ERA5 Reanalysis, Explained

ERA5 is the fifth-generation global atmospheric reanalysis produced by ECMWF (the European Centre for Medium-Range Weather Forecasts) and distributed publicly through the Copernicus Climate Data Store (CDS), run by the Copernicus Climate Change Service (C3S). It reconstructs the state of the global atmosphere — temperature, wind, humidity, radiation, precipitation, pressure, and dozens of other variables — on an hourly basis, going back to 1940.

A reanalysis is not a forecast and not a raw observational record. It's what you get when you run a modern numerical weather model over historical data and continuously correct it against every observation available for that period: satellites, weather balloons, ships, buoys, aircraft, ground stations. The output is a physically consistent, gap-free grid of the entire atmosphere for every hour since 1940 — something no single observation network could ever provide on its own, because no observation network has ever had uniform global coverage.

That combination — long record, global coverage, hourly resolution, dozens of simultaneous variables — is why ERA5 shows up everywhere: wind and solar resource assessment, crop modeling, hydrology, building energy simulation, insurance risk, and climate research. It is, by a wide margin, the most heavily used reanalysis product in the world.

ERA5 is built from a numerical weather prediction model — the same family of physics used for tomorrow's forecast — run backward over history and pulled toward reality using 4D-Var data assimilation. At every timestep, the model's short-range forecast (the "background") is compared against millions of real observations for that hour, and both are blended, weighted by their respective uncertainty, into a best estimate of the true atmospheric state. That estimate becomes the starting point for the next timestep, and the cycle repeats, hour by hour, back to 1940.

The practical consequence: some ERA5 variables are close to direct measurement, and some are almost entirely model output. Temperature, pressure, and wind are tightly constrained — dense global observation networks pull the model close to truth for these. Surface solar radiation is a different story. ERA5 doesn't observe cloud cover the way a satellite does; it models cloud fraction and cloud optical properties as part of the physics, then computes downward radiation through that modeled cloud field. Get the cloud model slightly wrong — and cloud microphysics is one of the hardest things in atmospheric modeling to get right — and every derived quantity downstream of it inherits the error. Sunshine duration, computed from radiation, is exactly such a derived quantity.

Modeled cloud cover, not observed cloud cover

The core issue in practice: ERA5's sunshine duration is a downstream product of modeled cloud fields, not a direct measurement of how much the sun actually broke through. We compute it from ERA5's surface solar radiation downwards (SSRD) and top-of-atmosphere incident radiation (TISR) via the Ångström–Prescott clearness-index method — standard practice, not a shortcut. The problem isn't the formula. It's what goes into it. When the underlying cloud model is slightly too optimistic about how much cloud actually blocked the sun on a given day — a systematic tendency in reanalysis surface radiation, well documented in the broader literature — every downstream sunshine-hours estimate inherits that optimism.

Which direction the error runs, and how big

In our data, raw ERA5 sunshine consistently overestimates actual observed sunshine — and the size of the overestimate scales with how cloudy the climate is. (The ground stations it's measured against carry their own instrument biases, which is a separate problem worth understanding first.) Against our 56-station reference set:

• Clear, dry climates (deserts, Mediterranean): +5–16% overestimation
• Mid-latitude temperate climates: +20–40% overestimation
• High-latitude, cloudy climates (northwest Europe): +40–73% overestimation

That gradient is the signature of a cloud-modeling bias, not random noise: the more cloud a climate has, and the more that cloud varies hour to hour, the more room there is for the model to get it wrong. Bergen, Glasgow, and Reykjavik look substantially sunnier in raw ERA5 than they are on the ground — which makes any ranking of "cloudy vs. sunny" cities meaningless without correction. We fit a linear correction against WMO/KNMI/ECA&D ground-station observations:

R² = 0.82 · fit on a 56-station subset · residual RMSE ≈ 180 hr/yr · full reference set 182 stations, IDW radius 500 km

The correction pulls the worst overestimates down hardest — which is exactly the behavior you want, since the error itself is worst exactly where cloud cover is highest.

The grid-cell problem, in practice

Every reanalysis cell covers a real chunk of terrain, not a point. A coastal cell averages sea and land into one number, smoothing out the sharp temperature and cloudiness gradient that actually exists at the shoreline. A cell over hilly terrain averages valley floor and ridge line into a value that describes neither well.

Scale decides how badly that bites. ERA5's atmospheric grid is 0.25°, roughly 31 km on a side. ERA5-Land — the product behind our sunshine fallback — is 0.1°, roughly 9 km, which resolves a great deal more coastline but still cannot see a valley. Our published per-city dataset is coarser again at 0.5° (~55 km), and its documented error bands show exactly what that costs: coastal microclimates run ±10–20% error and high-elevation locations ±12–15%, precisely because a 55 km cell cannot resolve a shoreline or a mountain range. Anywhere the real world changes character faster than the cell size, reanalysis hands you an average, not an answer.

