Improving wildfire occurrence modelling by integrating time-series features of weather and fuel moisture content

• Published in: Environmental modelling & software : with environment data news Vol. 170; p. 105840
• Main Authors: Quan, Xingwen, Wang, Wenli, Xie, Qian, He, Binbin, Resco de Dios, Víctor, Yebra, Marta, Jiao, Miao, Chen, Rui
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2023
• Subjects: China; computer software; Fuel moisture content; fuel moisture index; Machine learning; probability; time series analysis; Time-series features; topography; Weather; Wildfire; Wildfire probability; wildfires; China; Weather; Wildfire; Fuel moisture content; Time-series features; Wildfire probability; Machine learning
• ISSN: 1364-8152; 1873-6726
• DOI: 10.1016/j.envsoft.2023.105840

Wildfire occurrence is a non-linear process resulting from interactions between weather, topography, fuel, and anthropogenic factors amongst others. Modelling the probability of wildfire occurrence has commonly focused on the features of the dynamic conditions (i.e., weather and fuel variables) one-day before the fire. However, few studies have examined the time-series features of these variables during the days to weeks prior to the fire. Here, we tested whether wildfire probability modelling could be improved by integrating the time-series features of weather and fuel moisture content. We developed a wildfire probability model across southwest China and observed a clear improvement in wildfire probability modelling after including the time-series features. Analysis of the variable importance further confirmed that the time-series features played a crucial role in driving wildfire occurrence. Our modelling approach provides a novel framework for modelling wildfire probability by integrating time-series features of dynamic conditions leading up to the fire. •We combined static parameters with time series of conditions leading to the fire.•Time-series analyses included remotely sensed global FMC and weather variables.•Significant model improvements occurred after adding time series analyses.•Our approach presents a new generation of wildfire probability modeling.