Using Machine Learning to Accurately Predict Wildfire Danger
By Gurik Mangat
Intermediate Category (Grades 9-10)
Innovation | Big Data / AI, Environment
In the field of fire dynamics, there are a modest number of studies on wildfire analysis, but a lack of implemented modern computational methods which can produce accurate fire prediction models. The need for accurate wildfire prediction models is burgeoning as the substantial emissions of greenhouse gasses accelerates the rate of climate change. As a result of this, the severity and frequency of wildfires are drastically exacerbated year over year.
In regards to this growing threat, our project aims to make the danger prediction of wildfires universally faster and more accurate by deploying a neural network machine-learning algorithm integrated into an application interface intended for use by firefighters in the field. The procedure of this project was split into five main categories, (1) developing a methodology for which fire related conditions will be considered using the Canadian Forest Fire Weather Index System, (2) constructing a dataset using data from the Landsat-8 and Sentinel satellites based on current wildfire data including the aforementioned conditions, (3) utilising a K-means clustering algorithm to quantify fuel moisture in satellite imagery and inputting that into the dataset, (4) training a binary classification neural network on the previously completed dataset, and (5) integrating this trained machine learning model into an easy-to-use interface for firefighters to use on the field. This project has successfully predicted the presence of dangerous wildfire environments in satellite/environmental data at an accuracy rate of over 96%. Moreover, it should also be noted that during dataset construction, this project accounted for regional data bias by including data in the dataset from a multitude of different regions and terrains from across the United States and Canada. This is a distinct and crucial advantage over other wildfire prediction models as it exhibits the ability to dynamically generate accurate predictions even in foreign regions or conditions. By giving firefighters the ability to generate accurate predictions out on the field, we aim to make the process of firefighting much safer, easier, and drastically more efficient.
Keywords: machine learning, wildfire prediction, neural networks, k-means clustering, convolutional neural networks, algorithms, data-bias, climate change