· · · · · · · · · · — composed on one globe, from public data only
earth composes the world's open climate data on one globe: eighteen satellite layers, the ocean's interior from Argo floats and the GLORYS reanalysis, glaciers, biodiversity, facility-level emissions — plus dashboards for global temperature, Earth's energy imbalance, the AMOC and sea level, and a searchable catalog of 277 open datasets. Public data only; no keys, no accounts, no server.
In a hurry? The words in the tagline above are one-click scenes — "sea ice" puts the daily sea-ice field on the globe, "floats" the live Argo fleet, and so on. Each shows exactly one layer (and replaces whatever was on); the chips that appear on the globe undo it.
This Layers view is the globe's control room:
The other tabs: Temp — warming since 1880 · Energy — the planet's energy imbalance · AMOC — the overturning circulation this app is ultimately about · Sea level — the rise and its causes · Tides — the global tide animated live, with the moon · Cones — the space-time cone our own forecaster reads for one pixel, anywhere on Earth, drawn on the globe in kilometres, with a Data switch that fills it with the real values out of the global training tensor and lets you tap any dot to read it · Data — download the observation stores (satellite ocean colour, sea-surface temperature, cloud tops, ship, float, drifter, mooring and bottle measurements), the global tensor the forecaster reads, the atmosphere's temperature, humidity and wind at 13 heights from the ERA5 reanalysis, and derived maps (fishing effort, the model's climatology) for any period, month, box and resolution, as a file built in your browser · Catalog — all 277 datasets · About — background & further reading.
Daily NASA GIBS imagery. Aggregate averages any continuous layer per pixel over the past N days ending on the chosen date (1–730). Missing pixels (clouds, night) are excluded — each pixel divides by the days it was actually observed, so cloudy-day gaps fill in; for precipitation, dry pixels count as zero, so the result is a true mean rate. Independent of the comparison mode.
Which date it compares against. An offset (“vs 10 years ago”) is relative: move the Date and the comparison moves with it, keeping the same month on both sides — which is what makes a satellite comparison mean anything, since most of a raw month-to-month difference is the seasonal cycle. A specific date pins it instead, for reading everything against one reference. Typing in the comparison date field, or stepping it, pins it automatically; choosing an offset again hands it back to tracking.
Side-by-side: drag the divider (left = past, right = current). Computed change: per-pixel "now − then" for continuous layers (SST & anomalies, sea ice, snow, land temperature, salinity), and a ×-fold ratio of window means for log-distributed ones (precipitation, chlorophyll, aerosol).
While a window is active, layers that can't be averaged (true colour, 30-min precipitation) are hidden rather than shown as a misleading single day — a note appears below when that happens. Averaged views cap zoom to stay responsive and sample up to 12 days from the window.
The window can be set four ways, all the same number: drag the slider, nudge it a day at a time with −1d / +1d, type it outright, or take a preset. 12d is there because it is a full satellite repeat cycle — the window at which the swath union below closes over the whole planet.
Swath layers (Sentinel-2, Landsat, Sentinel-1, NISAR, the two surface-water maps) treat the window as a lookback instead: they show the union of every pass in it, newest on top — 12 days covers the whole planet for NISAR — at full zoom depth. A pixel no pass reached stays blank. For the two surface-water maps the union also looks through cloud (and the radar masks): those classes are not a measurement, so an earlier clear day shows through them. The legend strikes those classes through and lets you switch it off.
Composes everything the app can know about a clicked point: live weather and a 7-day forecast, all nine satellite fields at the chosen date, climate normals with the SST anomaly, and nearby observing/emitting context. When this is off, a click reads the value of the layer you're looking at instead — unless no layer is active, in which case the full card is the only sensible answer.
Three independent estimates of the same sea-surface temperature field, combined in the browser. The mean is a best guess; the spread shows where the analyses disagree — i.e. where the ocean is poorly observed and any single product is least trustworthy.
The world's largest individual sources of greenhouse gas — power plants, steelworks, oil & gas fields, shipping lanes — estimated facility-by-facility from satellites and public data by the Climate TRACE coalition. Dot size scales with annual CO₂-equivalent emissions; click a dot for details.
