Ground Truth for Maps: What Can Only Be Verified on Site

Ground truth is data collected through direct, on-site observation or measurement, used as the reference layer against which remotely sensed imagery, modeled geodata, and crowd-sourced maps are checked for accuracy. In GIS and remote sensing, the term has a precise, technical meaning: a pixel's classification, a geocoded coordinate, or a database attribute is only as good as the field measurement it was validated against data collected on the ground through direct observation or measurement, which is used as a reference to evaluate remote sensing products or GIS-derived models . Esri's own GIS dictionary defines it even more narrowly, as the accuracy of remotely sensed or mathematically calculated data based on data actually measured in the field .

That definition matters more than it used to. The geospatial analytics market was worth an estimated $103.75 billion in 2025, growing to $122.96 billion in 2026 at a compound annual growth rate of 18.5% , and Europe's GIS market specifically is projected to climb from roughly USD 3.0 billion in 2025 to USD 5.0 billion by 2034, a CAGR of 5.53% . Every one of those billions is spent assuming the underlying map is correct. Most of the time, for most features, it is close enough. But a meaningful share of what delivery platforms, navigation apps, and infrastructure operators need to know cannot be derived from any sensor, register, or model. It has to be seen, by a person, standing there.

This article is a reference for exactly that boundary: what imagery and open data can confirm, what they structurally cannot, and how field verification closes the gap in a way that holds up in the EU's address and spatial-data regulatory frame, with notes on where other markets diverge.

What Does "Ground Truth" Actually Mean in a Mapping Context?

The phrase gets used loosely in marketing copy, so it is worth being precise. Ground truth is information collected by direct observation at a known place and time, used as the reference standard that a sensor, map, or model is checked against . It is not a synonym for "accurate" or "official." A government register can be wrong. A satellite image can be current and still misleading. Ground truth is specifically the independent, on-location measurement used to test everything else.

This distinction has teeth in accuracy assessment. Researchers must visit the actual location and physically measure it at ground level, then compare this information with the data collected from remote sensing to determine its accuracy . The output of that comparison is typically an error matrix — overall accuracy, producer's accuracy, user's accuracy — not a single pass/fail score. A dataset can be 94% accurate for building footprints and 60% accurate for business names at the same address, because those are different claims requiring different verification.

Positional accuracy refers specifically to how closely a recorded coordinate matches the true location of a feature within a spatial reference system, a formal definition carried in ISO 19157 accuracy of the position of features within a spatial reference system . That is one of several quality dimensions — completeness, logical consistency, thematic accuracy are others — and ground truth is the only mechanism that can test all of them simultaneously, because it is the only data source not derived from the thing being tested.

Why Can't Imagery and Public Registers Settle This on Their Own?

Satellite and aerial imagery answer a narrow question well: what is the shape, extent, and spectral signature of a feature at the moment of capture? They cannot answer what is inside a building, whether a business behind a façade is trading, or which of several doors on a block is the actual customer entrance. A methodology guide for remote-sensing validation puts the limit plainly: a satellite product without ground truth is an unvalidated hypothesis: it may look plausible and still be wrong . Timing compounds the problem. Ground truth must correspond to the imagery's date and conditions — a vegetation point collected in March cannot validate an August scene . The same logic applies directly to commercial geodata: a photograph of a storefront taken eighteen months ago cannot validate whether that retailer is still operating today. Public open registers face a related but different limit. They are frequently excellent at resolving the question "does this address exist," and structurally weak at resolving "what currently occupies it." France's Base Adresse Nationale is the clearest documented example of this split.

How Do National Address Systems Differ Across the EU?

INSPIRE, the EU directive underpinning spatial data infrastructure across member states, defines the Addresses theme narrowly, as the location of properties based on address identifiers, usually by road name, house number, postal code . That definition itself signals the boundary: an address record is a locator, not a business directory, and INSPIRE's implementing rules require member states to report dataset quality using ISO 19157 categories such as completeness and positional accuracy Data quality information shall be provided in accordance with DIS 19157...including selection of appropriate default data quality elements, sub-elements: Completeness – Commission, Completeness – Omission...Positional accuracy . In practice, each country has built a different implementation on top of that frame.

Country	Core address/geodata layer	Openness & update model	Known structural limit
France	Base Adresse Nationale (BAN)	Open, government-maintained, with communes legally required to feed local address bases into the national one, published as CSVs updated monthly	Resolves to street number, not individual unit or business — BAN only handles addresses, it doesn't geocode business names, landmarks, or POIs; for those you need a commercial provider
Germany	Address/cadastral data (ALKIS/AdV framework)	Maintained at state (Länder) level, with varying open-data terms rather than one national open feed	Fragmentation across 16 states means coverage and refresh cadence differ by jurisdiction
Netherlands	BAG (Basisregistratie Adressen en Gebouwen)	National, open, building- and address-level registry	Strong on buildings and addresses; still does not record occupancy, trading status, or equipment condition inside a unit
Belgium	Regional address registers (e.g., CRAB in Flanders)	Open but organized regionally, not under one unified national schema	Cross-region matching requires reconciliation before use at national scale
Switzerland	Federal address and building register (Register of Buildings and Dwellings)	National, maintained by the federal statistics and cadastral authorities	Non-EU; not bound by INSPIRE, so comparability with EU neighbor datasets needs manual checking

The pattern holds across all five markets: public address infrastructure is a mature, well-governed layer for "where is this location," and a structurally incomplete layer for "what is happening there right now." A national address register is not a stylistic shortcut away from field verification — it is a different data product answering a different question.

