How Sunshine Hours Are Measured & Why They Differ

"Sunshine hours" sounds like a simple count of daylight the sun actually showed up for. In practice it's a proxy — a stand-in for direct solar irradiance strong enough to count as "bright." Four instrument families have produced the sunshine-hour figures now circulating for any given city, and they don't measure the same thing:

The instrument that still anchors a lot of "average sunshine" folklore was invented in 1858 and redesigned into the form still recognisable today by the physicist George Gabriel Stokes in 1879, as the UK Met Office documents in its own guide to sunshine measurement. It has no electronics. A solid glass sphere, roughly the size of a grapefruit, sits mounted at the centre of a metal bowl. It acts as a burning lens: whenever the sun is out, the sphere focuses direct sunlight into a small hot spot that scorches a trace onto a specially graduated card held in the bowl behind it. As the sun moves across the sky, the hot spot moves with it, burning a continuous line. At the end of the day, an observer measures the length of the burn and converts it to hours of sunshine.

It's an elegant piece of Victorian engineering, and it is still an official WMO-recognised instrument at many non-automated climate stations. It is also a poor proxy for "was it actually sunny." The sphere doesn't measure irradiance in watts per square metre — it measures whether the light was concentrated enough to char paper.

The WMO's Guide to Instruments and Methods of Observation (WMO-No. 8) names two principal error sources for the method: the card's sensitivity to temperature and humidity, and overburning. The second one dominates. A scorch mark doesn't stop the instant a cloud crosses the sun — the burn keeps extending for seconds to minutes afterwards, so every gap in broken cloud gets scored as sunnier than it was. Germany's DWD, comparing its reference stations directly, found the effect concentrated almost entirely on days with frequent sun-cloud alternation: up to four hours a day in summer, and one station over-reading by 189 hours across a single year.

Card handling adds its own noise. WMO's specification for reference-grade cards demands pasteboard "not affected appreciably by moisture," with a permitted moisture effect within 2 per cent — a numeric tolerance that exists precisely because the material's response is known to shift. And converting the burn into an hours figure stays a human judgement call: the WMO evaluation rules ask for two or more trained observers to score the same card independently, because a narrow, faint, or interrupted trace is genuinely ambiguous.

None of that is a criticism of the instrument on its own terms — Campbell-Stokes readings are internally consistent within a station's own long record, which is exactly why met services haven't torn all of them out. It's a problem only when a Campbell-Stokes normal gets compared, unadjusted, against a threshold-sensor or satellite figure for somewhere else.

Modern electronic sunshine recorders don't burn anything. A rotating photodiode array or a pyranometer measures direct solar irradiance continuously, and the World Meteorological Organization's Guide to Instruments and Methods of Observation (WMO-No. 8) defines "bright sunshine" formally: a period counts as sunshine only while direct normal irradiance exceeds 120 W/m². Below that line, no matter how bright it looks to the eye, the minute doesn't count.

That single, hard number is what makes electronic sunshine data comparable across stations and countries in a way Campbell-Stokes cards never quite were — the threshold doesn't care what the card is made of or who's reading it.

It's worth knowing where 120 came from, because it isn't a constant of physics. WMO arrived at it by measuring what irradiance actually made Campbell-Stokes cards start to burn, and those measurements ranged from 70 to 280 W/m² depending on the instrument. 120 is the mean of that spread, adopted as a convention, and WMO accepts a further 20 per cent tolerance on top. One published intercomparison measured its two Campbell-Stokes recorders scorching at 55 and 110 W/m² — both under the nominal threshold.

So a card can genuinely record haze and cloud-edge brightness that a correctly calibrated sensor scores as sub-threshold. That's real, and it contributes. But it is the smaller effect: the bulk of the gap between an old Campbell-Stokes climatology and a modern figure for the same city comes from overburning on broken-cloud days, not from the threshold offset.

Neither a burning glass nor a ground sensor exists everywhere — most of the planet has no sunshine recorder within hundreds of kilometres. Satellite retrieval fills that gap by working the problem backwards: geostationary weather satellites don't watch the ground, they watch cloud tops. The Heliosat method infers how much sunlight reached the surface from how reflective the cloud cover above it was at each moment, then converts that inferred irradiance into a sunshine/no-sunshine call using the same kind of threshold logic as a ground sensor.

GeoRank's own sunshine layer runs on SARAH-3 (Surface Solar Radiation Data Set — Heliosat), produced by EUMETSAT's Climate Monitoring Satellite Application Facility (CM SAF) from the Meteosat MVIRI and SEVIRI instruments, accumulated into a 1991–2020 monthly climatology at native 0.05° (~5 km) resolution. It's a genuine measurement of what happened at the cloud layer, not a model guess — which is why it needs no cloud-physics correction the way a reanalysis product does. But it's still an inference, one satellite pixel away from the ground, and it inherits its own edge cases: broken cloud at sub-pixel scale, low sun angles at high latitude, and coastal cells that blend sea and land all reduce fidelity. It also only covers the Meteosat full disk — roughly ±65° latitude and longitude centred on Europe and Africa. Outside that footprint, GeoRank's map falls back to ERA5 reanalysis at coarser resolution.