A fleet of ~3,800 robotic floats drifting through every ocean, each diving to 2,000 m every ten days and measuring temperature and salinity on the way up. Argo is how we know the ocean absorbs over 90% of global-warming heat; click a float for its live dashboard.
A loitering event is a vessel drifting at sea at low speed for hours: not steaming, not fishing, just waiting. Global Fishing Watch derives them from AIS — the Automatic Identification System, the radio transponder ships broadcast their position on — by finding vessels that stay inside a small radius below a speed threshold for long enough. It is the signature of transshipment at sea, where a catch is moved onto a refrigerated carrier so the fishing boat never has to come to port, and of waiting generally. Click or hover a dot for the vessel, its flag and MMSI, and how long it drifted. The snapshot is a rolling 30-day window; outside that window the layer draws nothing and says so, rather than showing you old points under a new date.
How much plastic floats in the Great Pacific Garbage Patch — the accumulation zone between Hawaii and California where ocean currents converge. In summer 2015 The Ocean Cleanup sent 18 vessels to tow nets across it (a fine-mesh net catches pieces down to 0.5 mm, a larger net the 5–50 cm pieces), and in October 2016 a C-130 aircraft photographed it from about 400 m to count objects over 50 cm. From these the paper estimated at least 79,000 tonnes of floating plastic (45,000–129,000) in about 1.6 million km², over three-quarters of the mass in pieces larger than 5 cm and at least 46% of it fishing nets, while 94% of the 1.8 trillion pieces are microplastics. Each dot is one tow or one photo mosaic; values are corrected for wind mixing (waves push light pieces below the net), and the card also shows what was actually caught.
The fixed sentinels of the climate system: the mooring arrays that measure the Atlantic overturning circulation (RAPID, OSNAP, MOVE, SAMBA — feeding the AMOC tab) and the observatories behind the iconic CO₂ records (Mauna Loa's Keeling curve, Jungfraujoch). Each marker links to its own data.
Every glacier on Earth (274,531) from the Randolph Glacier Inventory, joined with each glacier's measured 2000–2020 thinning rate (Hugonnet et al. 2021, from differenced satellite elevation models). Switch to melt-rate colouring to see 78% of glaciers thinning — and the anomalous stable Karakoram.
Red water stands above mean sea level, blue below (legend on the globe; tap the ocean for the exact cm) — the two tidal bulges sweeping the ocean, rotating around amphidromic points, spring and neap breathing over a fortnight. While this layer runs, the sun lights the globe on the sim clock, so daylight, ☀ and the solar tide agree. Five harmonic constituents (M2 S2 N2 K1 O1) per 1° cell, ~90% of the tidal signal.
Nearly four billion dated, located records of where life has been observed — museum specimens to citizen-science photos — from the Global Biodiversity Information Facility. Filter by kingdom, major group (birds, insects, mammals…), or a climate-indicator species; warming shifts many ranges poleward and uphill at a median of ~17 km per decade. Humans are in there too, but privacy-restricted to a tiny count.
Annual global surface-temperature anomaly vs the 1951–1980 mean. The land-only curve runs well above land+ocean — land warms faster than the sea because water stores and mixes heat. NASA GISS GISTEMP v4. Hover for values.
Land has warmed roughly twice as fast as the ocean surface. That contrast drives shifting climate zones, heatwaves, and the poleward species shifts in the Biodiversity layer — and it is why a "global" number understates what continents actually experience. For land temperature in space, turn on the Land surface temperature (MODIS) layer on the globe.
Source: NASA GISS GISTEMP v4.
The planet-scale cause of this warming — the energy imbalance driving it — has its own Energy tab.
The planet is running an energy surplus: averaged over the last decade of the ocean record charted below, Earth took in about – W/m² more energy than it returned to space — a measured number, and the direct physical cause of global warming. Everything else in this app (rising temperature, melting ice, rising seas) is this surplus, distributed. It is small in relative terms — ~0.3 % of the ~340 W/m² flowing through the system — which is why it can only be seen in careful global accounting, never on a map.
The numbers on this panel fit together like this: the big tile, – W/m², is what this record directly measures — heat accumulating in the ocean's 0–2000 m layer, as a 10-year average. Dividing by 0.9 adds the ~10 % that melts ice and warms land, air and the deeper ocean, giving the total above. And the often-quoted "about 1 W/m²" is where the most recent years of the second chart stand — the imbalance varies year to year and has roughly tripled over the record, so any single figure needs its date attached. NASA's CERES radiometers, measuring the same quantity completely independently from orbit, agree on both the size and the rise (their 2023 annual value touched ~1.8 W/m²).
