Novel machine-learning-based fire weather indices

In light of the increasing frequency of droughts and extreme fire weather events, in Europe and world wide, the need for accurate wildfire risk estimation is becoming more and more acute. Wildfire risk depends however on complex non-linear interactions between multiple factors such as fuel moisture content, winds, humidity, topography and others.
Traditional fire indices, currently used by weather services, are based on linear models and or empirical and statistical analyses. Consequently their performance is somewhat limited.  Here we propose a novel set of fire weather indices (FWIs), developed using machine learning (ML) – the MLFWI.  We find that the MLFWI significantly outperforms the traditional fire indices in predicting wildfire occurrence, achieving an AUC score of 0.99 compared to 0.62-0.80.
We also analyze the influence of the various factors and their interactions on the models, providing scientific insights and understanding of the mechanism by which the models work. Finally, we compare the performance of the MLFWIs to that of traditional indices in predicting the 100 largest wildfires in the dataset. We find that our models were able to predict the vast majority of these 100 extreme events.
In the talk we will present the Machine Learning methodology and examine the performance of the the MLFWI. The ultimate goal of this research is to allow implementation of  the MLFWI in actual wildfire warning systems. We propose to build upon this study to gradually replace the existing fire weather indices with ML-based indices, which have the potential of substantially improving fire weather alerts.