IBTrACS v04r01: all global tropical cyclone tracks (ALL file)

This is a Parquet mirror of the complete IBTrACS v04r01 ALL file from NOAA NCEI. IBTrACS is the International Best Track Archive for Climate Stewardship. It contains every tropical cyclone known to IBTrACS, from 1842 to the present, in every basin, with the best tracks from all reporting agencies merged into one record.

A best track is an agency's post-season, quality-controlled estimate of a storm's position and intensity over time. Agencies make these estimates independently. IBTrACS matches them to each other but leaves the agency values as they were reported.

The mirror has two configs:

config	one row =	rows
tracks (default)	one storm position at one time (a "track point"), with the merged position, the WMO values and every per-agency column	730,348
storms	one storm (IBTrACS SID), with identifiers, genesis and lysis, basins, agencies, and lifetime max wind and min pressure per agency, as reported	13,631

The only other copy on the Hub is yuiseki/ibtracs-v04r01, a frozen subset covering storms since 1980. This mirror carries the full ALL file and is re-synced from NCEI on a schedule.

Snapshot

Source file	IBTrACS.ALL.v04r01.nc
Source Last-Modified	Thu, 29 Oct 2026 09:17:13 GMT
Source sha256	fa5a3c79cdf054792ff86e1eab76f0c9eefec4270fc82a8a0ccb3f46e6d321fc
Retrieved	2026-10-29T16:34:30+00:00
Storms (SIDs)	13,631 (PROVISIONAL: 77, US-PROVISIONAL: 92, US-PROVISIONAL_spur: 1, main: 13,170, spur-merge: 89, spur-other: 69, spur-split: 133)
Track points	730,348
Seasons	1842–2027
Time span	1842-10-25 03:00:00+00:00 – 2026-10-29 00:00:00+00:00

(build.py regenerates the block above on every build.)

Quick start

from datasets import load_dataset
tracks = load_dataset("taziksh/ibtracs", "tracks", split="train")
storms = load_dataset("taziksh/ibtracs", "storms", split="train")

-- duckdb: North Atlantic storms with lifetime max wind >= 137 kt (Cat 5, US 1-min winds)
SELECT sid, name, season, usa_wind_max_kt, wmo_pres_min_mb
FROM 'hf://datasets/taziksh/ibtracs/data/storms.parquet'
WHERE genesis_basin = 'NA' AND NOT is_spur AND usa_wind_max_kt >= 137
ORDER BY season;

# Join a storm's track; keep only times where some agency actually reported a position
import polars as pl
t = pl.read_parquet("hf://datasets/taziksh/ibtracs/data/tracks.parquet")
haiyan = t.filter((pl.col("name") == "HAIYAN") & (pl.col("season") == 2013) & ~pl.col("position_interpolated"))

Source

• Product page: https://www.ncei.noaa.gov/products/international-best-track-archive
• File: IBTrACS.ALL.v04r01.nc from https://www.ncei.noaa.gov/data/international-best-track-archive-for-climate-stewardship-ibtracs/v04r01/access/netcdf/
• Docs: Technical Details (PDF), Column documentation (PDF), Change log, Q&A forum
• Version: v04r01. The snapshot date is shown above.

Why the ALL file. NCEI also publishes basin files (NA, EP, WP, NI, SI, SP, SA) and period files (since1980, last3years, ACTIVE). Each of these is a subset of ALL, and they overlap: a storm that crosses basins appears in more than one basin file. Using ALL alone avoids duplicates and anything missing.

Why NetCDF instead of the CSV. The same ALL data is published as IBTrACS.ALL.v04r01.nc (about 23 MB) and ibtracs.ALL.list.v04r01.csv (about 330 MB). The NetCDF has four advantages:

• It is about 14 times smaller to download on every sync.
• It is typed, and it declares fill values, units and scale factors.
• It has storm-level fields the CSV does not: numobs, main_track_sid and the 15 per-agency source_* provenance strings.
• It contains exactly the same track points as the CSV.

build.py verify checks that last point on every build: it compares all 127,080,552 track cells (730,348 rows × 174 CSV columns) against the official CSV and found 0 mismatches. The CSV also has two parsing traps. Its second line is a units row, and BASIN = "NA" (North Atlantic) becomes null in pandas and similar readers.