And the much larger numbers in IPCC reports? Assessments quote ~2.8 W/m² of radiative forcing from greenhouse gases (IPCC AR5, 1750→2011; AR6 puts the total net human forcing at ~2.7 W/m² by 2019, once the ~−1 W/m² of aerosol cooling is netted in). Forcing and imbalance are different quantities: forcing is how hard we push the climate away from balance, while the imbalance on this panel is the part of that push not yet answered. Because the planet has already warmed ~1.2 °C, it radiates roughly 2 W/m² more back to space than in 1750 — cancelling most of the forcing. The remainder — the ~0.8 W/m² measured here — is what still flows into the ocean. If forcing stopped growing, that remainder would decay toward zero as warming catches up; instead the second chart shows it rising: the push is growing faster than the planet can answer it.
Y axis: accumulated heat, in 10²² joules (= 10 ZJ), relative to a mid-record baseline — not a rate. About 90 % of the planet's excess energy ends up in the ocean, so this curve is the imbalance's bank statement: its slope is the imbalance itself. A steady imbalance would be a straight line; the visible steepening means the imbalance has been growing. Always shown raw — a running total needs no smoothing; the Smooth control below applies to its slope. The depth layers nest: 0–700 m is contained in 0–2000 m, so their difference — the 700–2000 m slab — shows heat arriving in the deep layer, a growing share of the total: warming is reaching ever deeper. Below 2000 m no yearly series exists (ships survey the abyss roughly once a decade); those repeat surveys add another ~5–10 % (≈0.06 W/m², Purkey & Johnson), which is part of why the total imbalance exceeds the 0–2000 m number. Hover for yearly values.
Y axis: watts per m² of the whole Earth's surface — the slope of the curve above, computed over the Smooth window (at 1y it is simply each year's heat gain; wider windows show the underlying trend), so yes: the imbalance varies over time. The wiggles are real climate variability, and the shading names it: amber bands are moderate-or-stronger El Niño winters (the ocean sheds heat to the atmosphere — the rate dips), teal bands La Niña (the ocean banks extra heat), deeper tint = stronger event — hues deliberately unlike either data line, which plot heat content, not ENSO. Only moderate+ events (|ONI| ≥ 1.0, ~30 % of winters) are shaded: at the official "weak" threshold of ±0.5 °C, seventy percent of winters qualify — ENSO's neutral state is genuinely the minority — and the chart read as if every year were an event. Weak years still show on hover. The ▲ triangles mark eruptions that loaded the stratosphere with sulfur, dimming sunlight for ~2 years — watch the rate sag after Pinatubo 1991 (Hunga Tonga 2022 is the oddball: mostly water vapour, a slight warming agent). Famous ash crises like Eyjafjallajökull 2010 don't appear because they don't belong: grounding flights takes ash in the troposphere; changing Earth's energy budget takes megatons of stratospheric sulfur, which that eruption never delivered. The climb from ~0.2–0.4 W/m² in the late 20th century toward ~1 W/m² now is the acceleration. Divide by ~0.9 for the total imbalance (the other ~10 % melts ice and warms land and air). CERES satellite radiometry, measured completely independently at the top of the atmosphere, shows the same magnitude and the same rise — one of the strongest closure tests in climate science (Loeb et al. 2021).
The two thin curves put the measured lines in context — they are the push, not the stored heat: the dashed line is total human forcing (greenhouse gases minus the aerosol offset, AR6/Forster et al. annual series), climbing from ~0.3 W/m² in 1955 to ~3 W/m² now, and the green line is everything nature adds — solar variability plus volcanoes. Note how small the green line is: it hugs zero (±0.2 W/m², the ~11-year solar cycle) except when a big eruption drives it briefly to −1…−2 W/m² — dips that land exactly on the ▲ markers. The gap between the dashed push and the measured imbalance below it is not natural cooling — nature has nothing that size on offer. It is the planet's own answer: the ~1.2 °C of realized warming radiates roughly 2 W/m² more to space, paying off most of the push; the measured lines are the part still unpaid.
