UK open data — the catalogue¶

The sources behind the atlas chapters. All are free and need no account. All were checked on 10 August 2026.

If a source is down, use the cached copy in data/external/ and record the date you used it.

starter/fetch_external.py is a finished fetcher for the first source. Read it before you build the others. The pattern is the same every time.

Read this first¶

Three chapters need an area code before they return anything. NaPTAN needs an ATCO area code. The PCT needs a region folder name. The Census needs a local-authority code.

None of these can be guessed, and all of them look as if they can. An assistant will give you a confident wrong answer.

Do not guess. Do not ask. Each one has a known source:

You need	Where it comes from
ATCO area code (chapter 1)	`data/external/atco_area_codes.csv` in this repository — all 150, by name
PCT region name (chapter 5)	the list in the PCT entry below — they are historic counties, so Bristol is `avon`
Local-authority code (chapter 4)	the deprivation file itself, which carries district names beside their codes
The ONS boundary service address (chapter 3)	copy it from the ONS entry below — do not let an assistant reconstruct it from memory

The last row matters most. That address is long and contains a service identifier nobody could work out. Copy it. Do not retype it and do not ask for it.

The other three chapters need only your bounding box. Once chapter 1 sets your patch, they work straight away.

This page is a list of the things an assistant will invent. In real work nobody gives you one. You write it yourself before you start. See agent_guide.md.

The scope rule¶

Real geospatial analysis is a term's work. The atlas uses three operations, and everything on this page can be reached with them:

Operation	Example
Key join	LSOA code → deprivation decile
Bounding-box filter	keep rows whose lat/lon fall in your study area
Distance	how far is this casualty from the nearest stop

Out of scope: polygons, coordinate-system transforms, GDAL, R, PostGIS.

If your assistant suggests geopandas, shapely, pyproj, ogr2ogr or an EPSG:27700 conversion, the problem has grown too large. Ask for the bounding-box or distance version instead.

These tools are not bad. But you cannot check what you cannot read, and a projection error produces a map that looks correct.

Tier 1 — a CSV at a URL¶

pandas.read_csv(url) and you are done. Start here.

STATS19 — road casualties¶

Every reported injury collision in Great Britain, with coordinates.

Gives you: where people walking and cycling get hurt in your patch.
How: https://data.dft.gov.uk/road-accidents-safety-data/dft-road-casualty-statistics-collision-2023.csv (and -casualty-2023.csv), joined on collision_index.
Worked example: starter/fetch_external.py.
The trap: casualty_type and collision_severity are numbers. Their meanings are in a separate document. 0 is a pedestrian, 1 a cyclist. Severity 1 means fatal, not slight. An assistant will guess these, and the guess will look reasonable.
Licence: OGL v3, © Crown copyright.

Index of Multiple Deprivation 2019¶

The standard English measure of relative deprivation, by small area (LSOA).

Gives you: how your patch sits in England's deprivation distribution.
How: File 7 (all scores, ranks, deciles) — 32,844 rows, one per LSOA.
The main trap: IMD 2019 uses 2011 LSOA codes. The boundary service in tier 2 returns 2021 codes. Most codes are the same in both, so the join appears to work. It drops the areas whose codes changed. Count your rows before and after.
Second trap: decile 1 is the most deprived, not the least. Many published charts get this backwards. Check yours twice.
Useful: every row has both the name and the code of its local authority district. So you can cut this file by name, such as "Leeds" or "York", without knowing any code. The code in those same rows (E08000035 for Leeds) is the one the Census chapter needs.
Licence: OGL v3.

DfT road traffic counts (AADF)¶

Annual average daily flows on the road network, by local authority.

Gives you: what the traffic in your patch's local authority is doing, and whether it is growing.
How: https://storage.googleapis.com/dft-statistics/road-traffic/downloads/data-gov-uk/local_authority_traffic.csv
The trap: these are annual average daily flows. You cannot get a peak hour from them.
Licence: OGL v3.

Census 2021 via Nomis¶

Households, car availability, method of travel to work, by small area.

Gives you: how many households in your patch have no car — the people for whom walking, cycling, and the bus are the whole transport system.
How: the Nomis API returns CSV directly, e.g. https://www.nomisweb.co.uk/api/v01/dataset/NM_2062_1.data.csv?geography=...&measures=20100
The trap: finding the right dataset id and geography code is most of the work. The car-availability table (TS045) is dataset NM_2063_1, and a district's small areas are requested as geography=<district code>TYPE151 — for example E08000035TYPE151 for every LSOA in Leeds. You do not need to hunt for that district code: it comes free with the deprivation file above.
Verify it: check the row count and one value against the Nomis web interface by hand.
Licence: OGL v3.

Tier 2 — an API, shaped with the assistant¶

More capable, more ways to be silently wrong. Do these second.

OpenStreetMap via Overpass¶

Schools, shops, crossings, cycle lanes, anything mapped.

Gives you: what is actually near the stops — trip generators, severance, the cycle infrastructure that does or does not exist.
How: POST an Overpass QL query to https://overpass-api.de/api/interpreter (mirror: https://overpass.kumi.systems/api/interpreter).
The trap: you must use POST. A GET with the same query returns an HTML error page. response.json() then fails with an error that looks like a parsing problem. Also, schools are often mapped as areas rather than points. Use out center; or you get no coordinates.
Be polite. This is a free service on donated hardware. Set a real User-Agent, send one query at a time, and save the result to disk at once.
Licence: ODbL, © OpenStreetMap contributors. Attribution is required.

ONS Open Geography — which LSOA is this point in?¶

Point-in-polygon, done on their server, so you never touch a polygon.