Configs and columns

tracks: one row per storm per time

The columns follow the official CSV in the same order, lower-cased, plus a few convenience columns. Each storm's NetCDF quadrant arrays (*_r34, *_r50, *_r64, usa_searad) are flattened into the CSV's _ne, _se, _sw and _nw columns, and verify confirmed the quadrant order. Units and NetCDF long_name are stored as Arrow field metadata (units, description).

group	columns	notes
keys	sid, point_idx, valid_time	point_idx is the 0-based position within the storm. valid_time is iso_time parsed as a UTC timestamp. (sid, valid_time) is unique.
storm info (repeated)	season, number, name, track_type	Same on every row of a storm, as in the CSV.
merged / IBTrACS-derived	basin, subbasin, iso_time, nature, lat, lon, dist2land, landfall, iflag, storm_speed, storm_dir	lat and lon are the mean of the available agency positions.
WMO	wmo_wind (kts), wmo_pres (mb), wmo_agency	Values from the agency officially responsible at that location. Not interpolated.
USA (NHC/CPHC/JTWC)	usa_* (agency, ATCF id, lat/lon, record, status, wind, pres, SSHS, R34/R50/R64 ×4 quadrants, POCI, ROCI, RMW, eye, gust, sea height/radii)	1-min winds
other agencies	tokyo_* (JMA), cma_*, hko_*, kma_*, newdelhi_* (IMD), reunion_* (Météo-France), bom_* (Australia), nadi_* (Fiji), wellington_* (NZ)	Each column is that agency's own value.
historical datasets	ds824_*, td9636_*, td9635_*, neumann_*, mlc_*	Pre-satellite compilations.
convenience (added)	wmo_wind_avg_minutes, position_interpolated, geometry	See below.

Units are used exactly as the source gives them:

quantity	unit
wind, gust, storm speed	knots (kts)
pressure	mb (= hPa)
radii (R34/R50/R64, RMW, ROCI, eye)	nautical miles (nmile)
lat / lon	degrees north / east
dist2land, landfall	km
sea height	ft
Dvorak CI / T numbers	dimensionless

The added columns:

• wmo_wind_avg_minutes: the averaging period implied by wmo_agency. It is 1 for the US agencies (hurdat_atl, hurdat_epa, atcf, cphc), 3 for newdelhi and 10 for everyone else. It is null when there is no WMO value.
• position_interpolated: true when no source reported a position at that time, meaning every iflag character is P or _. 366,544 of the 730,348 points (about 50%) are IBTrACS 3-hourly interpolations.
• geometry: a WKB Point (lon wrapped to [-180, 180), lat), WGS84. The file carries GeoParquet 1.1 geo metadata, so geopandas.read_parquet works.

