Methodology · v1.0.0

# How the numbers are made

Every result in this project is computed by an open, deterministic pipeline from public data. This page documents what was measured, what was calculated, what was modelled and what was assumed — and how far each can be trusted.

observed

calculated

modelled

assumed

scenario

## Question & scope

**Question.** How does walkable access to everyday essentials differ between Singapore and Johor Bahru, why, and how could future mobility and urban interventions change it?

**Study area.** Singapore: the 55 planning areas (URA Master Plan 2019), residents from the Census of Population 2020 by subzone. Johor Bahru: the ten state constituencies N.40–N.49 (Tiram, Puteri Wangsa, Johor Jaya, Permas, Larkin, Stulang, Perling, Kempas, Skudai, Kota Iskandar), covering Johor Bahru, Iskandar Puteri and Pasir Gudang, population from the Malaysian Population and Housing Census 2020. Analysis grid: H3 resolution 10 hexagons (≈0.015 km², ≈66 m edge).

**Not in scope.** Travel by car, cycling, quality or price of services, opening hours, heat and shade, perceived safety, and trips across the border inside the 15-minute measure (the border is treated separately as a regional scenario).

## Data register

Each source is classified and cited. Download checksums and retrieval times are written to data/provenance.json by the acquisition step.

| Source | Class | Licence | Used for |
| --- | --- | --- | --- |
| [OpenStreetMap planet snapshot 2026-10-05](https://planet.openstreetmap.org/pbf/) OpenStreetMap Foundation (planet mirror on AWS Open Data) | [observed] | ODbL-1.0 | walk network, drive network, points of interest, parks, car parks, transit stops, basemap |
| [High Resolution Settlement Layer — general population (v1.5.5)](https://dataforgood.facebook.com/dfg/tools/high-resolution-population-density-maps) Meta Data for Good & CIESIN, Columbia University | [modelled] | CC-BY-4.0 | origin population weights (both cities) |
| [High Resolution Settlement Layer — elderly 60+ (v1.5.5)](https://dataforgood.facebook.com/dfg/tools/high-resolution-population-density-maps) Meta Data for Good & CIESIN, Columbia University | [modelled] | CC-BY-4.0 | older-resident accessibility |
| [Kontur Population (H3 r8), 2023-11-01](https://data.humdata.org/organization/kontur) Kontur | [modelled] | CC-BY-4.0 | population sensitivity (alternative allocation) |
| [Population and Housing Census of Malaysia 2020 — by state constituency (DUN)](https://open.dosm.gov.my/dashboard/kawasanku) Department of Statistics Malaysia (DOSM) | [observed] | Malaysia Open Data Terms of Use (data.gov.my) — attribution required | Johor Bahru study area, calibration and validation of gridded population, neighbourhood statistics |
| [Malaysian electoral (DUN, parliament) and district boundaries](https://open.dosm.gov.my/) Department of Statistics Malaysia (DOSM) | [observed] | Malaysia Open Data Terms of Use — attribution required | Johor Bahru analysis boundary and neighbourhood units |
| [URA Master Plan 2019 Planning Area Boundary (No Sea)](https://data.gov.sg/datasets/d_4765db0e87b9c86336792efe8a1f7a66/view) Urban Redevelopment Authority (URA), via data.gov.sg | [observed] | Singapore Open Data Licence v1.0 | Singapore analysis boundary and neighbourhood units |
| [Census of Population 2020 — resident population by planning area / subzone](https://data.gov.sg/datasets?query=Census%20of%20Population%202020%20subzone) Singapore Department of Statistics (SingStat), via data.gov.sg | [observed] | Singapore Open Data Licence v1.0 | validation of gridded population, neighbourhood statistics |
| [Singapore Residents by Planning Area / Subzone, Age Group, Sex and Floor Area of Residence, June 2011–2020](https://www.singstat.gov.sg/find-data/search-by-theme/population/geographic-distribution/latest-data) Singapore Department of Statistics (SingStat) | [observed] | Singapore Open Data Licence v1.0 | residents aged 65+ by subzone (elderly accessibility) |
| [LTA Rapid Transit System Station layer (April 2025)](https://datamall.lta.gov.sg/content/datamall/en/static-data.html) Land Transport Authority (LTA) DataMall — Geospatial Whole Island | [observed] | Singapore Open Data Licence v1.0 (LTA DataMall) | audit of OSM rail stations |
| [URA Master Plan 2019 Rail Station layer](https://data.gov.sg/) Urban Redevelopment Authority (URA), via data.gov.sg | [observed] | Singapore Open Data Licence v1.0 | audit of OSM rail stations |
| [Singapore bus stops and services (from LTA DataMall)](https://datamall.lta.gov.sg/) Land Transport Authority (LTA) DataMall, compiled by the open-source sgbusdata project | [observed] | Singapore Open Data Licence v1.0 (LTA DataMall) | audit of OSM bus stops, transit layer |
| [ESA WorldCover 10 m 2021 v200](https://esa-worldcover.org/) European Space Agency (ESA WorldCover consortium) | [observed] | CC-BY-4.0 | basemap land / water / vegetation / built-up layers |
| [Overture Maps Places, release 2026-09-23.1](https://docs.overturemaps.org/guides/places/) Overture Maps Foundation (sources include Meta, Microsoft, Foursquare) | [observed] | CDLA-Permissive-2.0 (places); see Overture attribution guidance | independent check of OSM POI completeness, POI-source sensitivity run |
| [myBAS Johor Bahru GTFS Static](https://api.data.gov.my/gtfs-static/mybas-johor) Malaysia Official Open API (data.gov.my) — operator BAS.MY Johor Bahru (Causeway Link) | [observed] | Malaysia Open Data Terms of Use — attribution required | Johor Bahru bus access to the border hubs, audit of OSM bus stops |
| [KTMB GTFS Static (incl. Shuttle Tebrau JB Sentral – Woodlands)](https://api.data.gov.my/gtfs-static/ktmb) Malaysia Official Open API (data.gov.my) — Keretapi Tanah Melayu Berhad | [observed] | Malaysia Open Data Terms of Use — attribution required | baseline cross-border rail option |

