Simple random forest classification algorithms for predicting occurrences and sizes of wildfires

In order to formulate effective fire-mitigation policies, it is important to understand the spatial and temporal distribution of different types of wildfires and to be able to predict their occurrence taking the main influencing factors into account. The objective of this short communication is to assess the capability of a fast and easy-to-implement random forest algorithm to estimate cumulative probabilities fire frequency and burned area using a large dataset collected in the USA. The input variables of the algorithm are voluntary restricted to climate and land use factors, which are easy to obtain in practice. No input related to fire frequency, burned area, or to any other fire characteristic is used. After model selection and training, the performance of random forest is assessed using an independent dataset including 80,000 observations of fire occurrence and burned area. Results show that the score of our simple random forest algorithm is 9% higher than the score of the winner of the data challenge of Opitz (Extreme, 2022) revealing that, although this model has a good performance, it is not the best. However, the approach proposed here can be implemented using standard packages, does not require any fire monitoring system after training, and requires little specialized knowledge in machine learning, which makes it usable by a large diversity of stakeholders. The results of this study suggest that random forest should be part of the toolbox of engineers and scientists involved in wildfire prediction.