Postdoctoral Fellowship: EAR-PF: Exploring the significance of the spatial field from integrated DInSAR+GNSS time series for machine learning and volcano early warning applications

Leading up to volcanic eruptions, the ground deforms in response to the movement of magma or hydrothermal fluids at depth. These surface motion patterns are recorded using high-resolution satellite remote sensing techniques such as Differential Interferometric Synthetic Aperture Radar (DInSAR), Digital Elevation Models (DEMs), and Global Navigation Satellite Systems (GNSS).

Advanced computational models can then harness this data to determine volcanic source parameters and eruptive thresholds, track magmatic transport behavior, and create additional synthetic time series data for training machine learning algorithms. Furthermore, DInSAR, DEMs, and GNSS may be integrated to form a three-dimensional (east, north, up) time series with enhanced ground motion measurements.

Integrated results are delivered either as a plotted time series at a single pixel location or as deformation maps over large regions. The plotted time series only contain a temporal component, while the deformation maps contain both temporal and spatial elements. Machine learning algorithms will first be trained using only the time-sensitive input, then compared to algorithms trained with integrated deformation maps consisting of both spatial and temporal properties.

Analyzing the effects of spatial information within the machine learning algorithm’s training data will lead to essential awareness of crustal to subsurface dynamics, help classify each stage of a volcanic eruption, more accurately estimate locations and geometries of magmatic storage reservoirs or transport pathways, and better characterize or forecast environmental changes through time. This project contributes towards NSF’s mission to promote the progress of science, prosperity, and welfare.

Societally relevant outcomes will include but are not limited to the development of enhanced processing frameworks for hazard early warning and volcanic research, the potential to save human life or protect community infrastructure due to natural hazards, and to increase public awareness and knowledge of scientific methods.

The greatest challenges in the field of volcanic observation research involve how to prepare for, or when and where to anticipate eruptive or high-magnitude events. To better understand magmatic structure, eruptive behavior, and to improve hazard early warning systems, this project will support the development of a standard workflow in which machine learning approaches are used to model geodetic deformation data.

Multi-band Synthetic Aperture Radar (SAR), Global Navigation Satellite Systems (GNSS), and high-resolution Digital Elevation Model (DEM) data collected over eruptive volcanoes in Iceland, Hawaii, and the Canary Islands will be integrated to generate novel, high-resolution, three-dimensional (east, north, up) time series containing surface deformation measurements with improved precision. This project will provide researchers with free, complex geodetic products and processing routines, and will expand on existing volcanic models by merging various source geometries and locations to better constrain the physical parameters unique to each volcanic system.

Advanced numerical and physical models from collaborators at the USGS Volcano Observatories, University of Iceland, and the Spanish National Research Council (IPNA-CSIC and IGEO-CSIC), will use the satellite observations to quantify volcanic composition, resolve magmatic transport behavior over local to regional scales, invert for subsurface structures such as pressure bodies or sources of dislocation, and generate synthetic training data. Machine learning algorithms will be trained to detect anomalous motions, or gradients, between remote sensing images and time series for volcanic monitoring and forecasting applications.

Plotted time series will be streamed through Long-Term-Short-Memory (LSTM) algorithms to predict the next, most-likely position of a ground point. Convolutional Neural Network (CNN) image classification algorithms will be trained using 3D, high-resolution, cumulative ground deformation maps, which involve dual spatiotemporal components. These methods will determine how interacting surface signals may be used to evaluate volcanic unrest, how topographic change from neighboring pixels affects the ways in which machine learning algorithms consume or process knowledge, and how efficient and reliable they are at forecasting various phases of an eruption.

Ultimately, the end goal of this research is to build a scalable system capable of ingesting datasets from disparate sources and domains (i.e., seismic, gas emission, surface temperature, tide gauge, tiltmeters, thermal, etc.), and recognizing patterns across all signals to alert scientists to major natural hazard events. Doing so will advance geodetic technology and processing methods, support scientific analyses regarding the onset of hazardous events, and contribute towards eruption early warning for the safety of nearby communities.

This project is co-funded by the Directorate for Geosciences to support AI/ML advancement in the geosciences.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.