Conference Agenda

About 45 volcanoes are erupting every day worldwide, cumulating in ~16,500 eruption days per year. Reaching a daily coverage of all erupting volcanoes with high-resolution Synthetic Aperture Radar (SAR) imagery therefore requires pooling data from national, international space agencies, as well as commercial companies, effectively forming a global virtual constellation. To deal with the diversity of the images acquired by various satellites, it is necessary to develop new methods for automatically extracting quantitative information (ground deformation; extent and volume of topographic changes) from any SAR amplitude image no matter its band and its geometry.

Here, we present a novel method for (i) measuring ground deformation, (ii) mapping topographic changes and (iii) reconstructing them in 3D from a sparse set of multi-angular SAR amplitude images acquired during an eruption.

The forward problem consists in simulating synthetic reference images (with geometries similar to those of the syn-eruptive SAR images) using a pre-eruptive Digital Elevation Model (DEM) and a radiometric terrain model. (i) For each pair of synthetic and real SAR image, ground deformation is measured along slant range and azimuth by image correlation. Combining the different geometries allows to retrieve the three components of the displacement field. (ii) Besides, low correlation scores are used to map topographic changes such as lava flows. We apply the method to a dataset of six multi-angular Capella Space images of the Piton de la Fournaise volcano and validate the results against ground-truth data (GNSS and lava flow maps) from the OVPF-IPGP observatory.

The inverse problem consists in iteratively modifying the pre-eruptive DEM to make simulated synthetic images match the real syn-eruptive SAR images, finally producing an updated DEM containing the new structures that formed during the eruption (iii). The optimization of the DEM is done using Radar Fields, an inverse rendering approach extending neural radiance fields to SAR imagery. We evaluate the performances of the method by artificially removing the lava dome of La Soufrière de Guadeloupe from a ground-truth LiDAR DEM and reconstructing it from two TerraSAR-X Spotlight images. Then, we apply the method to track the lava dome growth of the April 2023 eruption at Shiveluch, Russia, from Sentinel-1 and Capella Space images. This example demonstrates the multi-sensor and multi-angular capabilities of Radar Fields, which will allow to produce daily to weekly DEM time series during eruptions and enable a better understanding of short-term eruptive processes.

Piton de la Fournaise is one of the most active basaltic volcanoes worldwide and provides an outstanding natural laboratory for testing advanced geodetic imaging strategies. Conventional InSAR time-series analyses generally combine ascending and descending line-of-sight (LOS) observations to retrieve quasi-2D surface displacement fields, typically resolving east–west and vertical components while assuming negligible north–south motion. While this approach is widely used, such simplifications may lead to significant biases in complex volcanic regions where three-dimensional displacement is substantial. To address this limitation, we present a methodological framework to derive a true 3D InSAR time series using an exceptional large multi-geometry dataset from ALOS-2 L-band SAR data acquired between April 2021 and December 2025, and we assess its performance relative to conventional quasi-2D approaches and independent GNSS measurements.

Our dataset comprises more than 1,600 ALOS-2 images collected over 22 distinct LOS including right-looking and left-looking geometries as well as ascending and descending orbits. The nearly daily acquisitions across all LOS over the 4.5-year study period ensures dense temporal sampling. The L-band wavelength and high spatial resolution of SpotLight acquisitions ensures high coherence even over vegetated areas allowing dense spatial sampling across the whole volcanic edifice. At last, the unprecedented diversity of viewing configurations provides the geometric redundancy required to invert the full 3D displacement field (east, north, and vertical components) without imposing assumptions on motion direction.

The InSAR mass processing was performed using the AMSTer Toolbox, generating over 7,700 interferograms. Following phase unwrapping with SNAPHU, all deformation maps were geocoded onto a common UTM grid, allowing the integration of all 22 acquisition LOS for the 3D MSBAS inversion. We show that LOS measurements at grazing incidence (>55°) are the most sensitive to atmospheric artifacts. To reduce stratified atmospheric artifacts and improve the reliability of the time series solutions, we implemented the MANGO toolbox into the automatic AMSTer processing workflow to compute and apply atmospheric corrections based on GNSS-derived Zenith Total Delay (ZTD) from continuous GNSS observations provided by the OVPF monitoring network.

