Assessing wildfire extents in Siberian forests using machine learning.

Area of Science:

• Forestry
• Ecology
• Computer Science

Background:

• Wildfires pose significant threats to global ecosystems and forest management.
• Siberian forests, particularly in Krasnoyarsk Krai, are susceptible to large-scale fire events.
• Accurate wildfire size estimation is crucial for effective management and ecological impact assessment.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) framework for estimating wildfire size in the Krasnoyarsk Krai region of Siberia.
• To identify key environmental and meteorological factors influencing wildfire size.
• To compare the performance of various ML models for wildfire size classification.

Main Methods:

• Utilized a dataset integrating meteorological variables (temperature, humidity, wind, precipitation), forest composition, detection data, and historical fire records.
• Trained and compared multiple ML models: XGBoost, Random Forest, K-Nearest Neighbors, Logistic Regression, and Decision Tree.
• Employed SHAP (SHapley Additive exPlanations) analysis to interpret model predictions and feature importance.

Main Results:

• XGBoost demonstrated the highest classification accuracy at 88.8%, outperforming other tested ML models.
• Feature importance analysis revealed that proximity to urban areas, wind patterns, and meteorological conditions affecting fuel moisture significantly influence fire size.
• SHAP analysis indicated that localized weather conditions correlate with smaller fires, whereas prolonged dry periods are associated with larger fire events.

Conclusions:

• The developed ML framework shows strong potential for wildfire size classification in the studied Siberian region.
• The findings highlight the critical role of meteorological factors and human proximity in determining wildfire extent.
• The framework is currently exploratory and region-specific, necessitating local data calibration for broader applications.