Conference Agenda

The exceptional meteorological sequence that hit the Emilia-Romagna region in May 2023 triggered an unprecedented hydrogeological crisis. This crisis was not characterised by a single isolated event, but by two consecutive extreme rainfall pulses that affected the central-eastern sector of the Apennines. The first event (Event 1) occurred between 1 and 3 May, with cumulative rainfall exceeding 200 mm in 48 hours. This is a value with a historical return period of over 100 years in several basins. It induced rapid soil saturation and an initial wave of landslides and river flooding. This initial phase was followed by a minor intermediate event on 10 May that prevented slope drainage and maintained high saturation levels. The sequence culminated in the catastrophic Event 2 on 16–17 May, which matched the intensity of Event 1 by delivering an additional 200–250 mm of rain onto already compromised terrain. This brought the cumulative effect of the sequence to an estimated return period of over 500 years. The geomorphological response was documented in the RER2023 inventory, which maps over 80,000 landslides through the manual interpretation of high-resolution (0.2 m) aerial orthophotos acquired after the emergency. While this inventory is spatially accurate, it does not inherently distinguish between landslides triggered during Event 1 or Event 2 and landslide reactivations. Temporal attribution is hindered by persistent cloud cover that rendered optical satellite monitoring unusable during the critical weeks. To overcome this limitation, the present study proposes a three-phase, hierarchical classification workflow integrating hourly rainfall data and soil surface moisture (SSM) variations derived from Sentinel-1 satellite data, with weighting based on landslide typology. In Phase 1, we analyse the dominance of rainfall intensity by comparing the maximum 48-hour cumulative rainfall recorded in the two time windows for each landslide centroid. This is based on the assumption that, for runoff-driven failures, the trigger coincides with the peak in precipitation. Phase 2 introduces a hydrological refinement using 500 m resolution SSM data from the Copernicus Global Land Service, calculating the anomaly relative to a dry baseline from April 2023 in order to detect areas where soils reached saturation during Event 1 without draining prior to Event 2. Phase 3 involves applying typological weighting by categorising landslides as either runoff-dominated (debris flows and debris slides), which respond rapidly to intensity peaks, or infiltration-dominated (earth flows and deep-seated slides), which are more sensitive to antecedent saturation and deep pore pressure. Applying this workflow to the 80,997 mapped failures reveals Event 2 to be dominant, triggering 52.5% of cases (~42,500 landslides) and consistent with the most catastrophic phase of the emergency. Event 1 triggered 25.8% of the landslides (approximately 20,900), concentrated in the western sector (the Sillaro and Idice basins), where the cumulative rainfall during the first storm was higher locally than during the second. 21.7% of the inventory (approximately 17,600 landslides) was classified as 'uncertain', which does not represent a modelling error, but rather the physical signature of slopes where instability was initiated by Event 1 (loading phase and reduction of suction forces) and finalised by Event 2. Integrating SSM data enabled the correction and reallocation of over 30% of landslides that had initially been classified ambiguously or incorrectly based solely on rainfall. This was particularly evident in regions where the intensity of Event 2 was lower than that of Event 1, yet the soils remained critically saturated. The analysis confirms that mechanical properties dictate the temporal response. 'Fast' landslides show strong synchronisation with Event 2 rainfall peaks, while 'slow' and deep-seated landslides exhibit a 'hydrological memory' that blurs the distinction between the two triggers. Validation through official ARPAE reports showed total concordance between model attributions and documented field activations in early May (e.g. Casamento and Monte Trebbio), confirming the algorithm's ability to recognise early activations despite subsequent heavy rainfall. In conclusion, this study shows that using rainfall thresholds alone is not enough for sequential events, and that integrating radar-derived soil moisture metrics is vital for reducing uncertainty and improving future regional susceptibility models.

