Volcanic Forecasting with Intermittent Infrared Images via Deep Learning

Abstract:

Volcanoes are globally distributed and pose societal risks at multiple geographic scales, ranging from local dangers to international disruptions. Despite this, many do not have continuous ground monitoring networks, meaning that satellite observations may provide the only record of volcanic behavior. Among these observations, infrared (IR) imagery is monitored daily by several volcanic observatories to locate and quantify thermal anomalies. IR imagery captured under ideal conditions is valuable because it allows experts to examine the spatial distribution and extent of the thermal activity at great detail; however, scenes are often obstructed by clouds, meaning that probabilistic forecasts must be made off image sequences whose scenes are available intermittently through time.

These forecasts are often generated by statistical models and expert opinions - both of which struggle to utilize this data and capture the relationships within large quantities of nonlinear data. We draw on recent advances in deep learning to propose and evaluate models that generate probabilistic forecasts off intermittent image series by analyzing a sequence of raster data along with the time between scenes. We will present experiments with both simulated data and real volcanoes from multiple continents and discuss the applicability of these methods for broader volcanic observations and remote sensing in general.