Predicton of wildfire in British Colombia

Abstract

Wildfires represent an escalating threat to ecosystems, infrastructure, and public safety, particularly in regions such as British Columbia (BC), Canada, where climate change has intensified both the frequency and severity of fire events. This thesis proposes a comprehensive, data-driven framework for predicting wildfire occurrences using advanced machine learning (ML) and deep learning (DL) techniques. It integrates five years of wildfire incident records from the Canadian Wildland Fire Information System (CWFIS) with ERA5 climate reanalysis data to create a geospatially and temporally aligned dataset tailored to the unique environmental conditions of BC. This research contributes to the field by constructing a high-resolution wildfire dataset, conducting a systematic evaluation of six feature selection techniques across multiple ML and DL models, and performing a rigorous analysis using diverse train-test splits and 10-fold cross- validation. The study also identifies and ranks the most influential environmental variables for fire occurrence. Among the evaluated models, ensemble methods—particularly CatBoost and Random Forest—demonstrated consistently superior performance across accuracy, recall, precision, F1-score, and ROC-AUC metrics. These results underscore the effectiveness of integrating robust modeling approaches with informed feature selection in developing reliable wildfire prediction systems for high-risk regions.