The instrument families
Four families of instruments feed this site, and they see different things. Understanding which family a number came from is most of understanding the number.
JMA surface observatories
The Japan Meteorological Agency operates four observatories in our study area: Tokyo (35.6895, 139.6917; instruments at 136.76m elevation on the Chiyoda Ward site), Yokohama (35.4437, 139.6380; 39m, Nishi Ward), Chiba (35.6074, 140.1065; 36m, Chuo Ward), and Saitama (35.9089, 139.6234; 17m, Urawa Ward). Hourly records run continuously from 1949. We use the 1991–2020 climatological normals as the baseline for every long-term figure on the site — the seasonal matrix's monthly values are anchored to this 30-year reference period. The JMA network is the backbone: calibrated, maintained, published, and boring in the way that only properly run observation networks get to be.
AMeDAS
The Automated Meteorological Data Acquisition System runs about 1,300 stations nationwide, eight of them within Tokyo prefecture — Hachioji (35.6559, 139.3234), Fuchu, Oshima, and others — sampling at ten-minute intervals. AMeDAS is how we see micro-scale variation the flagship observatories smooth over: the Hachioji station is our Tama Hills reference precisely because it sits in the foothill environment the JMA Tokyo site can't represent. The trade-off is siting. AMeDAS stations were placed for aviation and agricultural needs, not urban climatology, so they capture the plains, hills, and coasts well and the urban canyons poorly. Nobody's network is perfect; AMeDAS is perfect for exactly the zones the wards aren't.
Tokyo Metropolitan Government monitoring
The metropolitan government's heat island monitoring program — most recently its 2023 report — operates 120 surface temperature survey points across the ward area and publishes satellite-derived thermal analysis. We use its Landsat thermal imagery work for surface skin temperature: the 48–52°C asphalt figures, the rooftop heat mapping in Taito, the Toyosu fill-material anomaly. Satellite thermal sees what air thermometers can't — the actual skin of the city — and the government program is the only entity running it systematically over our study area.
Our own fieldwork
Between 2019 and 2022 we ran a mobile survey program: a vehicle carrying aspirated, shielded thermometers at 1.5 meters, logging at ten-second intervals over fixed transects repeated at fixed hours in all five zones. That's the source of the ward-by-ward bias figures, the sea-breeze penetration measurements, and the Haneda gradient. In Saitama we maintain forty cooperative rain gauges hosted by agricultural cooperatives — the dry-slot mapping is theirs. And the live conditions boxes on every page come from the Open-Meteo forecast API, queried fresh on each page load for our five measurement points: Tokyo, Yokohama, Hachioji, Saitama, and Chiba, all on Asia/Tokyo time. Open-Meteo provides model-interpolated point data, not station observations; we label it as a live feed and we never substitute static numbers when it's down.
How the heat island bias is calculated
The +2.8°C figure has a recipe, and here it is. We take simultaneous readings at the JMA Tokyo station — which sits adjacent to green space and carries low local bias — and at representative urban stations across the ward cores, average the differences over a five-year period to smooth anomaly years, then cross-check the result against the mobile survey vehicle's transect means from 2019–2022. The stationary comparison and the moving comparison agree within 0.2°C, which is why we publish the figure with confidence and also why we publish it to only one decimal place.
Rainfall verification runs through JMA's radar composite, XRAIN, at 250-meter resolution. XRAIN is how we verify baiu front position against the gauge record and how we map typhoon rain bands across the five zones — the Hagibis and Faxai zone-by-zone totals are radar-gauge blends, not single-station readings.
The five measurement points
| Station | Zone | Latitude | Longitude | Network |
|---|---|---|---|---|
| Tokyo | Center Core | 35.6895 | 139.6917 | JMA observatory |
| Yokohama | Bay Coast | 35.4437 | 139.6380 | JMA observatory |
| Hachioji | Tama Hills | 35.6559 | 139.3234 | AMeDAS |
| Saitama | Saitama Plains | 35.9089 | 139.6234 | JMA observatory |
| Chiba | Boso Peninsula | 35.6074 | 140.1065 | JMA observatory |
Limitations, stated
Every source above has edges, and we'd rather describe them than have you discover them. First: AMeDAS siting. The stations serve aviation and agriculture, so the dense urban canyon environments — the hottest places in the metropolis — are systematically underrepresented in the automated record. Our mobile transects exist largely to fill that hole, and they fill it at transect scale, not block scale.
Second: the JMA Tokyo station moved three times between 1949 and 2020. The record is homogenized, and homogenization is done carefully, but each relocation introduces uncertainty into the long trend that no amount of care fully erases. When we cite decadal warming in ward minimums, that caveat rides along.
Third: point versus field. Open-Meteo's live data interpolates a model grid to a point; the real surface within a two-kilometer radius of any of our five points can differ by a degree or more depending on shade, pavement, and park proximity. The live boxes are honest indicators, not street-corner promises.
