Advancements and Challenges in Volcanic Ash Detection and Classification Using Satellite UV/VIS Measurements and Deep Learning

Abstract

Volcanic eruptions can impact human populations and the environment, with volcanic ash decting air quality, agriculture, and aviation. Satellite observations for monitoring volcanic ash may face challenges, including di”culties in distinguishing ash from other aerosols, limited coverage, and sensitivity to weather conditions. This study explores a new method that uses deep learning to analyze Ultraviolet (UV) and visible (VIS) satellite data for detecting and classifying volcanic ash in the atmosphere. Neural network classifiers and parameter estimators were trained on simulated data from a radiative transfer model, covering diverse atmospheric conditions, and applied to spectral measurements from the TROPOspheric Monitoring Instrument (TROPOMI) aboard the Sentinel-5p satellite. The results suggest that volcanic ash can be distinguished from other aerosols and no aerosols with over 90% accuracy in the test dataset. In TROPOMI observations near eruption sources, volcanic ash was successfully identified, particularly during conditions of high ash concentration. Additionally, this study demonstrates the potential to retrieve size information from volcanic ash, particularly for broader particle sizes (1–10 µm), typical of volcanic ash in its initial stages. However, the approach faces limitations, such as reduced accuracy for low ash concentrations, challenges in distinguishing ash from aerosols with similar optical properties, dependencies on model assumptions and input parameters, and di”culties in adapting the simulation-based neural networks to real-world variability. Despite these limitations, this work o!ers insights and tools to retrieve volcanic ash and its microphysical properties that di!er from traditional methods.