Missing Intense Tropical Cyclones

The question

A lifetime-maximum intensity is a judgment, and judgments have eras

Outside the aircraft-reconnaissance Atlantic, nearly every tropical cyclone intensity since the 1970s is a satellite estimate: an analyst matching an infrared picture to the Dvorak technique's patterns and tables, later assisted by the objective ADT. The technique is good — but it was run on coarser imagery, with fewer satellites, by agencies with different habits, and its known failure is the one that matters most for records: the small, violent storm, whose pinhole eye is one or two pixels wide and reads cold, and whose peak is brief enough to fall between images. Landsea, Harper, Hoarau & Knaff (2006) put it bluntly — Hoarau's Dvorak reanalyses suggested at least 70 unrecognized Category 4–5 cyclones in the Eastern-Hemisphere basins in 1978–1990 alone — and every trend in intense-storm counts inherits the problem.

This page does the reanalysis objectively and at scale: every storm in the HURSAT satellite archive, 1978–2016, re-estimated from the imagery by an algorithm that is not the ADT, then argued with every independent line of evidence the site has. The output is a list of case files — storms the book probably holds too weak, graded by how much independent evidence agrees — and a corrected count of intense storms by decade.

A second satellite estimator

Trained on aircraft truth, scored against the ADT on the same images

The count, corrected

How many Category 4–5 storms each decade actually had

Where the book and the satellite part ways

The case files

The storms, one at a time

Each file lays out every voice on one axis — the book (black), our estimate with its 10–90% interval (blue), the re-mapped ADT, the 1984 objective Dvorak, the other agencies, ERA5, the median of the storm's seven aircraft-measured satellite twins, and any published reanalysis (red) — then the imagery at the satellite peak, the intensity history, and the environmental case. STRONG: both satellite algorithms at least 15 kt above the book, our interval clear of it, and a non-imagery line (another agency, ERA5, the twins or a published reanalysis) agreeing. PROBABLE: both algorithms at least 12 kt above. POSSIBLE: one algorithm at least 20 kt above.

Every candidate

The full list, sortable

Checking ourselves

Against the published reanalyses

The other direction

Storms the book holds too high

Method

What each line of evidence is, and is not

Imagery. HURSAT-B1 v06 (Knapp & Kossin 2007): ISCCP B1 geostationary data re-gridded to 0.07° (~8 km), 3-hourly, centred on every IBTrACS storm 1978–2016 — 3,946 storms, 78 GB, streamed storm by storm and reduced to per-image measurements (pipeline/hursat_b1.py). Matched to the merged best track by time and position, because IBTrACS identifiers changed between the v03 the archive was built on and v04 (Tip is 1979275N05159 in one and 1979275N06159 in the other). The best view per 3-hourly slot is kept (lowest view angle, <65°).

Measurements. A digital Dvorak reading of each image: the warmest pixel within 40 km of the objective centre (the eye temperature), the Dvorak "surrounding ring" (the warmest pixel on each ring 24–136 km out, then the coldest of those rings), their contrast, an eye radius and a pinhole index; cold-cloud fractions at −50 to −80 °C within 100, 200 and 300 km; azimuthal symmetry; the principal components of the azimuthal-mean IR profile (Kossin et al. 2007); ARCHER's objective centre, eye probability and radii and the 1984 objective Dvorak number that ride in every v06 file; each also 6, 12 and 24 h earlier.

Model. Gradient-boosted trees, trained only on images within 3 h of an aircraft centre fix in NHC's f-decks, Atlantic and eastern Pacific 1989–2016; five-fold cross-validation grouped by storm. A storm's satellite LMI is the maximum 9-hour running mean over open water, mapped to recon-era best-track LMIs by quantiles with a linear upper tail (and a hard ceiling of 190 kt). The ADT-HURSAT v01 record is re-mapped the same way: the storm ledger's map stops at its top quantile (~152 kt), which cannot call a Category 5 above that.

Anchoring. Every satellite voice is anchored to its basin's modern best track (2005–2016): the median satellite-minus-book lifetime maximum of that basin's hurricanes is removed before any storm is judged, so a basin effect (an Atlantic-trained model reading western Pacific storms high; the ADT reading the North Indian low) cannot masquerade as a book error. The audit therefore measures each basin's older book against the standard its modern book keeps.

Aircraft. Where an aircraft measured the storm the book is the measurement: no storm is called missed if a recon fix lies within 3 h of its satellite peak, and none overcalled if one lies within 6 h of its book peak (before the 1989 f-deck archive, Atlantic Category 4–5 peaks and western Pacific typhoons to August 1987 are treated as flown). An overcall also needs both satellite records to cover the book's peak.

Environment. SST at the satellite's peak (the book's peak when the satellite did not see the storm) and along the track (ERSSTv5), days over ≥28.5 °C water, the ONI, and a logistic plausibility model fitted to 2000–2024 — the probability a storm with this duration, track and environment becomes a Category 4–5 in the modern record — plus the share of modern storms peaking within 500 km and 30 days of year that did. It uses no intensity measurement of the storm it scores.

References