Inversion of slip-surface geometry and subsurface property of slow-moving landslides remains a significant challenge. Especially using SAR interferometry due to the underdetermined nature of surface measurements and the complex, nonlinear mechanics governing soil and rock mass behaviour. To address this gap, this research builds a hybrid inversion framework for surface deformation observed from INSAR that couples a Bayesian Physics-Informed Neural Network (PINN) with a finite-element (FEM) representation of visco-elastoplastic landslide mechanics. The FEM component enforces dynamic momentum balance using a Drucker–Prager viscoplastic constitutive law and a frictional basal slip interface. It computes the weak-form residuals of the governing equations over a 3D domain. The PINN parameterises the displacement field and unknown rheological parameters, enabling joint inference of kinematics and material properties while respecting the physics at every collocation point via automatic differentiation. A Bayesian treatment of the rheological parameters within the PINN allows for the quantification of uncertainty and propagation of observational noise through to subsurface estimates. I demonstrate the method on UAVSAR datasets and show that the hybrid approach accurately recovers displacement, yield zones, and posterior distributions of constitutive parameters. This framework seamlessly integrates multi-source SAR data (such as potential fusion of Sentinel-1 and NISAR) into a physics-consistent inversion, offering a path toward improved hazard assessment and mechanistic interpretation of landslide processes.

Multi-temporal (MT) Differential Synthetic Aperture Radar (SAR) Interferometry (DInSAR) [1-5] has become a cornerstone technique for monitoring ground displacements associated with tectonic processes, volcanic unrest, subsidence, infrastructure stability, and slope instabilities. By exploiting long-term sequences of SAR acquisitions, MT-DInSAR provides spatially dense deformation time series with sub-centimeter accuracy over wide areas. However, its performance in vegetated and geomorphologically complex environments remains constrained by decorrelation effects [6], which reduce coherent pixel density and degrade displacement measurement reliability, particularly at shorter radar wavelengths, such as C- and X-band. In this context, the increasing availability of L-band SAR sensors has opened new perspectives for extending coherent, long-term deformation analysis to challenging scenarios. In particular, owing to their longer wavelength (~23 cm), L-band systems exhibit enhanced vegetation penetration capabilities, enabling improved temporal coherence over long-term observation periods in vegetated or agricultural areas, wetlands, and environments affected by seasonal snow/ice cover, where surface scattering properties may significantly change between successive SAR images. Furthermore, L-band radar data are inherently less sensitive to phase unwrapping errors, a key advantage for the investigation of rapid and/or large deformation phenomena in geohazard risk scenarios, such as slope instabilities and landslides.

These intrinsic advantages have driven important international investments in L-band SAR constellations specifically designed for multi-temporal interferometric applications. Operational L-band missions include the Japanese ALOS-2 and ALOS-4 satellites, and the Argentine SAOCOM-1 constellation, while the recently launched NASA-ISRO NISAR mission [7], the forthcoming ESA ROSE-L system [8], and the "National L-band Radar Mission" of the Italian Space Agency (ASI) will further strengthen global capabilities with systematic, wide-coverage datasets for reliable MT-DInSAR analyses.

These missions mark a strategic step toward systematic, global, long-term L-band SAR acquisitions, opening unprecedented opportunities for robust ground deformation monitoring at national and continental scales. Among these, the twin-sensor, full-polarimetric SAOCOM-1 system [9] plays a prominent role in long-term deformation monitoring over the Italian territory, thanks to its flexible acquisition modes, systematic DInSAR-oriented acquisition program over the European territory, and favourable revisit times [10].

In this work, we investigate the capabilities of the SAOCOM-1 L-band SAR data for extended MT DInSAR analysis in densely vegetated and landslide-prone areas of the Italian territory. Specifically, we evaluate the coherence preservation over vegetated slopes and the temporal continuity of displacement time series in environments typically affected by rapid decorrelation at shorter wavelengths. To this aim, we apply the Parallel Small BAseline Subset (P-SBAS) processing chain [10-12] to large L-band (SAOCOM-1) and C-band (Sentinel-1) datasets acquired over densely vegetated, geomorphologically complex areas. In particular, the SAOCOM-1 P-SBAS results are systematically compared with Sentinel-1 products, which are widely used for operational monitoring, to assess spatial coverage, coherent pixel density, and large-scale deformation mapping capability, especially in areas where shorter-wavelength SAR systems exhibit significant limitations.

