Skip to content

Property-level results — feasibility note (Phase 17 spike S1)

Status: drafted 2026-09-28 for spike tracker #1463 (plan item S1, design record D18 in phase-17-revamp.md). The owner asked on 2026-09-26 to "investigate whether we can go down to property level" (market-scan comments C98, C99, C100, C105). This is desk research: no footprint ingest, no per-building compute and no UI is proposed for Phase 17. The build is the Phase 18 charter stub. Every licence, figure and resolution below carries its source URL or path:line (paths are relative to the repository root at origin/main f4567feb); anything unsourced is marked [ASSUMPTION]. The licence readings are readings of the quoted text, not legal advice.

Verdict

Property-level results are feasible in Phase 18, but not by pointing today's pipeline at building footprints. Three findings bind the build:

  1. Licence — every usable European footprint source (OSM, Overture buildings, EUBUCCO, Microsoft footprints, the JRC's DBSM) is ODbL. Maps and PDF reports are Produced Works (attribution plus an offer of the method). Any per-building data the platform derives from footprints and shows or exports to a tenant must itself be ODbL. The tenant's own values stay theirs only if they are kept as an independent layer that uses no footprint data.
  2. Hazard resolution — the finest hazard grid in the catalog is 3″ (≈ 93 m × 73 m at 38° N, JRC river flood). The worker gives every exposure point the value of the one pixel that contains it, so all buildings in a pixel get the same depth. A flood result is defensible per building only on building-scale hazard (England's official maps, one of the JRC's validation benchmarks, are built on DEMs of at least 5 m). On the 3″ grid a per-building flood figure is a pixel value and must be labelled so.
  3. Value — no ingested exposure carries a per-building value. LitPop is 150″ (≈ 4.6 km) and GHSL is 30″ (≈ 1 km) at best as ingested (3″ native). The defensible value source is the tenant's own schedule of values through the existing CSV exposure upload. The modelled fallback is footprint area × storeys × a sourced €/m² replacement cost (Huizinga 2017, already in the repo), labelled "modelled".

What "property level" means here

The owner's ask covers per-building results, editable homes and street-level photos. The Phase 17 prototype (claude.ai/artifact/1TXwPDAiGY3Pj1tqB9bBN5, js/views/lookup.js) shows the target as a labelled Phase 18 preview: Lookup → "Edit home" and "Street photo". Today an exposure is a POINT with a scalar value and a free properties bag (src/climate_lama/models/exposure.py:165-172). There is no footprint geometry and no per-asset impact-function binding. That binding already moved to Phase 18 (D21 in phase-17-revamp.md).

1. Licence outcome for footprints (OSM, Overture, EUBUCCO under ODbL)

The in-repo input is docs/DATA_SOURCES.html. It rejects NC/SA sources from the product path and records ODbL share-alike as "risk … deferred to an explicit license review" (docs/DATA_SOURCES.html:128). It parks "Microsoft ML Building Footprints + OSM/Geofabrik Greece" (ODbL) for that review (docs/DATA_SOURCES.html:361), and lists "ODbL propagation semantics for footprint-derived products" as an open question (docs/DATA_SOURCES.html:499, :506). This section answers that question for the property-level build.

