Storm Comparison
Add Storms to Compare
Search and filter storms, or use + Compare from Storm Detail.
Basin:
–
Sort:
—
—
best track
IR eye fix
GHOST — recon-independent intensity reanalysis
GHOST estimates tropical cyclone intensity — minimum central pressure and maximum sustained wind — from geostationary infrared imagery and the large-scale environment alone. No aircraft reconnaissance, and no best-track intensity, is used as input.
What goes in
- Merged global geostationary infrared imagery (NASA GES DISC MergIR, ~4 km, half-hourly, doi:10.5067/P4HZB9N27EKU; NOAA NCEI GridSat-B1 before 1998) — the same imagery the frame viewer on this page displays. Data credits.
- Large-scale environmental fields from reanalysis (thermodynamic and kinematic).
- Storm position from IBTrACS — position only, never intensity.
Because the estimate never sees reconnaissance or best-track intensity, it can be compared against them independently. That is what the GHOST − BT view is for.
How it is built (data version v5.3, the paper model)
- Pmin is the equal mean of two independent members read from the same imagery: a small first-principles ridge on the environmental pressure deficit (FPM), and a boosted-tree stack with an inner-core tier (the tree member). Where the tree abstains on a frame its last value is carried for six hours; where it never existed the ridge stands alone.
- Vmax is the tree member’s direct wind head with the Cat-5 corrections. In the Atlantic, above 130 kt it is blended with the wind implied by the published pressure, so wind and pressure stay consistent (v5.3). In the East and Central Pacific, once the infrared read has fallen well below its peak, the inner-core corrections may no longer raise the wind (post-peak decay gate, v5.2).
- On frames rejected by quality control the wind is filled and drawn dashed (v5.1); filled values are excluded from every published metric and from the storm peak.
- Every published value is held out. Atlantic and East Pacific members were fitted leave-one-year-out; West and Central Pacific are out-of-basin transfer. There is no separate frozen channel — the number shown is the held-out one.
- The same model runs live on the Real-Time Monitor’s Experimental tab (there, the Cat-5 wind corrections also run in the East and Central Pacific). Earlier data versions remain selectable below: v1–v2 (a single FPM channel per basin), v4–v5 (the two-member blend), v5.1 (wind fill), v5.2 (East and Central Pacific decay gate).
How it is built (data version v6)
- Vmax is the equal mean of two members read from the same imagery, each smoothed over the preceding four hours: the boosted-tree model’s wind (the tree member), and a vision transformer — a neural network fine-tuned on single infrared images, trained on aircraft-verified intensities where reconnaissance exists and on the best track elsewhere. Above 130 kt in the Atlantic, East and Central Pacific it is blended with the wind implied by the published pressure, so wind and pressure stay consistent.
- Pmin in the Atlantic, East and Central Pacific is half the first-principles ridge on the environmental pressure deficit (FPM) and a quarter each from the tree member and the vision transformer; in the West Pacific it is the equal mean of the ridge and the tree member. For the strongest storms, a wind–pressure guard deepens the pressure where it would otherwise read far weaker than the wind, and no frame is published above the environmental pressure minus 1 hPa.
- Only frames whose center is over water get the full wind estimate. Over land, or where the image is unusable, the wind is filled and drawn dashed, and the storm’s peak wind never rests on a filled frame (except on a storm with no unfilled frame, where the peak is marked as filled). The pressure there is still an estimate: the ridge, with the tree member’s last value fading over about six hours. A few East and Central Pacific frames the new model does not cover (all of Erick 2019 among them) keep the published v5.3 values and are marked as filled.
- Every published value is held out: each frame comes from models that never saw its year. Atlantic, East and Central Pacific members were fitted leave-one-year-out (the 1998–99 Atlantic seasons were never in training at all); the West Pacific is an out-of-basin transfer. There is no separate frozen channel — the number shown is the held-out one.
- The Real-Time Monitor’s Experimental tab still publishes the earlier model. Its GHOST v6 (ViT, experimental) overlay previews the new wind in real time; its pressure follows the earlier recipe until the real-time model is updated. Earlier data versions remain selectable below: v1–v2 (a single FPM channel per basin), v4–v5 (the two-member blend), v5.1 (wind fill), v5.2 (East and Central Pacific decay gate), v5.3 (Atlantic wind–pressure consistency above 130 kt).