The dashed curve would not drop to zero — it is attached to the stock of greenhouse gases already in the air, not to ongoing activity. If emissions stopped dead: the aerosol sunshade (~−1 W/m² of the offset inside the dashed line) washes out within weeks, so the first effect is a few tenths of a degree of extra warming; methane (~12-year lifetime) decays within decades; CO₂ draws down only over centuries to millennia. Model experiments (ZECMIP, assessed in AR6) find temperature then roughly plateaus for a century or more: the slow CO₂ decline and the ocean's slow release of banked heat nearly cancel. The imbalance charted here would taper toward zero — but the thermometer would not come back down on any human timescale. Stopping the cause freezes the effect in place; it does not rewind it.
Remove the dashed curve entirely and the green line is the whole story: a world pushed only by the ~11-year solar cycle and the occasional volcanic dip. In that world the late 20th century looks like the green line — flat to gently cooling. Reconstructions and attribution studies agree: natural forcing alone since 1850 gives roughly 0 to −0.1 °C — slight cooling, from the slow orbital drift toward the next glacial inception (tens of thousands of years away) plus episodic volcanism. Essentially all of the modern warming, and everything the ledger above records, sits in the gap between the green and dashed curves. The tempting arithmetic "human push is ~1.8–2.8 W/m², the imbalance is only ~0.8, so ~1–2 W/m² must be natural cooling" fails because the missing term isn't nature — it's the warming itself, already radiating the difference back to space.
1 W/m² planet-wide ≈ 16 ZJ per year — roughly 250 times humanity's annual energy consumption, absorbed as heat, every year. The famous comparison is arithmetically exact: today's imbalance equals detonating about – Hiroshima-sized bombs' worth of energy every second, around the clock (15 kt TNT ≈ 6.3×10¹³ J; James Hansen's 2012 figure was ~4/s — the number has grown with the imbalance). The image is worth keeping honest: nothing explodes — the same joules arrive as diffuse warmth spread through billions of tonnes of seawater, which is precisely why something this enormous stays invisible without careful accounting. The +224 ZJ the upper ocean has banked since 2005 resurfaces elsewhere in this app: as the steric (thermal-expansion) share of the Sea level tab's budget, as the "upper 700 m stored heat" line when you click the ocean with the pixel inspector, and as the subsurface marine heatwaves in the 300 m anomaly layer. The Energy balance (CERES) globe layer shows the ±100 W/m² geography whose near-cancellation hides this number: the tropics bank energy, the poles shed it, and the currents and winds between them are the delivery system.
Source: NOAA NCEI Global Ocean Heat Content (Levitus et al.); forcing curves from the Indicators of Global Climate Change ERF series (Forster et al., AR6 method, annual to 2024); framing per Loeb et al. 2021 & von Schuckmann et al.
Atlantic overturning transport at 26.5°N, 10-day means, Apr 2004 – Mar 2024. 1 Sv = 10⁶ m³/s. Hover the chart for values.
Data: RAPID-MOCHA-WBTS
(NOC / U. Miami / NOAA). Refresh with scripts/refresh_data.py.
The AMOC carries ~17 Sv of warm water north at this latitude on average. Statistical early-warning studies (Ditlevsen & Ditlevsen 2023) and model experiments (van Westen et al. 2024) suggest weakening toward a possible tipping point — this dashboard tracks the primary observational record. Roadmap: OSNAP & Florida Current series, SST-fingerprint indices, variance / autocorrelation early-warning statistics.
Observed global mean sea level (white) and its causes as separate lines; the grey dashed line is the summed budget — it tracking the observed line is budget closure. Blue dots are satellite altimetry. 1900–2018 budget from Frederikse et al. 2020; altimetry from NOAA. Hover for the year-by-year breakdown.
Sea-level rise is a physical accounting identity: the observed rise must equal thermal expansion of warming water plus the mass added by melting glaciers, Greenland, Antarctica, and changes in land water storage. That these independent measurements sum to the observed total ("closure") is one of the strongest cross-checks in climate science — and it decomposes why the ocean is rising: roughly a third thermal expansion, the rest added ice mass. A small residual between the observed line and the summed stack reflects remaining measurement uncertainty.
Sources: Frederikse et al. 2020, NOAA STAR altimetry.
