All figures here are after correction: the raw satellite estimate adjusted by a residual fitted against Defra finals and predicted from July rainfall. The fit absorbs, together, the harvest-period losses a satellite cannot see and the winter-wheat-versus- all-wheat basis — raw runs ~0.35 above Defra in a normal year — so corrected values are Defra-comparable by construction (section 3). Since 20 Aug the product headline is the raw, raster-faithful measurement (7.70 t/ha) and this fitted value is presented as the labelled “fitted route”; on that basis 2026 comes out at 7.75 t/ha, −0.6% against the corrected 2018–25 mean of 7.80 — a year below the best of the record, but not a collapse.
1Eight years of tracking the official figure
Across every settled harvest from 2019 to 2025 the satellite estimate has followed the official series: r = +0.93 with a slope of 0.90, meaning it captures the year-to-year swing almost one-for-one rather than reverting to an average. It called the 2019 and 2022 highs, and the 2020 and 2024 lows, without ever seeing a farm return.
That track record is what gives the 2026 number standing.
The 2026 forecast
The crop measures 7.70 t/ha raw; the fitted Defra-comparable route gives 7.75 t/ha — −0.6% against the corrected eight-year mean of 7.80. That places it above 2024 (7.25), 2025 (7.34) and 2020 (7.08), and below 2019 (8.85), 2022 (8.60) and 2023 (8.18).
We differ from the current AHDB in-season survey, which reports 6.7 t/ha at 85% cut. Our reading is that 2026 is below the strongest years of the record, but that there is no evidence of a significant collapse in yield. The gap between the two estimates is discussed in section 9.
Data — area-weighted yield by year
| Year | Raw | Published (corrected) | Defra final | Difference |
|---|---|---|---|---|
| 2018 | 7.54 | 7.48 | — | — |
| 2019 | 9.19 | 8.85 | 8.9 | −0.05 |
| 2020 | 7.39 | 7.08 | 7.0 | +0.08 |
| 2021 | 8.02 | 7.59 | 7.8 | −0.21 |
| 2022 | 8.55 | 8.60 | 8.6 | 0.00 |
| 2023 | 8.96 | 8.18 | 8.1 | +0.08 |
| 2024 | 7.68 | 7.25 | 7.3 | −0.05 |
| 2025 | 7.59 | 7.34 | 7.2 | +0.14 |
| 2026 | 7.70 | 7.75 | not yet published | — |
2018 has no Defra row: each Agriculture in the United Kingdom edition prints only about three years and a 2018 UK wheat table could not be sourced. It is excluded rather than estimated.
2An independent source that agrees
Across the seven years where Defra finals are verifiable, the satellite product correlates at r = +0.93 with a regression slope of 0.90 — it tracks the year-to-year swing almost one-for-one, rather than reverting to a long-run average.
The +0.35 t/ha mean bias is expected, not error. Defra reports all wheat, UK; this product reports winter wheat, Great Britain, and winter wheat outyields spring wheat.
3Correcting the residual bias
The +0.35 t/ha offset is a basis difference, and it can simply be subtracted. Doing so cuts the error by a third. But a second correction does far better — and explains what the offset is actually made of.
| Correction | Parameters | LOO RMSE (t/ha) | vs raw |
|---|---|---|---|
| none | 0 | 0.435 | — |
| constant −0.35 t/ha | 1 | 0.294 | −32% |
| linear calibration | 2 | 0.350 | −20% |
| July-rainfall aware | 2 | 0.151 | −65% |
Linear calibration is rejected: it fits marginally better in-sample yet performs worse out-of-sample than the single constant — a second parameter buying nothing but overfitting on seven points.
Why rainfall, and what the bias really is
The satellite measures the standing crop — canopy development and senescence timing. Defra records what reached the barn. Between them sits the harvest, and a wet July causes lodging, sprouting and physical losses that no satellite can see.
The residual tracks England’s July rainfall at r = +0.90. The two extremes prove it: 2023 had 122 mm (+84% on normal, the sixth wettest UK July on record) and gave our largest overshoot at +0.86 t/ha; 2022 had 23 mm (−65%) and is the only year we read below Defra.
Correcting for it lands 2023 at 8.19 against Defra’s 8.1, and 2022 at 8.60 against 8.6. The constant correction cannot do this — it pushes 2022 further from truth, because that year had no harvest losses to remove.
Applied to 2026, the constant correction gives a Defra-equivalent of 7.27 t/ha. The rainfall-aware correction gives 7.7 — because a record-dry July implies almost no harvest losses to subtract, so the raw estimate needs little adjustment.
What we publish
The July-rainfall correction is the published model. Every figure in this report is after it. It reduces out-of-sample error by 65% and, unlike a constant offset, it gets the dry years right as well as the wet ones.
