Changelog

What changed on the atlas and in the data, newest first — including the mistakes. Data corrections also flow to the open dataset and its mirrors.

30 July 2026

• US sunshine hours run about 213 hours low, and the methodology page now says so. Checking the model against the 52 US cities that publish a recorder-measured figure gives a bias of −213 hours and a mean absolute error of 230 hours, with 47 of the 52 reading low. The published +24-hour bias is a global average that hid it: outside the US the same model runs +69 hours. The miss scales with moisture (correlation −0.56 with annual rainfall), so the arid West is close (Boise −4, El Paso +13) while the humid East is not (Milwaukee −534, Miami −501, Boston −477, Cleveland −460). The tempting explanation — a recorder counting sunshine between summer cumulus that a coarse satellite cell averages away — does not survive the check that matters: the same climate bands outside the US barely show the effect, with humid-subtropical cities at −249 hours in the US against −18 hours elsewhere. Cloud physics does not stop at a border, so the asymmetry points at the reference data, most likely because the published US figures descend largely from NOAA's historical sunshine network, which was down to roughly 160 stations by the early 1980s and is no longer a current product. China (−175) and Japan (−133) are the next-largest one-sided gaps, which fits. Unresolved. Until it is, treat US sunshine hours as carrying roughly a 200-hour uncertainty and prefer the station figure where one exists. Rankings are unaffected, because the atlas uses one consistent method throughout, and the Europe-versus-US comparison survives being recomputed entirely from published station numbers. No values were changed.

29 July 2026

• The far north is wrong, and the methodology page now says so. Above the Arctic Circle the sunshine model returns a median of about 2,130 hours a year, more than it gives the band beneath it and more than much of the mid-latitudes, which cannot be right. Maritime places come out sensibly (Iceland ~1,320, coastal Norway ~1,425); the dry continental Arctic inflates badly (Canada ~2,280, Alaska ~2,060, Russia ~1,950), most likely because snow and sea-ice albedo defeat the satellite's clearness retrieval. The accuracy check barely reaches this group: one of the 73 Arctic destinations has a published figure to compare against, and the two high-latitude cities that do have one are both about 450 hours over (Tromsø, Iqaluit). The published 212-hour error is therefore a mid-latitude number and does not describe the far north. The methodology page now states this outright and tells you to treat inland sunshine hours above roughly 60°N as unreliable. Temperature, rainfall and the warmth term are unaffected. No figures were changed; a correction is a separate piece of work and this is the honest interim.
• Dataset edition 2026.11 adds one column, duplicate_of_iata (duplicateOfIata in the JSON). Nine metros are served by two airports and so hold two rows — Heathrow and Southend both filed under London, Glasgow and Prestwick, Seattle and Paine Field, and six more. That was documented in the dataset README but invisible in the files themselves, so anyone reading the CSV alone met two Londons with different sunshine totals and no way to tell why. The column now names the primary on each second row, so filtering it empty gives exactly one row per place: 3,824 of the 3,833. Both rows stay, because both are real — each carries the climate of its own coordinates, which is why the pair rarely agrees exactly. No existing value, column order or page URL changed.
• The published accuracy figures were out of date and are now corrected. The methodology page and every dataset mirror said the modelled sunshine hours had been checked against 404 met-service cities with a bias of +22 hours. Re-running the validation against the current data gives 410 cities and a bias of +24 hours — the correlation (0.92) and mean absolute error (212 hours a year) are unchanged. The old numbers were correct when written and drifted as destination names were corrected in 2026.09 and 2026.10, which let more cities match the reference list.

27 July 2026

• Dataset edition 2026.10 published to every mirror (GitHub release, a new Zenodo DOI version, Hugging Face, Kaggle, figshare). It carries the midday UV index into the downloads as midday_uv_index and peak_midday_uv_index, and disambiguates 22 destinations that shared a name with somewhere else — Portland (Maine) from Portland, Oregon; Victoria (Seychelles) from Victoria, British Columbia — plus two renamed to the place their airport actually serves (RUN to Saint-Denis, SMS to Île Sainte-Marie). No IATA code changed, so page URLs and dataset joins are unaffected, and the climate values are identical to 2026.09.

