The Sussex bridge atlas

104 of 676

Every structure carries a confidence tier, given to it by whether any independent open dataset agrees that something is there. A structure is shown as verified when it sits in the top tier and has not been flagged as a probable artefact. Everything else is a candidate, and candidate here does not mean nearly verified. It means the height model saw something and we have not established what.

103 of the 104 verified structures reach the top tier because a railway or a watercourse crosses the road there in Ordnance Survey open data, which is two datasets agreeing without either knowing about the other. The last 1 of them reaches it through a validation list whose reuse terms we have not resolved and which is therefore not published anywhere on this page. Disclosing that one is cheaper than quietly keeping it or quietly dropping it.

What a precision of 0.13 means

Those last two columns are the reason this page has the shape it has. 45 structures were sampled, fifteen drawn at random from each tier, and every one was rendered as a 200 m image chip and looked at individually: hillshade beside height-above-terrain, with the road and the detected run drawn on. A chip counts as real only if a road genuinely passes under a structure there.

In the medium tier, 2 of 15 were real. 3 were trees or hedges, 1 was a road carried over rather than passing under, and 9 were noise: an Ordnance Survey centreline sitting a few metres off the real carriageway and clipping a farm building or a wall. That is what 0.13 means. It is not a middling grade. Roughly seven in eight of that tier are wrong.

The low tier scores the same, and that equality is the finding. In the coastal strip a road crossing another road at a clear angle usually was a grade separation, and the medium tier scored better there. Inland it is usually two lanes meeting at a hedge line. The medium tier does not earn its name in this county, so it is not presented as a middle confidence band, and nothing in it is shown as verified.

15 chips per tier is a small sample and each figure carries roughly ten percentage points of sampling error either way. That is wide enough to make 0.80 and a slightly better figure the same number. It is not wide enough to make 0.13 and 0.80 the same number. Verdicts came from the imagery alone. No site was visited.

How a structure gets found without anybody going there

The detector produced 3,853 raw runs. Three geometric tests reduced those to 676 structures after clustering anything within 25 m into one record, so split carriageways and parallel links collapse rather than counting twice.

1. 01
   Sample every road

   Every Ordnance Survey Open Roads link in the box, sampled every metre. 13,195,722 samples over 13,144 km of centreline.
2. 02
   Read three height models at each sample

   Bare-earth terrain, last-return surface and first-return surface, from the Environment Agency 1 m LiDAR composites.
3. 03
   Find solid cover

   Runs where the last return sits between 2.3 m and 14 m above the terrain and the two returns agree to within a metre. Agreement is what separates a deck from a tree canopy, which the first return catches metres above the last.
4. 04
   Check it spans the road

   A cross-section test at 2, 4 and 8 m either side. A deck spans the carriageway. A hedge or a building clipped by an imprecise centreline does not. This one filter throws away most of what the detector finds.
5. 05
   Check the road goes under it

   On a road passing under a structure the terrain stays at road level on both approaches. On a road carried over one it drops away beneath. This is the difference between a bridge and a bridge seen from the wrong side.
6. 06
   Ask what it crosses

   Railways from Ordnance Survey Open Zoomstack, watercourses from Open Rivers, roads from Open Roads. Independent corroboration, and the whole basis of the confidence tiers.

The cross-section test is the expensive one in both directions. It buys most of the precision the pipeline has and it costs narrow structures: cattle creeps, depot accesses and single-track lanes get discarded along with the hedges. That trade is a choice, and it is the choice that makes the top tier worth publishing.

Two metres of deck, and only 1.2 of it is published

Aerial LiDAR sees the top of a structure. A driver needs the bottom of it. Between the two sits the deck, which no aerial survey can see through, so every clearance here is a measured height minus an assumed thickness. We carry that assumption as two numbers rather than one, because only one of them has a published basis.

The fitted part is deliberately generous, and deliberately biased low. Reading a bridge as taller than it is costs a vehicle roof. Reading it as shorter costs a survey visit. A pipeline built to stop bridge strikes should spend the cheap error, so the estimates here sit below the truth more often than above it, and they should be described as conservative estimates rather than best estimates.

We cannot show you that fit. The comparison set it was tuned against is a list whose reuse terms we have not resolved, so neither it nor the error table it produced appears on this page. That leaves only the weaker claim, and the weaker claim is the one to take: the allowance is a fitted constant, it is biased low on purpose, and you cannot check it from here.

None of these numbers is a height. Not a signed height, not a legal height, not a certified clearance. They are a survey-targeting aid, which is the whole of what they are for, and nothing here should ever be used to route a vehicle. Our routing does not read them and this page does not ask it to.

Against the map

The point of the exercise is the comparison. In the same rectangle, on 21 September 2026, OpenStreetMap held 433 elements carrying a height value. This pipeline found 676 structures, of which 348 are estimated under 4.5 m and 104 are defensible today.

Neither dataset is a subset of the other and neither is an inventory. The claim is not that the map is missing 243 bridges. The claim is narrower and more useful: open data and a method find structures a tag-based map does not hold, in numbers large enough to be worth surveying, and they find them without anybody driving anywhere first.

The county, as found

The ten a survey vehicle would visit first

Verified structures only, ordered by scan priority. The measured deck height is shown beside the estimate so the arithmetic is visible: the estimate is that height minus the 2.0 m offset, carrying a band of plus or minus 1.5 m.

Every one of them is an A or B road under a railway line. Read that as the pipeline getting the obvious cases right rather than as ten discoveries. A road under a railway is the easiest thing in this dataset to be confident about, which is exactly why these are the ten at the top and why the interesting question is what sits behind them.

What this does not do

It does not move the zero. The measured count on what we actually know still reads 0, and publishing this page does not change it. Measured, in our navigation app’s vocabulary, means two heights read out of one survey taken on one day on the ground. An estimate off a national LiDAR mosaic of unknown vintage is a different kind of thing and gets a different word. What this atlas does is tell a survey vehicle where to go, which is how that number eventually stops being zero.

It does not know what half of these cross. The crossing is unknown for 337 of 676 structures, which is most of the candidate set and the main reason it stays a candidate set.

Absence from it is not evidence a road is clear. 0.4 percent of samples had no LiDAR available at all, the composites are mosaics of surveys flown over many years so a recent structure may simply not be in them, and at 1 m resolution a footbridge, a sign gantry or a thin canopy can fall between samples. A road with no structure on this page is a road this method did not find one on.

An arch has no single clearance. A masonry arch is lowest at the kerb and highest at the crown, and the figure here is closest to the centreline value, which is the most generous point on it. Roads running through embankments rather than under a discrete deck can be wrong by metres in either direction, and both directions occur.

Check it yourself

Every figure above came out of the files below. They are the run’s own outputs, identified by hash, and the page is built from a reduction of them committed to this site’s repository rather than from anything typed in by hand.

Where the data comes from

Every height in this atlas is derived from Environment Agency LiDAR Composite data published under the Open Government Licence v3.0, and every road, railway and watercourse from Ordnance Survey OpenData under the same terms. Contains OS data, Crown copyright and database right 2026. Contains Environment Agency information licensed under the Open Government Licence v3.0. The comparison counts are from OpenStreetMap, made available by OpenStreetMap contributors under the Open Database Licence, and are never joined into the atlas itself.