One safeguard is applied. July 2026 was 6.5 mm in England — 10% of average, and less than half the previous record of 13.4 mm set in 1911. That is far outside the 23–122 mm range the relationship was fitted on, so the rainfall input is clamped to the fitted domain. Without the clamp the line, which has no physical floor, would keep adding yield as rainfall approaches zero (+0.19 t/ha at 6.5 mm, giving 7.89). With it, 2026 is treated as the driest harvest we have actually observed — 2022 — and the correction is +0.05 t/ha, giving 7.75.
The clamp is not cosmetic. Rain at harvest can only reduce what reaches the barn relative to the standing crop; it cannot raise it. Below the driest harvest on record there are no losses left to correct for, so the correction should flatten rather than keep growing.
4Area — and the 2020 tell
2020 is the exception that proves the method
Our area falls to 73% of Defra's that year — a far larger gap than any other. The flooded autumn of 2019 stopped winter wheat being drilled, and growers switched to spring wheat, which Defra counts and this product, by design, does not.
The product detected a national change in cropping. That is a capability, not a discrepancy — and it is the kind of shift that moves a balance sheet months before it appears in official statistics.
5The variation is between fields, not a uniform drop
A single national average hides what the field-level map shows plainly. Across 198,515 fields:
| Statistic | Value |
|---|---|
| Area-weighted mean | 7.70 t/ha raw (fitted route 7.75) |
| Standard deviation | 1.03 t/ha (CV 13.3%) |
| 5th – 95th percentile | 5.72 – 9.11 t/ha |
| Best fields vs worst | 1.59× |
| Area below 6 t/ha | 7.1% of the crop, 111 kha |
| Area above 9 t/ha | 7.6% of the crop, 118 kha |
So 2026 is not a season in which the crop failed everywhere. 131 kha is genuinely poor, below 6 t/ha — and at the same time 122 kha is yielding above 9. Reports of severe individual losses and our moderate national average are both true, and the field-level map is what reconciles them.
By region — where the damage sits
| Region | Median | p10 – p90 | Area (kha) | % below 6 | % above 9 |
|---|---|---|---|---|---|
| North East | 8.32 | 6.90 – 9.18 | 60 | 2.9% | 17.0% |
| Scotland | 8.32 | 7.14 – 9.18 | 124 | 1.6% | 16.1% |
| Yorkshire & Humber | 8.24 | 6.74 – 9.11 | 211 | 4.0% | 14.5% |
| West Midlands | 8.00 | 6.59 – 8.87 | 138 | 4.6% | 6.3% |
| South West | 8.00 | 6.66 – 8.79 | 140 | 3.8% | 4.9% |
| Wales | 7.92 | 6.59 – 8.87 | 18 | 4.6% | 6.5% |
| North West | 7.77 | 6.03 – 8.71 | 24 | 9.5% | 4.5% |
| East of England | 7.61 | 5.96 – 8.79 | 431 | 10.4% | 5.8% |
| South East | 7.61 | 6.03 – 8.63 | 198 | 9.4% | 3.5% |
| East Midlands | 7.45 | 5.80 – 8.63 | 303 | 13.2% | 3.9% |
The two lowest-yielding regions are also two of the three largest: East Midlands (303 kha, 13.2% of its crop below 6 t/ha) and East of England (431 kha, 10.4%). Together they hold 734 kha — 45% of the national crop — and they are where the losses are concentrated.
What holds the national figure up is Yorkshire & the Humber: a median of 8.24 on 211 kha, with 14.5% of its crop above 9 t/ha and only 4.0% below 6. Scotland and the North East have equally strong distributions but on 124 and 60 kha, so they lift the average far less. Wales, at 18 kha, barely moves it at all.
The gradient matches what growers are reporting
The South East median is 7.61 t/ha against Yorkshire & the Humber at 8.24 — and the difference is in the whole distribution, not just the average: 9.4% of the South East crop is below 6 t/ha against 4.0% in Yorkshire, and 14.5% of Yorkshire is above 9 t/ha against 3.5% in the South East.
That is the pattern coming out of the harvest itself: regional reporting has put East Yorkshire crops at 9.5 t/ha and the Lincolnshire coast at 10–12 t/ha, against 4.8 t/ha in Gloucestershire. Our independent estimate agrees with the growers, region by region.
6Where the losses fell
Yield rises +0.117 t/ha per degree of latitude (r = +0.68) and falls −0.108 t/ha per degree east (r = −0.60). Read as drought exposure rather than latitude, the ordering is coherent: the wetter north and west held up, the dry east did not — and the South West, southernmost but maritime, sits mid-table rather than bottom.
The three lowest-yielding regions hold 932 kha between them — over half the national crop. That is why a moderate yield anomaly still carries a material production effect.
