Time Series Analysis and Machine Learning Methods to Predict Wildfire Spread in California
Discipline: Mathematics and Statistics
Session: 2
Room: 4 - Hanover F
Gwen Squires - Southeast Missouri State University
Co-Author(s): Daniel Breininger, Florida Institute of Technology, FL Dr. Nezamoddin N. Kachouie, Florida Institute of Technology, FL
Wildfires are natural phenomena that are essential to fire-adapted ecosystems around the world. However, climate change and expanding urbanization have made these events more threatening, adversely impacting the environment, public health and safety, infrastructure, and the economy. Wildfires in the state of California have displayed increasing severity in recent years, with 2020 being a record-breaking year for the state with 4 million acres burned [1], but their complexity makes them difficult to understand, predict, and mitigate. The purpose of this study is to determine the flammability of California’s landscapes by predicting whether fire will spread given an ignition occurs at a specified time and location. To achieve this goal, predictive models were constructed using climate factors, vegetation data, and geo-referenced wildfire records for the months from January 1992 to December 2020 in southern California. Due to the temporal nature of the data, time series analysis was implemented first using climate variables with daily resolution. Upon investigating the stationarity of these variables, yearly differencing was used to address the cyclical nature of the seasonal series. Immediate improvement was observed in the autocorrelation function plots. However, the comparative performances of simple AR models for the original and difference series were insubstantial. With no evident best model, we moved on to ADL modeling using the original and lag-1 series for all variables to predict natural fire occurrences. Unfortunately, the model did not produce any reliable predictions due to the rarity of the response. Thus, logistic regression was employed next to predict a thresholded indicator of fire spread rather than occurrence. This updated response reflects the capabilities of the dataset which does not contain ignition or response time variables. Logistic regression was better than the AR and ADL models at producing reliable predictions, but there was still room for improvement regarding model accuracy and sensitivity. The best-fit models confirmed expectations regarding the most important predictors which included precipitation, wind speed, and vapor pressure deficit. Overall, logistic regression outperformed time series analysis in terms of predictive ability, but that came at the cost of disregarding the temporal nature of the research problem. Future work will focus on accounting for both spatial and temporal factors, testing alternative machine learning models, and including more detailed vegetation data such as biomass measures and vegetation continuity.
References:
[1]. Office of Environmental Health Hazard Assessment [OEHHA]. (2022). Wildfires. In Indicators of Climate Change in California. California Environmental Protection Agency.
https://oehha.ca.gov/media/downloads/climate-change/document/04wildfires.pdf
Funder Acknowledgement(s): Funding was provided by an NSF REU Grant to Dr. Nezamoddin N. Kachouie, Florida Institute of Technology, FL.
Faculty Advisor: Dr. Nezamoddin N. Kachouie, nezamoddin@fit.edu
Role: I processed and cleaned the data, conducted the time series analysis, and fit the logistic regression models in R.