned areas and huge casualties. Fire propagation potential (FPP) provides quantitative distinction between these two different circumstances. Machine learning techniques are more and more applied in fire management tools as they provide us with techniques for learning from the past data and predicting the future outcomes. Majority of previous work is focused on analysis of the large fire events, their causes and development. However, when modelling the FPP, we should consider situations on both ends of the outcome spectrum - situations when fire ignites and propagates and situations when fire ignites and does not propagate. If one uses only data on fires that propagate, without considering the alternative situations data, results that are achieved can be incomplete. In this paper we propose a novel and more full approach to fire danger assessment by analysing situations of both cases - high and low fire danger. We simplify the value of FPP and consider that in cases the fire propagates the value of FPP is one, and zero otherwise. We used data collected from the events of both cases. We obtained a balanced dataset and trained machine learning model with a data set having representatives of both ends of the FPP spectrum. The research is demonstrated in the study area of Split and Dalmatia County. We consider past fires that are sensed by satellite and recorded in the EFFIS system as situations when FPP had value 1. To assess the situations when FPP was 0 we analysed the fire intervention database maintained by fire departments. We filtered fire interventions related to forest fires that lasted less than 2 hours and engaged 2 or less firefighters since these records represent time and place of the fire that did not propagate. For these two cases of events, we collected Sentinel-2 imagery and weather data that consists of temperature and wind speed. Sentinel-2 imagery pixels were extracted for the area associated with both types of events. The dataset was split into train and test datasets, where classifiers were trained by using 80% of data and 20% of remaining data was used for testing the classifier performance. Experiments were conducted by training classifiers using commonly used classifiers - Decision Tree Classifier, K-Nearest Neighbors, Multi-layer perceptron, Random Forest Classifier, Naive Bayes Classifier and Logistic regression. The best performance, according to the R2 score and RMSE is measured on Decision Tree Classifier.