## Street network

The walking network is built from OpenStreetMap ways (snapshot 2026-10-05). Footways, paths, pedestrian streets, steps, corridors, cycleways (unless `foot=no`), service roads, residential streets and arterial roads are walkable; ways tagged `foot=no|private|use_sidepath`, `access=private|no` (unless foot is explicitly permitted), indoor rooms and areas are not. Components smaller than 200 nodes are dropped. Edge lengths are geodesic metres; walking time is length ÷ speed, so one search serves every speed.

**Strict variant.** Trunk and primary carriageways are walkable only where a sidewalk is tagged. Singapore maps its pavements as separate footways, so it loses little; Johor Bahru loses walks along arterials that may have no pavement. The variant bounds how much Johor Bahru's measured access depends on such walks.

**Snapping.** Origins and destinations join the network at the nearest point on the nearest edge (not the nearest node), reaching both end nodes with the along-edge offsets. The off-network leg is counted at 1.3 × its straight-line length; points further than 250 m from any walkable edge are treated as off-network. Parks are reached through points along their boundary, not their centroid.

## Destinations

Six essentials form the basket. Each is defined by OpenStreetMap tags in three nested variants — strict ⊂ core ⊂ broad — and the core definition is the published baseline. Places with lifecycle prefixes (`disused:`, `abandoned:`, `construction:`…) are excluded.