1.1 Licence per footprint source (verified 2026-09-28)

Source Licence Evidence (quoted) Greece fit
OpenStreetMap ODbL 1.0 "OpenStreetMap is open data, licensed under the Open Data Commons Open Database License (ODbL) by the OpenStreetMap Foundation (OSMF) … If you alter or build upon our data, you may distribute the result only under the same license." — openstreetmap.org/copyright Coverage uneven (see EUBUCCO row)
Overture Maps — buildings theme ODbL "Buildings — License for theme: ODbL © OpenStreetMap contributors. Available under the Open Database License. Esri Community Maps contributors. Available under CC BY 4.0. Global ML Building Footprints. Licensed by Microsoft under the Open Database License. Google Open Buildings. Available under CC BY 4.0." — docs.overturemaps.org/attribution Pulls in the Microsoft footprints, the best Greek source per docs/DATA_SOURCES.html:361
EUBUCCO v0.1 ODbL (46 of 50 input datasets; Prague CC-BY-SA, Abruzzo CC-BY-NC) "we decided to apply ODbL to all possible datasets … we could harmonize the license for 46 out of 50 datasets, which represent more than 95% of the buildings" — Milojevic-Dupont et al. 2023, Sci Data 10:147, doi:10.1038/s41597-023-02040-2 Weak: "in Greece, we have ~870 thousands buildings for a population of ~10 million … the coverage in Greece is likely too be overall very low" (same paper)
Microsoft GlobalMLBuildingFootprints ODbL "ODbL, stated in the repo license" — docs/DATA_SOURCES.html:361; Overture's attribution page above ("Licensed by Microsoft under the Open Database License") Best available Greek footprints (docs/DATA_SOURCES.html:361)
JRC DBSM R2025 (EU Digital Building Stock Model) ODbL 1.0 (authoritative inputs keep their own licences) "The data is distributed under ODbL-1.0 license, and the authoritative sources included retain their original licenses" — Martínez et al. 2025, JRC142133, doi:10.2760/0629989 "MSB [Microsoft Building footprints] area takes a predominant role in countries like Greece"; per-building height, use, construction epoch (same report)

No footprint source for Greece avoids ODbL. Google Open Buildings is CC BY 4.0 but has no European coverage (docs/DATA_SOURCES.html:361). The one licence-clean building-scale layer is GHS-BUILT-S E2018 at 10 m, a raster of built-up surface rather than per-building identity. Its directory GHS_BUILT_S_E2018_GLOBE_R2023A_54009_10 carries the same EU copyright notice that licenses the GHSL products under CC BY 4.0 (read live 2026-09-28, https://jeodpp.jrc.ec.europa.eu/ftp/jrc-opendata/GHSL/GHS_BUILT_S_GLOBE_R2023A/GHS_BUILT_S_E2018_GLOBE_R2023A_54009_10/copyright.txt).

1.2 What ODbL 1.0 requires — the clauses this outcome rests on

Quoted from the licence text, opendatacommons.org/licenses/odbl/1-0 (plain text: odbl-10.txt, read 2026-09-28):

  • "Publicly" — "means to Persons other than You or under Your control by either more than 50% ownership or by the power to direct their activities". A tenant of a managed deployment is a Person other than the operator, so serving a tenant is public use.
  • "Use" — "includes without limitation distributing, copying, publicly performing, publicly displaying, and preparing derivative works of the Database". Showing data in a web UI is Use.
  • "Produced Work" — "a work (such as an image, audiovisual material, text, or sounds) resulting from using the whole or a Substantial part of the Contents (via a search or other query)".
  • §4.3 Notice for using output — "if you Publicly Use a Produced Work, You must include a notice … that Content was obtained from the Database … and that it is available under this License."
  • §4.4 Share alike — "Any Derivative Database that You Publicly Use must be only under the terms of: i. This License; ii. A later version …; or iii. A compatible license."
  • §4.4 c — "A Derivative Database is Publicly Used and so must comply with Section 4.4. if a Produced Work created from the Derivative Database is Publicly Used."
  • §4.5 a / b / c — Collective Databases need not be ODbL "but this License still applies to this Database … as a part of the Collective Database". "Using this Database … to create a Produced Work does not create a Derivative Database". "Use of a Derivative Database internally within an organisation is not to the public".
  • §4.6 Access — "If You Publicly Use a Derivative Database or a Produced Work from a Derivative Database, You must also offer … a. The entire Derivative Database; or b. A file containing all of the alterations made to the Database or the method of making the alterations to the Database (such as an algorithm)".

