Small TPU Census Zones (211 zones)

Source

Census and Statistics Department (CSD / censtatd) + Planning Department (PlanD)

• Census 2021 data: https://www.census2021.gov.hk/doc/STPUG_21C.zip
• STPU boundaries: Esri ArcGIS FeatureServer (mirrored from GeoData Store, closed Aug 2023)
• Boundary FeatureServer: https://services3.arcgis.com/6j1KwZfY2fZrfNMR/arcgis/rest/services/Hong_Kong_Population_Distribution_by_Sex_by_Small_TPU_in_2021/FeatureServer/0

Format

CSV + GeoJSON One-off (Census 2021). No live API.

211

STPU Zones

7.47M

Total Population

51 KB

Centroid JSON

Census 2021

Vintage

Why STPU, not Districts?

Hong Kong has 18 districts, but within a district, population density can vary wildly. Tuen Mun has both high-rise estates (50K+/km²) and country parks (near zero). District-level averages smooth this out and produce garbage estimates.

Small Tertiary Planning Units split those 18 districts into 211 zones, each roughly uniform in character. This gives us ~20x better spatial resolution at the cost of only 51KB of data.

Granularity	Zones	Avg Population	Spatial Precision
District	18	~415,000	~5-15 km
Large TPU	72	~104,000	~2-5 km
Small TPU	211	~35,400	~0.5-2 km
DCCA	452	~16,500	~0.3-1 km

Tip

DCCA (452 zones) is more granular, but boundaries change every 4 years with elections. STPU boundaries are stable across censuses, making them better for longitudinal comparison.

Schema / Fields (Centroid JSON)

Each zone in stpu_centroids.json:

Field	Type	Example	Description
id	string	"111"	Zone identifier (matches Census CSV column 1 + “S” suffix)
cx	float	114.15108	Centroid longitude (WGS84)
cy	float	22.28485	Centroid latitude (WGS84)
pop	integer	60366	Total usual resident population
hh	integer	22668	Number of domestic households
hhSize	float	2.6	Average household size
income	integer | null	40480	Median monthly household income (HKD). Null for zones with suppressed data
activeIncome	integer | null	52850	Median monthly income of economically active households
floorSqm	integer | null	40	Median domestic floor area (m²)
workPop	integer | null	30188	Working population
ageU15	integer | null	7184	Population aged under 15
age2544	integer | null	17520	Population aged 25-44
age4564	integer | null	18640	Population aged 45-64
age65	integer | null	11690	Population aged 65+
areaSqkm	float	3.12	Zone area in km²

Derived from Census 2021 CSV (STPUG_21C.CSV, 200+ columns) — key column mappings: col 3=total_pop, col 125=domestic_households, col 132=avg_hh_size, col 148=median_active_income, col 200=median_floor_area.

How It Works in the API

The backend (functions/api/analyze.ts) uses a centroid-based radius approach:

1. For a query point (lat, lng), calculate haversine distance to all 211 zone centroids
2. Aggregate population in cumulative rings: 0-400m, 0-800m, 0-2km, 0-5km
3. Compute population-weighted median income across zones within 2km
4. Derive local density from the nearest zone: pop / areaSqkm

// Simplified from findNearbyPopulation()
const rings = [
  { maxM: 400,  pop: sumPopWithinMeters(400) },
  { maxM: 800,  pop: sumPopWithinMeters(800) },
  { maxM: 2000, pop: sumPopWithinMeters(2000) },
  { maxM: 5000, pop: sumPopWithinMeters(5000) },
];
const localDensity = nearestZone.pop / nearestZone.areaSqkm;
const weightedIncome = popWeightedAvg(zonesWithin2km, 'income');

Caution

Centroid vs polygon: because full zone boundaries are 17MB (too large for Cloudflare Worker’s ~1MB script limit), we use zone centroids (51KB). This means a zone is “within 800m” if its centroid is within 800m, even though parts of the zone may extend further. For the 211-zone resolution, this approximation is acceptable (avg zone radius ~0.5-1km).

Sample Zones

Zone	Area	Pop	Density	Income	Character
111	Sheung Wan	60,366	19,348/km²	HK$40,480	Dense urban, mixed commercial-residential
301	Mong Kok core	92,180	51,230/km²	HK$25,200	Extreme density, budget-mid market
602	Sai Kung Town	8,440	1,690/km²	HK$42,100	Suburban, weekend destination
203	The Peak / Mid-Levels	3,620	560/km²	HK$163,000	Ultra-low density, ultra-high income

Pipeline (How the Data Was Built)

1. Census CSV: Downloaded STPUG_21C.zip from censtatd (212 zones × 200+ columns)
2. Esri boundaries: Paginated download from ArcGIS FeatureServer (3 batches of 100, 211 MultiPolygon features)
3. Matching: Zone IDs mapped via suffix rule (GeoJSON "111S" → CSV "111")
4. Centroid extraction: Computed polygon centroid for each zone using average of boundary coordinates
5. Merged output: stpu_centroids.json (51KB) with centroid + all census stats per zone
6. Integration: Inlined as STPU_ZONES constant in analyze.ts

Full geometry available at /tmp/hk-census/stpu_combined.geojson (17.4MB) for offline analysis.

Used By

Model	How
Gravity Model	Real zone populations for exclusive radius rings instead of density × π×r² estimation
Catchment Analysis	STPU population for 5-min and 10-min walk zones
Regression Model	Local density and income as independent variables
Geodemographics	Zone-level income + age distribution for segment classification
Microsimulation	Seeds synthetic agents from actual zone demographics

Notes / Gotchas

• Data vintage: Census 2021 — now 5 years old. Population shifts (new NT housing estates, post-COVID expat changes) are NOT reflected. Treat as best-available baseline, not current truth
• COVID census: 2021 figures reflect pandemic conditions (lower population in expat-heavy areas)
• Null income: Some zones have suppressed income data (small sample size). Backend falls back to district median
• Daytime vs nighttime: Census counts residents at their usual address, not where they work. Central’s daytime population is 5-10x its resident population. This is the single biggest limitation for CBD locations — a restaurant in Central serving lunch crowds sees 10x the foot traffic that census data suggests
• 0-400m ring in CBD: Central and Wan Chai often show pop=0 in the innermost ring because STPU zone centroids are spaced 400m+ apart in commercial areas with few residents. This is expected — CBD locations rely on commuter inflow (gravity model), not local residents
• No building-level data yet: STPU zones average ~35K people. Future work: disaggregate to building level using footprints and floor counts from Lands Department GeoInfo Map API