SIH26011 · Open-source prototype

From a flat parcel number to an addressable floor A 3D vertical identity for every property

AeroNerds 3D Cadastre proposes a vertical linkage — ULPIN → building → floor → unit — so a citizen can locate their own floor in 3D, while a registrar can adjudicate conflicts with a complete, inspectable evidence trail behind every claim.

🏠Launch the 3D portal 👤Sign in as a citizen 🏢Registrar desk

⚠

Read this first. This is a competition prototype, not a government system. It does not create, amend or recognise an official ULPIN, and it asserts no legal title, tax or encumbrance status. Every linkage it produces is a proposal for review by the competent authority — never a legal determination.

Vertical identity stack — illustrative ● proposed

Unit / volumeU-402

Floor 4FL-4

Floor 3FL-3

Building footprintOSM footprint

Cadastral parcelOSM parcel

Determined by geometry or surveyed evidence

Proposed linkage — needs human confirmation

Blocked and shown as NOT_DETERMINABLE where evidence is absent

What this prototype actually does

Four commitments that shaped every design decision

The hard part of a 3D cadastre is not drawing cubes. It is being able to defend, or retract, any single vertical claim you make.

🔍

Evidence-linked, never guessed

Every floor, height and linkage carries its source, date, licence, processing step and model version. Where evidence is missing the answer is NOT_DETERMINABLE — not an invention.

⚖

Deterministic geometry, probabilistic AI

Containment, overlap, CRS checks and volume tests run as exact geometry operations. AI assists extraction and flags anomalies; it never adjudicates ownership.

💬

Human in the loop, always

Conflicts and low-confidence claims are surfaced to a reviewer with APPROVE, CORRECT, REJECT or MARK_UNRESOLVED — and every action is audit logged.

🔒

Role-scoped by construction

Access is enforced on the server, not hidden in the UI. A citizen can only ever read records linked to their own account, and sees only the fields they are entitled to.

Current build

What is actually loaded in this prototype

These figures are measured directly from the GeoJSON layers shipped in frontend/data/. A regression test asserts that the numbers on this page still match the data files, so they cannot silently drift away from reality.

Pilot regions

Bhopal, Bengaluru, Indore,
Navi Mumbai, Kalyan, Coimbatore

12,887

Building footprints

OSM outlines with 3D attributes

582

Cadastral parcels

Matched 2D parcel geometry

185

AI anomalies flagged

Isolation Forest, routed to review

ⓘ

Stated precisely: the full evidence-fusion output — AI anomaly scoring and the final verification gate — is present for 6,163 buildings (Bhopal 3,429 + Bengaluru 2,734). Within that set: 2,326 VERIFIED, 3,652 PROVISIONAL, 185 HUMAN_VERIFICATION_REQUIRED. The other four regions carry earlier-stage height and floor evidence only, so we do not claim anomaly or verification status for them.

Services

Four role-specific workspaces

Each desk exposes only the capabilities that role is entitled to. The list below is generated from the same capability matrix the API enforces — it is not a decorative list.

🏢

Registrar

Sub-Registrar office · final adjudication

Full read access to every building, floor and parcel. Adjudicates conflicting vertical claims, issues certificates, and owns the audit trail.

building:read:all parcel:read:all review:adjudicate review:annotate certificate:issue manifest:export auditlog:read

Open registrar desk →

🏗

Town Planner

Municipal planning office · zoning

Reads the full 3D stack for planning and zoning studies — floor-area potential, vertical density, and parcel-level development capacity.

building:read:all parcel:read:all zoning:audit review:annotate manifest:export

Open planner desk →

📜

SDM / Grievance Officer

Revenue administration · grievances

Owns the citizen grievance register: triages submissions, tracks resolution state, and escalates unresolved vertical-identity disputes for human review.

grievance:adjudicate review:annotate building:read:all parcel:read:all

Open grievance desk →

👨‍🏫

Citizen

Landowner / co-owner · own records only

Views the 3D stack of their own linked property, checks the proposed vertical identity, downloads a certificate, and files a grievance. Nothing else is reachable.

building:read:own parcel:read:own certificate:view:own grievance:file

Open citizen view →

How it works

From raw open data to a reviewable 3D vertical identity

The full build follows a 17-phase plan. These six stages are the ones that decide whether a vertical claim survives to the reviewer.

Discover & ingest

OSM footprints and parcels, Bhuvan/NRSC elevation, and public basemaps, each with a manifest recording licence, CRS, date and known gaps.

Real data

Match parcels to buildings

Deterministic containment and overlap tests in a validated CRS. Every match records CONTAINED, MAJORITY or an explicit failure.

