Loading active storms…
Check back during an active storm for satellite imagery.
Current MJO / BSISO phase + 60-day Wheeler-Kiladis-filtered OLR Hovmöllers. Active TCs are projected onto each Hovmöller as a vertical longitude trace so you can see whether a storm is riding a convective envelope or wave.
Pre-season and intra-season SST evolution from OISST v2.1, anchored against the 1991-2020 climatology and historical annual ACE (1982-present, NHC).
Cumulative Accumulated Cyclone Energy through the season, against the climatological spread for the same calendar day. The shaded bands are the 10th-90th and 25th-75th percentiles of 1991-2020 seasons; the heavy line is the live season built from operational NHC/JTWC best tracks (ACE = 10⁻⁴ Σ v², 6-hourly synoptic points at ≥ 34 kt while tropical or subtropical). Southern Hemisphere basins run July-June and are labeled by the year the season ends. The curve therefore advances in 6-hourly steps: a storm at 65 kt adds 0.42 per step (1.7/day), one at 130 kt adds 1.69 per step (6.8/day).
Scrub through any past TC season with the chosen ERA5 environmental field underneath and IBTrACS storm tracks materializing as the time slider advances. All daily-archive variables (shear, level winds, vorticity, divergence) read the 00Z synoptic snapshot — no daily-mean smearing of transient features. Storms remain on the map after landfall as long as they retain a true-TC nature (TS/TD/HU/TC/SS/SD); they fade once classified extratropical or dissipated. Anomaly mode subtracts the 1991-2020 climatology so you see how a given season's environment compared to normal at each time step.
Pixel-wise Pearson correlation between monthly SST and annual basin ACE (1982-present). Red = warm SST coincides with active seasons; blue = warm SST coincides with quiet seasons. Toggle raw vs detrended to separate shared long-term warming from year-to-year coupling.
Top-10 historical analogs to the target year/month. Grid-weighted by |r| (default and recommended) compares the full SST anomaly field pixel-by-pixel, weighting each pixel by its absolute correlation with the chosen basin's ACE — no region definitions and no overlap. The other two methods operate on the 14 fixed region-mean values (which have some overlap, e.g. Atl basin contains MDR); use them if you want a faster mental model or to compare.
| # | Year | Distance | NA ACE | Storms |
|---|
For the selected target year + month, each historical year's region-mean SST anomaly vector is compared to the target's vector. Smaller distance = more similar SST pattern.
Correlation-weighted distance:
d = √Σ wᵢ · (anomᵢ - target_anomᵢ)²,
where the weight wᵢ = |r(SST_region_i, ACE_basin)|
is the absolute Pearson correlation between region i's
monthly SST and the chosen basin's annual ACE
(1982-present). Regions with little linkage to ACE
(small |r|) contribute little. Weights are
re-computed per basin × month × raw|detrended.
Derived from the same OISST record as Panels A-E. ENSO horizontal markers at ±0.5 °C anomaly are the conventional El Niño / La Niña thresholds. AMM proxy = North Tropical Atlantic SST anomaly − Tropical South Atlantic SST anomaly (sign-of-the-gradient version of Vimont & Kossin 2007).
Aircraft reconnaissance flight-level data — USAF (30-s HDOB) & NOAA (1-s) tracks, dropsondes, and VDM center fixes — time series synced to an interactive map, updating in real time.
Every 2026 storm with aircraft reconnaissance, re-scored with the current SEAR model so the season is one framework. Pick a storm to replay its flights on the Live Flight map — step or play the mission in time, with the satellite backdrop, dropsondes, VDM fixes and SEAR 10-m estimates as they would have appeared.
Recent NOAA P-3 Tail Doppler Radar missions from the HRD real-time archive. Select a mission to open it in the TDR viewer.
Browse and visualize real-time Tail Doppler Radar analyses from NOAA P-3 reconnaissance missions.
Full-sortie in-situ (IWG1) & SFMR time series from the NOAA P-3 reconnaissance data stream. Select a mission to load its flight-level traces.
Decoded center fixes, minimum sea-level pressure, and peak flight-level / surface winds from NHC reconnaissance VDMs (REPNT2 / REPPN2). Select an active storm to load its messages.
How TC-ATLAS turns DeepMind's raw ensemble tracks into the genesis markers on this map
The genesis markers summarize Google DeepMind's experimental cyclogenesis ensembles — FNV3 (a 1,000-member set, plus an earlier-publishing 50-member companion) or, when the WN3 toggle is on, the 64-member WeatherNext 3 ensemble (experimental, not yet operational). The markers are produced by TC-ATLAS's own clustering of the raw ensemble tracks — they are model genesis scenarios, not observed disturbances — as follows.
Genesis definition. Each ensemble member's forecast tracks are scanned for the first point at which the member reaches tropical-storm strength (≥34 kt). We deliberately use TS strength rather than tropical-depression genesis: it filters out weak vortices whose closed-circulation status is ambiguous in an automated tracker, so these probabilities read slightly differently than NHC outlook percentages.
Clustering. First-genesis points are binned on a 3° grid, local density peaks seed the clusters, and each member genesis event is assigned to the nearest peak within 1,000 km whose mean genesis time agrees to within ±60 hours. A marker's formation probability is the count of distinct ensemble members in its cluster divided by the ensemble size — no member is counted twice within a marker. A member landing in two clusters closer than ~500 km is kept only in the nearer one, and a conservative merge step reconnects a single system that genesis-timing spread has split into adjacent fragments (the fragments' mean tracks must travel together at shared valid times, and their members must overlap in space like one population rather than two systems in tandem).
Wave families (linking). One physical wave whose members disagree about where genesis first occurs still appears as several clusters strung along its corridor. A linking pass groups those sibling scenarios into a family — without changing any per-cluster number — using two independent tests. The corridor test compares cluster-mean positions at the same valid times (so a second wave trailing on the same track never matches: at any instant it is far behind), backstopped by a member-spread ratio that separates one population from two systems traveling in tandem. The exclusivity test uses membership statistics: sibling scenarios compete for members, because a member that first develops in one place cannot also first develop in another — so shared membership between true siblings collapses far below the chance rate expected of independent systems. A third, kinematic test covers the cohort structure the first two cannot see — members that stay weak and develop far downstream, whose tracks never coexist with the early developers': anchored on an invest's position at the forecast's start, each cluster's genesis implies a propagation speed, and clusters whose implied speeds agree (within the physical easterly-wave band, along the same bearing, with anti-correlated membership) are genesis events on one phase line — one wave — while a cluster implying a clearly different speed is a distinct system even when it sits on the same corridor. Linked clusters are drawn with violet marker rings and a translucent violet ribbon that follows the family's deduplicated net mean track — the wave's overall corridor, which no single cluster's mean can show in a start/stop/re-develop scenario (the ribbon deliberately carries no intensity information, because a per-hour family average blends members at very different lifecycle stages) — and their combined probability counts each ensemble member once — which is why the per-cluster percentages must never simply be added: some members produce a genesis event in more than one cluster (develop, decay, and re-develop downstream), and the family union removes that double counting.
The clustering and linking are fully automated and their thresholds were calibrated against known same-wave and known-distinct cluster pairs; like any automated method they can occasionally split one system or group two. Questions and suggestions are welcome via the contact section on the home page.