OSMF community guidelines (board-endorsed) settle the grey zones:

  • Produced Work guideline (endorsed 2014-06-06): "If the published result of your project is intended for the extraction of the original data, then it is a database and not a Produced Work." Usually Produced Works: ".PNG, JPG, .PDF, SVG images and any raster image". The same page adds: "if you publish a produced work, the underlying database has to be published as well … according to section 4.6".
  • Collective Database guideline (endorsed 2016-06-17): a combination stays Collective, with no share-alike on the non-OSM part, when "the non-OSM and OSM datasets do not reference each other", or when "a non-OSM database replaces or adds a property of a primary feature, and uses either all OSM data or no OSM data for that property". A reference "by a database key or any other method of identifying a specific OSM … element that may be used with a database join" counts as a reference.
  • Attribution guideline: "© OpenStreetMap contributors" is acceptable. On an interactive map the credit goes "in a corner of the map". For "a printed map and similar media (that is ebooks, PDFs and so on) … The URL to openstreetmap.org/copyright must be printed out."
  • Substantial guideline: "Less than 100 Features" in a one-off extraction is insubstantial. A national building layer is far past that, so this exemption never applies.

1.3 What that means for results and reports

Output of the property-level build ODbL class Obligation
Map tiles / map canvas drawing footprints (UI) Produced Work (§4.5 b) Corner credit "© OpenStreetMap contributors" (plus Overture/Microsoft/DBSM credits when used) linked to the licence (§4.3; attribution guideline)
PDF report page or PNG map showing footprints or per-building results Produced Work Printed credit and the printed URL openstreetmap.org/copyright (attribution guideline). If it was produced from a derivative footprint database, that database is publicly used (§4.4 c) and must be ODbL, with §4.6 access
Per-building values or losses computed from footprint geometry or attributes (area, height, levels), shown in the UI, returned by the API or exported as CSV/GeoJSON Derivative Database, Publicly Used ODbL for that layer (§4.4), notices (§4.2), and an offer of the database or "the method of making the alterations … (such as an algorithm)" (§4.6 b)
The tenant's own schedule of values (sum insured, occupancy) held beside footprints Collective Database if the tenant's properties use no OSM data (Collective Database guideline) None on the tenant's layer; the footprint layer stays ODbL (§4.5 a)
A self-hosted deployment where the customer ingests footprints for its own staff Internal use (§4.5 c) Share-alike does not bite until the customer publishes

1.4 Licence outcome

Accept ODbL footprints in Phase 18 under four rules:

  1. Footprints form a separate ODbL layer. Footprints are never merged into the tenant exposures rows. They live in their own layer with license="ODbL-1.0" and the attribution recorded per artefact, as docs/DATA_SOURCES.html:128 already demands.
  2. Per-building figures derived from footprints are published as ODbL. Their derivation method is published too (§4.6 b is met by a public methods page). They carry no tenant data.
  3. Tenant values stay out of the footprint layer. Tenant losses are computed at the tenant's own geocoded point and value. Footprints only draw the building (a Produced Work), which keeps the tenant's data out of share-alike.
  4. Every map and report carries the ODbL credit. The UI map gets the corner credit, and PDFs print the credit and the copyright URL.

[ASSUMPTION] Rule 3 needs sign-off. It rests on a reading of the Collective Database guideline: a spatial join that picks the footprint for display only does not make the tenant's loss figure "OSM data". The licence-review session that docs/DATA_SOURCES.html:499 defers to must ratify it, or counsel must, before any tenant-facing per-building export ships.

2. Hazard resolution for defensible per-building results

2.1 What the catalog holds (verified from the ingest manifests)

native_resolution is stored per hazard dataset in degrees per pixel (src/climate_lama/models/hazard.py:211-214). The finest value declared by any ingest manifest is 3″:

Hazard source (as ingested) Native resolution Source
JRC / CEMS-EFAS river flood, historical RP maps 3″ (0.000833333°), "~90 m" scripts/historical_catalog_manifest.json:22, :24
WRI Aqueduct river flood, scenario RP maps 30″ (0.008333°), ≈ 1 km scripts/scenario_hazards_manifest.json:22, :24
USGS ShakeMap (Samos 2020 replay) 1/60° node spacing scripts/historical_catalog_manifest.json:60, :62
NASA FIRMS VIIRS active fire "Nominal 375 m fire pixels" scripts/historical_catalog_manifest.json:135
CLIMADA API tropical-cyclone windfields 150″ docs/DATA_SOURCES.html:312 ("150 arc-second")

Pixel size at 38° N (central Greece), computed from those declarations: 3″ ≈ 92.8 m × 73.1 m (≈ 6,800 m²); 30″ ≈ 928 m × 731 m (≈ 0.68 km²).

