Validation

How the datasets of the first runs with real data compare with independent references, what the numbers mean, and why ERA5-Land precipitation lies above the gauge totals in Poland. (Updated 6 October 2026)

Summary. Every workflow was run on real data for Poland and checked in two ways: internally (is the processing exact?) and against independent references (do the values agree with other products?). The processing reproduced independent computations to rounding precision in every workflow. Against external references, temperature agrees with ERA-NUTS to a mean absolute difference of 0.25–0.29 K, heating degree days with Eurostat to 2 % per region on average, and NDVI with MODIS with a correlation of 0.93–0.96. ERA5-Land precipitation is 20–39 % above the area totals of IMGW-PIB; this is a property of the reanalysis and of the gauges, not of the processing.

Two kinds of check

Internal checks test the processing. Parents are compared with their children, values with an independent computation outside the pipeline, outputs of two routes to the same data with each other. Their tolerances are tight, and the parent-child check is part of the quality record of every dataset.

External comparisons test the data against other products. Differences are expected, because the references use other models, stations, grids and boundaries; the question is whether they are of the size and pattern the literature and the design predict. These tolerances are not built into the service, because a job does not have the reference data.

ERA5-Land, monthly, NUTS and TERYT regions

Ten years (2015–2024) for the 98 NUTS regions and 2,875 TERYT units of Poland.

Internal. Parents agree with the area-weighted mean of their children to 6e-8 (relative); every region and month has a value; computed areas agree with the register areas (median difference 0.002 %).
Temperature against ERA-NUTS (ERA5 aggregated to NUTS 2016 by the JRC; monthly means 2015–2021): mean absolute difference 0.25 K at NUTS 0, 0.28 K at NUTS 1 and 0.29 K at NUTS 2; correlation of the monthly anomalies 0.9955–0.9965. The largest regional biases are in the mountainous south, where the finer orography of ERA5-Land was expected to differ.
Temperature against IMGW-PIB (station-based area means of Poland): annual values within −0.10 to +0.22 K; monthly values 2022–2024 with an RMSE of 0.37 K. ERA5-Land is about 0.3 K colder in winter and 0.3–0.5 K warmer in summer and autumn.
Against ERA5 through the same pipeline: annual precipitation within 4.1 % at NUTS 0–2; temperature with a mean absolute difference of 0.25–0.29 K.

The precipitation excess

ERA5-Land gives clearly more precipitation over Poland than the gauge-based figures: 20–39 % per year above the area totals of IMGW-PIB (mean of 2015–2024: 756 mm per year), more in winter and spring (winter and spring ratios about 1.45 in 2022–2024, summer 1.15).

The difference is not produced by Qumasc. Monthly totals computed independently from the hourly ERA5-Land data equal those of the workflow, and ERA5 aggregated by the same pipeline gives the same totals to within 0.5 % per year.

The comparison with the gridded gauge analyses of GPCC separates the causes. The mean ratio of 1.28 between ERA5-Land and IMGW-PIB in 2020–2024 splits into three factors:

the GPCC gauge analysis is 5–6 % above the IMGW-PIB area total; both use uncorrected gauges but differ in stations, interpolation and area;
GPCC estimates the systematic error of the gauges over Poland at 16 % of the annual total (30 % in winter, 9 % in summer): gauges catch less than falls, most for snow and in wind;
ERA5-Land is 3–5 % above the gauge analysis corrected for that error, most in spring.

Against the uncorrected GPCC analysis the excess is 14 %. Undercatch explains between a third and all of it; the rest, up to about 9 % of the annual total and largest in spring, is the wet bias of ERA5 documented in the literature (too many days with light precipitation). The shares cannot be fixed more precisely, because the size of the gauge correction is itself uncertain by a factor of about two.

Every dataset with ERA5-Land precipitation therefore carries a note: totals to be compared with station statistics need a bias correction; anomalies and the ranking of regions and months are not affected to the same degree.

Daily values and degree days

Day boundary. The monthly sums of the daily totals equal the monthly-means product to 0.03 %; the monthly means of the daily temperatures agree to 0.03 K at worst.
Against the CDS daily statistics: daily mean, minimum and maximum agree to 0.008–0.016 K, the precision of the two copies.
Heating degree days against Eurostat (2015–2024): +0.5 % for Poland, 2.0 % mean absolute difference per NUTS 3 region and year, 98.6 % of NUTS 3 region-years within 10 %, rank correlation of the regions 0.93. The largest differences are in mountain regions.
Cooling degree days are 9 % above Eurostat for Poland and differ by 10–60 % for single regions and years: they are sums over a handful of hot days, so small temperature differences near 24 °C give large relative differences.

NDVI 300 m

The regional values reproduce an independent computation from the same files to 0.0003; parents agree with their children to 6e-10. Against MODIS MOD13A1 (2019–2024) the monthly values have a bias of about −0.04, an RMSE of 0.047–0.055 and a correlation of 0.93–0.96; anomalies of well covered months correlate at 0.69–0.81. The bias is systematic and expected: MODIS is a maximum-value composite, which by construction lies above a mean state, and the monthly maximum of the dataset is indeed much closer to MODIS (−0.01). No regional NDVI statistics are published for Polish regions, so no spot check against official figures was possible.

GBIF occurrences with climate

48,887 records of three taxa were retrieved, exactly the number GBIF reports; cleaning kept 4,890. Every sampled value equals an independent extraction from the same cube; 47 coastal records took the nearest land cell and 6 have none within 15 km. The search route and both download formats gave the same records and values.

Regional statistics, OpenStreetMap, vector services, combined tables

Statistics: sums of powiaty equal the national values of the Local Data Bank exactly for population, area and registered unemployment; Eurostat and GUS differ by definition, not by processing.
OpenStreetMap: two independent readers agree; the Overpass and extract routes give the same values for Poznań; parents equal the sum of their children to 0.0008 %. Against official statistics the indicators are strongly correlated (r = 0.96–0.998) and systematically different, for reasons that are properties of OpenStreetMap.
Vector services: the number of features fetched equals the count of the service in every case with a count.
Combined tables: every one of 55,334 (NUTS) and 515,388 (TERYT) input rows is in the combined table with the same value, valid fraction and flags.