A deep learning framework for rapid inversion of ground deformation to model volcanic sources

Rapid detection of the locations and movements of magma within the crust is essential for tracking volcanic unrests. The pressure exerted on the Earth's crust by magma migration causes ground deformation that can be measured by a variety of geodetic instruments. Consequently, the inversion of deformation signals allows the geometry and the position of the magmatic source to be inferred.

In the field of volcanic monitoring, the high temporal resolution of continuous Global Navigation Satellite System (GNSS) measurements makes them widely used for near real-time applications. However, traditional inversion techniques are usually time-consuming, model dependent, and often require a dense, well-distributed GNSS network, which is available only in a few volcanoes worldwide.

To overcome these challenges, machine learning provides efficient tools for emulating direct deformation models, accelerating the inversion process while modelling sources with complex geometries. Taking advantage of generalization capabilities of deep learning algorithms, we present a station-independent deep learning-based inversion framework that can instantly reconstruct underground magmatic causative sources from as few as ten GNSS stations without any prior knowledge of the station configuration or the target volcano.

Trained and tested on hundreds of synthetic deformation patterns, the deep learning-based inversion proves its potential and robustness in the retrospective application to the May 2008 eruption of Mount Etna as well as to Iceland's intrusive sequence between December 2023 and August 2024.