Earth-observation (EO) archives grow by terabytes per day, far more than analysts can inspect. Computer-vision and geospatial foundation models turn imagery into maps and change alerts: what is built, farmed, forested or flooded, and what changed since last month. Organisations replace manual digitising and occasional field surveys with repeatable, auditable monitoring.
Problem & value
- Today: land-use maps are produced by manual interpretation or classical per-pixel classification and updated every few years. Authorities, utilities, insurers and agribusiness rely on field visits, periodic orthophoto flights and expensive analyst time.
- AI value: frequent (Sentinel-2 revisits every few days), consistent and scalable mapping; automatic change flags that focus humans on the few percent of the area where something happened; area statistics for reporting.
- Free data base: the EU Copernicus programme provides Sentinel-1/-2 data free and open, via the Copernicus Data Space Ecosystem, and the Copernicus Land Monitoring Service provides reference products such as CORINE Land Cover and High Resolution Layers [1][2]. Benefit figures vary by programme; publish your own baseline rather than quoting vendor claims.
Variants
- Public authorities: land-use/land-cover updates, sealing and urban sprawl monitoring, tree cadastre, illegal-dumping hints (compliance workflows: VIS-038).
- Agriculture: crop-type maps, field boundary delineation, yield and stress indices.
- Utilities and infrastructure: vegetation encroachment near lines, asset inventories (VIS-003).
- Insurance and finance: exposure mapping, building footprints, event-before/after (VIS-058, PRED-036).
- Environment and ESG: forest loss, wetlands, water bodies (VIS-049), methane (VIS-042).
- Real estate and construction: construction-progress and site-change monitoring.
- Private/NGO: citizen-science mapping with free tools (OpenStreetMap, QGIS plugins).
How it works
- acquire (STAC catalogue: Sentinel/Landsat/commercial/orthophoto)
- preprocess (atmospheric correction, cloud masking, co-registration, tiling, temporal stacking)
- model (segmentation / classification / change detection / embeddings)
- postprocess (vectorise, minimum mapping unit, merge with cadastre)
- quality control (confidence, sample-based accuracy assessment)
- human review of change polygons (GIS)
- publish (maps, statistics, alerts, API)
- Segmentation and classification: U-Net/Mask2Former-style networks, or lightweight heads on a geospatial foundation model for land cover, buildings, roads, fields.
- Change detection: bi-temporal Siamese networks or time-series models on stacks; SAR is used where clouds block optical sensors. Raw pixel differences produce many false alarms (season, shadows, sun angle), so classify the change type rather than only flagging difference.
- Foundation models and embeddings: pretrained EO models give few-shot mapping with few labels; precomputed annual embedding layers allow similarity search ("find more areas like this").
- VLMs: natural-language querying and captioning of imagery is useful for exploration, but not yet reliable for measurement or legally relevant findings.
- Human-in-the-loop: analysts verify change polygons before they trigger decisions; accuracy is checked on an independent, stratified sample.
Data preparation
Data needed: satellite or aerial imagery (Sentinel, Landsat, orthophotos, drone flights) plus own reference layers such as cadastre, parcel boundaries, asset registers and labelled polygons. Imagery is public or purchased and easy to get; the hard part is the own reference data. In an SMB this typically lives in the GIS, in ERP asset lists, in Excel sheets with coordinates and on file shares as shapefiles or PDFs of plans.
Typical gaps: inconsistent coordinate systems, outdated or overlapping parcel boundaries, no labelled training polygons, clouds and haze in optical scenes, scattered drone flights without metadata, and unclear access rights for third-party geodata.
Preparation steps: 1. Inventory layers, owners and licences (see PREP-001 and PREP-002). 2. Harmonise coordinate reference systems and clean vector geometries in QGIS or GeoPandas. 3. Fetch imagery through STAC or openEO, apply cloud masks and build time series. 4. Create a small, expert-checked label set from existing records such as subsidy applications or inspection reports (PREP-020, PREP-018). 5. Reuse structured tables via PREP-011 where coordinates sit in spreadsheets.
Effort: roughly half of the project goes into reference data, labels and pipeline set-up; the model itself is the smaller part.
