REWT Rivers of England and Wales, Temporally

Methodology

The finding that sets the order

A scoping exercise built this the obvious way: take the modern survey, filter out channels known to be later, splice on attested medieval courses, model the flow. Each stage layered on the last.

Almost every serious defect it found was a fault in the modern survey’s topology that the later stages inherited, amplified, and then hid. Rivers that stopped inland. Basins draining into a node with no outflow. Tributaries drawn flowing away from the river they join. Thousands of kilometres the water could never reach from the sea. Flow arriving at a confluence and vanishing.

None of that is a reconstruction problem. It is a survey-topology problem, and everything built above it stood on unstable ground. Hence the order below, which is a conclusion rather than a preference.

0 · A build that reproduces

One command from an empty checkout to the published outputs, producing the same result every time. Inputs immutable and checksummed; every stage idempotent; no cached intermediate ever preferred to the configuration that produced it; the audit run as part of the build rather than afterwards by hand.

This is first because the scoping exercise could not make that claim, and the failures were mundane and expensive: stages that silently overwrote each other, a cached intermediate preferred to a corrected setting for a month, a mistyped identifier that did nothing at all while the pipeline reported it as applied.

1 · Make the modern network traversable

OS Open Rivers does ship a topology — some 197,000 explicit nodes, with flow direction and names — and it is very nearly the only open British watercourse product that does. What it does not offer, and does not claim to, is traversability: that from any stretch the water can actually be followed downstream to the tide. It is a generalised cartographic derivative with a node structure attached, not a hydrological model, and nothing in it guarantees that a chain of links is unbroken or that every direction is right.

Making it traversable is the first phase’s work — as faithfully to the survey as the evidence allows. This phase reconstructs nothing. Its product is a functioning drainage network, and keeping that boundary sharp is what lets it use anything that helps.

Three shapes of fault account for nearly all of it, and one question separates them: which end of the stranded channel touches the network?

finding what it is remedy
downstream end has inflows and no outflow a real gap a connector
upstream end touches a draining line drawn backwards reverse it — costs nothing
a reach two basins need in opposite directions neither make it two-way
the end sits on another channel, unjoined no gap at all a junction: add no geometry

Getting that question wrong produces a fix that looks right nationally and is wrong at the place. Check at the place, never in the total.

And a fault found at one place is repaired everywhere or it is not repaired. Each of the remedies above is a general rule applied across the network, never a targeted mend at the junction somebody happened to look at; the rules are stated once, in order, with the reasoning that produced them and the rows they were raised from, on the rules page. That page is generated from the file the build reads, so it cannot claim a rule is applied when it is not.

Why start from OS Open Rivers at all?

Because the alternatives are worse, and for instructive reasons. The choice was made on a comparison rather than by default.

CEH’s 1:50 000 Digital River Network — no topology, by its publisher’s own statement, and available for educational use only. Neither problem is fixable.

OS OpenMap Local (water) — open, and far better geometry: real polygons from which a true medial axis and channel width can be derived. But it is cartography with no topology whatever; measured, only about 4% of its endpoints are shared between adjacent features. Four separate attempts to give it one failed. Its proper role is to contribute attributes — width, better centrelines — to a structure that comes from elsewhere.

The un-generalised OS water network (Features API) — the same data before generalisation, and the obvious answer but for its licence: Premium, and not redistributable. Nothing derived from it could ever be published, which disqualifies it for an open dataset however good it is.

A network derived from a terrain model — flow-routing over a digital elevation model produces a network that cannot have these defects: routing over a filled surface yields a directed forest, with no sinks and no cycles, by construction. It is genuinely tempting. But at 50 m resolution a cell spans roughly an eighth of a lowland meander wavelength, so the geometry is poorest exactly where this project cares most — the Fens, the Levels, the Humberhead — and it carries no names, so nothing in it joins to a documentary attestation. It is admissible as a check on connectivity and not as the network itself: it can say whether water can get from one place to another, and cannot say by which line.

So: OS Open Rivers supplies the topology and the names; other products contribute geometry and evidence to it. The alternative is to ask a cartographic product to supply a topology it never had, which has been tried.

What an unburned terrain run actually does

This qualification decides how far a terrain check can be trusted, and it cuts against the idea above.

Any hydrological conditioning burns the mapped network into the surface before routing, deeply enough that the burned line wins against 50 m cells over embanked and levelled ground where the true gradient is below the vertical resolution. Wherever the survey draws a channel through a culvert or under a bridge, the burn cuts a trench and the water follows it. Continuity there is supplied by the survey’s line, not by the terrain — which is precisely why a burned run is no use as independent evidence.

