What problem does it solve?
Property registers and local property taxes rest on what owners declare. A pool dug in the garden, an extension or a shed built without a declaration stays invisible until someone surveys the area. Field surveys are slow and expensive, and a national register of millions of buildings falls out of date between surveys. The tax and mapping offices that keep the registers hold exactly the records needed to see the gap: the cadastral plan, permits and declarations. What they lack is a cheap way to look at every plot.
Aerial and satellite images exist in bulk. Countries such as France, Italy, Germany and the Netherlands publish or hold orthophotos at high resolution. Using them to find unregistered construction raises its own questions: what counts as a taxable object, how to handle objects that look alike, and how to stop a model's guess from turning into a tax notice without a person looking at it.
- The French Senate's report on AI and fraud puts the potentially taxable pools in France at about 150,000, around 5 percent of the 3 million in ground pools in the country.IA, impôts, prestations sociales et lutte contre la fraude (rapport d'information n° 491) (2024)
How does it work?
- Get current imagery. Use the state's official orthophotos and, where available, surface height models. DGFiP uses the public aerial images of the national mapping agency IGN, Italy's revenue agency uses orthophotos supplied by AGEA, the Bavarian survey office uses true orthophotos and image based height models, and The Hague uses 2.5 centimeter pixel photos.
- Extract objects. A deep learning model labels pixels as building, pool or other and turns them into outlines. Height models help to spot changes such as an added storey.
- Compare with the register. The outlines are matched with the cadastral plan and, for tax, with declarations, planning permissions and completion notices. A shape without a matching record becomes a candidate.
- Verify. An officer looks at every candidate on screen, with the image and the register side by side, and confirms or rejects it. Italy's agency, The Hague and DGFiP all describe this step as part of the process.
- Act and learn. For a confirmed object the owner is invited to regularize, the register is updated and the tax is recalculated. Rejections go back to improve the model.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Revenue recovered | Not pooled | about EUR 10 million | 1 | 1 independent |
Value drivers: Compliance quality, Lower cost to serve, Employee productivity.
Indicative value
A regional tax authority responsible for 2 million residential properties
EUR 150,000 to EUR 7.5 million
Additional local tax from newly found objects per year
How this is calculated
Formula: properties * foundShare * taxPerObject. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Residential properties in scope properties, properties | 1,000,000 | 3,000,000 | Editorial assumption. Replace with the size of your own register. |
| Share of properties with a newly found taxable object foundShare, fraction of properties | 0.001 | 0.005 | Editorial assumption with no sourced count of properties. The French Senate report puts potentially taxable pools at about 5 percent of in ground pools, but it does not give a share of all properties, so this range is our own estimate. Source |
| Additional local tax per newly found object taxPerObject, EUR per object | 150 | 500 | The Senate report gives close to 10 million euros from more than 20,000 pools in the 2022 pilot, about 500 euros each; it does not say whether this includes back taxes. The low end is an editorial assumption. Source |
What it leaves out: The result is the extra tax assessed, not tax collected, and a first campaign may include back taxes that later years will not. It leaves out the cost of imagery, the model, the verification team and appeals, and the false candidates that officers reject.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Agenzia delle Entrate
Italy · Government and public sector · 2025
Italy's revenue agency compares high resolution orthophotos from AGEA with the cadastral maps to find buildings that are not declared in the cadastre. In some areas AI models that use semantic segmentation detect the building outlines automatically; the outlines are compared with the registry geometries to highlight discrepancies. The results appear on a photo interpretation platform where the agency's technical staff verify, confirm and classify each report proposed by the AI.
The agency's director told a parliamentary oversight committee in October 2025 that the work covered about 65 percent of the country, with the first phase due to finish in 2026. The same statement says AI is used only in the pre investigation phase, to identify taxpayers for checks, and that algorithms never decide alone. No yield figures are given.
No outcome disclosed.
Direction générale des Finances publiques (DGFiP)
France · Government and public sector · 2023
The French public finance directorate launched Foncier innovant in 2022. A deep learning model extracts the outlines of buildings and swimming pools from the public aerial images of the national mapping agency IGN. The outlines are compared with the cadastral plan and with DGFiP's own data, such as planning declarations and building permits, to find constructions that are not taxed.
A tax officer checks every potential anomaly in the case system, and a confirmed owner is invited to regularize through the tax portal. Only the public IGN images are processed on Google's cloud; the match with tax data runs on DGFiP infrastructure. The 2022 pilot in nine departments was extended to metropolitan France in 2023. After the first national campaign, more than 120,000 owners were contacted in June 2023 to regularize pools, according to a National Assembly report.
