Predicting Large Wildfires in Southern California with Advanced Machine Learning Techniques
Discipline: Data Science
Session: 1
Room: 5 - Embassy A
Daniel Breininger - Florida Institute of Technology
Co-Author(s): Gwen Squires, Southeastern Missouri State University, Cape Girardeau Missouri. Nezamoddin N. Kachouie, Florida Institute of Technology, Melbourne Florida
Climate change has resulted in wildfires becoming more destructive across the globe (Caldararo 2002). Similarly, more direct human impacts have altered fire regimes from starting and suppressing fires and developing natural land (Bowman et al. 2011). These impacts compound, creating areas prone to uncontrollable fires where in just the year 2020 over $12 billion worth of damage was caused and 33 lives lost to wildfires (California Wildfires & Fire Resilience Task Force, 2022). With so many random sources that can ignite a fire, modeling the regime has proven to be very difficult. The unpredictability of wildfires and their damage has also led to California having an unstable insurance market and has created a need to have a better understanding of the current and future regime.
To model the regime advanced methods must be performed to get at the complexity of the relationships between the predictors and fire. Neural networks have become popular in literature with the abundance of data and powerful computers which are required to get accurate results. Long short-term memory neural networks (LSTM) have grown popular because of their ability to remember both short and long-term impacts of predictors on sequential data. In this study a 10,000 square km extent in Southern California was split into 625 cells where a location embedded LSTM was fit to predict if a cell had a monthly burn area of at least one acre or not.
The data collected included the number of fires, the total fire area, the dominant habitat type within a cell, and 13 climate variables including the max temperature and the amount of precipitation from 1992-2020. Multiple data transformation techniques were applied to the data to find the best representation for each variable to be used in the model. The data was split into an 80% training and 20% testing set. Balanced accuracy, the average sensitivity of each class, was used as the performance metric because of the imbalanced data. The balanced accuracy on the testing set was 88.94% and associated fire with habitats experiencing sudden wet and windy from previously hot and dry conditions. This study is a good foundation for people to use and prepare for a fire. Future work includes adding human factors such as roads and population which can give the model better insights on locations where an unnatural fire may start.
References:
Bowman, D. M., Balch, J., Artaxo, P., Bond, W. J., Cochrane, M. A., D’antonio, C. M., Swetnam, T. W. (2011). The human dimension of fire regimes on Earth. Journal of biogeography, 38(12), 2223-2236.
Caldararo, N. (2002). Human ecological intervention and the role of forest fires in human ecology. Science of the Total Environment, 292(3), 141-165.
California Wildfire & Forest Resilience Task Force. (2022). California’s strategic plan for expanding the use of beneficial fire. https://wildfiretaskforce.org/wp-content/uploads/2022/05/californias-strategic-plan-for-expanding-the-use-of-beneficial-fire.pdf
Funder Acknowledgement(s): Funding was provided by an NSF REU Grant to Dr. Nezamoddin N. Kachouie, Florida Institute of Technology
Faculty Advisor: Nezamoddin N. Kachouie, nezamoddin@fit.edu
Role: I collected/processed the data and walked the undergraduate student through the modeling process