To better isolate subtle deformation signals, we compute two distinct 3D time series. The first corresponds to a “raw” solution, in which the MSBAS framework ingests the complete set of differential interferograms previously computed. The second is constructed after removing co-intrusive displacements from all interferograms. Specifically, displacement fields associated with dyke intrusions are estimated and reprojected into the LOS geometry of each interferogram, and the corresponding co-intrusive phase contribution is subtracted prior to time-series inversion. This strategy prevents the large-amplitude, sudden displacements related to dyke propagation from dominating the solution and hightlight lower-amplitude deformation signals.

The corrected 3D time series allows a more detailed analysis of pre-, post-, and inter-eruptive deformation that is otherwise obscured by large co-intrusive signals. By separating dyke-related displacements from other contributions, we better resolve long-term deformation patterns, distinguishing intrusive magma-driven movements from gravitational, tectonic, or relaxation processes. In particular, the analysis reveals that, at first order, the deformation field is dominated by eastward motion of the eastern flank of the volcano. Significant subsidence is also observed over the most recent lava flows, consistent with progressive compaction and cooling. Outside the Enclos Fouqué caldera, several gravity-driven slope instabilities are detected, the most prominent being located at the head scarp of the Rivière de l’Est. Although of smaller amplitude, additional, less well-documented deformation signals are identified within the northern and southern summit rift zones, affecting the summit edifice. Furthermore, transient post-intrusive displacements are observed along the eastern flank following dyke emplacement episodes. A change in deformation trend is also detected in early 2024, coincident with the pause in the eruptive cycle following the July–August 2023 eruption, which marks an approximately 2.5-year period without eruptive activity.

The comparison between the raw and corrected solutions demonstrates the strong impact of episodic dyke intrusions on time-series stability and highlights the necessity of isolating transient intrusive events when investigating subtle, longer-term deformation processes in highly active volcanic systems.

A comparison between the reconstructed full 3D displacement fields, conventional quasi-2D east–west/vertical solutions, and independent GNSS observations further demonstrates the added value of the complete 3D inversion. The 3D approach improves consistency with GNSS measurements, particularly by capturing north–south motions that are systematically underestimated by classical methods. We show that when the north–south component is neglected, approximately 10% of its signal energy leaks into the vertical component, leading to biased vertical displacement estimates.

Overall, our results show that combining multi-geometry ALOS-2 acquisitions, co-intrusive signal mitigation, and GNSS-based atmospheric corrections provides a robust framework for resolving complex 3D volcanic deformation at Piton de la Fournaise. This methodology not only clarifies the dynamics of dyke intrusions but also improves the understanding of long-term, volcano-wide deformation, offering key insights into the evolution of rift zones, flank instability, and gravitational processes in the lower slopes. Importantly, the unprecedented combination of high spatial and temporal resolution, enabled by the agile integration of right- and left-looking, ascending and descending acquisitions, provides a powerful monitoring capability, allowing near-real-time tracking of subtle and rapid deformation signals across the entire edifice.

The magmatic unrest at Mount Agung preceding the November 2017 eruption was identified by previous InSAR studies in September 2017, but the precursory signal had begun much earlier, masked by a concurrent tectonic signal. Applying Variational Bayesian Independent Component Analysis to the full Interferometric Synthetic Aperture Radar (InSAR) time series, we separate the magmatic deformation from atmospheric noise and regional tectonic subsidence driven by interseismic loading on the Flores back-arc thrust, and reveal that magmatic inflation began in May 2017, six months before the eruption. The separated signals are independently verified by Global Navigation Satellite Systems displacement and seismicity patterns. Signal separation of tectonic and magmatic deformation, demonstrated retrospectively at Mount Agung, provides a framework for early precursor detection at tectonically complex arc volcanoes.

