What National-Scale Analysis Reveals About UK Hydrology

Research question

What becomes visible when UK flood-hydrology evidence is examined across a national network rather than one catchment at a time? The practical question is not whether a large dataset automatically produces certainty. It is whether repeated comparisons reveal inconsistencies, validation leakage, sampling limits and temporal sensitivity that a one-site workflow can leave untested.

Data and bounded cohorts

The evidence comes from four distinct experiments. The network study area-scales QMED before comparing adjacent gauges paired by river name. The QMED revalidation freezes a national station cohort and compares random station validation with five-fold unbuffered KMeans spatial-cluster cross-validation. The growth artefact selected 874 gauged stations, each with at least 10 non-rejected AMAX years and per-peak stage for the rating fit. The temporal study changes the held-out periods for a case station and a control.

The growth ceilings were estimated by resampling years within each station, recomputing L-moments and GF100, and using within-station sampling variance to bound observable-target R². These cohorts answer different questions and are not pooled into a single uncertainty calculation. Their exact sample sizes, statistics and evidence identifiers are recorded in the distributions and evidence record on this page.

Validation design changes the claim

Random folds test interpolation among stations that may still have nearby examples in training. Spatial-cluster folds instead hold out centroid-based geographic groups, making the experiment closer to transfer into an ungauged area. The KMeans folds were unbuffered: they do not assert a protected distance around each evaluation station. Donor custody was also enforced by restricting donor selection to the training folds.

Information symmetry matters here. The frozen challenger used the statistical-method estimate as an input, while its frozen urban-extent field carried no variation. The comparison is therefore a test of that specific challenger against donor-adjusted SM2025, not evidence that a wholly separate estimator surpassed the benchmark.

What the national view exposes

A single-site calculation can be internally tidy while neighbouring records remain mutually difficult to reconcile. Adjacent-gauge comparison brings that inconsistency into view before a predictive model is introduced. It is most useful as a review trigger: check the pairing, catchment areas, record periods, rating evidence and local hydrological explanation before deciding what the disagreement means.

The diagnostic does not identify which gauge is closer to truth and does not allocate disagreement among measurement, catchment or record effects. That is why its distribution remains separate from model validation and site-specific uncertainty statements.

Scale, skew and growth are different quantities

QMED is the index-flood magnitude. L-CV is the second L-moment divided by the first, a dimensionless measure of relative spread. L-skew is the third L-moment divided by the second and describes asymmetry, which gives the sparsely observed upper tail greater leverage. GF100 is the fitted one-hundred-year quantile divided by QMED: a tail growth ratio, not a substitute for diagnosing either relative spread or skew.

Finite-record sampling constrains L-CV, L-skew and GF100 differently. The ceiling estimates bound observable-target R² using the measured within-station sampling variance. The published artefact does not report an uncertainty band on those ceiling estimates, and the within-record GF100 spread must not be promoted into a general allowance for design-flow uncertainty.

The temporal example is a case, not a national trend

Each partition uses 60% calibration and 40% validation after a 730-day spin-up. P-A asks early → late, P-B late → early and P-C middle → outer. Holding out these different periods asks whether a model transfers across different parts of a record, not merely whether it reproduces a random sample from the same mixture of years. The case and control illustrate why the chosen periods should be disclosed and why a single aggregate score can hide temporal sensitivity.

The experiment was designed as a methodological demonstration. It has neither the cohort nor the attribution design needed to estimate how widespread non-stationarity is or to assign a cause.

Withdrawn interpretation and ruled-out conclusions

The random-validation accuracy-win interpretation is the only retraction recorded here. It was withdrawn because it did not survive the geographic-transfer test. The other two failed outcomes are ruled-out conclusions, not retractions or tested hypotheses: paired-gauge disagreement does not justify a station-level interval, and within-record GF100 variability does not justify a universal design-flow allowance.

Operational implications

• Review neighbouring gauges and their evidence before treating one at-site estimate as an unquestioned reference.
• Match validation to the deployment claim, disclose the fold construction, and keep held-out information out of donor choices.
• Report index-flood magnitude, relative spread, skew and tail growth separately so that sensitivity is visible rather than compressed into one number.
• Keep network consistency, population cross-validation, sampling variability and temporal sensitivity as distinct uncertainty evidence unless a calibrated combination has been demonstrated.

Limitations and disclosure boundary

River-name pairing is heuristic. The spatial clusters are unbuffered centroid groups. The growth result applies near the studied record length, and the published artefact gives no uncertainty band for the estimated ceilings. The temporal result is deliberately a bounded case study. Each limitation constrains interpretation even though the underlying cohort is broader than a conventional single-site analysis.

We disclose cohorts, validation logic, exact result distributions, failed hypotheses, evidence identifiers and operational meaning. We withhold feature engineering, learned parameters, training recipes, deployable artefacts, private data mechanics, operational infrastructure, customer material and licensed outputs. This is enough to audit the claim without reconstructing the engine.

What we do not conclude

• The same-river diagnostic is not a site-specific uncertainty interval and does not identify a single cause of disagreement.
• The QMED experiment does not show that the challenger is separate from SM2025 or that random validation is adequate for ungauged transfer.
• The growth experiment does not provide a universal uncertainty percentage for a design flood.
• The temporal case does not estimate national prevalence and does not attribute the observed sensitivity to climate.
• None of these experiments establishes regulatory approval or the replacement of established flood-estimation methods.