| Essential | Strict | + Core (baseline) | + Broad |
| --- | --- | --- | --- |
| **Groceries** basket | shop=supermarket, shop=greengrocer, amenity=marketplace | shop=grocery, shop=convenience, shop=butcher, shop=seafood, shop=frozen_food, shop=health_food | shop=bakery, shop=deli, shop=dairy, shop=beverages, shop=department_store, shop=variety_store |
| **Healthcare** basket | amenity=clinic, amenity=doctors, amenity=hospital, healthcare=clinic, healthcare=doctor, healthcare=hospital, healthcare=centre | amenity=pharmacy, healthcare=pharmacy | amenity=dentist, healthcare=dentist, shop=chemist |
| **Education** basket | amenity=school | amenity=kindergarten, amenity=childcare | amenity=college, amenity=university |
| **Green space** basket | leisure=park, leisure=nature_reserve, leisure=recreation_ground, leisure=garden, landuse=recreation_ground, landuse=village_green | — (parks ≥ 0.5 ha; strict ≥ 1 ha) | leisure=playground, leisure=common, leisure=dog_park, leisure=fitness_station |
| **Public transport** basket | highway=bus_stop, amenity=bus_station, railway=station, railway=halt, railway=subway_entrance, railway=tram_stop | — | public_transport=platform, public_transport=stop_position |
| **Eateries** basket | amenity=food_court, amenity=restaurant, amenity=fast_food | amenity=cafe | amenity=ice_cream, shop=bakery |
| **Rapid transit** supplementary | railway=station, railway=halt, railway=subway_entrance | — | — |
| **Civic & community** supplementary | amenity=community_centre, amenity=library | amenity=post_office, amenity=townhall, amenity=social_centre | amenity=place_of_worship |
| **Car parking** supplementary | amenity=parking | — (parks ≥ 0 ha; strict ≥ 0 ha) | — |

The park threshold follows the WHO Regional Office for Europe (2016) indicator of green space of at least 0.5–1 ha within 300 m. Overture Places (release 2026-09-23, confidence ≥ 0.6) supplies an independent set of groceries, healthcare, education and dining places for the sensitivity run.

## Population

Residents are distributed to H3 cells by dasymetric mapping: each census unit's count (Singapore subzones; Johor Bahru state constituencies) is spread over the cells in proportion to Meta/CIESIN HRSL settlement weights, so totals match the census exactly. Where HRSL sees no settlement inside a populated unit (e.g. Punggol's newest estates), Kontur population weights are used instead; Kontur is also the alternative population model in the sensitivity analysis. Residents aged 65+ come from the official age tables (SingStat by subzone; DOSM by constituency) and are spread with the HRSL elderly layer.

| City | Model | Units | log-correlation | Units within ±25 % |
| --- | --- | --- | --- | --- |
| Singapore | HRSL (Meta/CIESIN) | 211 subzones | 0.45 | 30% |
| Singapore | Kontur | 211 subzones | 0.78 | 40% |
| Johor Bahru | HRSL (Meta/CIESIN) | 10 state constituencies | 0.92 | 80% |
| Johor Bahru | Kontur | 10 state constituencies | 0.92 | 90% |

Model counts summed by census unit before calibration, against the census. These figures describe where people are placed inside units; calibrated totals are exact.

**Who is counted.** Singapore's figures cover residents, citizens and permanent residents: 4.04 million in the Census of Population 2020. About 1.64 million non-residents, many of them work-permit holders living in worker dormitories, often in industrial areas, are outside that count. Johor Bahru's census counts all persons, including non-citizens, so the comparison sets Singapore's residents against all of Johor Bahru's population.

The effect can be bounded from the data. Even if none of Singapore's non-residents had all six essentials within reach, Singapore's share of all persons would still be at least **57.1%** (at least 3.25 million complete residents out of 5.69 million persons), against 12.0% in Johor Bahru. The direction of the finding cannot flip. The gap is at least 45 points, against 68 for residents alone.

Non-residents: Population in Brief 2020 (Singapore Department of Statistics and partners): non-resident population of 1,641.6 thousand in June 2020.

## Metrics

For each cell $i$ and essential $k$, a multi-source Dijkstra search from every destination of $k$ gives the network distance to the nearest one:

$$
d_{ik} = \min\bigl(D_k(u) + o_u,\; D_k(v) + o_v\bigr) + 1.3\, s_i
$$

$u, v$: end nodes of the edge the cell snaps to; $o_u, o_v$: along-edge offsets; $s_i$: snap distance; $D_k$: network distance from the nearest destination of $k$

With walking speed $v$ (baseline 1.2 m/s) and time budget $\tau$ (baseline 15 min), **completeness** is the share of the 6 essentials within reach:

$$
C_i(\tau) = \frac{1}{K} \sum_{k=1}^{K} \mathbf{1}\!\left[\, d_{ik} \le v\,\tau \,\right]
$$

Population-weighted indicators use the residents $w_i$ of each cell. The headline is the **share complete**, $\sum_i w_i \mathbf{1}[C_i = 1] \,/\, \sum_i w_i$; a resident is in a **low-access** area when $C_i < 0.5$. Distributions are summarised with weighted quantiles and a weighted Gini coefficient.