Why we demoted ERA5 from primary to fallback

Since July 2026, the sunshine layer on the GeoRank map is built primarily on SARAH-3 (EUMETSAT's Climate Monitoring SAF satellite radiation product), at native 0.05° (~5 km) resolution, covering the Meteosat full disk — roughly ±65° latitude and longitude centered on Europe and Africa. ERA5 now fills in everywhere SARAH-3 doesn't reach: the poles, the Americas, Asia-Pacific. A per-pixel quality band on our merged grid records which source fed each cell, so the switch is never hidden.

The reason for the swap is simple: satellite-observed radiation cells need no cloud-model correction, because they're not modeling cloud in the first place — they're looking at it. That's precisely why SARAH-3 replaced modeled ERA5 as the primary source within the Meteosat disk. ERA5 didn't get worse. Satellite observation was always the better source for this specific variable; it just wasn't available everywhere until we built the merge pipeline.

ERA5 still does real work in our stack outside the sunshine layer — it supplies dewpoint (humidity) and wind speed for the Outdoor Hours Index, where its coarser grid acts as a regional modifier layered on top of a higher-resolution temperature base from CHELSA. Demotion to fallback isn't retirement.

Neither source is strictly better — they're better at different things, and conflating them is where most casual write-ups on this topic go wrong.

Where reanalysis wins

• Coverage: truly global, poles included, no gaps from satellite scan geometry or a fixed viewing angle.
• Record length: a consistent, physically coherent hourly record back to 1940 — most satellite radiation products start decades later.
• Variable breadth: one internally consistent dataset gives you temperature, humidity, wind, pressure, and radiation together, which matters when your downstream model needs several of them to agree with each other physically.
• Well-observed variables: for quantities with dense ground and satellite observation networks feeding assimilation — temperature, pressure, wind — reanalysis is excellent, often indistinguishable from direct measurement at the grid-cell scale.

Where satellite retrieval wins

• Cloud and radiation: a satellite radiometer looks at reflected and emitted radiation directly. It doesn't need to model cloud microphysics to know a cloud is there — it can see it. That's the entire reason SARAH-3 needs no bias correction against ground stations where ERA5 needs a substantial one.
• Resolution: SARAH-3's 0.05° (~5 km) grid resolves coastlines, islands, and terrain that ERA5's 0.25° (~31 km) atmospheric grid — or ERA5-Land's 0.1° (~9 km) — simply averages over.
• Regional detail: within its coverage disk, satellite data captures local cloud patterns — sea breeze cloud, orographic cloud — that a global model at reanalysis resolution can't represent.

The honest summary: use reanalysis when you need every variable, everywhere, all the time, and can tolerate model-driven error in the harder-to-observe fields. Use satellite retrieval when the variable you care about is something a satellite can directly see, and you're inside its coverage footprint.

• Radiation and cloud fields are modeled, not observed. Any variable derived from cloud cover — sunshine duration foremost among them — carries systematic bias that scales with how cloudy and variable the local climate is. Treat raw ERA5 radiation-derived products as a starting point that needs ground-truth correction, not a finished number.
• Grid resolution averages over real terrain features. At ~31 km per cell (~9 km for ERA5-Land), coastlines, islands, narrow valleys, and mountain ridgelines get blended into a single value that can misrepresent every specific point inside that cell.
• Not a substitute for station-level precision. If a project genuinely needs the number for one specific site — a single farm, a single rooftop — a nearby ground station or a downscaled product will always outperform a reanalysis grid cell centered dozens of kilometers away.
• Land-surface and derived variables are weaker than core atmospheric ones. ERA5's temperature, wind, and pressure fields are pulled tight by dense observation networks. Fields further from direct measurement — soil moisture, surface radiation, precipitation type — lean more heavily on the underlying model physics and inherit more of that model's blind spots.
• Update latency. The initial ERA5T release for any given period is preliminary and gets superseded by a finalized version roughly two to three months later, once the full observation set for that period has arrived and been assimilated.

• Hersbach, H. et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049.
• ECMWF — ERA5 dataset overview
• Copernicus Climate Data Store — ERA5 hourly data on single levels
• ECMWF Confluence — ERA5 data documentation

See GeoRank's own use of ERA5 in full: the sunshine calibration methodology covers the correction formula and station set in detail. Related pages: SARAH-3, explained, the live sunshine map, and the climate & rain layer.