Deformation trend change detection for infrastructure monitoring

The Copernicus Sentinel-1 Synthetic Aperture Radar (SAR) mission has been a game-changer for Earth Observation. SAR-derived information is used to produce millimetric-precision ground surface deformation maps. Each information layer is made of hundreds of thousands of measurement points, and can be used for detecting significant instabilities on critical infrastructures. Big Sentinel-1 data availability combined with cloud-based solution allows TRE Altamira to deliver even nationwide InSAR databases, offering millions of measurement points with a time series of displacement at every new update.

Running the MATTCH project (ESA Open Call for Science, 2019), TRE Altamira has already implanted DL-based methodologies to big Sentinel-1 data to extract only displacement time series exhibiting anomalous behaviour for data screening purposes. “Hotspots” can be identified, but no indication about driving mechanisms of displacement is given to end-users. With this use case, we build upon MATTCH to create a commercial service to monitor critical infrastructure at large scales. We develop new DL architectures to obtain actionable information from large InSAR point clouds and classify points sharing some attributes or features that are key for monitoring risk over infrastructure.

DeepCube aims to link any hotspots to a possible reason for their motion behaviour. Specifically, using AI on InSAR data and other external information layers, we aim to classify each hotspot automatically detected and impacting urban infrastructure, as well as critical infrastructure such as highways, tunnels, bridges, and airports.

Reliability Prediction Model

So far, most of the research has tackled the removal of non-reliable measurements from large InSAR point clouds using a multimodal GNN architecture for the reliability prediction and the identification of points showing a change in their motion trend based on a spatiotemporal analysis of their displacement time series, progressing the MATTCH’s result with the design of a Spatio Temporal GNN architecture.

What was not foreseen initially, namely the development of the reliability prediction model, has turned out to be an impactful step to remove uncertainty from the measurements and focus only on reliable measurements with a positive effect on data confidence by operators.

The development of the reliability prediction model is a novelty introduced by DeepCube. The model combines InSAR point cloud information with other modalities - SAR amplitude images, Optical images from Sentinel-2, Land-Cover and Land-Use classification maps, and a Digital Elevation Model of the terrain - to filter out those points that do not satisfy a reliability threshold.

Model serving service on Hopsworks

For demo purposes, part of the reliability prediction model is now deployed on Hopsworks through a model serving service. The deployment of the model is made automatically using the GitHub action services of DeepCube. The scope of this repository is also to showcase how Hopsworks is exploited within an industrial environment where versioning and integration pipelines are fundamental.

You can access the repository on GitHub:


Federico Ricciuti, Alessio Rucci, Chiara Gervasi, Ioannis Papoutsis, Roberto Acerbis, “Multimodal InSAR Reliability Prediction using Graph Neural Networks (H2020 DeepCube project)”, Living Planet Symposium 2022

Use Case Leader
Tre Altamira

Interested in learning more? Contact us!
Chiara Gervasi,
Federico Ricciuti,