Search

Abstract Wildfire occurrence is attributed to the interaction of multiple factors including weather, fuel, topography, and human activities. Among them, weather variables, particularly the temporal characteristics of weather variables in a given period, are paramount in predicting the probability of wildfire occurrence. However, rainfall has a large influence on the temporal characteristics of weather variables if they are derived from a fixed period, introducing additional uncertainties in wildfire probability modeling. To solve the problem, this study employed the weather variables in continuous nonprecipitation days as the “dynamic-step” weather variables with which to improve wildfire probability modeling. Multisource data on weather, fuel, topography, infrastructure, and derived variables were used to model wildfire probability based on two machine learning methods—random forest (RF) and extreme gradient boosting (XGBoost). The results indicate that the accuracy of the wildfire probability models was improved by adding dynamic-step weather variables into the models. The variable importance analysis also verified the top contribution of these dynamic-step weather variables, indicating the effectiveness of the consideration of dynamic-step weather variables in wildfire probability modeling.