Abstract

Wildfires are one of the most dangerous natural disasters that strike California and the United States, destroying homes and forest resources. Prescribed burns are a critical technique for wildfire mitigation. However, prescribed burns hold the risk of going out of control and starting their own wildfires. 

In this project, machine learning models (ANN, Random Forest, SVM, KNN) are developed to predict the risk and behavior of prescribed burns. The models utilize environmental and geographic data to predict wildfire sizes.
Wildfire data from the California Fire Perimeter database is provided by CalFire. The analysis focuses on the period between 2000 and 2023. The weather data is gathered from Open Meteo. Vegetation data is computed from LandSat Databases in the form of NDVI and NDMI. Multiple machine learning models were developed and tested to identify better prediction techniques of fire area for the prescribed burns.

Two existing wildfire simulators were used to model a selected set of fires. The predicted fire sizes from the simulators are compared with the actual data for testing. 

The experimental results show that the classifier models reached accuracies of ~70-75% with binary classification, with ANN performing best. The burn areas estimated by simulators were significantly underestimating, likely due to mapping limitations for input variables and output range. 

This project shows that machine-learning models can be an effective tool to manage prescribed burns. Further work is needed to expand the models with historical climate and geographical data to improve accuracy and explore other metrics for fire behavior.