How the score works

Why a count, not an average

A single blended average would hide exactly what this project exists to show: neighborhoods where several different problems pile up at once. A neighborhood with one severe problem and two mediocre ones can end up with the same average as a neighborhood with three moderately-bad ones — but those are not the same situation, and a smoothed average can't tell them apart.

So instead of averaging, Underlaid counts: for each neighbourhood, how many of the 4 categories below place it among the most affected quarter of neighbourhoods (the 25% most affected of the 2,752 neighbourhoods of Paris and its inner suburbs for that category). That number, from 0 to 4, is the cumulative score.

What this score doesn't measure

This is called an exposure score, deliberately not a "vulnerability score." Climate researchers usually split vulnerability into three separate things: exposure (is a place physically subject to heat, pollution, bad housing?), sensitivity (how much does that actually harm the people there — age, health?), and adaptive capacity (can they do something about it — the means to act, a home large enough to find a cooler room, the option to leave during a heatwave?). The exposure score only measures the first one. Adaptive capacity is shown separately (below, and in the map's "Residents' resources" theme) and never folded into the score; sensitivity isn't measured.

That matters in practice. The income-vs-exposure chart on the map shows that exposure, as measured here, doesn't reliably track income — every score band spans nearly the whole income range, including several of the metro area's wealthiest neighborhoods. A wealthy household facing the same physical exposure (heat, noise, an old apartment) usually has more ways to cope with it than a low-income one does. Two neighborhoods can share the exact same score and still face very different real odds of coping — which is precisely why this page never calls that score "vulnerability."

The ranking page's split into "Group A" and "Group B" exists for the same reason: it keeps a lower-income neighborhood with a real service shortfall visually distinct from a wealthier one whose score comes from dense, older housing stock instead — same exposure score, different starting point for coping with it.

Means to cope: a separate axis

Next to the exposure score, the map shows a second, separate measure: the resources residents have to cope (the "Residents' resources" theme, and a filter that crosses it with highly exposed neighbourhoods). It combines three INSEE figures (income 2021, 2022 census): median income, the share of overcrowded homes (fewer rooms than the household needs — no cooler room to retreat to), and the share of secondary residences (the option to leave during a heatwave). Each neighborhood is ranked on each figure, the three ranks are averaged, and the 2,752 neighborhoods are split into three equal groups: lowest, middle and highest third of the metro area.

It is never added to the exposure score. Kept apart, the two can be read side by side: two neighborhoods with the same exposure can have very different means to cope with it. INSEE doesn't publish income for 223 small neighborhoods (mostly woods, parks and business areas); they have no means class and are shown as such, never estimated.

What it shows: highly exposed neighborhoods don't face the same exposures depending on their means. Those with the highest means are mostly in Paris and stack heat and old, inefficient housing; those with the lowest means are mostly in Seine-Saint-Denis and Val-de-Marne and stack mainly heat, air pollution and noise. Overall, the link between exposure and means is weak — and slightly positive only because of dense, older central Paris; in the three inner-suburb departments there's almost no link. Where high exposure and low means do meet, it's concentrated: in Seine-Saint-Denis, 62% of highly exposed neighborhoods (at least 2 of the 3 exposures: heat, air and noise, housing) are in the metro area's lowest third of means; in Paris, 4.5%.

What it doesn't show: it measures means, not what households do with them — nothing here says whether a home has air conditioning or whether someone can take time off during a heatwave. The groups are relative to this metro area, not an absolute threshold of "enough."

The four categories

Every category is built from several raw indicators, each converted to "how exposed" so they combine consistently, then averaged together with equal weight — there isn't yet a principled reason to consider one indicator more important than another.

Access to care: taken out, rebuilt, back in the score

At launch, the score counted a fourth category, access to services. Crossing the score with residents' means showed it wasn't measuring what it claimed, and a series of checks confirmed it for each of its four indicators. It was taken out of the score in September 2026, then rebuilt; those earlier indicators stay on display for information in the detail panel.

The social position index of nearby schools reflected families' resources, already measured separately on the means axis: keeping it in the score counted them twice. It accounted for most of the link between the access category and residents' means (correlation −0.52).

Wheelchair-accessible entrances come from a crowdsourced register, missing for 39% of neighbourhoods. Neighbourhoods without it landed in the least well-served quarter 29.6% of the time, against 17.6% for the others: the same missing-data bias the project had already corrected once.