On August 2, 2025 (16:38 UTC), Krasheninnikov volcano began erupting within four days of the great M8.8 Kamchatka earthquake (on July 29, 23:24 UTC) and more than 400 years since its last eruption [Girina et al., 2025]. The eruption is on-going, mostly effusive, as of this writing. Krasheninnikov is comprised of double strato-volcanoes emerging from an older caldera, with the younger northern cone featuring nested craters at its summit, the location of the 2025 eruption. Here we present an analysis of the eruption deformation sources, preliminary estimation of lava flow volume as of late September, and whether the eruption was triggered by the great M8.8 Kamchatka earthquake, by looking at time series of interferometric synthetic aperture radar (InSAR) and diffuse thermal anomalies from NASA MODIS data. Soon after the M8.8 earthquake the Japanese Aerospace Exploration Agency (JAXA) acquired ALOS-2 SAR data over the Kamchatka peninsula spanning most of the peninsula, including Krasheninnikov located approximately 200 km north of the northern extent of the earthquake rupture zone. The European Space Agency (ESA) managed Sentinel-1 (S1) and ALOS-2 InSAR observations closely bracket the eruption with S1 observations on July 30, 31, August 1, 2, 3, 5, 6, 7, from a mix of ascending and descending data. Deformation is not evident on S1 data through July 31. InSAR observations on August 1, 2, and 3 show upward migration of deformation, with the data from August 2 descending track showing clear buried dike patterns over the summit hours before the eruption start while the August 3 ascending track shows stronger deformation but with loss of coherence at the summit, likely due to eruption deposits. S1 observations from August 5-7 reinforce the August 3 observations.

We model the InSAR observations through a combination of Bayesian inference using a Markov chain Monte Carlo (MCMC) estimation of dike source parameters for combinations of interferograms during the first week of the eruption. The August 1 and 2 interferograms constrain a buried, SW dipping dike, striking from the north craters to the NW beneath the flank of the northern cone, and a nearly N-S vertical dike bisecting the summit of the northern cone. The summit dike geometry is consistent with field observations [Gorbach et al., 2026]. We use the basic geometries of each dike from the August 2 InSAR to model the distributed opening of each dike through a least-squares inversion with opening regularization (smoothing), computing Greens functions for each dike patch using the boundary element code POLY3D to include the effects of topography following Lundgren et al. [2015]. We find that initial opening of both dikes on August 2 remained mostly buried on the SW dipping dike but increased in magnitude with time beneath the summit where lava effusion vents were concentrated. Intruded dike volumes ranged from 20-35 million cubic meters, increasing progressively from August 1-3 (see figure below). Models constrained by observations on August 1 and 7 from S1 Path 162 show that the SW dipping dike was mostly active prior to the eruption with the summit N-S dike opening focused directly beneath the northern cone craters once the eruption began.

InSAR time series from S1 data show no precursory inflation in the 10 years prior the eruption. Analysis of MODIS radiance data for the past 20 years (2006-2026) shows no significant thermal anomalies suggesting low-temperature long-term heating through the edifice, which has been found at other volcanoes and considered precursory to volcanic eruptions [Girona et al., 2021].

We also examine elastic coseismic stress changes at Krasheninnikov due to the M8.8 earthquake using the coseismic slip distribution computed by Liu et al. [2026] and the boundary element software CutAndDisplace [Davis, 2017]. Liu et al. [2026] examined the pressurization of the crust at upper crustal depths and found that Krasheninnikov underwent compression. We will examine crustal pressure changes along with principal stresses to explore preferred dike orientations.

Finally, we will examine our findings considering possible models for volcano source processes to understand both triggering mechanisms and source properties such as depth and volume since we do not see evidence of a deflationary source feeding the co-eruptive dikes.