Implementation paths
Which path for whom
- SMB: Path B, because farms, forestry or planning offices need results tied to their own parcels and layers, and no off-the-shelf product covers that cheaply.
- Enterprise: Path B, since cloud APIs on EU-hosted Copernicus data scale to many assets and integrate with the existing GIS.
- Public: Path C, since authorities need sovereign, auditable processing on German or EU infrastructure and open Sentinel data keeps costs low.
- Developer: Path C, because open geospatial foundation models and STAC tooling give full control over training and pipelines.
Path A — Buy (off-the-shelf)
- Platforms with analytics: Planet (analytics feeds, change detection), Airbus Defence and Space (EU, imagery + OneAtlas), Satellogic, ICEYE (SAR), Maxar/Vantor; EU/GDPR-friendly: Airbus, EOMAP, Sinergise (Sentinel Hub, now part of Planet), Starion, GAF AG, Orbital EOS, Kayrros (EU, analytics), Pixalytics, and national mapping agencies' orthophoto services.
- GIS suites with built-in deep learning: Esri ArcGIS Pro (pretrained models), QGIS with plugins.
- Choose A when: a standard product (building footprints, crop map, vegetation encroachment) exists and there is no in-house geospatial ML team.
Path B — Build on cloud APIs (recommended default)
- Data access: Copernicus Data Space Ecosystem (openEO, Sentinel Hub APIs, STAC, EU-hosted) [2]; Microsoft Planetary Computer and AWS Open Data for global archives; Google Earth Engine for planetary-scale analysis incl. the AlphaEarth satellite embeddings dataset [3].
- Models: managed training/inference on cloud GPUs (EU regions), foundation-model heads (Prithvi-EO-2.0, TerraMind, Clay) fine-tuned with a few hundred labelled polygons.
- Orchestration: STAC-based pipelines, scheduled per new acquisition; results to PostGIS / GeoParquet / cloud-optimised GeoTIFF and a web map.
Path C — Self-host (open source)
- Stack: GDAL, rasterio, xarray/stackstac/odc-stac, TorchGeo, PyTorch, Segment Anything (SAM/SAM 2) for interactive labelling, YOLO-class models for object detection, PostGIS, QGIS, openEO.
- Open-weight EO models: IBM/NASA Prithvi-EO-2.0, TerraMind (IBM/ESA), Clay, SatMAE and others published on Hugging Face [4].
- Hardware: HW2 for inference on regional extents; HW3 for large continuous mapping and fine-tuning ViT-L-class models; tiling and storage (tens of TB) usually dominate over GPU.
Path D — Custom training
Worth it when national or regional specifics (crop calendars, local building styles, rare classes) are not covered, or when production accuracy must be certified. Needs reference labels from field surveys, cadastre or existing authoritative maps; start from a foundation model and fine-tune. Mind spatial autocorrelation: split train/test by geography, not randomly.
State of the art (2026)
- Geospatial foundation models (Prithvi-EO-2.0, TerraMind, Clay, DOFA, AlphaEarth embeddings) are now the default starting point; benchmarks such as GEO-Bench and PANGAEA compare them, and gains over a well-tuned U-Net on a specific task are often modest [4].
- Open data at scale: Sentinel-2 10 m bands, Landsat, Dynamic World-style near-real-time land cover, ESA WorldCover and EU CORINE/HRL products serve as labels and baselines [1].
- Very-high-resolution commercial sensors (sub-metre optical, SAR) enable object-level inventory; satellite tasking and rapid revisit improve timeliness.
- Limits and failure modes: clouds and shadows, seasonal phenology mistaken for change, domain shift between regions and sensors, mixed pixels at 10 m, label noise, inconsistent definitions of "land use" (use vs. cover), and overconfident maps without uncertainty. VLM answers can hallucinate objects. Always report accuracy per class and area-weighted.
Shared resources & combinations
- Shares remote-sensing + segmentation + STAC data pipeline with VIS-038 (CAP monitoring), VIS-049 (deforestation), VIS-058 (disaster mapping), VIS-061 (InSAR), VIS-042 (methane), VIS-046 (PV roof analysis) and drone inspection VIS-003; one GPU pool (HW3) and one tile store serve all.