2.2 How an exposure point meets a hazard cell today

The worker uses cell membership: "each exposure's pixel (col, row) is computed by inverting the affine transform", and that pixel's intensity column is the exposure's (src/climate_lama/core/centroid_assignment.py:7-14). So every building whose point falls in the same 3″ pixel receives the identical depth. The hazard grid, not the footprint, sets the spatial detail of a per-building flood result. Risk-surface cells stop at H3 r8, and finer-than-r8 truth "stays in the hazard COG and is read via a windowed lookup at query time" (src/climate_lama/models/risk_surface.py:40-43). Per-building reads would therefore use the COG path, not the DB cells.

2.3 The minimum — sourced

  • Grid resolution in urban flooding is bounded by the buildings themselves: "critical length scales related to building dimensions and building separation distances exist in urban areas that determine maximum possible grid resolutions for hydraulic models of urban flooding". Source: Fewtrell, Bates, Horritt & Hunter 2008, Hydrological Processes 22(26):5107–5118, abstract at fathom.global.
  • The paper behind the repo's own 3″ grid describes it as continental-scale. Dottori et al. 2022, ESSD 14:1549–1569, essd.copernicus.org/articles/14/1549/2022, on its "high-resolution (100 m) hazard maps":
    • "the flood maps developed do not account for the influence of local flood defences";
    • one validation benchmark, England's official maps, "are constructed using DEMs of at least 5 m resolution";
    • the findings "highlight the need for a high-resolution DEM fed with local-scale information to achieve adequate performance in lowland areas".

Minimum for a defensible per-building flood result: a building-scale hazard grid, meaning the ≤ 5 m DEM class of England's official maps, one of the JRC's validation benchmarks. The catalog's 3″ grid is roughly 15–19× coarser on a side (73–93 m against 5 m). On it, a per-building flood number is defensible only as "the value of the ≈ 90 m cell this building sits in", with the cell size shown next to it.

[ASSUMPTION] For earthquake, heat, windstorm and wildfire, the non-flood hazard grids (section 2.1 lists 375 m to 150″) are taken to vary slowly at building scale. On that premise, per-building differentiation for those perils comes from vulnerability (building type, height, epoch) rather than from a finer hazard grid. No source was fetched for this in the spike, so Phase 18 must source it per peril before a per-building result for those perils is labelled anything above screening grade.

3. Exposure source that would carry building value

3.1 Why the ingested exposures cannot

  • LitPop is a 150″ grid ("LitPop's 150 arcsec cell is ~4.6 km", src/climate_lama/core/display_names.py:38-39). It is a wealth-weighted population proxy, not a building value.
  • GHSL is ingested at 30″ (≈ 1 km) at best (section 4). GHS-BUILT-V is m³, "not currency", with no sourced €/m³ factor (docs/DATA_SOURCES.html:344).
  • The CSV upload already takes per-point values: lat, lon and value are required (src/climate_lama/core/exposure_service.py:43), and every other non-empty column lands in properties (src/climate_lama/core/exposure_service.py:306-309). What it lacks is footprint geometry (POINT only, src/climate_lama/models/exposure.py:165-167) and a per-asset curve.

3.2 Candidates

Candidate What carries the value Licence Verdict
Tenant schedule of values (insurer portfolio, public-sector asset register) via the existing CSV upload The tenant's own sum insured / replacement value per location Tenant-owned; no third-party licence Primary. Defensible because it is the tenant's own figure; ADR-077's real-ingest rule holds
Modelled value = footprint area × storeys × country €/m² Footprint area (DBSM / Overture / OSM), storeys from DBSM height or OSM building:levels, and Huizinga et al. 2017 max structural damage (Greece 440.58 €/m², Cyprus 446.996 €/m², 2010 prices) ODbL for the footprint-derived part (section 1.4, rule 2); Huizinga factor per scripts/packs/country_profile.py:151 (Greece), :209 (Cyprus) Fallback for citizens and portfolios without values, labelled "modelled"
GHS-BUILT-S × Huizinga €/m² (already built by scripts/packs/ghsl_exposure.py) Built-up surface m² per ≈ 1 km output cell CC BY 4.0 Not per building (cell aggregate). Stays the licence-clean area-level exposure
EUBUCCO attributes Height, type, construction year ODbL Greek coverage "likely too be overall very low" (section 1.1). Superseded by DBSM R2025 for Greece

