Machine learning to predict final fire size at the time of ignition

Area of Science:

• Ecology
• Forestry
• Climate Science

Background:

• Boreal forest fires in Alaska are increasing due to climate warming.
• Understanding fire size at ignition is vital for effective management, especially with frequent ignitions.
• Current fire management strategies need enhanced predictive capabilities.

Purpose of the Study:

• To investigate the key factors controlling final fire size immediately after ignition.
• To develop a predictive model for classifying fire ignitions into size categories (small, medium, large).
• To assess the utility of this model for fire management and resource allocation in Alaska.

Main Methods:

• Utilized decision tree classification algorithms to predict fire size.
• Input variables included vapor pressure deficit and fraction of spruce cover near ignition.
• Compared decision trees with other machine learning models like random forests and multi-layer perceptrons.

Main Results:

• A decision tree model achieved 50.4% accuracy in classifying ignitions as small, medium, or large.
• Vapor pressure deficit and spruce cover were the most significant predictors.
• The model indicated 40% of ignitions would become large fires, contributing 75% of the total burned area.
• Simpler decision tree models outperformed more complex machine learning algorithms.

Conclusions:

• A straightforward classification system based on ignition-time variables can predict final boreal forest fire size.
• This predictive capability can inform optimal resource allocation for fire management.
• The approach supports efforts to maintain historical fire regimes and protect Alaskan ecosystems.