How it is built (data version v6.1)
- Vmax blends two members read from the same imagery, each smoothed over the preceding four hours: the boosted-tree model’s wind (the tree member), and a vision transformer — a neural network fine-tuned on single infrared images, trained on aircraft-verified intensities where reconnaissance exists and on the best track elsewhere. Below 50 kt the two count equally; from 50 to 80 kt the vision transformer’s share rises to three quarters. Above 130 kt, in every basin, the wind is blended with the wind implied by the published pressure, so wind and pressure stay consistent.
- The tree member reads the vision transformer’s own estimates as two of its inputs (one image network), and its environment uses observed sea-surface temperature (NOAA OISST) and a thermodynamic potential intensity rather than climatological ocean fields.
- Pmin, in every basin, is half the first-principles ridge on the environmental pressure deficit (FPM) and a quarter each from the tree member and the vision transformer. For the strongest storms, a wind–pressure guard deepens the pressure where it would otherwise read far weaker than the wind, and no frame is published above the environmental pressure minus 1 hPa.
- Since v6.1.1, frames whose center is over land are scored like any other, and their wind decays from the last over-water estimate, faster the more of the storm’s core (100 km) is over land, so a storm scraping a coast weakens slowly and one moving fully inland at the full rate. The decay was fitted on 672 landfalls (1998–2025). Every error statistic (each storm’s agreement with best track, the list’s QC flag and the headline numbers) and each storm’s peaks use over-water times only.
- Only frames whose center is over water get the full wind estimate. Over land, or where the image is unusable, the wind is filled and drawn dashed, and the storm’s peak wind never rests on a filled frame (except on a storm with no unfilled frame, where the peak is marked as filled). The pressure there is still an estimate: the ridge, with the tree member’s last value fading over about six hours. The few frames the new model does not cover keep the published v6 values and are marked as filled.
- Every published value is held out: each frame comes from models that never saw its year. Atlantic, East and Central Pacific members were fitted leave-one-year-out (the 1998–99 Atlantic seasons were never in training at all); the West Pacific is an out-of-basin transfer from models trained on the other basins. There is no separate frozen channel — the number shown is the held-out one.
- The Real-Time Monitor still publishes the earlier model; it has not yet been updated to v6.1. Earlier data versions remain selectable below: v1–v2 (a single FPM channel per basin), v4–v5 (the two-member blend), v5.1 (wind fill), v5.2 (East and Central Pacific decay gate), v5.3 (Atlantic wind–pressure consistency above 130 kt), v6 (equal-mean wind with the vision transformer; West Pacific pressure from the ridge and the tree member).
How to read it
- Outside the recon stratum, best-track pressure is itself largely Dvorak-derived — so a disagreement there is model-vs-model, not error. Use Recon-proximate only for the decisive comparison.
- OOB out-of-basin: the model is trained on Atlantic / East-Pacific reconnaissance labels, so West and Central Pacific storms are an out-of-basin application — zero frames from those basins were in training. That is a stricter condition than the held-out channels offered for the Atlantic and East Pacific, not a weaker one: nothing about the basin, in any season, was fitted. It is also why no held-out column is published there — there is nothing left to hold out.
- QC flagged for large trace-vs-best-track disagreement; these sink to the bottom of the default ordering.
- On the frame, the blue cross is the best-track position the scene is cut about; the red cross is the IR eye fix — where the pipeline’s centre-finder placed the eye before computing the frame’s predictors (accepted only within 50 km of the track; otherwise the best-track centre was used). On the lifecycle chart, hatched spans are stretches the best track calls extratropical or subtropical, and the recon strip marks frames with an aircraft pressure fix within ±3 h (ticks: individual aircraft fixes from the f-deck).
- ? the storm’s deepest frame is also its largest disagreement with the best track, with no reconnaissance nearby — a disagreement is not by itself an error, but that minimum has no independent support. 17 of the 724 storms in v4.
- Known cost of the mean at the extreme tail. Where the tree member saturates on the very deepest storms the mean halves the ridge’s deeper read: Patricia 2015 (best track 872 hPa) is 893 hPa in v4 (ridge 886, tree 900). Below 920 hPa the Atlantic estimates run about +6 hPa weak on average; pooled over everything the Atlantic RMSE is 6.6 hPa (slope 0.92), East Pacific 8.0, West Pacific 10.8 against Dvorak-era JTWC labels.
Manuscript in preparation. GHOST (Geostationary-based Hurricane Objective Strength Technique) is an unpublished research method; this page is provided for scientific transparency and is not an official forecast or analysis. Please contact Dr. Michael Fischer ([email protected]) before citing or redistributing these estimates.