Two robustness measures avoid the hard cut-off: **cumulative opportunities** (destinations within 5, 10, 15 and 20 minutes) and a **gravity index**

$$
G_{ik} = \sum_{j \in k} e^{-\beta\, d_{ij}}, \qquad \beta = \frac{\ln 2}{h\, v}
$$

$h$: half-life, 10 min (the weight of a destination halves for every 10 minutes of walking)

The explorer recomputes completeness in the browser from the published distances for any speed, threshold, definition or population group, using the same formulas (tested against the Python implementation).

## Urban form

- **Intersection density**: nodes with three or more connections per km² of populated land.
- **Dead-end share**: nodes with a single connection.
- **Pedestrian-way share**: walkable length on footways, paths, pedestrian streets, steps and corridors.
- **Arterial exposure**: share of trunk/primary length with no sidewalk tagged.
- **Pedestrian shed ratio**: residents reached within 15 minutes on the network ÷ residents inside the 15-minute circle (1.08 km), for 1,500 population-weighted sample homes per city.
- **Circuity**: network ÷ straight-line distance for local trips of 300–1,200 m.
- **Parking land**: mapped car-park polygons (surface and multi-storey), m² per resident.

## Scenarios

### 1 · RTS Link: door to platform (regional)

Minutes from a Johor Bahru home to the first Singapore MRT platform, decomposed so that each term is either calculated or an explicit assumption:

$$
T = T_{\text{access}}(\text{home} \to \text{hub}) + 2\,T_{\text{clearance}} + T_{\text{crossing}} + T_{\text{connection}}
$$

[calculated] $T_{\text{access}}$ on real networks: walking, or walking plus myBAS buses with frequency-based waits (half the scheduled headway in the 07:00–09:00 weekday window from the official GTFS, capped at 15 min, plus a 2-minute boarding margin), or a shared autonomous feeder on the road network. [assumed] Everything after the hub.

| Parameter | Value | Basis |
| --- | --- | --- |
| Border regime “low” | 5 min per country · Causeway 8 min | Assumption bracketing automated off-peak clearance to peak congestion |
| Border regime “central” | 15 min per country · Causeway 20 min | Assumption bracketing automated off-peak clearance to peak congestion |
| Border regime “high” | 30 min per country · Causeway 45 min | Assumption bracketing automated off-peak clearance to peak congestion |
| Causeway bus (today) | every 6 min · +12 min to MRT | Assumption |
| KTM Shuttle Tebrau (today) | 5 min ride · 10 min planned wait · +15 min to MRT | Ride: KTMB GTFS; wait and transfer: assumptions |
| RTS Link | 5 min ride · every 3.6 min · +4 min to TEL | Ride, headway: official/reported design values; transfer: assumption |
| Autonomous feeder | wait 3/6/10 min · peak speed ×0.6 · pooling ×1.25 · pickup ≤ 3 min walk | Assumptions |

Not modelled: fares, capacity limits and queues at the RTS station, the possible retention of the KTM shuttle, and car or motorcycle crossings. Peak-day car queues of several hours are beyond the “peak” regime.

### 2 · Parking to Proximity (local)

Candidate sites are mapped car parks of at least 2,000 m² (surface, multi-storey, rooftop or untagged; underground excluded). Converting a site adds a hub providing groceries, healthcare, education, dining, transit at its access points; sites of 1 ha or more also provide green space. Sites are chosen greedily (maximal coverage): at each step, the site whose conversion makes the most residents newly complete. Results are reported for 10, 25, 50 sites per city. This is an exploration of where land could matter, not a forecast that autonomous vehicles will free it.

### 3 · One footpath (local)

Which single short public footpath would bring the most Johor Bahru residents at least one more everyday essential within a 15-minute walk?