Pedestrian-path density, from OpenStreetMap, mostly measures how volunteers map sidewalks: Paris draws 53% of its sidewalks as separate paths, Seine-Saint-Denis 17% — elsewhere they're usually noted as an attribute of the street, which wasn't counted. Checked on aerial imagery: the "Klock" neighborhood in Clichy (53,000 residents per km², sidewalks clearly visible) counted 52 m of path per km²; Saint-Ambroise 4 in Paris's 11th, with the same kind of streets, 165,000.

Travel time to the nearest school, clinic or transport stop is 1.5 minutes or less for 98% of neighborhoods in this dense area, so it barely tells them apart. Hence the rebuild: a measure that accounts for how many people share each service, not just how close it is.

Sidewalks in OpenStreetMap: counted properly, still uneven

To see whether the pedestrian figure could be repaired, sidewalks were counted both ways OpenStreetMap records them — drawn as separate lines, or noted on the street itself (sidewalk=both/left/right) — each sidewalk counted once. That fixes the Klock example: 1.87 m of sidewalk per metre of street, against 1.76 in Saint-Ambroise 4. But the deeper problem remains: for many streets OpenStreetMap says nothing about sidewalks at all, and how often depends on the department, even between neighborhoods of the same density.

Share of street length with any sidewalk information in OpenStreetMap (extract of September 30, 2026), by population density of the neighborhood

Density bands are fifths of the inhabited neighborhoods (50 residents or more). * Fewer than 20 neighborhoods of that department in the band.

Unknown doesn't mean absent: in Firmin Gemier (Aubervilliers), OpenStreetMap records no sidewalk on any street, while aerial imagery shows them. What matters in a wheelchair is recorded even less evenly: sidewalk width for under 1% of sidewalks everywhere; surface for 72% in Paris and 21% in Seine-Saint-Denis; kerbs about 119 times per 100 pedestrian crossings in Paris, 33 in Seine-Saint-Denis.

Conclusion: no sidewalk figure from OpenStreetMap can compare neighborhoods across departments. The rule, set before computing, allowed a gap of at most 15 points between departments at equal density; the gap reaches 71 points. So none enters the score, and the old footway density no longer appears on the map. Access to services will be rebuilt from official sources (Île-de-France Mobilités accessibility data, health directories), with walking routed on the full street network rather than on sidewalks.

The access measure: without a car

Access to care is measured with a method that shares each service's capacity among everyone who can reach it, not just the distance to the nearest one. It counts what residents can reach on foot and by public transport, not by car. That choice follows from what Underlaid is about: equal access to everyday services for everyone, including people who can't drive or don't have a car — children and teenagers, many older people, people with disabilities or a pushchair, and the many households without a car in dense areas.

So a neighborhood where doctors are easy to reach by car but far on foot or by bus ranks low on this measure: that's what it's meant to show. It is also why it differs from the official DREES indicator (APL), which uses car travel times between communes. The two answer different questions.

How access to care was rebuilt

Liberal GPs and GPs employed by health centres (RPPS directory, the same field as the DREES APL indicator) and pharmacies (FINESS), shared among the population of each 200 m square (INSEE, 2021), weighted by age with the DREES weights. Travel times are computed on foot and by public transport on the OpenStreetMap street network and the Île-de-France Mobilités timetables, for a Tuesday between 10:00 and 11:00. The whole of Île-de-France is included, so neighborhoods on the edge of the inner ring also see services across the boundary.

Structures without local, everyday consultations were left out after checking by hand every address with at least 15 GPs: teleconsultation platforms, on-call and home-visit services, hospital emergency departments, services reserved to one population (students, airports). The list, with the reason for each exclusion, is published in SCORING.md.

Before crossing this measure with residents' means, the hypothesis and its criteria were written down and dated: are neighborhoods with the lowest means at a disadvantage in access to care, clearly and in several departments? Result: yes. In both Hauts-de-Seine and Val-de-Marne, about 41% of neighborhoods in the lowest third of means are in the least-served quarter of neighborhoods, against 24% in the highest third; at equal population density, the gap reaches 23 points. Seine-Saint-Denis couldn't be assessed by the rule set in advance: only one of its neighborhoods is in the metro area's highest third of means.