The picture is on the globe — opening this tab switches the Tide height (live) layer on (also in the Layers list); this tab is its control room. ☾ marks the sub-lunar point (the moon directly overhead), ☀ the sub-solar point.
The curve runs from six hours ago to three days ahead, so the blue now line sits on the water and you can see whether the tide is rising or falling; dashed lines are local midnights, dots are high and low water (dimmed ones have already happened). The vertical axis scales to this point — the globe's colour scale is fixed for every point, so a shelf sea and mid-ocean look very different here and identical there. Tap another ocean point, or search a place, to move this readout.
This is the actual shape of the global tide, reconstructed each frame from the five main harmonic constituents of the EOT20 ocean tide model — the moon's twice-daily tide (M2), the sun's (S2), the moon's elliptic-orbit correction (N2), and the two daily constituents (K1, O1). Red water stands above mean sea level, blue below.
Two things the equilibrium cartoon never shows: the bulges cannot keep up with the moon — continents are in the way — so the tide instead rotates around fixed hubs (amphidromic points, the still spots the red/blue pattern pivots on); and the range varies enormously, from centimetres mid-ocean to metres on wide shelves (switch the Tidal range layer on, in the Layers tab, to see where). When ☾ and ☀ pull together (new/full moon) M2 and S2 align: spring tides. At quarter moons they oppose: neap. Watch the pattern breathe over a fortnight at 1 day/s.
Fine print: harmonic constants fit to 1992–2019 satellite altimetry (EOT20, Hart-Davis et al. 2021, CC-BY); 1° grid, five constituents (~90% of the signal in most of the ocean; shallow-water overtides and the 18.6-year nodal modulation, ±10–15% on lunar amplitudes, are omitted). The clock starts at real time and the moon marker is a low-precision ephemeris (~1°) — this tab runs on its own clock, not the date selector.
In data mode the family follows the channel you pick below — change the channel, not the family.
What the codec reads at lag 0: the 3×3 patch, one row per channel, nine cells across.
In plain words. A pentad is a fixed five-day bin — the tensor has one frame per pentad, counted from 1982-01-01, so "lag 3" means three bins, fifteen days, earlier. The codec stencil is what the encoder reads to compress one pixel: the 3×3 patch at lag 0 plus a sunflower of dots at lags 1–6. The stage-2 stencil is what the forecaster reads on top of that: a ring of 24 dots per lag from lag 7 out to lag 143, over the compressed codes rather than the raw field. Raw is the measurement in its own unit; anomaly is that value minus its own calendar month's average (computed on training years only) and then divided by the channel's spread — which is the number the model is actually given, and why it is roughly −3…+3 rather than 15 °C. A hollow dot is one the model cannot read: it falls off the grid the cone stands on, and the model treats it as missing rather than wrapping it around the globe. On the global default — family 7, live or exported — there are almost no hollow dots at all: that tensor covers the whole planet and its columns close, so a dot east of the dateline comes back round at longitude −180° instead of falling off an edge, and the only way left to leave the grid is to reach past a pole. A dimmed dot is one the model does read, but which was never observed there — cloud, land, or a float that did not profile.
These are the geometry's own constants. Move one and the dots on the globe, the three tiles above and the cross-section below are all recomputed from it — but the trained codec read the cone at the values the reset button restores, so a changed slider is a question ("what would it read if…"), never a picture of the model. The badge above says when you are in that state.
The picture is on the globe — tap anywhere on Earth to move the anchor. Red ring: the pixel being predicted. The grid underneath is the whole planet at 0.25°: its columns close at the dateline, so a dot that reaches past ±180° comes back round the other side, and the only way off the grid is past a pole — those few dots are drawn hollow, because the model reads them as missing. Choose the family-4 comparison source under Data and a dashed box appears: that earlier tensor covered only the North Atlantic window (0–70° N, 100° W–20° E), and every dot outside it is hollow.
Two animations, and they move different things. Sweep lags holds the date still and walks the lag from 0 to 143, so you watch the cone open out from one 3×3 patch to a 4,444 km envelope. Advance time holds the lag still and walks the anchor date forward one pentad (five days) at a time, so you watch the cone travel along the time axis. Switch follow the cone on and the globe's date selector follows each lag's own date — put sea surface temperature or ocean currents on in the Layers tab first, and you are looking at the field the forecaster actually reads there. A lag whose date lands in 2009, 2017 or 2023 is flagged red: those three years are held out of training, so the model has never been shown them.
