Improving Satellite-based Wildfire Smoke Plume Detection with Deep Ensembles

Annabel Wade, Rey Koki, Christina Kumler-Bonfanti, Dale Durran

Abstract: With increasing frequency and severity of wildfires, there is an urgent need for wildfire and smoke detection tools that can effectively and rapidly monitor smoke at a large scale. Recent advancements in computer vision have demonstrated the potential of machine learning to automatically label regions of high-resolution images with high accuracy. However, single-model approaches can struggle with generalization and accuracy in diverse conditions, which is necessary for operational smoke detection. To address these challenges, we propose using an ensemble of deep learning models to produce more accurate annotations of wildfire smoke plumes and their relative density (light, medium, heavy) in satellite imagery. Our results indicate that deep ensemble techniques improve performance compared to using a single model. This approach aims to provide a more reliable satellite-based tool for real-time smoke monitoring, thereby aiding fire and hazard management efforts and improving the modeling of wildfire behavior and air quality.

Submission Number: 4