A comprehensive observational database of deformation at global volcanoes for machine learning applications

A key indicator of potential and ongoing volcanic activity is deformation of a volcano's surface due to magma migrating beneath. The European Sentinel-1 radar archive now contains a large number of examples of volcano deformation, yet the vast majority of subaerial volcanoes are not well monitored. We therefore aim to systematically extract all deformation signals at volcanoes globally, including smaller scale signals associated with processes such as landslides and local changes in hydrothermal systems, which can provide a basis for machine learning approaches to automatically classify and potentially forecast deformation.

We have developed an approach to automatically derive high-resolution displacement time series centred on each volcano. To avoid the loss of decorrelated signal in areas of glacial coverage, winter snow and heavy vegetation, we build a highly redundant small baseline network of interferograms tailored to each volcano using coherence tests. We implement an improved phase unwrapping algorithm that separately unwraps signals at different spatial scales, to improve results in decorrelating areas. To mitigate the effect of phase propagation through the atmosphere, we provide multiple atmospheric correction methods, including a spatially-varying scaling method that uses interferometric phase to refine the interpolation of a weather model in time and space.

The processed products, stored in a database with annotated metadata (VolcNet), are available for the further interpretation. We show here the volcanic unrest at a large number of volcanoes taken from the database, detected using a machine learning algorithm LiCSAlert. We also show a statistical analysis based on the processed time series for the assessment of volcanic risk.