Reach (km) against lag, both axes √-scaled so the six inner pentads stay readable beside 143 outer ones. Solid: each family's inner reach — A dies with its 10-day memory, C drops at lag 2 as the atmospheric stirring goes out of it. Dashed: stage 2's outer envelope, flat from lag 34 where the 4,444 km cap binds. The vertical rule is the lag you are looking at.
A pixel's future depends on a cone in space-time: a driver moving at speed v reaches the anchor from within dx ≤ v·(dt + l) after l pentads, and only lags inside that driver's memory are worth reading. Reach is floored by a correlation length — below it the field is one number, not a gradient. E-069 splits that cone in two: the inner cone (lags 0–6) goes into the codec as raw values, so the embedding can carry local motion; the outer cone (lags 0–143, minus the near field the codec already read) stays in stage 2 over embeddings.
Why six lags, and why these speeds: a displacement is resolvable once it exceeds one cell (0.25° = 28 km meridionally, 21 km zonally at 40° N). At pentad cadence that is 0.06 m/s per lag: eddy advection (0.1–0.3 m/s) moves 2–6 cells per pentad and is resolved at lag 1; Rossby propagation (0.03 m/s) needs ~2 pentads per cell, so 6 lags give 3 cells; the atmospheric forcing decorrelates in 1–2 pentads (family A, lags 0–1); SST anomalies persist 3–6 months, longer than the inner window, and are deliberately left to stage 2.
Lag 0 is not dots. The codec keeps the 3×3 patch there — one token per channel — so every archived comparison stays like-for-like; the dots are what is new. Each lag then gets the anchor's own column plus a golden-angle sunflower over its disc, elliptic (0.71 north–south) because the flow is, and with as many points as the disc's area warrants: 6 bearings at 130 km, 24 at 907 km. The depth channels get the column only — Roemmich–Gilson is one live pentad a month, so a sunflower there would be ~70 tokens that are structurally missing.
Sources: the E-069 plan · ml/cone.py, which is the definition this tab draws — the numbers here are exported from it, never restated.
Pick a store — the fine observation stores (satellite maps at 2–4 km and ship, float, bottle and mooring measurements at their own points), the global tensor the forecaster reads with its own point observations, the atmosphere on pressure levels from the ERA5 reanalysis, or a derived map — choose a period, months, a box and a resolution, and the file is built here, in your browser, from the store's own files. The estimate line says what it will cost before anything is read.
The data reader failed to load, so this tab cannot read any store. Reload the page; if it persists, the files are still there to download by hand from the public data store.
These are the global tensor's monthly normals, composed here from its per-year monthly sums and counts: pick the YEARS the average runs over (and any to leave out), the calendar months, and either one normal per month over the period or the by-year stack — each year's monthly mean side by side. Every value comes with the number of five-day bins behind it.
This store also keeps precomputed monthly sums — per year and calendar month, the sum, the count and the squared deviations of every native map. A monthly mean, one mean over the selection or a normal per calendar month over the period made of whole months (no day-of-month or hour filter) is read from those: a handful of requests where the native maps are thousands. The estimate says which way it reads, and why. Climatology: set the years (and any to leave out), then pick the time step Climatology — mean per calendar month (12 maps) or, where offered, Climatology — mean per day of year (one map per calendar day, 366 for a whole year; use the months and the days slider to take fewer).
This store is a climatology: one map per calendar month, averaged over all the years 1982–2024, so it has no years and no days — the months below pick which calendar months.
Where the numbers come from. Each store is a set of files on the project's public data store (a Hugging Face dataset). A satellite map store keeps one file per five-day bin, every daily (or 3-hourly) map in it cut into 256 × 256 tiles that are compressed one by one, so a box is a few tiles and a period is a list of bins. A point store keeps its reports sorted by time, so a period is one stretch of each column. The page asks for exactly those byte ranges — nothing is computed on a server — and refuses any answer that is not the range it asked for, so a failed read stops the whole download with the address in the message, rather than leaving a silent hole in the file.
The cap. At most 600 MB read or 400 MB of result, whichever comes first. One five-day bin of the whole globe at 4 km is about 108 MB, so a year of the whole globe (about 8 GB) is refused and the estimate links the store's files instead; the same year over the North Atlantic is a few hundred megabytes.
Coarser is not cheaper to read. A coarser resolution or a time mean is computed here after the native bytes are read: it shrinks the file, not the download. That is why the estimate prints both numbers. Every mean skips missing values and comes with a per-cell count of the observations it averaged.
Point stores read the whole period. A box filters the reports here, after their positions have been read, so a short period is cheap and a long one is not, whatever the box. The row count in the estimate is exact — the store records where every five-day bin starts. The days fields are how to make a period short: a few days of the densest stores (the SWOT swaths, the altimeter tracks) is tens of megabytes where a month is gigabytes.
Drawing the box. draw on the globe, then click one corner and the opposite one (or press and drag; on a phone, tap two corners or drag with one finger) — a rectangle follows the pointer, the four fields fill in, rounded to 0.01°, and the box keeps eight handles (corners and edge middles) to drag. Two longitudes are joined the shorter way round, so a box across the dateline is drawn as one; the other way round swaps it. Esc or the Cancel chip stops drawing. While drawing, a click is a corner: the pixel inspector and the value read-out wait.
Waiting. Any read that takes more than a moment — the estimate, the first look a store opens on, the preview, the download — shows a bar, the requests and megabytes done of the total, the time so far and, once the rate is steady, the time left, with a Cancel. Where nobody can know the total yet (the store's indexes are still being read, or the first look is still trying months) the bar runs without a percentage and counts instead. A slow read says why: many small requests, each a round trip, or simply many megabytes.
Cheap long-period averages. Eight map stores — the daily OISST, GLORYS, NCEP and OC-CCI stores (family 7.2d) and the four ERA5 stores (family 1.2) — also keep per-year monthly sums, counts and squared deviations of every native map. A monthly mean, one mean over the selection, or a normal per calendar month over the period made of WHOLE months is read from those — January 2015 of ERA5 at 500 hPa over the North Atlantic is 2 requests and 56 kB instead of 255 requests and 24 MB — and gives the same means to float32 rounding, with the same counts. A day-of-month range or (for ERA5) an hour filter cuts months, so then every native map is read; the estimate says which, and why. A frame belongs to the calendar month of its own start. The standard deviation is offered as an extra variable: the spread of the native maps behind each mean, in population form.
The paper's climatology. Store Monthly normals of the global tensor (the group with your channel: 0.25° ocean surface, 1° fluxes and weather, ocean colour, or Argo at depth); years 1982 to 2020; leave out 2009, 2017; all twelve months; time step Climatology — mean per calendar month over the years. The paper split chip sets exactly that; development split is 1982–2024 leaving out 2009, 2017 and 2023, and all years is 1982–2024 with nothing left out — the forecaster's three published climatologies (clim/paper, clim/dev, clim/all), reproduced to float32 rounding (about 10⁻⁶ in the channel's unit). Only these stores reproduce them: they use the model's own rule — a five-day bin belongs to the month it opens in — and the tensor's own standard scores; the daily stores' calendar-month normals are close but not identical. Size: a whole-globe 0.25° twelve-month normal is far over the 600 MB cap (about 2.3 GB of sums and counts per channel), so take one month at a time (about 190 MB for one channel of the whole globe) or a box — 1° cells shrink the file, not the read, of a 0.25° group; a 1° group's whole-globe twelve-month normal of one channel is about 150 MB. Or download the published whole-globe files:
Three kinds of store. The fine observation stores keep each source at its own resolution. The global tensor (the forecaster's own input, 0.25° and 1° five-day maps since 1982) is stored as one uncompressed file per group with every channel side by side, so a box reads a band of rows across ALL of a group's channels whatever you tick — the estimate counts it — and its values are stored as standard scores, which the page turns back into physical units with each channel's mean and spread. The monthly normals are per-year sums and counts, averaged here over the years you choose.
The atmosphere is a reanalysis, not an observation. The ERA5 stores (family 1.2) hold temperature, humidity and wind at 13 pressure levels, from 50 hPa (about 20 km up) to 1000 hPa (the surface), every six hours since 1982. ERA5 is ECMWF's reanalysis: a weather model run forward and pulled toward every available observation, at about 31 km, averaged here onto a 1° grid. It fills places and heights nobody measured, which is what makes it global — and why it is a model's best estimate, not a measurement. Each map is an instant at 00, 06, 12 or 18 UTC, not a six-hour mean. Humidity is in grams of water vapour per kilogram of air (1000 × ERA5's kg/kg). All 13 levels sit side by side in every tile, so one level reads as much as thirteen — the estimate counts it. Contains modified Copernicus Climate Change Service information, CC BY 4.0; every file carries the full attribution.
What the values are. Sea-surface temperature from
satellites (sst_acspo02) has gaps wherever there was cloud — it is what the
satellite saw, not a gap-filled analysis, and not a finer version of the 0.25° daily
analysis (OISST) the rest of this site uses. Cloud-top brightness temperature
(irtb) is kept every three hours, although the satellites see every half
hour, and is stored as a whole number of kelvin minus 160; the file carries kelvin. Ocean-colour chlorophyll is the base-10
logarithm of the concentration, not the concentration itself. Each file states its
units and, in its attributes, the store and the selection that made it.
Guide: the Data tab, control by control · the plan (E-084: this tab and its reader)
The picture is on the globe — this tab animates the date, so whatever you have on the globe right now is what plays. Set the view up in the Layers tab, come back here, choose a span and press play: the date walks from start to end and every dated layer follows it, exactly as if you were holding down the +1d button.
Which means the comparison and the aggregate come along for free. Play with a side-by-side comparison on and you get a wipe of a moving present against a pinned past; play in computed difference and you animate the anomaly rather than the field; play with a 30-day window and you get a running mean instead of daily weather noise. Nothing here overrides those — it just moves the date underneath them.
Nothing on the globe follows the date yet, so there is nothing to play. Switch a dated layer on in the Layers tab first — surface temperature or sea ice are good ones to watch — and its frames will show up here.
Choose a start and an end date — or one of the presets — and press ▶. What the step resolved to, and how fast the frames are really arriving, is reported here while it plays.
The step follows the data, not the calendar. On auto it takes the finest cadence among the layers you have on — a day for daily products, five days for the 5-day sea-surface-height epochs, a month for the monthly ones, a year for the annual ones. A monthly product has nothing new to show on the 2nd of the month, so stepping it a day at a time would mean thirty identical pictures and thirty identical requests to NASA. Frames that would look the same are dropped; you can still force a step by hand.
The requested speed is a speed limit, not a promise. Frames are satellite tiles arriving over the network from NASA, and the player waits for the picture to be on screen before advancing — so asking for 8 frames/s on a slow link gets you whatever the link gives. The status line above says which of the two is binding, because a player that claimed 8 while showing 1.4 would be lying about the thing you are looking at.
A long range coarsens the step; it never shortens the range. Thirty years at one day is twelve thousand frames, which is neither watchable nor polite to a public NASA service. Past a few hundred frames the step widens a notch — day to 5 days, 5 days to a month — and the panel says so. The span you asked for is always the span you get; only its resolution gives way.
Not offered, deliberately: saving the animation as a GIF or a video. What you are watching is the live globe, with everything that implies — you can still spin it, click a pixel and read its value mid-play.
earth is a prototype for exploring the world's open climate data on a 3D globe — and, eventually, for generating predictions from it (e.g., will the AMOC collapse, and when?). Everything runs from public data with zero API keys and no server of its own.
On the globe: eighteen NASA GIBS satellite layers (temperature, ice, rain, vegetation, forest loss, water storage, energy balance …), the ocean interior from Argo floats and the GLORYS reanalysis (currents, mixed-layer depth, subsurface heat), the Randolph glacier inventory with melt rates, GBIF biodiversity, Climate TRACE emitters, and the AMOC monitoring network. Behind it: a machine-readable catalog of 277 open climate datasets and dashboards for temperature, the planet's energy imbalance, the AMOC and sea level. The full guide lives at the top of the Layers tab.
Analysis notes: combining datasets · the per-pixel state vector · species & climate
Data catalog (readable) · catalog.json · source on GitHub
ML status & training curves — live runs of the Earth-State Embeddings codecs, their AMOC probes, and the paper.
Build a34edd54