Finally, we present a quantitative validation of SAOCOM-1 P-SBAS time series accuracy by exploiting independent GNSS measurements across diverse geodynamic test sites, thus evaluating reliability, robustness, and performance under varying geological, land-cover, and deformation conditions.

REFERENCES

[1] Ferretti, A., C. Prati, and F. Rocca, “Nonlinear subsidence rate estimation using permanent scatterers in differential SAR interferometry,” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 5 I, 2000, doi: 10.1109/36.868878.

[2] Berardino, P., Fornaro, G., Lanari, R., and Sansosti, E., “A new Algorithm for Surface Deformation Monitoring based on Small Baseline Differential SAR Interferograms”. IEEE Trans. Geosci. Remote Sens, 40, pp.2375-2383, 2002.

[3] Werner, C., U. Wegmüller, T. Strozzi, and A. Wiesmann, “Interferometric Point Target Analysis for Deformation Mapping,” in International Geoscience and Remote Sensing Symposium (IGARSS), 2003. doi: 10.1109/igarss.2003.1295516.

[4] Lanari, R., Mora, O., Manunta, M., Mallorquí, J.J., Berardino, P., and Sansosti, E., ”A small baseline approach for investigating deformations on full resolution differential SAR interferograms”. IEEE Trans. Geosci. Remote Sens., 42, 1377-1386, 2004.

[5] Hooper A. J., “A multi-temporal InSAR method incorporating both persistent scatterer and small baseline approaches,” Geophys Res Lett, vol. 35, no. 16, 2008, doi: 10.1029/2008GL034654.

[6] H. A. Zebker and J. Villasenor, "Decorrelation in interferometric radar echoes," in IEEE Transactions on Geoscience and Remote Sensing, vol. 30, no. 5, pp. 950-959, Sept. 1992, doi: 10.1109/36.175330.

[7] Kellogg K., et al. (2020). NASA-ISRO synthetic aperture radar (NISAR) Mission. in Proc. IEEE Aerosp Conf., Big Sky, MT, USA, 2020, pp. 1–21, doi: 10.1109/AERO47225.2020.9172638.

[8] Rostan F., K. Mak, M. von Alberti, A. Bauleo and N. Gebert (2024). Rose-L SAR Instrument Development Status and Mission Performance Prediction. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, 2024, pp. 6618-6620, doi: 10.1109/IGARSS53475.2024.10642943.

[9] Delgado, F., T. Shreve, S. Borgstrom, P. León-Ibanez, J. Castillo, and M. Poland, “A global assessment of SAOCOM-1 L-band stripmap data for InSAR characterization of volcanic, tectonic, cryospheric, and anthropogenic deformation,” IEEE Trans. Geosci. Remote Sens., vol. 62, 2024, Art. no. 5216821, doi: 10.1109/TGRS.2024.3423792.

[10] De Luca, C, et al. "SAOCOM-1 L-band DInSAR Time Series generation through the P-SBAS approach: algorithm extension and products analysis." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18 (2025): 2680-2703.

[11] F. Casu et al., “SBAS-DInSAR parallel processing for deformation time-series computation,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 7, no. 8, 2014.

[12] M. Manunta et al., “The Parallel SBAS Approach for Sentinel-1 Interferometric Wide Swath Deformation Time-Series Generation: Algorithm Description and Products Quality Assessment,” IEEE Trans. Geosci. Remote Sens., vol. 57, no. 9, pp. 6259–6281, 2019.

