Project Summary

The research activities held during the visiting period at the ESA Φ-lab were divided in two main sub-projects:

• Development and testing of a self-supervised LSTM-based framework for the classification of InSAR Displacement Time Series.
• Downscaling of GRACE-derived groundwater estimations through the adaptation of an unsupervised U-Net based framework.

Development Tools

• For the first sub-project, EGMS displacement time series, distributed in .csv files, were exploited. Pytorch was used to construct the LSTM, while the HDBSCAN library was employed for clustering.
• For the second sub-project, ERA-5 Land Reanalysis data and GRACE-derived products were downloaded as NetCDF files, and the xarray library was used to process them. The U-Net algorithm was constructed using Pytorch.
• All the code was developed in Python using Visual Studio Code.

Development Outputs

• Conference presentation: the work developed under the first sub-project was presented on the 11th of August at the 2026 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Washington D.C., United States (US), under the title “SELF-SUPERVISED CLASSIFICATION OF INSAR DISPLACEMENT TIME SERIES: A COMPARISON BETWEEN AN LSTM-BASED AUTOENCODER AND ICA”. It will soon be published in the corresponding IEEE X’Plore Conference Proceedings.
• Conference presentation: part of the results from the second sub-project were presented as a poster at the Φ-Innovation Summit, which took place in Rome, Italy, from the 23rd to the 25th of June.
• On-going: results obtained from the development of the second sub-project were submitted for publication to the MDPI Remote Sensing Special Issue “Advancing Earth Observation Through Artificial Intelligence: From Foundation Models to Intelligent Retrieval Systems”. The submission is currently under review.
• Dataset: the downscaled datasets over France were uploaded to Zenodo under the doi: 10.5281/zenodo.22235907

Project Description

Sub-Project N° 1: Self-supervised Classification of InSAR Displacement Time Series through an LSTM-based Autoencoder

The main aim of this first project was to investigate the potential of a self-supervised framework that combines a Long Short-Term Memory (LSTM) Autoencoder (AE) with the Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm for identifying and clustering InSAR deformation time series. The approach was tested in two areas near Bologna, Italy, using five years of European Ground Motion Service (EGMS) InSAR time series covering the period 2018–2022, and the results were compared with those obtained from a classical statistical analysis (Independent Component Analysis – ICA).

The LSTM-AE was used to learn compact latent representations of the displacement time series, allowing the model to capture their main temporal characteristics while reducing the influence of noise. For this objective, a noise-aware reconstruction loss was adopted, and the spatial relationships between neighboring measurement points were also considered through a graph-based regularization term. After the training phase, the latent representations were clustered using HDBSCAN, which separates the noisy time series and identifies the main clusters in the data.

The proposed framework was able to identify distinct displacement behaviors in both study areas. In the first area, three main clusters were recognized, all characterized by predominantly linear deformation trends but with different mean displacement velocities. The cumulative LOS displacement over the five-year period reached approximately −80 mm for the strongest deformation pattern, while the other clusters exhibited lower cumulative displacements of approximately −50 mm and −22 mm. Conversely, four clusters were identified in the second study area, characterized by more complex, non-linear deformation behavior. Three clearly seasonal patterns with different amplitudes were retrieved, in addition to a less seasonal negative signal, interpreted as a linear trend.

An important advantage of the proposed methodological framework is that HDBSCAN does not require the number of deformation patterns to be predefined, since the algorithm determines the number of representative clusters directly from the data and also separates noisy time series. However, this also represents a limitation, as a substantial fraction of the available observations can be classified as noise, particularly in areas where the deformation patterns are less clearly defined. In the analyzed areas, 72% and 24% of the points were classified as noise, respectively. Beyond pattern recognition, the LSTM autoencoder also provided reconstructed versions of the original displacement time series. These reconstructions produced smoother and less noisy representations of the deformation signals, potentially providing a useful basis for subsequent analyses. In particular, the learned representations and reconstructed time series could be further exploited for applications such as detailed displacement analysis at finer spatial scales, time series forecasting, and predictive subsidence analysis.

Sub-Project N° 2: Unsupervised Downscaling of GRACE-derived Groundwater Estimations

Groundwater is one of the most important freshwater reservoirs on Earth, but its monitoring remains challenging because it largely relies on in situ measurements, which are often sparse and sometimes unavailable. The GRACE and GRACE-FO satellite missions provided a unique opportunity to monitor changes in terrestrial water storage at regional and global scales. Still, their coarse spatial resolution remains a challenge. With the rapid development of Machine Learning (ML) and Artificial Intelligence (AI) techniques, many approaches are being investigated to improve the spatial resolution of these products, while maintaining the information contained in the original satellite observations.

In this context, the objective of this second project was to investigate the potential of an unsupervised deep learning approach to downscale the GFZ Global Gravity-based Groundwater Product (G3P) from its original spatial resolution of 0.5° (approximately 55 km) to 0.1° (approximately 11 km). The methodology was based on a High-to-High (H2H) U-Net convolutional neural network framework, originally developed for downscaling next-generation satellite gravimetry missions (NGGM, GRACE-C, MAGIC) from 3° to 1° resolution. The processing was performed across France and validated in the Paris Basin area using piezometric level data from the open-access IGRAC Global Groundwater Monitoring Network (GGMN) groundwater wells database for the period from January 2003 to September 2023.

In contrast to conventional supervised approaches, the H2H U-Net CNN tested in this project does not require a high-resolution reference dataset during training. Instead, the network generates a high-resolution groundwater storage anomaly (GWSA) prediction, which is subsequently aggregated back to the original 0.5° resolution and compared with the G3P product. The difference between the aggregated prediction and the original G3P observation is then used to optimize the network parameters. This configuration allows the model to learn high-resolution spatial patterns while being constrained by the information available from the original GRACE-derived product.

The model was trained using only globally available open datasets to maximize the methodology's transferability to other regions, including areas where dense groundwater monitoring networks are not available. Therefore, thirteen climatic variables from the ERA5-Land reanalysis were used as dynamic predictors, including precipitation, evaporation, runoff, soil moisture at different depths, snow-related variables, surface pressure and temperature. In addition, elevation information from the ETOPO 2022 digital elevation model (DEM) was incorporated as a static predictor. Since groundwater storage responds to hydrological conditions over time rather than instantaneously, antecedent information from previous months was also included for selected climatic variables (“lag months"). Seasonal information was further incorporated through a sinusoidal and cosinusoidal temporal encoding.

A strong consistency between the high-resolution and original products was observed in both the spatial and temporal dimensions. When the 0.1° predictions were compared with the corresponding 0.5° G3P pixels, the two datasets showed an almost 1:1 relationship, with a spatial correlation coefficient of r = 0.99, an RMSE of 4.21 mm EWH, and an MAE of 2.82 mm EWH. Their statistical distributions were also highly similar, with mean values of approximately −6.2 mm EWH for the original G3P product and −6.5 mm EWH for the downscaled product.

The results were validated by comparing the HR GWSA and the original G3P products against in situ measurements over the Paris Basin. Although the HR product outperforms the original G3P data at the basin scale (r = 0.67 vs. r = 0.62, respectively), both products show similar correlations at the well scale (r = 0.39 vs. r = 0.38). These results highlight an important aspect of the methodology: while the H2H framework allows high-resolution predictions to be generated without requiring high-resolution groundwater observations during training, favoring applications in data-scarce regions, the information available to the network may be insufficient to capture the highly localized variability observed at individual groundwater wells.