storms: one row per SID

column(s)	meaning
sid, season, number, name	IBTrACS identifiers. name is the raw value: UNNAMED is kept, and combined names such as A:B occur in 106 storms.
storm_id	lower(<name>_<genesis_basin>_<season>), matching OCHA-DAP's ocha-lens. Null when UNNAMED.
track_type, main_track_sid, provisional, is_spur	See quirks below. provisional is true for any track type containing PROVISIONAL.
n_points	Number of track points (the NetCDF numobs). The sum over storms is 730,348.
genesis_time, lysis_time, duration_hours, genesis_lat/lon, lysis_lat/lon	Times and positions of the first and last track points. These can include disturbance and extratropical stages.
genesis_basin, basins, subbasins, natures	Basin at the first point, then the distinct values along the track in order.
atcf_id, usa_atcf_ids	The first ATCF id (ocha-lens atcf_id) and all ATCF ids. A storm that crosses basins gets several.
wmo_agencies, usa_agencies, agencies	Responsible WMO agencies, US source datasets, and all 15 source datasets with any data (taken from iflag).
wmo_wind_max_kt, wmo_pres_min_mb, wmo_wind_max_agency, wmo_wind_max_avg_minutes	Lifetime WMO max wind and min pressure, and which agency and averaging period produced the max.
<agency>_wind_max_kt, <agency>_pres_min_mb	Lifetime max and min as reported by each agency for usa, tokyo, cma, hko, kma, newdelhi, reunion, bom, nadi, wellington, ds824, td9636, td9635, neumann and mlc. Nothing is converted.
usa_sshs_max	Max Saffir–Simpson category, using the codes below.
min_dist2land_km, has_landfall	has_landfall means any landfall == 0.
source_usa … source_mlc	Per-agency provenance strings from the NetCDF, for example bal102026.dat or BOTH_tcvitals.txt:Line=1755.

How this lines up with ocha-lens. The storms columns include ocha-lens's get_storms() fields: sid, atcf_id, number, season, name, genesis_basin, provisional and storm_id. There are two deliberate differences:

• name keeps UNNAMED rather than null.
• provisional also covers US-PROVISIONAL. ocha-lens only tests == "PROVISIONAL", which misses the 93 US-PROVISIONAL storms in this snapshot.

ocha-lens's get_tracks() picks one "provider" value per point and drops interpolated points. This mirror keeps every point and every agency column instead, so you can reproduce that view:

• Filter on ~position_interpolated.
• Take wmo_* for best tracks, or usa_* for provisional tracks.