Unburned, the two cases behave very differently on a bare-earth model:

So the terrain’s verdict is sound on connectivity — can the water get from here to the sea at all — and unsound on route — by which line. That is still the question this phase needs answered, because a dead end is a connectivity failure. But it means the terrain may flag and rank, never draw. Any continuation it proposes is a place to look, not a line to adopt.

Three cheap things sharpen it, in rough order of value:

  1. Run it both ways and diff. A discontinuity that appears in the burned run as well is a real gap in the survey; one that appears only unburned is an embankment artefact. That single comparison separates the two populations with no extra data.
  2. Record breach depth per carve. A trench cut several metres through ground is telling you it crossed an embankment; a shallow one is following a valley.
  3. Mask the embankments using the open road and rail networks — a culvert candidate is a road or railway crossing a modelled flow path. Most work, least certain, and it should follow the first two.

The sea is a test, not a definition

Every rule above ends in the same place — and then it reaches the tide — and until now the tide was not modelled at all. A river mouth terminated in a patch of nowhere, each basin in its own separate patch, and “reaches the sea” was true by definition rather than by test. That is also why an estuary is the hardest thing here: the survey draws a river mouth as an area and offers the network no line to end on. Measured against this build, 511 of the 1,281 outstanding in-scope dead ends sit within 250 m of mean high water and more than 250 m from any tidalRiver link — they arrive at the coast and stop, because there is nothing there for them to join.

The share has risen, and that is the part worth noticing. When the construction was first specified the measurement was 245 of 968: the denominators belong to different builds and are not comparable as counts, but the proportions are — a quarter has become two fifths. So the case for giving the sea a geometry is stronger than the sentence that used to argue it, which had been quoting the older pair. A stale figure that supports the conclusion it was written for reads exactly like a current one, and nothing about it invites a check.

So the sea is being given a geometry, in this phase rather than a later one. The trunk is not drawn. It is what remains where the routes agree.

  1. A cost surface over the sea, from bathymetry: water shallower than a stated clearance is impassable, and deeper water costs less as it deepens. One parameter, carrying a number and no adjective.
  2. Least-cost paths run between the river mouths, not out to open water. This is the step the construction turns on. A path seaward would take each mouth to the nearest edge of the survey by the cheapest route and nothing would converge — east-coast mouths would run east and west-coast mouths west. A coastal route exists only as a relation between the mouths, so the mouths are what the paths connect.
  3. The cheapest crossings between mouths give a network over them, each of its edges the deepest available channel between two river mouths — which is what “follow the deepest channel” turns out to mean once you stop drawing it and start measuring it.
  4. The trunk is the segments that many of those routes share. If no continuous trunk emerges, that is a finding about the bathymetry and is reported as one, not repaired by lowering the threshold until a line appears.

What this buys is a test. No node has inflows and no outflow, except at the sea becomes except where it joins the trunk — checkable rather than definitional. A river that cannot reach the trunk is a finding and is named; today it is an assumption. It also makes port to port computable, down one river, along the coast and up another, which is the journey the navigation evidence is about and which no purely fluvial network can express.

The coastal half of that journey is built on a different rule and has its own page. Where this section gives the sea a geometry so that a river mouth has something to end on, the routing surface decides which water a passage may cross at all — and it decides it on sight of land, which is geometry rather than history and can be re-derived by anyone from a height and one formula.

And it changes what a published number means, so both readings go out together — but they are not two grades of one test. Measured against this build, 93.58% of in-scope length reaches tidal water and 96.26% reaches the sea, and the second is the larger because 2,890 km drains to the coast without ever touching a watercourse the survey calls tidal. A mouth discharging through a sea wall is not a defect. The readings cross rather than nest: 58 km reaches tidal water that the sea will not take, and 3,894 km reaches neither, which is the defect and the work.

The warning that goes with this has inverted, and is the more necessary for it. While the sea could only make reachability harder, the risk was a number quietly worsening under a stricter test. What happened instead is that the stranded total fell, from 6,785 km to 3,894 km — and not one kilometre of it was repaired. Coastal drainage that always did reach the sea is now counted as reaching it. A number improving because of a change in what is counted is the flattering direction, and therefore the dangerous one. Every figure measured before the change belongs to the old definition and is not comparable with one measured after it.

It carries no date, and the naming is what keeps it that way. The artefact is named by its parameter and not by its purpose: it is the connectivity structure of water deeper than a stated clearance. The words plausible and sailing are excluded deliberately — a vessel implies a draught, a draught implies a period, and a period is a date. A route optimised for a particular hull in a particular season is a historical claim and stays out of scope; it would look much like this on a map, and building it while calling it this is the drift the project is arranged to prevent.