Stack the three methods up and the disagreement stops looking mysterious. A Campbell-Stokes climatology, a WMO-threshold electronic normal, and a satellite-derived figure for the same city are three different physical quantities wearing the same label. A tourist board quoting a decades-old Campbell-Stokes normal, a national met service publishing current electronic-sensor data, and a satellite dataset like SARAH-3 or a reanalysis product like ERA5 can — and routinely do — produce three different numbers for the same place, and none of them is lying.

ERA5 deserves its own callout because it's the base layer under a lot of gridded climate data, GeoRank's included, and it isn't a sunshine sensor at all — it's a physics-based atmospheric model reconstructing radiation from assimilated observations, and it systematically overestimates sunshine. The size of the overestimate depends on climate regime: our own comparison against ground-station observations (see the full methodology) finds roughly +5–16% in clear, dry climates like deserts and the Mediterranean, +20–40% in mid-latitude temperate climates, and +40–73% in cloudy, high-latitude climates like northwest Europe. Uncorrected, that makes Bergen, Glasgow, and Reykjavik look substantially sunnier than they are — the cloudier the real climate, the bigger the fiction.

There's no single "correct" sunshine number — only numbers fit for different purposes. For today's forecast or this week's holiday, trust your national met service's live electronic-sensor reading; it's threshold-based and current. For comparing long-run climate across cities and countries — the question anyone relocating is actually asking — a raw brochure figure or an uncorrected satellite/reanalysis pull is the wrong tool, because the systematic biases above don't cancel out between cities; they compound in exactly the direction that makes cloudy places look better than they are.

A worked example: GeoRank's own calibration

GeoRank corrects for this rather than publishing raw model output. We compare the gridded sunshine data against 182 reference stations — sourced from WMO climate normals, KNMI, ECA&D, and national met-service archives — and fit a linear correction on a 56-station subset:

R² = 0.82 · n = 56 stations · Residual RMSE ≈ 180 hr/yr — the remaining 126 stations serve as cross-validation.

That correction is then spread across the full grid with inverse-distance weighting, so a station's correction fades out with distance rather than applying uniformly. Bergen — one of the 56 fitting stations — has an observed long-run mean of 1,413 hours a year. Before correction, the raw model grid cell sits well above that, comfortably inside the +40–73% overestimation band documented for high-latitude, cloudy climates above; the correction pulls it back down to what the ground station actually recorded. Where SARAH-3 satellite coverage is available, we use it as the base layer instead of modelled ERA5 precisely because it needs a smaller correction — it's an observation, not a reconstruction. The full station list and formula derivation are on the methodology page; the underlying dataset ships openly on npm and PyPI.

Calibration narrows the gap between methods; it doesn't erase it.

• Station density is uneven. 182 reference stations is enough to fit a global correction, not enough to catch every local microclimate — coastal cells, valleys, and sparsely-instrumented regions carry more residual error than the R²=0.82 headline figure implies for any single point.
• SARAH-3 only covers the Meteosat full disk. Outside roughly ±65° latitude/longitude — the Americas, Asia-Pacific, the poles — GeoRank's map falls back to coarser ERA5, and accuracy drops accordingly.
• Campbell-Stokes normals aren't wrong, just different. A station's own multi-decade Campbell-Stokes record is internally consistent and useful for tracking trend at that station. It's the cross-station, cross-method comparison — Campbell-Stokes city A against threshold-sensor city B — that produces misleading rankings.
• None of this is a live feed. Every figure discussed here, GeoRank's included, is a climatological normal — a multi-decade average, not today's forecast or this year's total. A single unusually cloudy or clear year can miss the normal by a wide margin.

• World Meteorological Organization. Guide to Instruments and Methods of Observation (WMO-No. 8), Part I, Chapter 8 — "Measurement of Sunshine Duration." The 120 W/m² threshold, its 70–280 W/m² empirical origin, card specifications, and the two named error sources.
• Sanchez-Romero, A. et al. (2015). The historical sunshine duration record of Barcelona (1885–2013). Atmospheric Measurement Techniques 8, 183–194. Measured Campbell-Stokes burn thresholds of 55 and 110 W/m², and quantifies overburning under intermittent cloud.
• Kaspar, F. et al. (2019). Statistical analysis of sunshine duration measurements at German reference stations. Advances in Science and Research 16, 175. The DWD parallel-instrument comparison, including the 189 hr/yr single-station over-read.
• UK Met Office. How we measure sunshine. Instrument history and current UK practice.
• EUMETSAT CM SAF — the SARAH-3 satellite record. ECA&D — European station archive used in our calibration set.

Each of the sources named above has its own quirks, and each gets the same treatment as this page:

• SARAH-3 — EUMETSAT's satellite sunshine record, and why observing cloud beats modelling it.
• ERA5 — how reanalysis reconstructs the atmosphere, and the bias that leaves in raw sunshine figures.
• CHELSA v2.1 — the ~1 km temperature and rainfall climatology, and what downscaling can and cannot recover.