Mapping wildfire ignition probability and predictor sensitivity with ensemble-based machine learning

Wildfire ignition models can help in identifying risk factors and mapping high-risk areas, which addresses an urgent issue as wildfires become increasingly destructive. Despite advancements in data collection and data analysis, challenges persist as ignitions depend on numerous interconnected factors at a fine spatial resolution. Predicting wildfire ignitions from data is an imbalanced classification problem, given the vast number of non-ignition data compared to ignition data. To address this issue, this study proposes an ensemble-based model for binary classification of wildfire ignitions. The data are collected for a 24,867 km2 area in northern California from January 2014 to May 2022 and includes 76 predictors covering topographic, land cover, anthropogenic, and climatic data. Different base classifiers are evaluated and the random forest is found the most performant, yielding a recall of 0.67 and a specificity of 0.87. Feature importance analysis shows that the Topographic Wetness Index is the most important climatic predictor, while population density and land cover development are also highly rated. Comparison of yearly average of computed daily probabilities with ignition data shows that the model accurately captures the spatial pattern of ignitions, which can reveal high-risk areas. The model is then used to assess how climatic and anthropogenic factors impact wildfire ignition frequency. The projected scenarios show that the number and spread of ignitions would significantly increase with an increase in population in sparsely populated areas, while climatic factors have secondary effects in isolation but in combination may compound the risk. As current land development and climate change trends are expected to increase the frequency and severity of wildfires, data-based models can provide insights to inform policy and mitigate risk.