Abstract
Predicting wildfires using Machine Learning (ML) models is relevant and essential to minimize wildfire threats to protect human lives and reduce significant property damages. Mixed results have been found in this domain, potentially because of dataset manipulations to enable multi-class classification. This is because two or more classes are used in wildfire prediction modelling, where non-fire labels are created artificially, leading to an unbiased dataset for non-fire data. This thesis aims to discuss research that built wildfire prediction models using One-class classification algorithms. The significant features that influence wildfire ignition were derived from One-class ML models using the Shapley values method which is a novel contribution to the wildfire prediction domain. Elevation, vapour pressure deficit and dew point temperature were among the most influential features that were derived using the Shapley values method. The One-class algorithms used were Support Vector Machine, Isolation Forest, Neural network-based Autoencoder and the Variational Autoencoder models. The input features to the models were grouped based on topography, weather, plant fuel moisture, and population. Outcomes were validated using 5-fold cross-validation to avoid bias in the training and testing dataset selection on the ML models’ performances. These One-class models resulted in a high mean accuracy ranging from 98-99%, exceeding multi-class models’ performances in similar environmental conditions. The findings of the research have the potential to influence the state-of-the-art methods in wildfire prediction. Finally, a web-based tool to predict wildfires is presented as a proof of concept to show the usability of ML models for wildfire predictions.