Deterministic

Reconstruct 3D massing

DEM gives terrain, DSM gives surface. Their difference yields an above-ground height signal — explicitly not a floor count.

Modelled

Extract floors with AI

A U-Net / Mask R-CNN segmentation stage proposes building masks; a tree-based model predicts height, level and footprint quality.

AI-assisted

Fuse evidence & flag conflict

Independent signals are combined into a single status. An Isolation Forest surfaces anomalies for review rather than silently scoring them away.

Human review

Propose the vertical ULPIN

A deterministic identifier chain links parcel → building → floor → unit, marked as a proposal with its verification status attached.

Proposal only

Data provenance

What is real, what is modelled, what is refused

Separating these three categories is the difference between a useful prototype and a misleading one. Anything we could not verify is reported as NOT_DETERMINABLE and is not displayed as a fact.

Layer / attribute	Origin	Standing	Notes
Building footprints	OpenStreetMap (ODbL)	REAL	Outlines and building:levels tags where contributors recorded them.
Cadastral parcels	OpenStreetMap (ODbL)	REAL	Open cadastral geometry. Not a certified survey copy.
Terrain elevation (DEM)	Bhuvan / NRSC	REAL	Ground elevation only.
Surface elevation (DSM)	Bhuvan / NRSC	REAL	Surface elevation. Height signal derives from DSM−DEM.
Building height	Model prediction + evidence fusion	MODELLED	Reported with an explicit confidence state, never as a surveyed value.
Floor count / floor entities	Inference from height, levels tag and imagery	MODELLED	Flagged PROVISIONAL unless independently corroborated.
Floor plans, GNSS/CORS control	Not available in pilot regions	NOT DETERMINABLE	Explicitly out of scope. Not simulated as if present.
Legal title, tax dues, encumbrance	Outside prototype authority	NOT DETERMINABLE	Never inferred. Requires the competent authority.

Limitations

What this prototype cannot do

A prototype that hides its limits is worse than no prototype. These are the boundaries we hold ourselves to.

🚫

It is not a government system

It neither creates nor recognises an official ULPIN. A proposed vertical linkage is a technical suggestion for a competent authority to accept, correct or reject.

🏫

Height is not floor count

DSM−DEM gives an above-ground height signal where the datasets are compatible. It is never presented as an exact floor count. Uncorroborated floors stay PROVISIONAL.

📈

Not all regions are equally complete

Only Bhopal and Bengaluru carry the full AI-anomaly and verification-gate output. The other four regions have earlier-stage height and floor evidence, and are reported as such.

🔑

Demo accounts, demo credentials

The sign-in screen shows its own demo passwords on purpose, for judging. This is not a production identity system, and the stored data is not a government record.

⚖

The one rule that outranks every feature

If the evidence is absent, insufficient or contradictory, the system returns NOT_DETERMINABLE and hands the case to a human. It never fills a gap to make a demo look complete — because a cadastre that guesses is worse than one that admits it does not know.

Frequently asked

Questions we expect from an evaluator

Why can't you just compute the floor count from the building height?

Because storey height is not a constant — it varies with use, structure, era and local building by-laws, and ground floors are often double height. A 12 m building might be three floors or one. We use height as a signal, corroborate it with recorded building:levels tags and imagery evidence, and otherwise report the floor count as provisional or not determinable.

Can a citizen see a neighbour's property?

No. Ownership links are stored server-side and the API filters every read by the signed-in account. The capability check happens in the backend, so hiding a button in the interface is not the security control — it is only a convenience. Attempting to request another building's identifier returns a not-found response rather than the record.

Where does the AI actually help, and what does it refuse to do?

The AI helps with building segmentation, height and level prediction, and anomaly detection. It does not adjudicate ownership, legal rights or disputes — those are deterministic topology checks plus a human reviewer. When the model and the geometry disagree, the system records a conflict instead of picking a winner.

What happens when two sources disagree?

Nothing is silently overwritten. The disagreement is preserved with both sources, both dates, and a CONFLICT or HUMAN_VERIFICATION_REQUIRED status, and it appears in the reviewer's queue with the provenance needed to settle it.

Is the Bhiwandi pilot area in the original charter?

The original charter named Baramati ULB, Pune district, as the pilot because it appears in published NAKSHA material. The Bhiwandi and other regional datasets were added later as separate dataset versions; the original pilot area was retained. Each dataset carries its own manifest and licence.

Can this run offline?

Yes. The portal and every layer are served locally and no public tile server is overloaded. The web fonts degrade gracefully to system fonts if you are disconnected, and the full evidence trail is inspectable in the repository rather than hidden behind a vendor API.