Predictability of global wildfire risk with transformer-based model

Extreme weather conditions, such as heatwaves and droughts driven by climate change, have led to a surge of large wildfires across the globe. This trend is exacerbated by rapid urban expansion and increasing interactions between human societies and wildlands, making the prediction of wildfire risk an urgent research priority.

While the Fire Weather Index (FWI) has been widely employed to evaluate fire risk, it primarily considers meteorological factors including wind, precipitation, temperature, and relative humidity. Some classic machine learning algorithms, such as Random Forest (RF), and deep learning approaches, represented by Convolutional Neural Networks (CNN), have been utilized to better capture nonlinear characteristics of wildfires. However, Transformer models, although proven efficient in multiple tasks varying from natural language processing to weather forecasting, remain largely unexplored in the context of wildfire risk prediction. Moreover, few studies have attempted to predict fire regimes at a global scale.

Therefore, our research aims to predict the next day’s global wildfire danger with high accuracy. We first established a comprehensive global wildfire database covering years from 2001 to 2020. The database contains historical burned areas records, as well as 50 key variables influencing occurrence and spread of wildfires, categorized as ignition source, fuel availability, weather condition, human activity, and topography. We then employed the Earthformer model, a transformer-based model incorporates a space-time attention block, to effectively capture the complex interplay of factors affecting wildfire regimes. By utilizing the daily dynamic variables (e.g. relative humidity) for days t-1, t-2, …, t-10 and constant variables such as land cover type, we predicted the probability for wildfire on day t. Our results indicate that Earthformer performs well with an F1-score for the positive sample (which represents high fire risk) greater than 0.85, which outperformed RF and XGBoost according to the confusion metric. Additionally, we implemented explainable AI (xAI) techniques to rank the importance of each factor contributing to fire risk.

Our study re-evaluated and generated global fire risk maps since 2020, providing essential insights for resource allocation in fire prevention strategies. By enhancing the understanding of wildfire dynamics, we aim to facilitate a better coexistence between communities and wildfires, ultimately contributing to improved resilience and mitigation efforts in the background of climate change.