Fourth: anomaly years. A single typhoon-heavy season distorts one-year rainfall figures badly. That's why every annual total on this site is a 1991–2020 normal or a multi-year mean, and why we flag single-year events — Hagibis, Faxai, the 2018 yamase — as events, not climate.
Why our number differs from the official one
The Tokyo Metropolitan Government's Heat Island Report 2023 puts the official annual mean bias at +2.6°C. We publish +2.8°C. The 0.2-degree gap is not a disagreement about Tokyo; it's a disagreement about baselines. The official figure references a different comparison station set than our Tama Hills baseline, and both choices are defensible — theirs tracks ward-against-prefecture structure, ours tracks ward-against-foothill exposure. We state ours prominently and theirs here, on purpose: the reader should see both numbers and the reason they coexist. Precision without provenance is just confidence.
That's the whole method: four instrument families, five measurement points, a five-year averaging window, a radar cross-check, and a standing list of everything we can't see. The method notes carry the operational detail, and the corrections page carries the record of every time the method caught itself being wrong.
Instruments we deliberately don't use
Completeness requires naming what we exclude. Consumer-grade weather stations — the backyard networks — are excluded from zone figures: their siting is uncontrolled, their radiation shields vary, and their drift is undocumented. They occasionally inform a hunch, never a published number. Reanalysis datasets are excluded from the long-term figures for the opposite reason: they're models constrained by observations, and our figures need to be observations constrained by method. And single-station civic displays — the thermometer boards on buildings — are ignored entirely; they're unshielded, uncalibrated, and frequently mounted in exhaust flow. The discipline of exclusion is half of what makes the included record trustworthy.
The homogenization question
The JMA Tokyo station's three relocations deserve a fuller accounting than the summary above. Each move shifted the instruments' immediate environment — different exposure, different surroundings, different distance to the park edge. JMA's homogenization adjusts the series statistically, and we accept the adjusted series as the best available, but we handle it under two standing rules. First: long-term trend claims are stated to one decimal place and always carry the relocation caveat. Second: no conclusion on this site rests on the Tokyo series alone — every trend we publish is corroborated against at least one other network with a cleaner siting history. The rule costs us some dramatic numbers. It's supposed to.
Sampling windows and their arithmetic
The five-year averaging window for bias figures is a chosen compromise, and the choice is documented here. Shorter windows — two or three years — are hostage to anomaly seasons: one strong typhoon year or one yamase-heavy August shifts a two-year mean by tenths of a degree. Longer windows smooth better but lag real change; a decade-long window would still be showing the pre-2020 city. Five years damps the anomaly years to under 0.1°C of window noise while remaining short enough to track genuine drift. When the mobile survey and the station comparison disagreed during 2019–2022, the disagreement almost always traced to an anomaly the five-year window correctly refused to chase.
From instrument to page
The publication pipeline is deliberately short. Field logs and network data go into a plain archive; analysis scripts read the archive directly; the resulting figures are entered into the essays by hand, by the analyst who produced them; and a second team member re-derives every figure from the archive before publication. There is no dashboard, no automated pipeline pushing numbers to the page, and no figure on this site that a human didn't type while looking at the derivation. It's slower than automation. It's also why, when a reader challenges a number, we can answer from the record in hours rather than from a system in weeks. The site's ethic — measured or nothing — ends at the keyboard, and the keyboard is staffed.
Requests for data
Researchers and journalists can request underlying figures via the contact page. We share aggregated transect means and gauge-network summaries freely for non-commercial work with attribution. Raw logs are shared case by case, because some contain location details of cooperating households. We don't charge for academic use, and we don't sign exclusivity. The archive exists to be checked; a measurement nobody can examine is just an assertion with a serial number.
Cross-checking the live feed
The Open-Meteo boxes on this site are the only figures we publish that we don't generate or archive ourselves, so they get a verification regime of their own. Weekly, the feed's current conditions for our five points are compared against the corresponding JMA and AMeDAS observations; persistent divergence beyond a degree would trigger a note on the affected pages. So far the model-interpolated feed tracks the stations within its expected tolerance, with the predictable exception of the sharpest urban-canyon hours, where any grid-based product smooths what the street actually does. That exception is why every live box sits next to a link to this page: the feed is useful context, clearly labeled, and explicitly not the instrument record.
The 2019-2022 survey archive
The mobile survey program deserves its archive noted: three years of transects, fixed routes in all five zones, pre-dawn, midday, and evening windows, each route repeated at least eight times per season per year. The archive underlies the ward bias figures, the sea-breeze penetration record, the Haneda gradient, and the river-corridor cooling claims. It is also closed — the survey concluded in 2022 — and the figures derived from it are now being refreshed against the stationary networks, which continue. Where the archived transect record and the current station record ever diverge materially, the page in question will say so, and the corrections log will carry the reconciliation.