- Downstream: DATA-011 (geospatial analytics), PRED-036 (physical climate risk), PRED-027 (flood forecasting) consume the maps as features.
- Defence/intelligence use of the same technology: see VIS-051 (separate policy treatment).
Effort, cost & KPIs
- Path A: weeks; subscription or per-km2 pricing; very-high-resolution imagery is the main cost.
- Path B: 2-4 months for a production pipeline; compute and imagery licences dominate.
- Path C/D: 3-9 months incl. labelling and validation; ops for storage and GPUs.
- KPIs: overall and per-class accuracy / F1 (area-weighted, independent sample), change detection precision/recall, false-alarm rate per 100 km2, analyst minutes per verified change, revisit latency (acquisition to alert), cost per km2 per update.
People & change
Who and when: management sets the goal and budget; the department lead (planning, agriculture, asset management) defines which changes matter; data owners of cadastre and GIS layers release data early; IT builds the pipeline; the DPO checks only when high-resolution imagery can show people or private plots. Involve end users (analysts, inspectors) from the first label workshop.
Works council: usually no co-determination, because the system analyses land and assets, not employee behaviour (§87(1) no. 6 BetrVG). It becomes relevant if outputs are used to monitor staff, for example vehicle or crew tracking. The works council should still be informed in good time under §90(1) no. 3 BetrVG when work processes change; public bodies involve the Personalrat.
Data protection: a DPIA is rarely needed, but check imagery below roughly 0.3 m and drone footage; conclude an AVV (Art. 28 GDPR) with every cloud provider.
AI literacy: support staff with short training on map errors, confidence values and seasonal effects (Art. 4 AI Act, GUIDE-006).
Communication and pilot: start with one region or asset class, show analysts side-by-side maps and accept their corrections (GUIDE-009, GUIDE-008).
Typical concerns: fear of wrong sanctions or claims decisions. Address them by treating outputs as alerts, never as decisions: a human reviews every change polygon before action.
Risks, regulation & compliance
- EU AI Act: typically minimal risk. It becomes high-risk if outputs feed Annex III decisions (e.g. critical infrastructure safety components, access to essential public services or benefits such as subsidy sanctions on individuals); keep a human decision and document it. Surveillance of identifiable persons is out of scope here; see VIS-051 and the Art. 5 limits.
- GDPR: 10 m satellite data is not personal data; sub-metre aerial/drone imagery can show people, vehicles, and private plots, and linking parcels to owners creates personal data. Apply data minimisation, legal basis (often public task, Art. 6(1)(e)), blurring where possible, and a DPIA for systematic monitoring of individuals' land.
- Data licences and INSPIRE: respect imagery licences (commercial use, redistribution) and the INSPIRE directive / national open-data rules for authority data; Copernicus data are open under the EU Copernicus data policy [2].
- Quality risks: wrong maps used for enforcement or taxes cause legal exposure; mitigate with human verification, accuracy statements and audit trails.
- Dual use and export controls: very-high-resolution commercial imagery may carry licence restrictions in some regions.
Getting started
- Define the decision the map supports, class legend, minimum mapping unit and required accuracy.
- Pull a free Sentinel-2 time series for a pilot area via the Copernicus Data Space or STAC.
- Baseline against an existing product (CORINE, ESA WorldCover) before training anything.
- Fine-tune one open foundation model (Prithvi-EO-2.0 or TerraMind) on 200-500 labelled polygons; split by geography.
- Run a 3-month shadow pilot with analyst review; measure false alarms and review effort.
Sources
- Copernicus Land Monitoring Service (CORINE, High Resolution Layers) - https://land.copernicus.eu/
- Copernicus Data Space Ecosystem - https://dataspace.copernicus.eu/
- Google Earth Engine - Satellite Embedding dataset (AlphaEarth Foundations) - https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL
- IBM/NASA Prithvi-EO-2.0 on Hugging Face - https://huggingface.co/ibm-nasa-geospatial
- TorchGeo - https://github.com/microsoft/torchgeo
- ESA WorldCover - https://esa-worldcover.org/
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj