Earth
Loading globe…
Year and variables
200020052010201520202025
2025
Prefecture dashboard
Prevailing wind direction by year
Temperature, precipitation, humidity, wind speed, snowfall: Japan Meteorological Agency annual station records (JMA past weather data) for each prefecture's main observatory, named under the prefecture title. Snowfall is the cold-season total (August of the previous year to July). Years JMA flags as incomplete are left blank.
Solar radiation, cloud cover, wind direction: NASA POWER (CERES and MERRA-2) at the station location, 0.5° to 1° grid. Wind direction is the speed-weighted vector mean of monthly means.
Vegetation: MODIS Terra MOD13A1 v6.1 16-day NDVI at 500 m from the Microsoft Planetary Computer, averaged over every good or marginal quality land pixel inside the prefecture, the 20 composites from day 49 to day 353 of each year. Four 2022 composites and one 2023 composite missing from the Terra archive are filled from the neighbouring Aqua MYD13A1 composites. 2002, 2024 and 2025 are blank because the archive lacks a complete set of composites for those years.
Source: NASA Global Imagery Browse Services (GIBS), layer MODIS_Terra_L3_SurfaceReflectance_Bands143_8Day: true-colour surface reflectance from the MODIS instrument on NASA's Terra satellite (product MOD09A1), requested through the GIBS map service for every year from 2000 to 2025. Imagery courtesy of NASA EOSDIS GIBS.
How each year's image was chosen: a single day or week over Japan is mostly cloud in summer, so each year is a composite. The 16 eight-day images from 2 June to 30 September are stacked, and every pixel takes the value at the 25th percentile of brightness among its valid observations. Cloud is bright, so this keeps a clear view without picking the darkest value, which tends to be shadow. The same window is used every year so that years are comparable. The 8-day surface reflectance layer was preferred to the daily corrected reflectance layer because it is already corrected for the atmosphere and has no gaps between satellite passes.
Limits: the sea is a flat fill, drawn with the Natural Earth 1:50m land outline, because this product covers land. Land with no summer observation, such as Antarctica in polar night, is pale grey. The world image is 4096 by 2048 pixels, about 10 km per pixel at the equator, and the Japan image is about 1 km per pixel, so zoom stops being sharper beyond that. The images are stored with this page, each wrapped in a small script file so the page also works when opened straight from disk; it does not stream live tiles.
Repairs to the world image: the world composites had three faults, which showed on the globe as a ring near the Arctic and a hole at the South Pole. Speckle over the sea between about 60°N and 72°N is replaced with the flat sea fill. A strip of land from about 65°N to 69°N across northern Eurasia was missing from the mosaic; it is rebuilt by mirroring and blending the imagery immediately north and south of it, so that strip is an estimate, not an observation. The imagery ends at 85°S, so the Antarctic interior south of that takes the ice colour of the rows above it. The Japan images and the East Asia map are not affected.
Elevation: NASADEM, NASA JPL's reprocessing of the Shuttle Radar Topography Mission, from the Microsoft Planetary Computer. The 1 arc-second tiles are averaged to 30 arc-seconds and sampled onto a grid of roughly 3 km, so sharp peaks read lower than their true summit: Mount Fuji appears near 3,000 m, not 3,776 m. Heights are drawn 15 times taller than true scale so that relief is visible at the scale of the whole country. The model covers Hokkaido, Honshu, Shikoku and Kyushu; Okinawa is outside it.
Rendering: the globe and terrain are drawn with three.js; maps and charts with D3.
Prefectures: Global Map Japan by the Geospatial Information Authority of Japan (an MLIT agency), via dataofjapan/land, simplified. Small islands are omitted from the drawing. World outlines on the globe: Natural Earth 1:110m, via world-atlas.
Biomes: the 14 terrestrial biomes of the WWF Terrestrial Ecoregions of the World (Olson et al., 2001), from a dissolved and simplified Esri layer, simplified again here. They show the natural vegetation the climate supports, not present-day land use, so farmland and cities on the North China Plain still appear as forest.
Mountain ranges, plateaus, deserts, plains, basins, coastlines, borders, rivers, lakes: Natural Earth 1:50m physical and cultural layers. Range outlines are generalised. Ranges inside Japan, Korea and Taiwan, and the peaks, were placed by hand as dotted ridge lines and triangles.
Satellite: the same NASA GIBS MODIS Terra summer composites as the globe, cropped at about 10 km per pixel. The sea is a flat fill.
Extent: the map covers only Japan's neighbourhood, from the Tibetan Plateau (about 84°E) to Kamchatka and from the northern Philippines (15°N) to central Siberia (60°N): the regions whose winds, currents, mountains and dust reach Japan. Places further away, such as India and Central Asia, and five biomes that cover less than 2% of this area (tropical dry forest, tropical grassland, subtropical conifer forest, mangrove, rock and ice) are left out.
Winds, fronts, currents, trenches: drawn by hand from standard climatology to show typical positions. Real paths shift from day to day and year to year. The short explanations on that page are general background, not measurements.
Names: the sea between Japan and Korea is labelled Sea of Japan; it is called the East Sea in Korea. Borders follow Natural Earth and do not imply a position on any dispute.
Ordinary least squares slope with a two-sided t-test, Theil-Sen slope, Mann-Kendall test with tie correction, paired and Welch t-tests, Pearson correlation. Significance threshold p < 0.05. "Japan" is the unweighted mean of the 47 prefecture series.
Each forecast extends a straight line fitted to the 2000–2025 record. The Sen's slope forecast uses the Theil-Sen line (median pairwise slope, median intercept). The least-squares forecast uses the ordinary least-squares line, and the 95% prediction interval is the standard interval for one new observation: forecast ± t × s × √(1 + 1/n + (year − mean year)² / Sxx), with n − 2 degrees of freedom. The interval assumes independent, normally distributed scatter about a straight line and does not allow for the trend itself changing. Forecasts are held within each measure's possible range (for example, snowfall cannot fall below zero). When the Mann-Kendall test finds no significant trend, the page says so and gives the long-term mean as an equally fair forecast. These are extrapolations for teaching, not climate projections.
Learn: every worked example, detective-case answer and demo is calculated in the browser from the same series and the same statistical functions as the dashboard. The shuffle demo is a permutation test: it reorders the years at random 2,000 times and counts how often the Mann-Kendall score S is at least as far from zero as in the real record. Solved cases are remembered in the browser only.
Zen: a rule-based helper, not a language model. It looks for prefecture names, measure key words and years in a question, chooses one of a fixed set of answers (value, trend, comparison, ranking, projection, correlation, definition, quiz) and fills it with numbers from the functions above. It runs entirely in the browser and sends nothing anywhere. It will misread questions phrased in ways its rules do not cover.
Zen and the East Asia map: Zen reads the same explanations of winds, currents, fronts, places and biomes that the East Asia map shows (one shared file, eastasia.js). For each influence it then runs a fixed test on the prefecture records: the 2000–2025 mean of one measure in a group of prefectures the influence should affect, against a comparison group or Japan as a whole (paired t-test on the yearly group means), the group's trend (Sen's slope, Mann-Kendall), and that trend extended to 2040 as on the Predictions page. The groups are named in each answer. These tests show whether the records are consistent with an explanation; they cannot prove it, because latitude, altitude and station siting differ between groups too. Yellow dust and plate boundaries have no matching measure, and yearly values cannot isolate seasonal effects such as yamase; Zen says so.
Carry the 2000–2025 trend forward. The page fits a line to the record with two methods from the dashboard, Sen's slope and least squares, and extends it to the year you choose.
How to read it. The Sen's slope column is the main forecast: it is the trend line that one freak year cannot drag around. The least-squares column is the classic straight-line fit. The 95% prediction interval is the range a single future year would be expected to fall in 19 times out of 20, if the straight-line pattern of 2000–2025 holds. It widens the further ahead you look.
What it cannot know. A straight line assumes the past rate of change simply continues. It knows nothing about emissions, El Niño, volcanic eruptions or city growth. Treat these numbers as "if nothing changes" scenarios, not as weather or climate forecasts.
Arctic Visualizer holds 26 years of climate records (2000–2025) for all 47 prefectures of Japan. You can see them on a map, test whether they are changing, compare places and periods, and project the trends forward.
Every prefecture has one value per year for each measure. Most come from the prefecture's main weather station; vegetation is averaged over the whole prefecture from satellite.
Each idea below is explained in plain words, then worked out live on real data. Change the example to see the numbers move.
A fair comparison changes one thing at a time and asks whether the gap is bigger than ordinary year-to-year wobble. The Stats comparator has three modes, one for each kind of question.
Three habits of a careful comparer
Six open cases. Each asks a question the data can answer. Open the evidence, investigate with the site's tools, then give your verdict. The answers are worked out from the live data, so no guessing from memory.
Ten questions on the ideas above. Zen, the site's helper bot, gives a hint if you ask and explains every answer.
Zen is a small helper built into this page. It spots key words in your question (a place, a measure, a year) and works out the answer from the site's own data.
Zen can:
Zen is not a large AI model. It follows fixed rules, so it will miss questions phrased in ways it does not expect, and it only knows this site's data. It runs entirely in your browser: nothing you type is sent anywhere.