References:

Davis, T. (2017). A new open source boundary element code and its application to geological deformation: Exploring stress concentrations around voids and the effects of 3D frictional distributions on fault surfaces (M.Sc thesis. Aberdeen University).

Girina, O.A., Melnikov, D. V., Romanova, I.M., Manevich, A.G., Krasheninnikova, Yu.S., Sorokin, A.A., Kramareva, L.S., Marchenkov, V.V. (2025) The first historical eruption of Krasheninnikov volcano (Kamchatka) in 2025 according to satellite monitoring in the VolSatView information system. Sovremennye problemy distantsionnogo zondirovaniya Zemli iz kosmosa, 22(4), 397-404.

Girona, T., Realmuto, V., & Lundgren, P. (2021). Large-scale thermal unrest of volcanoes for years prior to eruption. Nature Geoscience, 14(4), 238-241.

Gorbach, N. V., Ozerov, A. Y., Rogozin, A. N., Tolstykh, M. L., & Ovsyannikov, G. N. (2026). First Historical Eruption of Krasheninnikov Volcano (Eastern Kamchatka): Field Observations and Composition of Lavas Erupted in August‒September 2025. Journal of Volcanology and Seismology, 20(1), 1-14.

Liu, C., Bai, Y., Lay, T., He, P., Wen, Y., Xiong, X., & Taymaz, T. (2026). Simple unilateral rupture of the great Mw 8.8 2025 Kamchatka earthquake. Science, 391(6787), 812-817.

Lundgren, P., A. Kiryukhin, P. Milillo, & S. Samsonov (2015), Dike model for the 2012-2013 Tolbachik eruption constrained by satellite radar interferometry observations, J. Volcanol. Geotherm. Res., 307, 79-88.

Over the past three decades, InSAR has transformed volcano monitoring by providing unprecedented observations of deformation patterns and the diverse behaviors of volcanoes worldwide. Spatial deformation signals captured in interferograms reveal variations in magmatic reservoir geometries, while time‑series analyses track magma migration and illuminate complex subsurface processes. Yet, the variability in volcanic behavior—both across different systems and within the same volcano under different eruption triggers—continues to challenge global forecasting efforts.

Despite important methodological advances, achieving global‑scale volcanic deformation forecasting remains difficult. Most studies still focus on individual volcanoes or small regional subsets, largely because long, clean, and temporally consistent deformation time series are scarce at many sites. Real InSAR data often contain irregular acquisition intervals, variable coherence, temporal gaps, and strong site‑specific characteristics, which complicate the assembly of large and homogeneous datasets needed for machine‑learning–based forecasting. For this reason, synthetic datasets have become essential: they allow controlled generation of diverse deformation scenarios under known physical assumptions, providing balanced inputs for models designed to generalize beyond specific volcanoes. Building on this motivation, we adopt a forecasting framework based on Convolutional Long Short‑Term Memory (ConvLSTM) networks.

To address the challenge of forecasting at a global scale, we generate a decade‑long InSAR time series for more than 1,000 volcanoes using Sentinel‑1 observations and pair these data with numerical simulations to train deep learning models. The wide range of deformation behaviors detected—from rapid inflation and deflation to subtle, long‑term signals—highlights the need to incorporate physically diverse scenarios directly into the training process.

We develop advanced dynamical models that integrate viscoelastic and poroelastic rheologies, enabling realistic simulations of high‑temperature, fluid‑rich crustal environments responding to magma intrusion, reservoir pressurization, and evolving subsurface conditions. These physically grounded simulations provide robust examples that enhance the network’s ability to forecast deformation across varied volcanic settings.

By integrating global InSAR observations, dynamic modeling, and deep learning, this work advances the development of a robust and scalable framework for forecasting volcanic deformation, utilizing data driven insights to refine monitoring systems and enhance predictive accuracy, ultimately improving early warning capabilities of volcano observatories and supporting safer, more resilient communities living near active volcanoes.