One sentence travels with the bathymetry. Research bathymetry is published DO NOT USE FOR NAVIGATION — a constraint on purpose rather than on redistribution, since it is surveys of varying age merged into a surface rather than a charted product. A topological terminus is not navigation. But nothing built on it may be offered as a route a vessel could follow, and the naming rule above is how that is kept rather than merely acknowledged.

And here continuous coverage beats resolution. The national hydrographic holdings are far finer and are published as thousands of individual surveys with gaps between them — and an accumulation cannot flow across a hole. It is the same trade as the Welsh leats: the better-structured source answering a different question.

Exit condition. No node carries more than a nominal flow into nowhere; the unreached residue is explained rather than merely counted; and the audit runs clean or every exception is named with a reason.

2 · Recover the old courses

Only once the modern network is sound, because a traced course must attach to something.

From the maps. Victorian six-inch sheets label some channels outright as superseded — Old Course, New Cut — and GB1900, a volunteer transcription of every label on them, has made those statements searchable. Each is an assertion by the surveyor, not an inference. Trace them at the resolution the sheet supports, snapping to the printed ink.

From the sheets themselves, and not only from their labels. A caption is the surveyor volunteering an observation. Most abandoned courses never got one: Old Course is written where somebody thought the fact worth recording, not wherever a river has moved. The labels are a lower bound on what the six-inch record holds — the cheapest findings in it, not the complete set.

So the sheets should also be read systematically: every river course inspected along its whole length against the printed channel, on the first edition where it exists and the second where it does not, flagging every significant departure of the historic channel from the modern line. That finds the courses no volunteer transcribed because no surveyor captioned them, which is most of them.

And on the sheets printed in colour, the water is already separated for you. Where an edition prints watercourses in blue — many of the earlier county sheets do, and the mill leat and new cut illustrated elsewhere on this site are examples — the channel can be lifted from a georeferenced sheet by colour alone, before anybody looks at it. That is a cheap national first pass: a raster mask of everything the surveyor drew as water, which the tracer then snaps to and a person adjudicates. It does not date anything and it does not decide which of two channels is the older, but it converts find the ink from a per-sheet human task into a per-sheet computation, and leaves the judgement for the part that needs one. It applies only to the editions and printings that carry the colour; the monochrome sheets still have to be read.

Two things keep this bounded rather than open-ended. The first pass is mechanisable: the modern network can be overlaid on the georeferenced sheet and the divergence measured, so a person judges only the places where it exceeds a threshold. And the threshold has to sit above the modern product’s own generalisation — OS Open Rivers is a cartographic derivative, and a sweep tuned finer than its own smoothing will report the difference between two drawing conventions and call it a change in the river.

It is work that scales with the length of the network rather than with the number of labels, which puts it in a different class of cost from the labelled assertions above. It also divides cleanly — one reach, or one sheet, at a time — which makes it the most natural thing in the whole programme to invite help with.

From under the reservoirs. Channels drowned since the mid-nineteenth century are recoverable from sheets made before the dam — which often means the first series, not the second, because much of Victorian water supply predates it.

From the documentary record. Drainage and navigation works were legal acts over specific ground, and a scheme had to be described against what was already there: land taken had to be identified, and a new cut located relative to the old channel. So the records behind such works — deposited plans, awards, estate and enclosure maps — sometimes show the river as it ran before the work, and were drawn for a purpose that required them to be accurate about it.

How often is not known, and should not be guessed. Survival is uneven, coverage is patchy before the nineteenth century, and many plans show only the works proposed and not the ground they replaced. The expectation is reasonable rather than established, and the first task in this strand is to test it on a sample and find out — a scheme at a time, recording what the record actually yields.

Exit condition. Every Old Course and New Cut label adjudicated — traced, rejected with a reason, or recorded as illegible — with provenance per vertex; every matched reservoir either traced from a pre-dam sheet or recorded as having none; and every reach swept against the earliest available edition, with each flagged departure traced, dismissed as generalisation, or recorded as illegible. How many that is.

Go to the sources, not to the synthesis

The deeper point about this phase, and it reframes what Oksanen’s GIS of pre-1348 inland navigation is for.

A synthesis of medieval navigation is a digitisation of other people’s evidence — documentary attestations of shipping, archaeological finds, place-name landing places, and courses fossilised in parish boundaries — published at the resolution a national synthesis needed. Where a course is boundary-derived it sits a median 236 m from the modern channel.