The same report quotes DGFiP saying that more than 60 percent of automatic detections are false positives, which the checking agent rejects. The Senate report gives a project cost of 24 million euros, half of it financed by the public transformation fund, and judged it already paid back; the National Assembly report gives 35.17 million euros. The Senate report names garages, sheds and verandas as possible next targets.
- Revenue recovered: about EUR 10 million, pilot in nine departments, 2022
"L’expérimentation menée sur 9 départements pilotes en 2022 a permis de confirmer le caractère taxable de plus de 20 000 nouvelles piscines, représentant près de 10 millions d’euros de recettes supplémentaires pour les communes (taxe foncière)."
Claimed by: independent
Gemeente Den Haag
Netherlands · Government and public sector · 2021
The municipality of The Hague uses ChangeMapp, which the algorithm register describes as AI that analyzes aerial photos for changes to buildings and the outdoor space around them.
The application compares the registered building geometry in the national address and building register (BAG) and the large scale topography register (BGT) with aerial photos at 2.5 centimeter pixel resolution and draws a colored area on the photo where something differs from the map, for example a shed placed in the past year. An employee always checks the results and decides whether the geometry in the registers must change.
The municipality says the tool is a large efficiency gain in an existing process that used to be manual; it publishes no figures. The municipality lists a legal duty to keep these registers current, and in consultation with the privacy officers it was decided that no data protection impact assessment was needed because individuals cannot be recognized in the photos.
No outcome disclosed.
Bavarian State Office for Digitization, Broadband and Surveying (LDBV)
Germany · Government and public sector · 2022
Bavaria's state surveying office funded the Technical University of Munich to build a system that finds buildings missing from the Digital Cadastral Map and changes to recorded buildings, such as added storeys, across the whole state. A machine learning workflow learns building patterns from the state's official true orthophotos and image based surface height models, then compares its results with the cadastre to produce change candidates. Outputs are designed to be interpretable so that people can verify them.
According to the university project page, the system is integrated into the office's internal production environment and runs in a semi annual monitoring workflow, so that surveyors inspect targeted candidates instead of reviewing the whole state by hand. No accuracy or volume figures appear on the project page.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Orthophotos with a known capture date, ideally with a surface height model
- Cadastral and building register geometry in digital form
- Permit, declaration and completion records linked to plots
- A set of objects confirmed or rejected by staff, for training and measurement
Systems to integrate
- Cadastre and tax case management system
- GIS or photo interpretation workbench for verification
- Letter or tax portal channel to invite owners to regularize
Complexity: Medium
The build needs current orthophotos, the cadastral geometry in digital form, the tax roll with permits and declarations, a place to run the model and a verification team that can look at several thousand candidates. DGFiP processes only public IGN images on a commercial cloud and runs the match with tax data on its own infrastructure.
- 1
Start with one object type
DGFiP started with swimming pools, a clear shape that is easy to confirm on a photo, tested the approach in 9 departments in 2022 and extended it to metropolitan France in 2023. Add sheds, garages and verandas once pools work.
- 2
Fix the verification rule before the first run
Decide who checks a candidate, on which screen and with what evidence. DGFiP agents check every potential anomaly in the case system, Italy's agency staff verify, confirm and classify each report, and The Hague employees review each result before a register changes.
- 3
Keep sensitive data out of the cloud
Send only public imagery to external compute. Match shapes with tax or personal data inside your own infrastructure, as DGFiP does.
- 4
Measure precision by area and object type
Count candidates that officers reject. In press coverage of the pilot a finance workers' union claimed errors of up to 30 percent in some places and DGFiP replied that the rate varied by department and the tool was still in test. A 2025 National Assembly report quotes DGFiP saying that more than 60 percent of automatic detections are false positives, which agents reject. Track rejection reasons and retrain.
- 5
Plan coverage and refresh cycles
Italy's agency reported about 65 percent of the national territory covered in 2025 and aims to complete the first phase by 2026. Bavaria runs its monitoring every six months. Set a cycle that matches how often imagery is refreshed.
Guardrails
- Every candidate is verified by a person before a register or a tax notice changes
- The owner can contest before any tax is assessed
- Imagery resolution does not allow recognition of individuals, and the use is documented
- Model outputs are stored as candidates with image, date and version, not as findings
- Scope stays on objects; ranking owners by fraud risk needs a separate assessment
KPIs to instrument
- Share of candidates confirmed by officers, by area and by object type
- Additional tax assessed per confirmed object, and the share contested
- Officer minutes per candidate
- Territory covered per cycle and age of the imagery
- Rejection reasons, such as temporary or lookalike objects
Human in the loop
Tax officers or surveyors review every candidate against the image and the register, decide on the change and send the invitation to regularize. The model proposes shapes; it does not assess tax or name an owner.