[ASSUMPTION] The fallback's storey multiplier assumes Huizinga's "Max Damage Structure (EUR/m2)" applies per m² of floor area. Today's pack applies it to built-up surface (footprint) (scripts/packs/ghsl_exposure.py module docstring). Which base the €/m² refers to must be read from JRC105688 before any footprint × storeys value ships. The 2010 price level is not inflation-adjusted (scripts/packs/country_profile.py:152-153).

3.3 Exposure-source outcome

The exposure source that carries building value is the tenant's own schedule of values, uploaded as points with building attributes in properties. It needs Phase 18's footprint reference and per-asset curve binding (D21). The modelled footprint × storeys × €/m² value is the labelled fallback, published as ODbL.

4. GHSL's ingested resolution — verified

Verified 2026-09-28 from the ingest code and config, then against the provider:

Ingest path Product and native grid Aggregation Ingested point spacing Source
scripts/packs/ghsl_exposure.py (BUILT-S, BUILT-V, POP; the production recipe) R2023A E2020 _4326_3ss: 3″ ≈ 100 m agg_factor default 10 30″ ≈ 1 km scripts/packs/ghsl_exposure.py:141-159 (URLs), :497 and :657 (default 10), :523 ("turns the ~100m native grid into ~1km output cells"); run with no --agg-factor per docs/plan/phase-9-ops-plan.md:771-775
scripts/ingest_historical_catalog.py ingest-exposure, Greece GHS-BUILT-V _4326_30ss: 30″ ≈ 1 km grid_aggregation_factor: 8 ≈ 7 km (4′) scripts/historical_catalog_manifest.json:331, :333, :338; option passed at scripts/ingest_historical_catalog.py:1244
Same, Cyprus GHS-BUILT-V _4326_30ss: 30″ ≈ 1 km grid_aggregation_factor: 2 ≈ 2 km (1′) scripts/cyprus_historical_catalog_manifest.json:128, :131

Both paths run through the same adapter. It block-sums "agg_factor native cells per block edge" and says "the GHSL pack uses 8" (src/climate_lama/core/ingest/raster_exposure.py:25-28, :72-74). agg_factor=1 would emit one point per native 3″ cell; no ingest path uses it today.

  • Provider documentation. The GHSL Data Package 2023 (JRC, GHS P2023, PDF) describes the native grids: "GHS-POP product is produced in World Mollweide at 100 m, and then aggregated at 1 km. These two datasets are then warped to WGS 1984 coordinate system, at spatial resolutions of 3 arcsec and 30 arcsec respectively". It adds: "Most of the datasets will be provided also in a warped version to WGS84 coordinate system, at 3 arcsec and 30 arc sec resolutions."
  • Live check (2026-09-28). The JEODPP directory https://jeodpp.jrc.ec.europa.eu/ftp/jrc-opendata/GHSL/GHS_BUILT_V_GLOBE_R2023A/ lists both GHS_BUILT_V_E2020_GLOBE_R2023A_4326_3ss and …_4326_30ss. The 3ss zip answered HTTP 200, application/zip, 4,362,815,348 bytes. Its copyright.txt reads "Any copyright and/or sui generis right on the dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence."
  • Production cross-check. A Greek GHSL dataset holds "~216 k rows" in production (src/climate_lama/db/repositories/exposure_dataset_repository.py:241-242). That fits the ≈ 1 km pack; the ≈ 7 km manifest path yields 7,621 points (docs/DATA_SOURCES.html:343).

