Abstract
Social media platforms have become important sources of information during disaster events, and prior crisis informatics research has demonstrated their value for situational awareness, emergency communication, and understanding public responses to hazards such as wildfires. Despite these advances, scalable approaches are still needed to systematically characterize large-scale public discourse and broader societal concerns related to wildfires over time. This study presents a hybrid machine learning framework that combines BERTopic with K-means clustering, thematic analysis, and zero-shot text classification to analyze wildfire-related discussions on Reddit. We examined 214,595 comments collected from Canada- and USA-based subreddits between 2020 and 2024. The discovered topic terms from the topic model were consolidated into human-interpretable themes through thematic analysis and subsequently used as candidate labels in a majority voting ensemble of five zero-shot text classification models to classify Reddit comments. Across the two countries, the analysis identified major themes including climate change, fire bans, pollution, forestry management, governance, wildfire alerts and evacuation, employment of firefighters, hospitality impacts, and insurance-related concerns. The findings contribute to understanding long-term public discourse surrounding wildfires and demonstrate the utility of combining topic modeling with zero-shot text classification for large-scale social media analysis.