Event Navigation
Dissertation Information
Title: Machine Learning Approach to Forecast Human-Caused Wildfires at Actionable Scales Across the Western Us
Program: Doctor of Philosophy in Computing
Advisor: Dr. Mojtaba Sadegh, Civil Engineering
Committee Members: Dr. Matt Williamson, College of Innovation and Design and Dr. Michael Perlmutter, Mathematics
Abstract
Human activities are a major contributor to the escalating wildfire hazard in many regions. In the United States, for example, humans are responsible for igniting 84% of all fires. To effectively reduce the wildfire hazard, a deep understanding of the patterns, drivers, and growth factors of human-caused fires is essential. This knowledge can inform cost-effective mitigation strategies and limit the adverse impacts of these fires. This dissertation aims to lay the foundation for a decision-support system that maximizes the benefits of fire management and prevention by addressing knowledge gaps and limitations of previous modeling efforts. By employing advanced machine learning techniques, the proposed model will incorporate both physical and social characteristics of a given area. It will also capture the intricate relationships between various predictor variables that influence the occurrence and spread of human-caused fires. Multiple Machine Learning and Deep Learning models with various structures using the Keras and PyTorch architecture were trained based on 752,461 observed ignitions and 59M points without an ignition. To tune Machine
Learning models, GridSearchCV was used to find the hyper parameters that results in higher overall accuracy till the accuracy changes less than 0.01%. Each of these models were evaluated based on the test portion of data and the model with highest accuracy was selected for further investigation. Using an innovative sampling technique based on the physical understanding of wildfire ignition processes, my model displayed a robust behavior both for the cases that observed an ignition and those that di not, exceeding an overall accuracy of 98.88% (F1 = 96.75%, Kappa coefficient = 96.07%). Through this dissertation, I will further refine my models to predict human ignitions based on the 12 different ignition causes/sources recognized in Federal documents, and investigate their potential for growth. I therefore begin by outlining my motivation for this project as well as the findings of other researchers. Then, I go explain my contribution to improve findings. After that, I broaden the approach I intend to employ in order to achieve my research objectives. Lastly, I discuss my current findings and assess my outcomes.
