- 1Machine Intelligence and Slope Stability Laboratory (MISSLab), Department of Geosciences, University of Padua, Padova, Italy
- 2Department of Earth Sciences, University of Cambridge, Cambridge, UK
- 3Department of Geography, University of Cambridge, Cambridge, UK
- 4Centre Tecnològic de Telecomunicacions de Catalunya (CTTC/CERCA), Castelldefels, Spain
- 5Department of Geodesy and Geoinformation, Technische Universität Wien, Vienna, Austria
Ground deformations, such as landslides, subsidence, and mining-related deformations, pose significant risks to communities and infrastructure. Accurate classification of these deformations is crucial for hazard management and land use planning. Existing classification methods primarily rely on thresholding or traditional machine learning models, failing to fully capture the rich temporal and spatial information available from spaceborne remote sensing data.
This study proposes a deep learning method that integrates both ground motion time series (European Ground Motion Service - EGMS) and geospatial data (spaceborne optical imagery, and morphological features) to classify ground motions. The method employs a dual-branch model, where 1D CNNs extract temporal features from ground motion time series, and 2D CNNs capture spatial characteristics from corresponding satellite imagery and topographic data. The features extracted by both branches are fused and fed to a multilayer perceptron to classify deformation processes, i.e., landslides, deep-seated gravitational slope deformations (DSGSD), subsidence, and mining-related deformations. To inform the model, we used a dataset over 26,000 Active Deformation Areas (ADAs), defined with the ADA finder tool(Navarro et al., 2020). We annotate each ADA by crossing it with existing inventories such as the Italian Landslide Inventory (IFFI) and CORINE Land Cover map. Corresponding time series data and imagery were subsequently extracted for each and fed to the model. Results, using cross-validation, show that the model achieves an overall accuracy of over 90%. This demonstrates its effectiveness and robustness in handling diverse deformation types. We finally deployed the validated model and classify all the ADAs generated for the entire Italy.
This research provides a scalable and automated framework for ground motion classification, and the classification achieved can lead to better-targeted risk mitigation strategies, and improved ground motion forecasting and early warning systems.
References: Navarro, J. A., Tomás, R., Barra, A., Pagán, J. I., Reyes-Carmona, C., Solari, L., Vinielles, J. L., Falco, S., & Crosetto, M. (2020). ADAtools: Automatic Detection and Classification of Active Deformation Areas from PSI Displacement Maps. ISPRS International Journal of Geo-Information, 9(10), 584. https://doi.org/10.3390/ijgi9100584
How to cite: Dong, Y., Nava, L., Palama, R., Monserrat, O., Festa, D., Floris, M., and Catani, F.: Integrating Temporal and Spatial Data for Deep Learning-Based Classification of Slow-Moving Ground Deformations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15532, https://doi.org/10.5194/egusphere-egu25-15532, 2025.