Slow mass movements, including landslides and permafrost-related slope processes, are characterized by diverse mechanisms depending on the materials involved and the geomorphological setting. Alpine environments are extensively affected by deep-seated rock and debris slides and active periglacial features, whereas fluvial-dominated mountain ranges are typically characterized by rockslides and long-lived earthflows. These processes span a wide range of volumes (10³–10⁸ m³), deformation patterns, and displacement rates (from mm/yr to m/yr), and pose risks either through slow movements or catastrophic collapse impacting lives and infrastructure. Enhanced capabilities for rapid, wide-area detection, classification, and monitoring of mass movements are therefore essential to support land-use planning, protect life and property, and reduce disaster risk—particularly in mountain regions increasingly subjected to anthropogenic pressures under climate change.

Traditional regional-scale approaches for detecting and characterizing active mass movements rely on manual geomorphological mapping, supported by field observations and remote sensing data. Although accurate and process-oriented, these methods are time-consuming and difficult to update over wide areas. Synthetic Aperture Radar (SAR) sensors provide systematic acquisitions and millimetre-scale deformation measurements; however, the massive data volume and technical complexity have so far limited their full regional-scale exploitation. Furthermore, mass movements are often too fast to be captured by widely-used multitemporal InSAR products, yet too slow to be detected through optical imagery or SAR amplitude analysis. Dual-pass satellite wrapped DInSAR products offer a valuable alternative through the analysis of interferometric fringes, without requiring sophisticated multitemporal analysis. Despite still affected by noise and artefacts, these products don’t need phase unwrapping, thereby circumventing errors typically occurring in steep topography and decorrelated regions.

To exploit the information content of wrapped DInSAR interferograms, we propose MIRAGE*, a deep learning framework for the automated detection and classification of active mass movements, using wrapped SAR interferograms derived from free, routinely available Sentinel-1 imagery. We adopted a YOLO-based convolutional neural network architecture, designed to detect and classify mass movements from deformation fringes within multi-scale spatial contexts. To emulate expert-based geomorphological interpretation, the model input layers include: (i) the DInSAR wrapped phase, (ii) a SAR signal reliability index integrating geometric sensitivity and coherence, and (iii) a composite geomorphometric variable derived from principal component analysis (PCA) of relevant morphometric parameters.

Raw filtered and geocoded SAR interferograms were generated from Sentinel-1 ascending and descending acquisitions with temporal baselines ranging from 6 to 365 days, capturing mass movements across multiple spatial and temporal scales. To train the deep learning model, we constructed a geomorphologically constrained library of approximately 5000 labeled wrapped-phase signals, associated with different mass movement types identified by expert interpretation across two wide areas, representative of contrasting environmental and geomorphological conditions. A 1900 km² sector of the Italian Central Alps was selected to represent high-relief, para-periglacial environments with widespread rock glaciers, rockslides and debris slides. In contrast, a 1200 km² sector of the Italian Apennines represents lower-elevation, fluvial-dominated terrain developed on clay-shale and flysch lithologies, where earthflows, earth slides, and rockslides predominate. These areas encompass diverse mass movement types, topographic conditions, land cover characteristics, and deformation rates, thereby enabling robust training and testing aimed at model generalization.

The network was trained using data augmentation and random dataset-splitting strategies. The trained model outputs objectness scores, bounding box offsets, and class confidence scores. The final bounding boxes are predicted using thresholds. The network demonstrates the ability to intercept mass movement signals directly within raw interferograms across multiple scales. Detection accuracy exceeds 0.75–0.8, and classification performs well even when interferogram components usually regarded as noise are included. In alpine environments with rugged high-relief topography and complex overlapping mass movement processes, detection performance decreases and classification becomes more challenging. However, in these settings, geomorphological assessment of false positives often reveals incompleteness in the training dataset, which negatively affects quantitative metrics.

Although further developments are required to enhance model generalization and classification robustness in complex geomorphological contexts, the results demonstrate the significant potential of exploiting raw InSAR products to streamline the rapid detection of mass movements over a wide range of sizes and velocities. This approach supports inventory updating, local monitoring strategies, and integration into landslide modelling frameworks.

* Funded by the European Union-Next Generation EU, Mission 4, Component 2, CUP H53D23001660006 (PRIN22 Project "MIRAGE: Mass movement Investigation and prediction through geomorphology, Remotesensing and Artificial intelligence").