EGU23-11453, updated on 27 Sep 2023
EGU General Assembly 2023
© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Machine learning model to assess the spatial probability of Sentinel-1 based deformation trend changes

Camilla Medici1, Pierluigi Confuorto1, SIlvia Bianchini1, Matteo Del Soldato1, Ascanio Rosi2, Samuele Segoni1, and Nicola Casagli1,3
Camilla Medici et al.
  • 1Earth Sciences Department, University of Firenze, Firenze, Italy (
  • 2Department of Geosciences, University of Padova, Padova, Italy
  • 3National Institute of Oceanography and Applied Geophysics-OGS, Sgonico, Trieste, Italy

Satellite ground deformation monitoring is now a well-established reality and the MTInSAR (Multi-Temporal Interferometry Synthetic Aperture Radar) techniques have widely demonstrated their feasibility for detecting a wide range of slow-moving phenomena, e.g., landslides and subsidence at different scales. The launch of the ESA’s Sentinel-1 constellation has allowed acquiring massive quantities of radar images with a worldwide coverage and a short revisiting time. These characteristics, combined with the increasing computation capabilities and advanced processing techniques, have opened the opportunity of implementing a continuous monitoring service of ground surface deformations at regional scale. Tuscany, Veneto, and Valle d’Aosta regions (Italy) have benefited from this service exploiting ground deformation maps, periodically updated, and identifying the so-called anomalies of movement of radar targets, i.e., trend changes (e.g., accelerations) in the time series of displacement. However, the continuous monitoring system only has the ability to detect the anomalies of movement without being able to assess the propensity of a territory to be affected by them. Therefore, an approach for assessing the spatial probability of trend changes of InSAR-based ground deformations occurrence has been proposed. The occurrence probability of anomalies is determined by a Machine Learning (ML) algorithm, Random Forest, and the data used for the application of the model are the anomalies database and the predisposing factors (PF). The selected PFs can be split into two groups, indeed, in addition to the classical morphological and geological features, even five variables related to the radar system have been integrated. The latter parameters are two radar visibility indexes (C-index and R-index), the horizontal, along East-West direction, and vertical component of the velocity of displacement and the standard deviation of the satellite line of sight (LOS) velocity. These two groups of PFs can be considered a synthesis of the main factors that lead to the generation of anomalies. The procedure has been tested on the Tuscany region for assessing the spatial probability of anomalies occurrence related to landslides and subsidence. The outcomes of the procedure are two maps of the spatial probability of occurrence of landslides and subsidence anomalies. A cross-validation procedure has also been performed to verify the reliability of the final maps by exploiting anomalies collected in a different timespan from the input data and the official landslide and subsidence inventories. The resulting information, periodically updated, can represent a useful instrument for the regional authorities to identify the main driving forces leading to ground deformation anomalies and the areas where site investigations are to be carried out to assess the preliminary risk.

How to cite: Medici, C., Confuorto, P., Bianchini, S., Del Soldato, M., Rosi, A., Segoni, S., and Casagli, N.: Machine learning model to assess the spatial probability of Sentinel-1 based deformation trend changes, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11453,, 2023.