What Can Only Be Verified On Site?

Once the address layer is solved, the remaining questions fall into a short, recurring set that no remote source answers reliably. Each one shows up constantly in delivery routing, navigation platforms, and infrastructure datasets.

• Operational status. Whether a business behind a given address is currently trading, permanently closed, or relocated — status changes faster than any refresh cycle on a satellite tile or a government register.
• Entrance geometry. Which of several doors, gates, or loading points is the actual usable entrance for a delivery driver or pedestrian, a detail no footprint polygon encodes.
• Multi-occupancy resolution. One street number can conceal several legally distinct tenants. On forums documenting BAN's limits, users have pointed out that an address is listed only to the granularity of the street number, even though that number corresponds to at least seven different structures, five of them apartment buildings . No address database, however current, disambiguates which tenant a delivery or inspection actually needs.
• Floor level and vertical position. Ground-floor versus upper-floor retail, or which unit sits on which level of a mixed-use building, is invisible to 2D positional data and is a separate problem from the indoor-accuracy question we cover in Indoor Positioning Without GPS: The Real Accuracy Numbers.
• Functional condition. Whether equipment — a charging connector, an ATM, a self-checkout kiosk — is physically present and actually works, rather than merely listed, requires someone testing it.
• Access restrictions. Gated courtyards, private roads, and seasonal closures that change a route's feasibility are rarely encoded in any public layer at all.

On-site verification is not a stylistic choice layered on top of good public data. For these six categories, it is a structural requirement: the information simply does not exist anywhere until someone records it in place.

A Worked Example: Verifying Infrastructure That Changes by Location, Not by Date

Consider a national EV charging network. A public register might list a station's coordinates, operator name, and nominal connector type, all correct on the day it was entered. What it typically cannot confirm is whether that specific connector still matches the listed standard, whether the advertised charging capacity is what the hardware actually delivers, or whether pricing posted at the unit matches what the app charges.

Teampl ran a nationwide field program covering more than 1,500 EV charging stations across the Netherlands, spanning urban, suburban, and rural areas rather than sampling major cities alone. For each location, the field team recorded station name, operator, address, number of charging points, connector types, charging capacity, and pricing as a structured dataset, backed by on-site photographic evidence for every record. That photographic layer is what turns a claim into ground truth — a record someone can audit later, not just trust. That evidentiary discipline is the same chain-of-custody logic covered in Photo and Video Evidence in Mystery Visits: Verification That Holds Up, applied to infrastructure instead of customer experience.

How Does a Defensible Field Verification Methodology Actually Work?

A single site visit is an anecdote. A defensible ground-truth program is a sampling design. The methodology literature on remote-sensing validation is explicit that reference samples must be independent of whatever trained or populated the dataset being checked: these sample sets must be independent — if you assess accuracy using points that also trained the model, you measure how well the model memorised, not how well it generalises, and the reported accuracy is optimistically biased . The same logic applies to POI and address verification: checking a dataset against the same crowd-sourced edits that built it tells you nothing about real-world accuracy.

Three elements separate a rigorous program from a spot check. First, a statistically sound sampling frame that covers edge cases — rural locations, multi-tenant buildings, recently opened or closed businesses — not just easy urban samples. Second, temporal matching: a verification visit answers a question about conditions at the time of the visit, and claims about freshness decay fast enough that the gap itself becomes data, a topic covered at length in Mapping Debt: How Fast Indoor Maps Decay, and What Fixing Them Costs. Third, documented chain of custody — timestamped photographs, geotagged coordinates, and auditable field notes — so a reviewer two years later can see exactly what was observed and when. Our own breakdown of how that sampling and routing gets built in practice is in Field Survey Methodology for Geodata: Sampling, Routing and QA.

What Happens When Verification Gets Skipped?

The cost shows up downstream, not at the point of neglect. A navigation app routes a driver to a demolished entrance. A delivery platform's estimated arrival window fails because the "address" is actually a loading dock three streets away. A government dataset reports completeness and positional accuracy metrics under ISO 19157 that look respectable on paper while the attribute that actually matters — is this place open, is this hardware functioning — was never independently checked at all. INSPIRE's quality framework is explicit that completeness, logical consistency, positional accuracy and usability are separate measurable dimensions; a dataset can score well on three and fail silently on the fourth if nobody ever walked the site.

None of this argues against public and crowd-sourced geodata. France's BAN, Germany's cadastral layers, the Netherlands' BAG, and the open-data commitments behind them are genuine infrastructure, and for pure address resolution they are often better than anything a private vendor could justify building from scratch. The honest framing is complementary, not competitive: open registers answer where, field verification answers what and whether.

The Structural Verdict

Ground truth is not a quality-assurance add-on to a mapping pipeline. For status, access, multi-occupancy, functional condition, and anything that changes faster than a dataset's refresh cycle, on-site verification is the only mechanism that produces the fact at all. Field verification is not a stylistic choice but a structural requirement for any geodata product that claims to describe the present, not just the past.

None of that is complicated to say. It is, however, genuinely hard to run well — recruiting field staff who show up, routing them efficiently across rural and urban terrain alike, checking their photos against a schema, catching the ones who faked a visit. Your next hire isn't a vendor. It's a data team.

Want the operational detail behind how programs like this actually get run — recruiting, quality control, chain of custody? That's what we do day to day.

Frequently Asked Questions