23 July 2026

• Head-to-head comparisons — two destinations side by side, month by month, with a verdict for each season and the months where it is genuinely too close to call.

22 July 2026

• A midday UV index for every destination, every month, on the WHO scale — in the facts grid and the monthly table on each destination page, and on the comparisons. It comes from the same NASA POWER climatology as the rest of the atlas, rescaled from POWER's 24-hour mean to the midday value forecasts report; the methodology page gives the rescaling and how it was checked.
• A Snow layer on the globe — paint the world by monthly or year-round snowfall, alongside Sunny, Warm, Cold and Rain. Same guarded NASA POWER snowfall model as the ❄ figure on the destination cards, so the two always agree.
• Two new ranking families: where is it rainy? — wettest and driest, every month — and where is it snowy?, running on the Snow layer's data. Both come with year-round editions: the wettest places in the world and the snowiest. In July the snow list belongs to the Andes and Patagonia; in January, to the Alaska panhandle.
• The browse pages — index, continents, countries — now show photo cards with each destination's picture and sunshine verdict, instead of bare text lists.
• Every ranking page now navigates like the globe: a chip bar at the top switches between Sunny, Hot, Rainy and Snowy and between months in one tap, ranked photo cards lead each list, and destination pages gained jump links to their sections. A de-clutter pass to match: the footer's link walls folded into two tidy drawers, intro paragraphs slimmed down, the six per-continent tables became one-line accordions. Year-round editions now exist for all four rankings — reliably sunny, hottest, wettest, snowiest — and the homepage's new Browse page puts every ranking one tap away.

19 July 2026

• Dataset edition 2026.09: 118 destination names corrected where the record carried the airport's host municipality instead of the place it serves — Gaziemir became Izmir, Årø became Molde, Ciudad de la Costa became Montevideo — plus typo fixes and island conventions such as La Gomera and Elba. 146 populations corrected to describe the named place itself rather than an inherited catchment figure. Page URLs and climate values unchanged.

18 July 2026

• A lifetime sunshine calculator — add the places you've lived and the years, get your total lived sunshine hours and a shareable card.
• Every page without a photo of its own now gets a generated share card, so links pasted into chats and feeds preview properly. And an RSS feed, for following new studies and dataset editions.
• Dataset edition 2026.08 published to every mirror: GitHub release, a new Zenodo DOI version, Hugging Face and Kaggle.
• A pass over the site copy for tone, and crisper favicons (96/192 px).
• This page, and an about page — overdue.

15 July 2026

• The climate layer is now NASA POWER end to end (2001–2020). Earlier builds anchored temperatures to CRU CL v2.0, a 1961–1990 normal — mixing two reference periods was wrong, so CRU is gone entirely.
• Fixed: the continent column. It used to come from per-airport metadata, was missing for a quarter of destinations, and silently defaulted to Asia — Fiji, Greenland and the Bahamas were all filed under "Asia". It's now derived from geography, with explicit rules for the six transcontinental countries.
• Replaced rainy-days-per-year with annual rainfall in mm: POWER measures precipitation depth, not wet-day counts, and the old column wasn't worth guessing.

14 July 2026

• Best-time-to-visit guides for 64 countries and 5 island regions.
• Every destination page gained an embeddable sunshine badge.

13 July 2026

• Destination pages gained computed "How sunny is X?" verdicts — every claim traceable to the data.
• Globe dots moved from the airport onto the city they serve.
• Homepage payload halved (1.43 MB → 0.71 MB on the wire).

12 July 2026

• The MCP server went live — any AI assistant can query the atlas, no key needed.
• The dataset got a DOI and public mirrors.
• Published launch-week crawler notes.

9 July 2026

• Sea-surface temperatures for coastal destinations.
• Launched on Hacker News.

8 July 2026

• Sunshine Map goes live: 3,833 destinations, ~3,850 pages, one globe.

Back to the sunshine globe →