1. **Candidates.** Every dead end of the Johor Bahru walk network is paired with each network node within 80 m in a straight line that is at least 400 m further away on foot (a bounded search from the dead end): 9,893 pairs.
2. **Desk filters.** A pair is dropped when the straight link crosses another walkable way (so a shorter link exists); an end is on a service road, track or corridor; passes within the clearance of a building; crosses a river, canal, stream, railway, major road, wall or fence; crosses a school, hospital, cemetery, military or prison site; an end is on a bridge or in a tunnel; crosses open water. Links over a mapped drain or ditch are kept and flagged as needing a footbridge. 1,488 pairs pass; 594 of them would bring someone a new essential.
3. **Exact evaluation.** A shortest path uses a new edge at most once, so with the link $(a, b)$ of length $\ell$ the distance from cell $i$ to the nearest destination of essential $k$ becomes

$$
d'_{ik} = \min\bigl(d_{ik},\; d_{ia} + \ell + D_k(b),\; d_{ib} + \ell + D_k(a)\bigr)
$$

$D_k$: baseline distance from every node to the nearest destination of $k$; $d_{ia}, d_{ib}$: baseline walk from the cell to each end (bounded searches). A unit test checks this against a full recomputation on random graphs.

Homes and destinations keep their current access points and nothing else changes. Every candidate is scored for every Johor Bahru cell; the one that gives the most residents at least one more essential within 15 minutes is chosen (ties: residents newly reaching all six). Alternatives listed below are at least 400 m apart.

[calculated] Recomputing the baseline this way reproduces the published distances for all 54,095 Johor Bahru cells (largest difference under 1 cm), and each sample route, traced node by node, has the same length as the engine's distance. [scenario] The link itself; [assumed] A public footpath about 2.5 m wide, open at all hours, step-free and lit, following the straight line between the two street ends. Where the line crosses a mapped drain or ditch, a small footbridge carries it. Homes and destinations keep their current access points; no shop, clinic, school or stop opens or closes. Walking at 1.2 m/s on the mapped network, as everywhere else in this study. The population of each hexagon is the modelled 2020 census population; nobody moves.

| Place | Between | Link | Walk today | Residents gaining an essential | Newly all six |
| --- | --- | --- | --- | --- | --- |
| **Bukit Indah** (Perling) | Jalan Penglin – Jalan Bestari 10/1 | 75 m | 1,402 m | 2,200 | 2,200 |
| Bina Park (Johor Jaya) | Jalan Gunung 21 – Persiaran Seri Alam | 31 m | 990 m | 1,600 | 0 |
| Taman Nusa Bestari Jaya (Skudai) | Jalan Makmur – Jalan Jati 11 | 64 m | 1,488 m | 1,600 | 100 |
| Taman Perindustrian Plentong (Johor Jaya) | residential – Jalan Masai Jaya 2 | 33 m | 1,513 m | 1,600 | 0 |
| Taman Dato Penggawa Barat (Kempas) | Jalan Mewah Ria 2/2 – Jalan Mewah Ria 2 | 53 m | 944 m | 1,500 | 0 |
| Taman Johor Jaya (Johor Jaya) | Jalan Bakawali 10 – Jalan Dedap 4 | 72 m | 1,791 m | 1,500 | 0 |
| Taman Skudai Baru (Kota Iskandar) | Jalan Jasa 51 – Jalan Hang Jebat | 67 m | 902 m | 1,500 | 500 |
| Kampung Ulu Pandan (Puteri Wangsa) | Jalan Daya – Jalan Bertam 23 | 80 m | 926 m | 1,400 | 0 |
| Taman Rinting (Permas) | steps – Jalan Cendana 13 | 61 m | 2,704 m | 1,300 | 70 |
| Taman Sri Bahagia (Kempas) | Jalan Aliff Harmoni 6/9 – Jalan Sri Bahagia 15 | 40 m | 1,768 m | 1,200 | 0 |

Street names are from OpenStreetMap; "residential" or "steps" means the way has no name. This is a desk study: ownership, unmapped walls and gates, drain widths, informal paths and residents' views can only be checked on site.

