Machine-Learning-based evaluation of the time-lagged effect of meteorological factors on 10-hour dead fuel moisture content

Abstract

Ten-hour dead fuel moisture content (DFMC10) is an important determinant of wildfire risk, as it provides a reasonable proxy to vegetation flammability. Thus far DFMC10 prediction by physical models has shown limited accordance with empirical measurements. DFMC10 can be measured using fuel-sticks, but to date these are only used in part of the meteorological stations. The goal of this paper is to improve the prediction accuracy of DFMC10 in the Mediterranean climate of Israel, where we collected four-year data from fuel sticks located in six meteorological stations. We show that Machine Learning (ML) models have the potential of accurately predicting DFMC10 based on historical values of temperature, relative humidity, rain, and additional meteorological factors. Our best ML model predicts low (<28%) DFMC10 values with an accuracy of 0.67% (MAE). The model even obtained a good MAE score (1.53%) when predicting data in stations it was never trained on.