Advancing wildfire fuel mapping and burn severity assessment in Alaskan boreal forest using multi-sensor remote sensing

Abstract

Wildfires in Alaska have been increasing in frequency, size, and intensity putting a strain on communities across the state, especially remote communities lacking firefighting infrastructure to address large scale fire events. Advances in remote sensing techniques and data provide an opportunity to generate high quality map products that can better inform fire managers to allocate resources to areas of most risk and inform scientists how to predict and understand fire behavior. The overarching goal of this thesis is therefore to build insight into methods that can be applied to create highly detailed fire statistic map products in Alaska. To address this overarching goal we tested several methods for generating fire fuel, burn severity, and wildfire hazard maps that were validated using data collected in the field. Applying the Random Forest classifier on Airborne Visible/ Infrared Imaging Spectrometer Next-Generation (AVIRIS-NG) hyperspectral data we were able to produce a fire fuel map with an 81% accuracy. We then tested two supervised machine learning classifiers, post fire standard spectral indices, and differenced spectral indices for their performance in assessing burn severity. We found that supervised machine learning classifiers outperform other algorithms when there is an adequate amount of training data. Using the support vector machine and random forest classifiers we were able to generate burn severity maps with 83% accuracy at the 2019 Shovel Creek Fire. Lastly, we looked for a relationship between burn severity and environmental conditions prevalent during the Shovel Creek and Nugget Creek fires. Overall, these products can be used by fire managers and scientists to assess fire risk, limit the damages caused by wildfires through adequate resource allocation, and provide the guidelines for creating future high quality fire fuel maps.