Assessing wildfire extents in Siberian forests using machine learning

Abstract

Abstract Wildfires significantly impact ecosystem dynamics and forest management strategies globally, including in Siberian forests. This study develops a machine learning (ML) framework to estimate wildfire size by integrating meteorological variables, forest composition, detection techniques, and historical fire records within the Krasnoyarsk Krai region of central Siberia. The dataset includes temperature, humidity, wind speed, precipitation, geospatial coordinates, and proximity to human settlements, which are used to train multiple predictive models, including XGBoost, Random Forest, K-Nearest Neighbors, Logistic Regression, and Decision Tree. XGBoost achieved the highest classification accuracy of 88.8%, outperforming other methods. Feature importance analysis highlights the influence of urban proximity, wind patterns, and meteorological conditions related to fuel moisture on fire size prediction. SHAP (SHapley Additive exPlanations) analysis indicates that smaller fires are associated with localized weather conditions, while extended dry periods correspond to larger fire events. While these results demonstrate the potential of ML for fire size classification in this specific region, the framework should be considered exploratory and region-specific. Future applications to other areas will require local data calibration.