TO BURN OR NOT TO BURN? MACHINE LEARNING THE DRIVERS OF LARGE WILDFIRE OCCURRENCE IN SASKATCHEWAN'S BOREAL FOREST
Date
2021-03-04
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0000-0002-0664-8209
Type
Thesis
Degree Level
Masters
Abstract
Previous attempts to characterize the drivers of large wildfire occurrence have considered the probability that an ignition event will happen in addition to the probability that it will grow into a large wildfire. Including the very low probability of an ignition event happening at a given time and place can obscure the understanding of what governs the transition from ignition to sustained combustion. The purpose of my thesis was to investigate the biotic and abiotic factors and conditions driving large wildfire occurrences in the boreal forest of Saskatchewan. My objectives were to a) analyze historical wildfire occurrences and patterns of what burned and when; b) identify and select variables that have been used to explain wildfire behavior; c) determine which variables were best suited to explaining wildfire occurrences by building a model; and d) apply the model as a lens through which to view historical wildfire occurrences to understand what will drive future occurrences. To investigate objective a, I analyzed historical large (> 200 ha) wildfire burn footprints occurring between 2008 and 2018 in Saskatchewan’s boreal forest by associating each with fuel types and ecozone, and comparing area burned with area available of each respective fuel type and ecozone. This analysis showed that historical wildfire activity in the study area varied considerably by ecozone, fuel type, and time throughout the year. To investigate objectives b - d, I assembled a dataset of historical (1990 to 2018) wildfire ignition locations, associated them with weather and Fire Weather Index (FWI) component variables from the same day and location, and identified the Fire Behavior Prediction System (FPB) fuel type that each fire started in. I then built a random forest classifier using grid search hyperparameter tuning on the dataset to determine when conditions were suitable for a large wildfire (eventual fire size > 200 ha) to occur or not (eventual fire size < 0.5ha). Predicting unseen test data with the model resulted in a recall (a.k.a. sensitivity) of 0.788. The most important variables in the model were relative humidity and fuel type. Applying the model to average conditions in historical high and low wildfire activity months showed that suitable conditions for large wildfires can vary greatly within and among years, and among fuel types. Applying the model to historical daily weather observations demonstrated that different fuel types have different threshold conditions for large wildfires, and the threshold conditions for some fuel types are observed more frequently than for others. This study helps advance our understanding of wildfire activity, and it can be used to inform wildfire management decisions to reduce the risk wildfires pose.
Description
Keywords
Wildfire, occurrence, prediction, drivers, random forest, large fire, boreal forest, weather, fire weather, abiotic and biotic regulation, sustained ignition probability
Citation
Degree
Master of Science (M.Sc.)
Department
Geography and Planning
Program
Geography