Inclusive mobility, for information

The same measure was recomputed for a wheelchair trip: step-free, using only the stops and journeys declared accessible. The hypothesis doesn't hold for this measure: the loss of access doesn't weigh more on the neighborhoods with the lowest means. Metro inaccessibility mostly penalises wheelchair journeys within Paris, whatever the neighbourhood's level of resources. Counted in GPs within reach, wheelchair access remains lower in the inner suburbs.

It is shown in the detail panel but isn't part of the score: it compares two ways of travelling in the same place, while the score compares places. That decision was made after seeing the crossing with residents' means; it doesn't change the result above.

The "not enough data" rule — and the bias it fixed

A category is only scored for a neighborhood if data exists for a strict majority of its underlying indicators. Below that, the category is marked "insufficient data" for that neighborhood and left out of its cumulative score entirely — never silently treated as "fine" or averaged from whatever scraps happened to be available.

This rule exists because of a real bias found in the original Paris-only data, not a hypothetical one. Two small, non-residential neighborhoods — a Périphérique interchange and a dense commercial block near a department store — were scoring the maximum at the time (4 out of 4, when access to services was still counted), driven partly by an "access" category built from just one or two of its four indicators. Checked across all 992 Paris neighborhoods: among those whose access category relied on a single indicator, 47% landed among the most affected quarter, nearly double the 25% you'd expect by chance. Sparse-data neighborhoods were being penalized for lacking data, not for having worse access.

After the fix, both of those examples dropped out of the top score — and the same kind of bias, found again in a milder form at launch, took the earlier access indicators out of the score, before access came back in a rebuilt form (see below).

The thresholds that delimit the most affected quarter are set from inhabited neighbourhoods only. The 62 neighborhoods with fewer than 50 residents — parks, river banks, rail yards, stations, business blocks — are placed against those thresholds afterwards: they keep a score on the map, marked "very sparsely populated", but they no longer take up places in the most affected quarter that belong to neighbourhoods where people live. 74% of them were among the most affected quarter for air and noise, since they often sit along rail lines and expressways.

Adding electrical thermosensitivity to housing

Housing originally used one indicator: the share of homes rated F or G on energy performance. A neighborhood can now also be flagged by how sharply its residential electricity use swings with winter temperature (data from Enedis, the grid operator) — a proxy for electric-heating-dependent homes with no cheap way to compensate for heat either, not just cold.

Both indicators are now required, not just one — a real tightening, not a bug. The two Paris examples above moved again as a result (to 2 out of 4 at the time): both still had enough housing data, but their thermosensitivity turned out close to the citywide average rather than among the most affected quarter, which took their housing category out of that quarter.

Expanding beyond Paris — and a new example

Underlaid covers the whole inner ring of the Greater Paris metropolis: Paris, Hauts-de-Seine, Seine-Saint-Denis and Val-de-Marne, 2,752 neighbourhoods instead of 992. Most sources were already national or regional. Two heat indicators (cool places, cool green space) existed for Paris only: they were replaced by regional sources (the public inventory of pools, museums and libraries, and a regional survey of green and wooded areas), applied the same way everywhere.

A fourth heat indicator was added at the same time: the share of each neighbourhood's surface that is sealed (built or paved, from the regional land-use survey), a structural cause of heat retention. Two Paris-only context facts (tree age, street lighting) have no regional equivalent and stay flagged as Paris-only; associational density had one and covers the 143 communes.

4 neighborhoods currently reach the maximum, 4 out of 4: Flachat I in Asnières-sur-Seine, Nonneville 3 and 4 in Aulnay-sous-Bois, Quatre Cités 3 in Champigny-sur-Marne. 81 inhabited neighborhoods reach 3 out of 4 or more, spread across Paris (22), Hauts-de-Seine (23), Seine-Saint-Denis (18) and Val-de-Marne (18); they are listed on the ranking page. Any example on this page is provisional — the score moves with each methodological correction.

This same territory already has a regional precedent worth naming directly: ORS Île-de-France, Ineris and L'Institut Paris Region published "Cumuls d'expositions environnementales en Île-de-France" (2022), a 5-year, 500m-grid study of composite exposure across the whole region, including a case study on Aubervilliers — a commune inside this project's own coverage. It confirms the same underlying idea from an independent source: environmental burdens in this region do stack up in identifiable places. Underlaid's own difference is granularity (IRIS rather than a 500m grid) and a count of the least favourable categories rather than a blended index (see "About" on the home page for the full comparison).

What we don't yet know

Naming these limits openly is what makes the score checkable — hiding them would only delay the moment someone else finds them.

