Modeling Naturally Occurring Wildfires Across the US Using Niche Modeling

Author

Brandon Polk, Southern Illinois University Carbondale

Date of Award

5-2017

Degree Name

Master of Science

Department

Geography and Environmental Resources

First Advisor

Schoof, Justin

Second Advisor

Wang, Guangxing

Third Advisor

Remo, Jonathan

Abstract

Wildfires can cause significant damage to an area by destroying forested and agricultural areas, homes, businesses, and leading to the potential loss of life. Climate change may further increase the frequency of wildfires. Thus, developing a quick, simple, and accurate method for identifying key drivers that cause wildfires and modeling and predicting their occurrence becomes very important and urgent. Various modeling methods have been developed and applied for this purpose. The objective of this study was to identify key drivers and search for an appropriate method for modeling and predicting natural wildfire occurrence for the United States. In this thesis, various vegetation, topographic and climate variables were examined and key drivers were identified based on their spatial distributions and using their correlations with natural wildfire occurrence. Five models including General Linearized Models (GLM) with Binomial and Poisson distribution, MaxEnt, Random Forests, Artificial Neural Networks, and Multiple Adaptive Regression Splines, were compared to predict natural wildfire occurring for seven different climate regions across the United States. The comparisons were conducted using three datasets including LANDFIRE consisting of thirteen variables including characteristics of vegetation, topography and disturbance, BIOCLIM containing climate variables such as temperature and precipitation, and composite data that combine the most important variables from LANDFIRE and BIOCLIM after the multicollinearity test of the variables done using variance inflation factor (VIF).

This results of this study showed that niche modeling techniques such as MaxEnt, GLM with logistic regression (LR), and binomial distribution were an appropriate choice for modeling natural wildfire occurrence. MaxEnt provided highly accurate predictions of natural wildfire occurrence for most of seven different climate regions across the United States. This implied that MaxEnt offered a powerful solution for modeling natural wildfire occurrence for complex and highly specialized systems. This study also showed that although MaxEnt and GLM were quite similar, both models produced very different spatial distributions of probability for natural wildfire occurrence in some regions. Moreover, it was found that natural wildfire occurrence in the western regions was more influenced by precipitation and drought conditions while in the eastern regions the natural wildfire occurrence was more affected by extreme temperature.