Ground and structures move slowly: above old mines, along tunnels, on rail embankments, around dams and in groundwater-depleted basins. Synthetic-aperture radar (SAR) satellites such as Copernicus Sentinel-1 re-image the same area every few days, and interferometry (InSAR) turns phase changes between acquisitions into millimetre-scale displacement. ML then separates real deformation from noise, classifies motion patterns and ranks assets for inspection, replacing manual surveying of whole networks with a prioritised, evidence-based list.
Problem & value
- Today: settlement and slope movement are checked by levelling, GNSS, inclinometers and periodic walk-over or drone inspections. These give precise but sparse point data, are expensive per kilometre and miss slow regional trends. Mining authorities and geological surveys rely on periodic campaigns; rail and road owners inspect on fixed cycles.
- AI value: satellite InSAR covers entire networks with no access to the asset and no on-site hardware. Millions of measurement points per region cannot be reviewed by hand, so ML is used for denoising, phase-unwrapping support, atmospheric correction, motion-pattern classification (linear, seasonal, accelerating, step), clustering of coherent deformation zones and ranking of assets by risk.
- Free base data: the Copernicus European Ground Motion Service (EGMS) provides harmonised Sentinel-1 ground-motion products for the Copernicus participating states, free of charge, with regular updates [1][2]. This removes the heaviest processing step for many users and shifts the work to analysis, integration and decision support.
Variants
- Rail and road networks: embankment, cutting and bridge-approach settlement, landslide hazard along corridors.
- Mining and post-mining: subsidence over active and abandoned mines (e.g. German coal regions, salt and potash), mine-water rebound uplift.
- Urban settlement: building and tunnelling-induced settlement, groundwater extraction, construction-site impact.
How it works
- Sentinel-1 SLC stack / EGMS products (+ optional commercial SAR: TerraSAR-X, COSMO-SkyMed, ICEYE, Capella)
- InSAR processing (coregistration, interferograms, PS/DS time series; or download EGMS L2/L3)
- correction (atmosphere, orbit, DEM, reference-point selection, GNSS calibration, ascending+descending
- vertical/E-W)
- ML analysis (denoising, motion-pattern classification, change-point detection, clustering of deformation zones)
- asset overlay (GIS buffers per rail km / dam / building) + context (geology, groundwater, mining plans)
- risk ranking and alert rules
- expert review (geotechnical engineer)
- on-site inspection / instrumentation / drone survey
- Time-series ML: per-point trend and seasonal decomposition, change-point and acceleration detection, and unsupervised anomaly detection against neighbouring points. Deep models (U-Net-style segmentation, CNN/Transformer encoders on displacement maps and interferogram patches) are used for deformation-zone detection, unwrapping aids and atmospheric-noise reduction. Foundation models for SAR and EO are emerging but still research-grade for InSAR.
- Geometry matters: Sentinel-1 measures only along the radar line of sight; combining ascending and descending tracks gives vertical and east-west components. Point density is high on buildings and rock, low on vegetation and snow.
- Human-in-the-loop: a geotechnical or surveying expert validates every alert, checks against reference measurements and decides on action. Satellite data is a screening layer, not a safety-certified sensor.
Data preparation
Needed data: Sentinel-1 SLC stacks or ready EGMS/national ground-motion products, optional commercial SAR, plus asset data (rail or pipeline geometry, dam and building outlines), geology, groundwater and mining plans, and reference measurements (GNSS, levelling). These live in GIS systems, asset-management software (SAP EAM), survey archives, file shares and paper plans.
Typical gaps: asset geometries without consistent coordinate reference systems, outdated or scanned plans, missing GNSS reference points, low point density on vegetation, and no history of confirmed events to label.
Preparation steps: harmonise asset layers and projections in QGIS/PostGIS; digitise legacy plans (PREP-006); choose a stable reference point and validate against GNSS; export data from asset systems (PREP-005) and tidy tables (PREP-011); define categories for motion patterns before ML classification (PREP-019); harvest confirmed damage cases as labels (PREP-020). Start with a gap analysis (PREP-001, PREP-002).
