Deep learning for hazard quantification from imagery

Fast and accurate quantification of a dangerous chemical release can reduce human exposure and environmental contamination. We demonstrate that deep learning models can provide quick and practical source term estimates from images of chemical releases, using a dataset compiled from volcanic plume observations. We nd that the best performing deep learning models are able to predict "large," "medium," or "small" SO₂ releases from unseen images of volcanic plumes with over 80% accuracy, and can be trained on imagery taken from the ground, air, or space. We test a range of model architectures, training strategies, and optimization approaches to determine what combination of properties produce a model that is robust and operates in the widest possible variety of situations. We evaluate how well the model generalizes to images of industrial SO₂ plumes and find that the best-performing volcanic plume model achieves 70% accuracy on images of industrial SO₂ plumes, demonstrating the potential of using deep learning for image-based hazard quantification.