Semantic segmentation of explosive volcanic plumes through deep learning

Tracking explosive volcanic phenomena can provide important information for hazard monitoring and volcano research. Perhaps the simplest forms of monitoring instruments are visible-wavelength cameras, which are routinely deployed on volcanoes around the globe. Here, we present the development of deep learning models, based on convolutional neural networks (CNNs), to perform semantic segmentation of explosive volcanic plumes on visible imagery, therefore classifying each pixel of an image as either explosive plume or not explosive plume. We have developed 3 models, each with average validation accuracies of >97% under 10-fold cross-validation; although we do highlight that, due to the limited training and validation dataset, this value is likely an over-estimate of real-world performance. We then present model deployment for automated retrieval of plume height, rise speed and propagation direction, all parameters which can have great utility particularly in ash dispersion modelling and associated aviation hazard identification. The 3 trained models are freely available for download at https://doi.org/10.15131/shef.data.17061509.