Rough share of effort: medium. With bought products it is mostly asset-data harmonisation and reference validation; own processing adds expertise and compute effort, but little labelling.
Implementation paths
Which path for whom
- Enterprise (infrastructure, mining, utilities): A, because quality-assured InSAR products from EU providers or the free EGMS are mature and owners rarely need in-house InSAR expertise.
- Public (surveys, agencies): C, because open tools (ISCE, MintPy, SNAP) and free Copernicus data on HW2 servers give transparent, reproducible processing and sovereignty over national monitoring.
- Developer: C, because open-source processing plus ML analysis on HW2 is the natural stack for building own analytics or research.
Path A — Buy (recommended for asset owners)
- EU InSAR service providers: TRE Altamira (Italy/Spain, ground-motion analytics and products), SkyGeo (Netherlands), Detektia (Netherlands), Gamma Remote Sensing (Switzerland, software and services), plus national surveys such as BGR (Germany) which run InSAR-based ground-motion services (e.g. the German Ground Motion Service) [3]. EU space/data providers: ICEYE (Finland, commercial SAR), Airbus Defence and Space (TerraSAR-X/PAZ).
- Integration: deliveries as GeoPackage/GeoTIFF/WMS, or into asset-management systems (Esri ArcGIS, QGIS, SAP EAM).
- Choose A when: you need quality-assured deformation products for an asset base without building InSAR expertise.
Path B — Build on cloud APIs
- Data: EGMS download service and Copernicus Data Space Ecosystem (EU-operated; Sentinel-1 archive, openEO, STAC API) [1][4]; ASF DAAC and HyP3 on-demand InSAR processing (ASF, USA) for custom interferograms [5].
- Compute and orchestration: openEO / Sentinel Hub (Planet), Google Earth Engine for ancillary data; EU cloud (OVHcloud, Scaleway, STACKIT, T-Systems) for processing; Python with geopandas, xarray, scikit-learn, PyTorch.
- Choose B when: you want custom risk scoring on EGMS data joined to your own asset and sensor data.
Path C — Self-host (open source)
- Processing: ISCE2/ISCE3, ESA SNAP with the Stamps/StaMPS PS approach, MintPy for SBAS time series, GMTSAR, COMET LiCSAR/LiCSBAS, ASF HyP3 open code, SARvey [5][6].
- ML: PyTorch or scikit-learn on displacement tensors; U-Net, ConvLSTM/Transformer time-series models, isolation forest or autoencoders for anomalies; STAC + Dask or Zarr for scalable stacks.
- Hardware: HW1 for analysing EGMS tiles; HW2 (one GPU with 24-32 GB and ample storage) for regional deep-learning workflows; raw SLC stack processing is CPU, RAM and storage-bound rather than GPU-bound (many TB per region).
Path D — Custom training
Worth it for a region-specific deformation-zone detector (e.g. mining subsidence bowls, landslide scarps) or to replace hand-tuned thresholds. Needs labelled deformation inventories (geological survey landslide and subsidence maps), reference GNSS and levelling, and multi-year stacks. Synthetic-deformation simulation helps with scarce labels.
State of the art (2026)
- Data supply: Sentinel-1 continues with the C and D units alongside A/B heritage, keeping 6-12-day revisit in Europe; EGMS provides annual-updated, harmonised, free products for Europe [1][2]. Higher-resolution commercial SAR (ICEYE, Capella, TerraSAR-X) and the NASA-ISRO NISAR L-band mission [7] improve vegetated-area coverage and reduce decorrelation.
- ML trends: deep-learning unwrapping, atmospheric phase screen estimation, automatic detection of deformation fronts and landslides from velocity maps, and cluster-based anomalous-motion detection on EGMS points; foundation models for SAR are being benchmarked. Operational practice still relies largely on rule-based and statistical screening with ML as support.
- Limits and failure modes:
- decorrelation over vegetation, snow, water and farmland yields no or poor points;
- atmospheric delay and unwrapping errors create false signals; fast motion (above roughly half a wavelength per acquisition) cannot be tracked;
- line-of-sight geometry hides some motion directions;
- single-point anomalies may be sensor artefacts, structural thermal expansion or re-surfacing work;
- not a replacement for in-situ monitoring of safety-critical structures such as dams.
