Hong Kong road traffic estimates

Summary

This page estimates daily traffic (AADT, both directions) and annual vehicle-kilometres travelled (VKT) for every road section in Hong Kong, using the Transport Department's 2024 Annual Traffic Census core-station counts together with the official road network, calibrated to TD's published territorial total.

The territory-wide total is calibrated to the official figure, so the total itself is not an independent result. The analytical value lies in the distribution across the network — which roads carry the load, how traffic is structured by road class, and how well the model reproduces TD's regional split.

1. Data sources

2. Method

2.1 Station matching (measured)

Appendix C of the census report gives the AADT of all 1,685 counting stations; 191 stations also have directional detail. Matching rules:

• Directional matching: directional AADT from the 191 workbooks is matched to carriageways within 250 m whose bearing agrees (≤75°); two-way single carriageways take the station's two-way total.
• Two-way matching: remaining stations (Appendix C) are matched to segments within 300 m on the same named road; each carriageway of a divided road takes half the two-way flow, a two-way single carriageway takes the full flow.
• Class consistency: the TD road class of the station (expressway / trunk / primary distributor / district distributor / local distributor / rural road) must be compatible with the segment's class, so trunk-road flow is not assigned to a neighbouring lane.

Result: 29.7% of the 51,498 segments carry a directly measured value.

2.2 Interpolation for un-surveyed sections (two stages)

1. Same-road corridor interpolation: a road's measured values are extended along the corridor (exponential decay, λ = 1,500 m) so long corridors (Tuen Mun Road, Tolo Highway) are not dragged down by class averages.
2. Network-distance interpolation: remaining sections decay from the nearest measured section measured along the road network (λ = 1,200 m, capped at 6 km) — not as the crow flies, so values are not borrowed across the harbour or through hills. Floor: 15% of the class median.

2.3 Vehicle-kilometres

Daily VKT = Σ(AADT × section length); annual = daily × 365. Tunnels and bridges are included.

2.4 Calibration (measured values unchanged)

Measured sections keep the published census AADT with no scaling at all (17.62 M veh-km/day). Interpolated sections are calibrated to TD's published regional control totals (Hong Kong Island 5.46 M, Kowloon 9.01 M, New Territories 25.22 M veh-km/day):

A final normalisation brings the territory total to 39.69 million veh-km/day (14.49 billion veh-km/year). In other words the interpolation was over-estimating by 30–53%, while measured sections are preserved exactly.

2.5 Why is measured coverage only 29.7% when TD's survey network is 87.1%?

The two percentages measure different things:

In short: this is a definition difference in how much road one station represents, not missing data. This analysis deliberately uses a conservative 250 m matching radius (a station's flow is applied only to the same road near the station); the rest is handled by the two-stage interpolation in 2.2. Our network length (3,451 km of carriageway = 2,224 km of road) is within 0.8% of TD's 2,241 km, confirming the road network itself is comparable.

3. Expert assumptions

These assumptions underpin the estimates; departures from them change individual road figures.

3.1 Calibration sensitivity (disclosure)

Three options were tested on the same matched dataset:

Option (B) fails because measured stations cluster in urban areas: a single factor applied only to interpolated sections loads the whole adjustment onto the New Territories (63.5% official vs 59.5% modelled, -4.0 pp). Option (C) calibrates per region (interpolated factors: HK Island 0.53, Kowloon 0.47, NT 0.70), preserving measured values exactly while matching both the territory total and the regional split. The trade-off: regional totals become inputs (control totals), so the regional split is no longer an independent check; what remains checkable is that the required correction is consistent across the three regions (0.47-0.70).

4. Results

4.1 Regional split vs TD published figures

TD published (ATC 2024 report): 39.69 M veh-km/day total — Hong Kong Island 5.46 M, Kowloon 9.01 M, New Territories 25.22 M. The modelled regional split differs by no more than ±1.1 percentage points — the key validation of this estimate.

4.2 By road class

Expressways are only 693 km of the network but carry 38.7% of all traffic; district and service roads together account for about a third.

4.3 Roads with the highest traffic (top 20)

Unnamed/connector roads (slip roads, interchanges) carry a further 4,754,112 veh-km/day (12.0%, 428.6 km) and are excluded from the table above.

4.4 Busiest sections

The highest measured AADT is Harcourt Road (149,710 veh/day) — a measured, not modelled, value.

5. Traffic map

Click the image to download the high-resolution version (3000 × 2100 PNG)

Colour and line thickness = estimated daily two-way flow (vehicles/day). Bands: <5,000 (light blue), 5,000–10,000, 10,000–20,000, 20,000–40,000 (yellow), 40,000–80,000 (orange), ≥80,000 (dark red). The inset shows Kowloon and the harbour core. Tunnels and bridges are included.

6. ATC station network (reference)

Click the image to download the high-resolution version (3000 × 2100 PNG)

All 1,694 traffic-census counting stations (1,685 of them with a 2024 AADT). Colour = coverage type, marker size = 2024 AADT (two-way flow). Coverage A are the core stations (on major roads, surveyed every year); B and C make up the rest of the survey network.

By road class

This distribution explains the coverage difference in section 2.5: stations cluster in urban areas (about 60% of the urban network has a station) but only about 30% of the New Territories network does, which is why the NT relies most on interpolation (section 2.4: NT interpolated factor 0.7033 versus 0.47–0.53 in urban areas).

7. Limitations

• Total is calibrated: the territorial total is fitted to the official figure and is not independent evidence; error shows up in the distribution between sections.
• Un-surveyed sections carry more error: 29.7% of sections have direct measurements (denser in urban areas). Urban core ±20%; rural and village roads up to ±50%.
• Regional split used as a control: interpolated sections are calibrated to official regional totals, so "the regional split matches" is no longer independent evidence; what is checkable is that the required correction is consistent across regions (0.47-0.70, i.e. interpolation was 30-53% too high).
• Units differ: the 3,451 km quoted here is carriageway length (both sides of a divided road counted separately), equal to about 2,224 km of road versus TD's 2,241 km (0.8% apart).
• Service roads: the least certain class (see A13).
• Class as a proxy for function: OSM classes substitute for TD functional hierarchy; expressway/trunk boundaries may shift.
• No vehicle mix or time-of-day split.
• Single base year: 2024; post-pandemic and new-road trends are not captured.

8. Reproducing this analysis

Everything was computed in Python (GDAL/rasterio, geopandas, scipy):

• parse_atc.py — parse ATC station workbooks into directional AADT
• vkt_model.py — station matching, IDW interpolation, VKT, calibration
• make_vkt_map.py — traffic flow map
• make_station_map.py — ATC station network map
• extract_regions2.py — fetch the three region boundaries via Overpass

Data revisions: network 2026-08-27, traffic census 2024.