Gives you: the LSOA code for any point in your patch, which unlocks IMD and Census.
The address — copy this, do not retype it and do not ask for it:

https://services1.arcgis.com/ESMARspQHYMw9BZ9/arcgis/rest/services/Lower_layer_Super_Output_Areas_December_2021_Boundaries_EW_BSC_V4/FeatureServer/0/query

with these parameters:

geometry=<lon>,<lat>
geometryType=esriGeometryPoint
inSR=4326
spatialRel=esriSpatialRelIntersects
outFields=LSOA21CD,LSOA21NM
returnGeometry=false
f=json

Verified working: the point -1.5491, 53.7965 returns E01033016 (Leeds 111E). Use that as your first test — if you get that answer back, your request is correctly formed and you can move on to your own points. - Why copy it. The address contains a service identifier (ESMARspQHYMw9BZ9) that nobody can work out. Ask an assistant for it and you get the right shape with the wrong identifier. - The trap: longitude comes first. Swap them and you get a silent NO MATCH for every point, or worse, a match somewhere in the North Sea. - Licence: OGL v3, © Crown copyright and database right.

NaPTAN — the real bus stop register¶

Every public transport access point in Great Britain. Chapter 1's source.

Gives you: real stop names, locations and codes for your patch.
How: https://naptan.api.dft.gov.uk/v1/access-nodes?dataFormat=csv&atcoAreaCodes=450 — where 450 is the ATCO area code for the region your patch sits in (450 is West Yorkshire, the Leeds demonstrator's area).
Finding your area code: data/external/atco_area_codes.csv lists all 150, by name. Look yours up there. Do not guess and do not ask the assistant. The codes look guessable and are not: West Yorkshire is 450, York is 329, Bristol is 010.
The trap: the code must be three digits. Keep the leading zero on codes like 010. The API says the format is wrong but not which code you wanted.
Check it worked: tens of thousands of stops come back. Filter to your bounding box and count them. If you get zero, the area code is wrong. Check the lookup table again before changing anything else. Also remove the stops marked inactive.
Licence: OGL v3.

Propensity to Cycle Tool — cycling potential¶

DfT-funded model of how much cycling each road segment could carry.

Gives you: where cycling would happen under a given scenario, against what happens now.
How: https://media.githubusercontent.com/media/npct/pct-outputs-regional-notR/master/commute/lsoa/west-yorkshire/rnet_full.geojson — swap west-yorkshire for the region containing your patch.
The trap: the regions are historic counties, not cities. Bristol is avon. Manchester is greater-manchester. Newcastle is north-east. Hull is humberside. There are forty-five:

avon, bedfordshire, berkshire, buckinghamshire, cambridgeshire, cheshire,
cornwall-and-isles-of-scilly, cumbria, derbyshire, devon, dorset,
east-sussex, essex, gloucestershire, greater-manchester, hampshire,
hereford-and-worcester, hertfordshire, humberside, isle-of-wight, kent,
lancashire, leicestershire, lincolnshire, liverpool-city-region, london,
norfolk, north-east, north-yorkshire, northamptonshire, northumberland,
nottinghamshire, oxfordshire, shropshire, somerset, south-yorkshire,
staffordshire, suffolk, surrey, wales, warwickshire, west-midlands,
west-sussex, west-yorkshire, wiltshire

A wrong name gives a 404 rather than a wrong answer. - The trap: it models 2011 Census commuting only. No shopping, no school, no leisure. The baseline is fifteen years old. It is a model of a scenario, not a measurement, and your sentences must say so. In some cities it does not match observed cycling at all. - Licence: OGL / CC-BY. See pct.bike.

open-meteo — historical weather¶

Hourly weather for any coordinates, back decades. Chapter 7's source.

Gives you: a year of rain and temperature for your patch, as JSON.
How: https://archive-api.open-meteo.com/v1/archive?latitude=53.80&longitude=-1.55&start_date=2025-01-01&end_date=2025-12-31&hourly=precipitation,temperature_2m — no key, no account. Point the coordinates at the centre of your patch.
The trap: the response is JSON containing parallel lists. One list of timestamps, one list per variable, matched by position. Describe the table shape you want, then check the first and last rows against the website's own charts.
Licence: CC-BY 4.0, attribution required. See open-meteo.com.

Tier 3 — real, and out of scope here¶

Named so you know they exist and know why you are not using them this term.

Source	Why not
OS Open Roads	~1 GB GeoPackage, needs GDAL. The right tool for road classification, wrong tool for a 3-hour studio.
Cycle infrastructure Level of Service (`osmactive`)	Needs an R toolchain. This is how the Coventry work classified cycle provision — ask if you want to see it.
Bus Open Data Service (live/GTFS-RT feeds)	Needs an API key and a feed parser. This is where real AVL data comes from.
EPC register	Bulk download sits behind a sign-in.
OS Open UPRN / Code-Point Open	Useful for address-level work, needs projection handling.

What to record for every source you use¶

Record these for every source you use:

Where it came from — the URL, and the date you pulled it.
The licence line — OGL v3, ODbL, and CC-BY all require attribution.
Row counts — national, after your patch filter, after every merge.
The approximation you accepted. A nearest centroid instead of true containment. A rectangle instead of a boundary. 2011 codes joined to 2021 codes. All of these are acceptable if you say you made them.

Real places¶

Everything in your atlas is real: real casualties, real deprivation, real streets where people live.

Two rules follow.

State what each dataset is and when you fetched it, so a figure cannot be mistaken for something else.

Keep description separate from judgement. "This area is in England's most deprived tenth" is data. Conclusions about the people who live there are not yours to draw.