Exploring Multi-Model Ensembling Methods to Improve on Deep Learning Based Smoke Detection

With increasing occurrences of wildfires in recent years, effective and fast-delivering wildfire and smoke detection tools are needed to support hazard management and reduce risks to human health and environmental resources. The NOAA Geostationary Operational Environmental Satellites (GOES) provide high spatial and temporal resolution imagery of North America that can be used to detect the presence and density of smoke plumes. Recent advancements in computer vision have shown the capabilities of deep learning models to automate the semantic segmentation of high resolution images by training encoder-decoder networks on large labeled datasets. Additionally, ensemble methods can improve model generalization and prediction accuracy by combining the results of multiple models. We present an ensemble of deep learning models that produces representative predictions of where wildfire smoke plumes exist and their relative density (i.e. low, medium, high) in GOES imagery. Our results demonstrate that ensemble techniques can improve performance compared to using a single model. This multi-model data-driven ensemble is expected to support fire and hazard management by being able to automate the monitoring of smoke in real-time from satellite imagery.