• Median income is masked by INSEE for small neighborhoods (privacy threshold): 223 of 2,752 have no published figure. Income is shown for context only — it is never part of the score.
• The heat vulnerability index covers 2,618 of 2,752 neighborhoods; the source dataset uses a slightly coarser grouping that misses full coverage. The other heat indicators (cool facilities, green area, artificialization) do cover all 2,752.
• Electrical thermosensitivity is a winter-heating signal — Enedis doesn't publish a summer-cooling equivalent. Using it as a proxy for "can't compensate for heat" assumes a poorly-insulated, electric-heating-dependent home is also hard to keep cool, which is physically reasonable but not independently checked against real summer electricity data.
• The artificialized-surface classification (54 of 79 land-use categories) is modeled closely on France's official artificialisation definition but isn't a certified reproduction of it — there's no single published crosswalk from this specific regional land-use legend to the legal nomenclature. It also doubles as a rough urban-biodiversity signal — sealed ground is lost habitat — though it isn't scored separately for wildlife, since no animal mortality or stress data comparable to what's available for humans exists at this grain.
• The score is an estimate, not a fixed ground truth. Every methodological correction moves the line — treat any single flagship example as provisional until checked against the raw indicators, the way every example on this page was.

Current distribution — 2,752 neighborhoods (Paris + inner suburbs)

4 neighborhoods currently reach 4 out of 4 and 79 others 3 out of 4 (81 inhabited in all). Treat them as provisional evidence, not a settled verdict: the score moves with each methodological correction.

Data snapshot as of October 1, 2026 — the score is re-computed periodically and can shift; a specific example above should always be read against this date, not treated as permanent.

Data sources & licenses

Everything here is built from public open data. Each source keeps its own license, checked against what the publisher itself states. Underlaid's code is under the MIT license; the published data (the score and every indicator behind it) is under the Open Database License (ODbL), because two sources — OpenStreetMap and the Ville de Paris datasets — are ODbL, which requires derived data to be shared under the same terms.

• INSEE — BPE, Filosofi 2021, 2022 censusLicence Ouverte 2.0 (Etalab)
• IGN & INSEE — IRIS boundariesLicence Ouverte 2.0 (Etalab)
• CSTB — Sat4BDNB (urban heat islands)Licence Ouverte 2.0 (Etalab)
• L'Institut Paris Region — green spaces, MOS land useLicence Ouverte 2.0 (Etalab)
• Airparif & Bruitparif — air-noise mapOpen data, mandatory citation
• DEPP — Ministry of Education (IPS)Licence Ouverte 2.0 (Etalab)
• ADEME — DPE energy certificatesLicence Ouverte 2.0 (Etalab)
• Enedis — electricity consumption by IRISLicence Ouverte 2.0 (Etalab)
• AcceslibreLicence Ouverte 2.0 (Etalab)
• ANCT — priority neighborhoods (QPV)Licence Ouverte 2.0 (Etalab)
• Ministry of the Interior — RNALicence Ouverte 2.0 (Etalab)
• Ville de Paris — street trees, public lightingODbL
• © OpenStreetMap contributors — footways, street networkODbL
• ANS — RPPS directory of health professionalsLicence Ouverte 2.0 (Etalab)
• Ministry of Health — FINESS (pharmacies)Licence Ouverte 2.0 (Etalab)
• DREES — APL age weights and decayLicence Ouverte 2.0 (Etalab)
• Île-de-France Mobilités — public transport timetables (GTFS)Licence Mobilités
• IGN — BD TOPO (staircases)Licence Ouverte 2.0 (Etalab)
• © OpenStreetMap contributors — Seine and Marne drawn on the mapODbL
• IGN, DINUM — API Adresse (BAN)Licence Ouverte 2.0 (Etalab)

Mandatory citations — INSEE: "Source: Insee". Air-noise map: "Source des données : Cartographie air-bruit établie par Airparif et Bruitparif – http://carto.airparif.bruitparif.fr". OpenStreetMap: "© OpenStreetMap contributors". Transport timetables: "Contains information from ‘Horaires prévus sur les lignes de transport en commun d'Île-de-France (GTFS Datahub)’, made available by Île-de-France Mobilités under the terms of the ‘Licence Mobilités’." To cite Underlaid, see the repository's CITATION.cff file, and state the data snapshot date (shown at the bottom of every page).