GHSL's ingested resolution is 30″ (≈ 1 km) at best, from a 3″ (≈ 100 m) native grid. The older GHS-BUILT-V catalog entry is ≈ 7 km (Greece) and ≈ 2 km (Cyprus). Neither is per-building.

5. Phase 18 build items

Each item names its proof. Order follows dependency.

  1. Licence ADR — footprints. Ratify section 1.4 (the four rules, including the rule-3 reading) in the licence-review session docs/DATA_SOURCES.html:499 defers to. Then add the ODbL row's decision to docs/DATA_SOURCES.html §10. Proof: an ADR in docs/DECISIONS.md citing ODbL §4.3–§4.6 and the two OSMF guidelines.
  2. Footprint pack (Greece, Cyprus). Add DBSM R2025, or Overture buildings, as a separate ODbL footprint layer. Stream it by bbox (architecture rule 4), record the licence and attribution per artefact, and keep it out of tenant exposures. Proof: the layer's rows carry license="ODbL-1.0", and a query shows zero tenant rows reference it.
  3. Footprint reference on exposures. Add a migration so an exposure can carry a footprint id or polygon. Today it is POINT only (src/climate_lama/models/exposure.py:165-167). Compute stays at the tenant's point (section 1.4, rule 3).
  4. Per-asset impact-function binding (D21). Pick the curve per building from use, height and epoch. Proof: two buildings in one hazard pixel with different types get different losses.
  5. Modelled building value. Compute footprint area × storeys × country €/m², after the [ASSUMPTION] in section 3.2 is resolved against JRC105688 and a sourced construction-cost index lifts 2010 prices. Proof: value_unit="EUR", a "modelled" label, and the value published as ODbL with its method page (§4.6 b).
  6. Building-scale flood hazard, or honest labelling. Either source ≤ 5 m-class official flood maps for Greece and Cyprus (licence to verify per ADR-059), or ship per-building flood as "value of the ≈ 90 m cell" with the cell size shown. Proof: every per-building flood figure in the UI and API carries the hazard dataset's native_resolution.
  7. Non-flood perils. Source the section 2.3 [ASSUMPTION] per peril. Until then, per-building EQ, heat, wind and fire results stay screening-grade.
  8. Editable homes (Lookup → "Edit home", C98–C100). A citizen's edits (floors, epoch, value) are stored as tenant-private non-OSM properties, which keeps them in the Collective Database pattern. Proof: an edited home recomputes, and the footprint layer is unchanged.
  9. Street photos (C105). Mapillary imagery is "subject to the Creative Commons Share Alike (CC BY-SA) license" (mapillary.com/terms §3(b), read via fetch 2026-09-28). That is share-alike again, so it needs the same licence review as item 1. The alternative is user-uploaded photos only.
  10. Attribution everywhere. Add the map corner credit, the PDF credit with the printed openstreetmap.org/copyright, and a license field on every per-building API/CSV export. Proof: a report-studio PDF with a footprint map contains the URL string.

Assumptions

  • [ASSUMPTION] Section 1.4 rule 3 (display-only spatial join keeps tenant losses out of share-alike) needs ratification by the licence review or counsel.
  • [ASSUMPTION] Section 2.3: non-flood hazard fields vary slowly at building scale — unsourced in this spike.
  • [ASSUMPTION] Section 3.2: the base (floor area vs footprint) of Huizinga's €/m² for a footprint × storeys value — to read from JRC105688.

Sources

Decided / Parked / Killed

  • Decided (for owner ratification): property level is a Phase 18 build on ODbL footprints under the four rules of section 1.4. The tenant's schedule of values is the value source, and per-building flood shows its hazard cell size.
  • Parked: building-scale (≤ 5 m) flood hazard sourcing, non-flood per-building defensibility, and street photos (share-alike) — all Phase 18 items 6, 7 and 9.
  • Killed: deriving per-building value from LitPop (150″) or GHS-BUILT-V (m³, no €/m³ factor). Also killed: EUBUCCO as the Greek footprint source (coverage), in favour of DBSM R2025 or Overture.