Integrating machine learning for enhanced wildfire severity prediction: A study in the Upper Colorado River basin

Area of Science:

• Environmental Science
• Computer Science
• Ecology

Background:

• Wildfires present a substantial global threat, necessitating precise forecasting for effective mitigation strategies.
• Accurate prediction of wildfire severity is crucial for resource management and public safety.

Purpose of the Study:

• To forecast wildfire severity in the Upper Colorado River basin using machine learning.
• To identify key environmental and land-use indicators influencing wildfire severity.

Main Methods:

• Utilized machine learning techniques, including Random Forest and Artificial Neural Network, on historical data (1984-2019).
• Analyzed key indicators such as weather conditions (temperature, precipitation, solar radiation), vapor pressure deficit, NDVI, fuel moisture, SPEI, and evapotranspiration.
• Validated model performance against actual wildfire severities from 2016-2019.

Main Results:

• Random Forest model achieved 72% accuracy in predicting wildfire severity, outperforming Artificial Neural Network.
• Identified air temperature, vapor pressure deficit, NDVI, and fuel moisture as highly influential predictors.
• Other significant contributing factors included solar radiation, SPEI, precipitation, and evapotranspiration.
• Validation demonstrated a mean prediction error of 11.2%, confirming model reliability.

Conclusions:

• Machine learning models, particularly Random Forest, are effective tools for predicting wildfire severity.
• Understanding the influence of various environmental factors is key to improving wildfire management.
• The findings offer valuable insights for wildfire mitigation in vulnerable regions like the Upper Colorado River basin.