Shared resources & combinations
Effort, cost & KPIs
- Path A: weeks to a few months to onboard; costs scale with area and revisit and are typically project or subscription priced.
- Path B: 2-4 months to a working screening workflow on EGMS data with GIS and asset-system integration.
- Path C/D: 6-12 months incl. building InSAR expertise; main cost drivers are specialists and storage.
- KPIs: share of asset network covered by usable scatterers, detection precision versus GNSS/levelling reference, false-alert rate per inspection, lead time between first detection and confirmed damage, inspection cost per km, number of avoided unscheduled closures.
People & change
Involve management for scope and budget, the asset or geotechnical lead as process owner, end users (inspectors, surveyors) who act on alerts, IT for GIS and asset-system integration, the data owners of asset and survey data, and legal/compliance for liability: satellite data is a screening layer, not a safety-certified sensor.
Works council: no in the normal case, because the system monitors ground and structures, not employees. Check if alerts feed into individual performance data; then §87(1) no. 6 BetrVG would apply. Under §90(1) no. 3 inform the works council if inspection routines change materially; public sector: Personalrat. Data protection: building-level resolution can touch personal data of residents in some cases (depends); ask the DPO, document the legal basis, and aggregate or limit resolution where possible. A DPIA is usually not required.
AI literacy: support (AI Act Art. 4) inspectors with training on line-of-sight geometry, atmospheric noise and false alarms. Communicate clearly that alerts trigger an inspection, not an automatic decision. Pick a pilot section (one rail line or dam) and compare against existing instruments to build trust. Typical concern: alert fatigue; address with ranked alerts and thresholds agreed with engineers. A geotechnical expert validates every alert and decides on action.
See GUIDE-009, GUIDE-011 and GUIDE-001.
Risks, regulation & compliance
- EU AI Act: minimal risk for screening and prioritisation. If the output is a safety component in the management or operation of critical infrastructure (e.g. road traffic or energy supply, Annex III point 2), classification as high-risk may apply; keep a human decision and independent in-situ confirmation. Art. 4 AI literacy applies. No Art. 5 issue.
- GDPR: 10-20 m scale products are rarely personal data, but building-level displacement of private properties can be linked to owners; use aggregation, access control and purpose limitation when sharing with insurers or lenders.
- Open data and licensing: Copernicus data and EGMS are free and open under Copernicus terms; commercial SAR carries restrictive licences [1].
- Sector rules: dam safety, mining law (e.g. German Bundesberggesetz), rail safety and EU Floods Directive duties still require certified surveys; InSAR supports but does not replace them.
- Mitigations: reference-point calibration with GNSS, ascending and descending cross-checks, documented uncertainty, expert review, version-controlled processing chain, audit trail of alerts and decisions.
Getting started
- Define assets and hazards (e.g. 200 km of rail on embankments) and check scatterer coverage in the EGMS viewer [1].
- Download EGMS tiles for your area and overlay them with asset GIS layers in QGIS or Python.
- Compute velocity statistics and anomalies per asset segment; compare with known problem sites and GNSS or levelling data.
- Pilot a ranked inspection list with the maintenance team and track confirmed versus false alerts.
- If coverage is poor or you need custom stacks, commission a provider (Path A) or build the processing chain with MintPy/ISCE (Path C).
Sources
- Copernicus European Ground Motion Service (EGMS) portal and viewer — https://egms.land.copernicus.eu
- Copernicus Land Monitoring Service — EGMS product page — https://land.copernicus.eu/en/products/european-ground-motion-service
- BGR — Bodenbewegungsdienst Deutschland (German Ground Motion Service) — https://bodenbewegungsdienst.bgr.de
- Copernicus Data Space Ecosystem — https://dataspace.copernicus.eu
- ASF HyP3 on-demand InSAR processing docs — https://hyp3-docs.asf.alaska.edu
- MintPy: Miami InSAR time-series software in Python — https://github.com/insarlab/MintPy
- NASA-ISRO NISAR mission — https://nisar.jpl.nasa.gov
- ESA Sentinel-1 mission overview — https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-1
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj