Machine Learning Approach in Fire Risk Estimation

Abstract

Forest fire weather index (FWI), empirical model for fire risk estimation based on temperature, humidity and wind speed and precipitation, is used by European Forest Fire Information System (EFFIS) and by Croatian Meteorological and Hydrological Service (DHMZ). EFFIS also estimates fires based on satellite images, which is not precise enough for Croatia forest fires because the number actual fires were several times higher number of real fires in southernmost part of Croatia (Dalmatia). The aims of the paper is to examine the possibility of Machine Learning (ML) in identification of forest fire risk by using only meteorological measurements and to compare ML in forest fire risk identification to widely used Fire Weather Index (FWI). Three different ML models have been built and compared - logistic regression, random forest and artificial neural network on two different sets of features on period 2017-2018. The contribution of the paper is the process of ML and FWI models comparison based on actual fires and generated no-fires. The actual fire database have been manually collected from Croatian Fire Brigade website. Our results show that the best model according to F-score is multiple layer perceptron (MLP) with temperature, humidity, wind speed and ultra violet index (UVI) (F-score is 0.78), while FWI model that seems to be good according to F-score is FWI > 11.2 (F-score is 0.7). The disadvantage of both FWI and ML models is that they report ’Fire’ in high percentage of situations in Dalmatia during fire season.