Independent reporting corroborates the pattern: commentary that England carried the worst of the moisture stress while Scotland and Wales fared better, and regional harvest reports of East Yorkshire at 9.5 t/ha and the Lincolnshire coast at 10–12 t/ha against Gloucestershire at 4.8 t/ha.
7Official figures are forecasts too
The satellite forecast updates weekly from May. The first official England provisional appears in October; the UK final in December or January — four to eight months after this product first reports.
Those official figures are themselves estimates, and they move. Defra's own publications show it:
| Revision | From | To | Change |
|---|---|---|---|
| 2022 UK wheat area | 1,809 kha | 1,813 kha | +4 kha |
| 2023 UK production | 13,890 kt | 13,980 kt | +90 kt |
| 2025 UK production | AHDB provisional | Defra final | +122 kt (+1%) |
The first two are revisions between successive Agriculture in the United Kingdom editions — that is, after the harvest year had closed. The 2024 final was published on 7 January 2025, months after the October provisional.
This matters for reading any single in-season number, including ours and including AHDB's.
8Every field, and the variation inside it
200,161 parcels are mapped individually, and the variation within each field is resolved — which is the scale at which agronomic decisions are actually made. The live portal carries this for every week of the season.
9The open disagreement
AHDB's harvest survey at 85% cut reports 6.7 t/ha, roughly 15% below its ten-year average. Our raw forecast is 7.70 (fitted route 7.75). The gap is 1.05 t/ha, and we publish it rather than reconcile it.
The correction widens this gap rather than closing it, and that is the honest consequence of applying it consistently: a record-dry July means there were almost no harvest losses to subtract, so the raw estimate needs almost no adjustment down. We apply the same model to 2026 that we apply to every other year.
Four things bear on it, none settled:
- Our cutoff is 1 August. July 2026 was the driest on record in England (6.5 mm, 10% of average). Damage after the cutoff cannot be in this forecast, and the correction models harvest losses, not in-season drought stress — a failure mode this model has never been trained on at this severity.
- AHDB is a survey in progress. Its figure moved 6.8 → 6.7 as coverage rose from 54% to 85%, and provisional estimates have been revised before — upward, in 2025.
- The moisture conventions differ. Our calibration targets are standardised (≈14.5–15% grain) yields; the survey reports dry delivered tonnage. With this harvest at 12–13% moisture that is worth about 0.15–0.2 t/ha of the gap: on a common basis our 7.70 reads 7.48–7.57 against the survey's 6.7.
- The basis is already accounted for. The corrected figure maps to the all-wheat UK basis via the Defra-fitted correction, so the winter-wheat premium explains none of this gap — it cannot be invoked twice.
Later forecast issues will show whether the model converges toward the survey as end-of-season observations accumulate. That test is public and falsifiable.
10Method, limits and data
Sentinel-2 optical time series over the GB arable belt are classified weekly by a temporal sequence model into winter wheat / not winter wheat, then refined to 5 m field parcels. A second model forecasts yield per 10 m pixel; the published field value is the parcel mean.
Published yields are then corrected for harvest-period losses using England July rainfall, fitted against Defra finals 2019–2025 (section 3). The correction is national and applied uniformly, since the relationship is fitted against a national series; regional figures carry the same shift. Rainfall is clamped to the fitted 23–122 mm domain.
Area is a direct count of classified 5 m pixels on the deduplicated field map — the area of the published picture, so the number and the map cannot disagree (they reconcile to −0.022%). Where Sentinel-2 tiles overlap, each location is attributed to the tile whose valid data it sits deepest inside, so the belt never double-counts seams.
Limits, stated plainly
- Area uncertainty is transferred, not measured in-season. The published area carries a 90% CI of ±1.0% (±16 kha) — converted to the product-wide 90% level from the ±1.2% (95%) reported by an Olofsson accuracy assessment of this pipeline's 2025 hindcast against settled reference labels: 50,009 points; wheat user's accuracy 0.94, producer's 0.91. Every published interval on this page is now a 90% interval. The same assessment measures an omission-led adjustment factor of +6.3% which is disclosed, not applied: the published number stays a direct pixel count of the map, so the number and the map cannot disagree. Both the CI and the unapplied factor assume 2025-measured accuracy carries to 2026.
- Cloud gaps under-report. A week processed during a cloud gap reads low and steps up at the next clear window. Per-week observation support is published with each issue.
- Yield uncertainty is dominated by a systematic term (0.35 t/ha) that does not average down nationally. Field validation against observed combine yields gives r = 0.85 over 270 fields.
- 2018 is absent from the Defra comparison because a 2018 UK wheat table could not be sourced. It is excluded rather than estimated.
Data and provenance
Live portal: wheat.rayx.co.uk — weekly maps at field and within-field level, with per-region statistics.
External benchmarks are recorded append-only with provenance: a revised official figure is added as a new record naming the one it supersedes, never overwritten. That discipline is what made the revision analysis in section 5 possible.