Abstract Title

A Machine Learning Approach to Predicting Wildfire Shape

Additional Funding Sources

The research described was supported by Boise High School.

Abstract

In recent decades, wildfires in the United States have increased in frequency and intensity. As global temperatures continue to rise as a result of climate change, wildfires are projected to burn more acres of land each season. The ability to predict the burned area of wildfires can minimize the destructive impact of wildfires on communities. Research efforts have aimed to use several Artificial Intelligence techniques, such as reinforcement learning, neural networks, naïve bayes, and random forests, to simulate wildfire growth. The goal of this study is to examine the latter method, random forests, by utilizing a classifier to predict the shape, or burned area, of a wildfire given the fire origin’s geographic coordinates. The method used in this study analyzes high resolution satellite imagery, and considers each image as a 2-dimensional matrix of pixels. By combining and stacking multiple spectral bands, numerical values that represent environmental variables are obtained. For each element in the matrix, values are assigned representing temperature, vegetation density, elevation, slope, aspect, and wind direction. These characteristics are fed to the root nodes of individual decision trees within a random forest ensemble; the output determines if each pixel is burned or not. The classifier was trained with data from the Carr Fire in July/August 2018 near Redding, California. Experimental results yielded a moderately high prediction accuracy (89.33%). This study indicates that random forest classification is a promising method of predicting wildfire shape, and utilizing this method can have beneficial implications in wildfire control and prevention.

Comments

W103

This document is currently not available here.

Share

COinS
 

A Machine Learning Approach to Predicting Wildfire Shape

In recent decades, wildfires in the United States have increased in frequency and intensity. As global temperatures continue to rise as a result of climate change, wildfires are projected to burn more acres of land each season. The ability to predict the burned area of wildfires can minimize the destructive impact of wildfires on communities. Research efforts have aimed to use several Artificial Intelligence techniques, such as reinforcement learning, neural networks, naïve bayes, and random forests, to simulate wildfire growth. The goal of this study is to examine the latter method, random forests, by utilizing a classifier to predict the shape, or burned area, of a wildfire given the fire origin’s geographic coordinates. The method used in this study analyzes high resolution satellite imagery, and considers each image as a 2-dimensional matrix of pixels. By combining and stacking multiple spectral bands, numerical values that represent environmental variables are obtained. For each element in the matrix, values are assigned representing temperature, vegetation density, elevation, slope, aspect, and wind direction. These characteristics are fed to the root nodes of individual decision trees within a random forest ensemble; the output determines if each pixel is burned or not. The classifier was trained with data from the Carr Fire in July/August 2018 near Redding, California. Experimental results yielded a moderately high prediction accuracy (89.33%). This study indicates that random forest classification is a promising method of predicting wildfire shape, and utilizing this method can have beneficial implications in wildfire control and prevention.