Upperblue Fast ·
Temperature methods and research
Upperblue’s experimental point correction adapts raw 2 m model temperature to the point’s ground height using the elevation model issued by the National Land Survey of Iceland (now part of the Icelandic Institute of Natural History). It works without a weather station. The forecast source details identify whether it was applied.
Adjustment = 0.0065 × (model ground height − point ground height) °C, limited to −5 to +5 °C. Missing terrain, missing DEM coverage and predominantly marine cells retain raw temperature. Missing values remain gaps.
IG
Upperblue starts from raw IG and terrain matched to its run. The existing IMO corrected field is a separate source product.
ECMWF
Upperblue uses native terrain from the exact sampled 0.25° cell. It uses the same point formula as IG.
DINI
Uses DINI’s own 2 km terrain and projection. Independent correction skill has not yet been established.
GFS
Uses NOAA’s native terrain from the sampled 0.25° cell. Independent correction skill has not yet been established.
CARRA1
Historical reanalysis uses its 2.5 km terrain. The provisional point option differs from the separate station-trained research method.
CARRA2
Historical reanalysis uses its own polar grid and 2.5 km terrain. Missing terrain in masked cells leaves temperature unchanged.
WAM Arctic
Wave fields only; there is no 2 m air temperature to correct.
What the tests show
Exploratory IG tests on 386 autumn cases gave average absolute error 1.08 °C for Upperblue and 1.25 °C for IMO. Upperblue missed 26 of 89 observed freezes versus IMO’s 54, with 13 false freezing predictions versus zero. The separate ECMWF sample had 194 cases: 0.99 °C after correction versus 1.37 °C raw. These short samples use unverified station observations and official station heights; they do not validate winter or points without observations.
A simple calm/clear switch did not reliably improve the tested method. January confirms weak calm/clear performance for fixed terrain. Wind, cloud and signed temperature profiles remain research.
Fast selects IG followed by ECMWF. DINI, GFS and CARRA methods are available through Maps/API. Country-map symbols and colour fields retain their source temperatures; the point correction applies to selected-place forecasts and meteograms when identified in the product.
January 2026 · IG winter comparison
All 248 January cycles were sampled at predeclared leads 1, 6, 12, 24, 48 and 72 hours. These are 70,633 accepted validation-site forecast cases at 50 station IDs over 31 days: 38,391 observed freezes and 32,242 above-freezing cases. Repeated forecasts share 24,323 distinct accepted station-hours.
| Method | MAE (°C) | Missed freezes | False freezes |
|---|---|---|---|
| Raw IG | 1.749 | 3141 | 6797 |
| Upperblue · DEM height · current | 1.653 | 3128 | 6067 |
| 2–100 m profile · DEM · research | 1.588 | 3265 | 5757 |
| Profile with freezing-class guard · research | 1.590 | 3128 | 6067 |
The current DEM-height method reduces average absolute error by 5.5%. Missed-freeze rate is nearly unchanged: 8.18% → 8.15%. False-freeze rate falls from 21.08% to 18.82%. Every tested lead improves: MAE is 1.375 °C at 1 hour, 1.631 °C at 24 hours and 2.118 °C at 72 hours.
Forecast windy/overcast cases improve from 1.341 to 1.137 °C MAE with fixed DEM terrain. Calm/clear cases worsen from 3.069 to 3.133 °C, and calm/overcast also worsens. The fixed calm/clear guard adds almost no overall gain.
The profile challenger preserves the signed gradient between verified 2 m and 100 m air temperatures and interpolates only inside that physical column. It supports 28.1% of cases; the others retain fixed DEM terrain. MAE reaches 1.588 °C but it misses 137 more freezes than the current method, while removing 310 false freezing predictions. It is not a served method; lower MAE alone does not establish freezing safety.
A follow-up research guard uses the profile only when it preserves the current terrain forecast’s freezing class; otherwise it retains terrain. January MAE reaches 1.590 °C, 3.8% below current terrain, with exactly the same 3128 missed freezes and 6067 false freezes. Calm/clear MAE improves from 3.133 to 2.813 °C. Nine validation stations improve, five worsen slightly and 36 are unchanged; below-column valleys retain terrain. January had already been viewed before proposing this guard, so these are exploratory results. Fresh November checks are defined but have not passed. This method is not served.
Evidence snapshot: 9 October 2026. The new IMO corrected field is absent from all 1,488 sampled January files, so it has no January score. Truth is current unverified AWS data under existing QC; 777 suspect validation cases are excluded from headline scores. As-issued preservation is not independently established. 2026 remains development; final sites stay excluded. One month does not certify all-season or unobserved-point accuracy.