## Validation

Every input with an independent reference is checked against it. Results:

| Check | Reference | Result |
| --- | --- | --- |
| Bus stops, Singapore | LTA DataMall bus stops (5,119) | 92.0% matched within 30 m · 97.2% within 60 m |
| Bus stops, Johor Bahru | myBAS Johor GTFS stops (831) | 82.9% matched within 30 m · 94.9% within 60 m |
| Rail stations, Singapore | LTA Rapid Transit System Station layer (Apr 2025) (189) | 93.7% within 200 m · 94.7% within 400 m |

### OpenStreetMap vs Overture Places

| City | Category | OSM | Overture | OSM found in Overture (100 m) | Overture found in OSM (100 m) |
| --- | --- | --- | --- | --- | --- |
| Singapore | groceries | 1,557 | 1,372 | 71% | 71% |
| Johor Bahru | groceries | 487 | 675 | 55% | 40% |
| Singapore | healthcare | 831 | 1,278 | 68% | 55% |
| Johor Bahru | healthcare | 271 | 853 | 72% | 31% |
| Singapore | education | 885 | 1,581 | 67% | 40% |
| Johor Bahru | education | 384 | 1,227 | 61% | 25% |
| Singapore | dining | 8,748 | 18,042 | 97% | 90% |
| Johor Bahru | dining | 2,120 | 8,237 | 96% | 56% |

## Sensitivity

The analysis is repeated for every combination of walking speed (0.9, 1.2, 1.4 m/s), time budget (10, 15, 20 min), taxonomy variant (strict, core, broad), network variant (permissive, strict; the strict network with the core taxonomy only) and population model (HRSL, Kontur): 72 runs. A finding is **robust** if it holds in ≥ 90 % of runs, **conditional** in 50–90 %, **fragile** otherwise.

| Finding | Holds in | Verdict |
| --- | --- | --- |
| Singapore has a higher share of residents with complete 15-minute access than Johor Bahru | 100% | robust |
| The completeness gap between the cities exceeds 20 percentage points | 97% | robust |
| Most Johor Bahru residents lack at least one essential within the threshold | 100% | robust |
| Most Singapore residents have every essential within the threshold | 79% | conditional |
| Johor Bahru has a larger share of residents in low-access areas | 100% | robust |
| In Singapore, the essential most often out of reach is healthcare | 100% | robust |
| In Johor Bahru, the essential most often out of reach is healthcare | 94% | robust |

## Limitations

- **Mapping completeness differs.** OpenStreetMap coverage of shops, clinics and footways is not uniform; the Overture run and the stop/station audits bound the effect but cannot remove it.
- **Presence is not service.** A clinic is counted whether it is a polyclinic or a private GP; capacity, price, opening hours and quality are not measured.
- **Walkability is more than connectivity.** Heat, shade, rain, crossing waits, gradients, pavement condition and perceived safety are not in the network weights.
- **Population inside census units is modelled.** HRSL can miss new estates and misplace residents in mixed areas; calibrated totals are exact, the pattern inside each unit is an estimate.
- **Different census definitions.** Singapore counts residents (citizens and permanent residents, 2020); Malaysia counts all persons (2020). Neither is current to 2026.
- **Scenarios are not forecasts.** The RTS Link has not opened; border clearance times vary by hour and day; autonomous-feeder and parking-conversion parameters are illustrative.

## Reproduce

The full pipeline runs on a laptop. Raw data are downloaded with checksums; everything else is derived.

```
make setup            # Python env + web dependencies
make data             # download sources (writes data/provenance.json)
make pipeline         # networks → access → form → scenarios → summary → export
make web              # build the static site into web/dist
make test             # pytest + vitest
```

Stages can be run individually: `python -m sgjb.pipeline networks features population access access_overture form scenarios summary export basemap`.

## Architecture

**Static-first.** A Python pipeline (pyosmium, GeoPandas, H3, numba-compiled Dijkstra) writes compact columnar binary tables, JSON and a tiled street graph; the site is plain static files that any web server or CDN can host. There is no runtime server and no API key.

- **Cells**: per-cell residents and quantised distances (uint16 metres) for each essential and variant, decoded into typed arrays in the browser.
- **Street graph**: 0.02° tiles with tile-relative coordinates and delta-encoded geometry, loaded on demand for in-browser isochrones (the same edge snapping and Dijkstra as the pipeline).
- **Basemap**: a self-hosted PMTiles archive built with tippecanoe from OpenStreetMap and ESA WorldCover; MapLibre renders it, and deck.gl draws the analysis layers interleaved in the same WebGL context.