Common failure modes
- Lookalike and temporary objects
- Objects that look like a pool or a building but are not taxable, such as temporary structures, produce false candidates. A finance workers' union cited error rates of up to 30 percent in some places in the DGFiP pilot, which DGFiP disputed. Verification capacity and a rejection log keep the cost of a false candidate low.
- Imagery older than the permit
- A building that was permitted after the photo was taken looks undeclared. Match capture dates to permit dates before a candidate is raised.
- Scope creep to people
- A system that finds objects can be extended to score owners. That changes the legal basis and the risk class. Keep the object use case separate.
- Verification queue bigger than the team
- A model that flags tens of thousands of candidates creates a backlog. Release candidates in batches that the team can verify within the cycle.
What are the risks and rules?
EU AI Act
Minimal risk
Finding buildings and pools in images and checking them against a register is not a use listed in Annex III, and Recital 59 says systems used by tax and customs authorities in administrative proceedings should not be treated as high risk law enforcement systems. The tier rests on that scope, not on the human check. GDPR still applies to owner records and to the tax data used in the match, and the tier would change if the system were extended to rank natural persons by fraud risk.
Rules that apply
Guidance
- IA, impôts, prestations sociales et lutte contre la fraude (rapport d'information n° 491) (Sénat, délégation à la prospective, Europe). Describes how Foncier innovant works, who verifies the output and where data is processed, with the cost and results of the pilot, in French.
- Algoritmeregister entry for ChangeMapp, Gemeente Den Haag (Gemeente Den Haag (Algoritmeregister), Europe). Shows the impact assessment reasoning, the human review step and the legal duty to keep registers current for a municipal deployment.
- Rapport d'information n° 1594 sur les dysfonctionnements dans la gestion des impôts locaux et leurs conséquences (Assemblée nationale, commission des finances, Europe). Evaluates Foncier innovant, including the first national campaign in 2023, the false positive rate DGFiP reports and the extension beyond pools, in French.
Controls to put in place
- Documented rule for human verification of every candidate
- Rejection log reviewed each cycle with precision by area and object type
- Separation of public imagery processing from tax and personal data
- Model version and image date stored with each candidate
- Impact assessment before any extension to a new object type or to ranking owners
When it went wrong elsewhere
- Union claims of high error rates in the French pool detection pilot. Reports that the CGT union said the tool had errors as high as 30 percent in some places, and DGFiP replied that rates varied by department and were still in the test phase.
Frequently asked questions
- Which public bodies use AI to find undeclared buildings?
On this page, France's DGFiP runs it for pools, Italy's revenue agency compares orthophotos with the cadastre and uses AI segmentation in some areas, the Bavarian survey office runs a state wide monitoring system for its cadastre, and the municipality of The Hague uses ChangeMapp for its building and topography registers.
- How much money did it bring in?
The French Senate reports close to 10 million euros of extra local tax from more than 20,000 newly confirmed pools in the 2022 pilot in nine departments. Another passage of the same report says 23,000, and it estimates the national rollout at 43 million euros, which is a forecast.
The ministry's press release on the 2023 activity report lists the detection of more than 120,000 undeclared taxable pools among its 2023 results, without saying whether the count is cumulative, and a National Assembly report says more than 120,000 owners were contacted in June 2023 after the first national campaign. Neither number is money collected.
The Senate judged the project, which it puts at 24 million euros, already paid back, while the National Assembly report gives a total cost of 35.17 million euros. Other agencies on this page publish no yield figures.
- Does a person check the results?
Yes in the three deployments that state it. DGFiP agents verify each potential anomaly, Italy's agency staff verify, confirm and classify each report proposed by the AI, and The Hague employees check every result before a register is changed. The Bavarian survey office describes its outputs as supporting human verification and giving candidates for targeted inspection, so a check is implied there rather than described.
- Does it need satellite images?
Mostly not. The deployments on this page use aerial orthophotos: The Hague's photos have 2.5 centimeter pixels, and Bavaria combines true orthophotos with image based height models. The French Senate report does say its algorithm extracts building outlines from satellite images in its deep learning paragraph, while describing the images DGFiP uses as public aerial images from IGN.
How to cite this page
Blits.ai AI Use Case Library, "AI detection of undeclared buildings and pools from aerial imagery for property tax and cadastre updates", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/undeclared-building-detection-from-aerial-imagery. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published