That 236 m is not error. It is the historical signal: the boundary remembers where the river was, and it is exactly the quantity a higher-resolution trace would sharpen. So the route to a better course is the material behind the synthesis, retraced and georeferenced more finely than the synthesis needed — not the synthesis resampled.

Why an existing synthesis is not spliced in

An earlier plan had a further phase: connect the courses carrying no map label and no traced replacement, so that the synthesis formed a network rather than a set of drawn assertions. It is dropped, and the reasoning is worth keeping.

A boundary-derived course sits a median 236 m from the modern channel. A traced course sits where the ink is. Merging two geometries of different orders of accuracy means the reconciliation gets made by the pipeline, silently, at every junction — in exactly the places nobody is looking. Splicing a national synthesis into a traced network would degrade it.

The synthesis keeps its two proper roles: an index saying where to look, and a check saying whether a trace is corroborated. Both are served by keeping it beside the network as an evidence layer rather than in it.

This also removes the hardest unsolved problem of the scoping exercise. Holding two graphs over one geometry was entirely a consequence of keeping a synthesised line and the modern network in a single table. If it is never spliced, the problem cannot arise.

3 · Model the flow

Last, because it is cheap to re-run and meaningless on an unsound network. A per-cell water balance routed over the network burned into a terrain model (OS Terrain 50), validated against modern gauged flow — and reported as a relative index of channel capacity, never as a reconstruction of premodern discharge.

What the flow model is for, beyond drawing the river

Power available at a site is flow times fall, and both are quantities this network will hold: a modelled discharge for every reach, and a drop along it taken from the terrain. So the same model that says how much water moved says where water could be made to work — and that is a different map from the river, with a different audience.

Two questions follow, and they belong to different centuries. Corn: how much grinding capacity a district could raise, set against how many people it had to feed — the constraint that put a mill on nearly every workable fall in medieval England. And power for tools: the fall-and-flow that drove hammers, bellows, fulling stocks and finally machinery, which is where a hydrography stops being a landscape question and becomes an industrial one.

It is a potential surface, and must be published as one. Three limits, stated here so they travel with it:

This is also what gives the mill channels a use even if they are never dated: some four thousand labelled leats are a validation set for a potential model, and a model that puts power where the surveyors found none, or none where they found a mill, has told you something either way.

Three questions the same model answers, and why they must not be asked together

Flow, fall, known mill sites and population are four quantities. Fix any three and the fourth becomes a prediction, which gives three genuinely different studies:

Each excludes the others in the same study. Population cannot both calibrate the model and be tested by it; a mill distribution cannot be both the training set and the validation set. The temptation to run the loop once and quote all three results is the failure mode here, and it produces figures that are internally consistent and evidentially empty.

Two further cautions belong with it. The labelled mills are c.1900 attestations, so calibrating a medieval model on them imports a Victorian distribution — one shaped by four centuries of consolidation onto the best sites, and by the mills that steam had already killed. And the model’s fall comes from a national terrain surface that cannot see a weir and a leat, so it under-reads exactly the sites where a modest stream was engineered into a usable head.

The comparison this makes possible for the first time

Water power and navigable water are usually studied apart, because no dataset has carried both. This one is built to: the same network holds a modelled discharge, a fall, and a dated navigability, reach by reach.

That allows a question worth asking directly. Before coal, the determinant of a settlement’s success may have been the power available to it rather than the water route out of it — grinding, fulling, forging and sawing are things a place did for itself, while navigation mattered most once there was bulk to move and, above all, coal to bring in. Which of the two better explains where places grew, and when the answer changes, is an empirical question that has wanted a dataset rather than an argument. This project does not settle it; it is the first thing to put both variables on the same lines.

Principles

Every line carries what it rests on. Source, date bounds and their certainty — never free text, never an unattributed assertion, and never a single confidence figure mixing how sure we are of a date with how sure we are of a modelled flow.

Absence is drawn, not hidden. Where the evidence runs out the map shows that it has run out. A channel missing from the picture must never be readable as no river here.

Nothing is deleted to correct it. A superseded line is retired with its reason and stays in the record. This is a research dataset; the audit trail is part of the product.

Judgements live in diffable files. A decision about a particular river belongs in a committed text file with its evidence, not in code.

The date is a parameter. Nothing in the method fixes a single year.

Where this is worked out in detail

   
The rules every general rule the build works to, in the order they apply, with what each one waits on
The rebuilt landscapes the regions where the modern channel is worthless, and the sources that record what was there before
The scale of the work how much is bounded, how much is not, and who does it
Temporality which dates get published, and why an epoch is a build rather than a filter
Evidence what each source can and cannot say