Known quirks and caveats

1. Wind speeds are not comparable across agencies. The US agencies report 1-minute mean winds, CMA 2-minute, IMD 3-minute, and JMA, BoM, Météo-France Réunion, Fiji, NZ, HKO and KMA 10-minute. IBTrACS applies no conversion. A rule of thumb is that 1-min ≈ 1.12 × 10-min, but procedures differ by agency and era (Knapp & Kruk 2010; Harper et al. 2008, WMO TD-1555).
   • Example: Super Typhoon Haiyan (2013) has tokyo_wind_max_kt = 125 (JMA, 10-min), usa_wind_max_kt = 170 (JTWC, 1-min), hko_wind_max_kt = 155 and cma_wind_max_kt = 151.
   • wmo_wind changes agency, and so averaging period, when a storm crosses an area-of-responsibility boundary, notably 180° between the US agencies and JMA. 375 storms have more than one WMO agency.
   • Use wmo_wind_avg_minutes or a single agency's column. Pressure is much more comparable across agencies.
2. WMO values are sparse and not interpolated. wmo_wind is filled on 170,316 of the 730,348 points, and older storms often have none.
3. 3-hourly interpolation. Most agencies report every 6 hours. IBTrACS adds 3-hourly points:
   • Positions are interpolated with splines and other values linearly, flagged per source in iflag (O original, P position interpolated, I intensity interpolated, V some other variable interpolated, _ no data).
   • The 15 iflag characters correspond, in order, to: usa, tokyo, cma, hko, kma, newdelhi, reunion, bom, nadi, wellington, ds824, td9636, td9635, neumann, mlc.
   • ACE and PDI are conventionally summed over 6-hourly points, so sums over all 3-hourly rows need adjusting (roughly halving them, or using only the 00/06/12/18 UTC rows).
4. Provisional data. track_type values are main, spur-*, PROVISIONAL and US-PROVISIONAL (plus _spur variants).
   • PROVISIONAL is real-time "working best track" data, mostly from US sources, that has not been reanalysed.
   • US-PROVISIONAL, new in v04r01, means the US columns and the IBTrACS-merged fields (lat/lon, nature, landfall, storm_speed, …) are provisional while other agencies' columns are final best tracks.
   • Provisional storms in this snapshot run from late 2024 to now. Expect positions, intensities and even storm existence to change for up to about 2 years. A row's values can change between syncs.
5. Spurs. These are alternate or uncertain tracks attached to a main track, with main_track_sid pointing at it. Exclude spurs when counting storms (NOT is_spur, 292 storms). main_track_sid is not always self-referential: one main storm points to a different SID.
6. Storm counts are inhomogeneous over time because observing systems changed. Satellite coverage started in the 1960s, and NCEI calls "since 1980" the modern geostationary era. Basin practices for depressions and subtropical storms also vary, and storms were missed in early years. See the Technical Details §5.3 and §6.9.
7. Longitude can exceed 180. The raw lon stays continuous across the antimeridian, ranging from −179.8 to 266.9 (10,658 points above 180). Some agency longitudes also exceed ±180, for example ds824_lon reaches −211.4. Use geometry or ((lon + 180) % 360) - 180 for mapping.
8. Season. season is the year the storm began, except that Southern Hemisphere storms starting after 1 July count toward the next year. This is why the snapshot contains season 2027.
9. Out-of-range and odd values are kept exactly as distributed (the CSV has them too). netCDF4's default masking would silently drop the ones outside the declared valid_min/valid_max.
   • reunion_gust = 999 on 264 rows is a missing-value sentinel.
   • bom_gust_per equals bom_gust on all 15,989 non-null rows. The documentation says it should be a gust period in seconds (3 s for Australian data), so this looks like an upstream bug. Do not use it as a period.
   • usa_sshs codes: −5 unknown, −4 post-tropical, −3 misc. disturbance, −2 subtropical, −1 tropical depression, 0 tropical storm, 1–5 hurricane category.
   • reunion_tnum, reunion_ci, reunion_r64_* and td9636_pres are entirely null. newdelhi_ci is always 0. usa_seahgt is always 12 (ft).
   • usa_agency contains both tcvitals and tcvightals (33 points). Both are separate labels in IBTrACS's USA source priority list.
10. Gaps by agency. Some agencies submit best tracks with a lag. For example, ocha-lens notes that RSMC La Réunion best tracks were missing for 2022 onward until recently; the March 2026 change log adds the 22–23 to 24–25 seasons. Check the change log for each agency's coverage.
11. Coordinates were cleaned of float32 noise. The NetCDF stores lat/lon as float32 on a 0.1° grid, and they are widened to float64 rounded to 4 decimal places, so 10.9 stays 10.9. Dvorak CI and T numbers have scale_factor = 0.1 applied.
12. The NetCDF date_created attribute is stale (it read 2026-10-06 in a file published 2026-10-27). The sync relies on HTTP Last-Modified/ETag and the file sha256 instead, all recorded in manifest.json.
13. Agency name fields and nature are combinations of the sources (MX = agencies disagree on the storm type).

How it's kept up to date

NCEI republishes the entire ALL file about three times a week, usually Sunday, Tuesday and Thursday. There is no incremental feed, so build.py handles a sync like this:

1. uv run build.py update --push first runs HEAD on the NetCDF and compares ETag, Last-Modified and Content-Length with the manifest.json stored in this repo. If they match, it stops.
2. Otherwise it downloads the file. If the file's sha256 is unchanged, it stops.
3. Otherwise it rebuilds both tables and hashes their content (Arrow IPC, so the hash does not depend on Parquet writer details). If the content hash matches the manifest, nothing is pushed.
4. Otherwise it uploads data/*.parquet, README.md (with the snapshot block refreshed), manifest.json and build.py in a single commit through huggingface_hub.

A scheduled GitHub Actions workflow (.github/workflows/sync-ibtracs.yml) runs this on Mondays, Wednesdays and Fridays at 12:00 UTC, the morning after each NCEI update. It needs an HF_TOKEN secret, and the repo must already exist because the script never creates repos.

Other modes:

• uv run build.py full does a fresh download and build.
• uv run build.py verify reconciles the output against the official CSV cell by cell (it downloads about 330 MB).

Storage growth

HF keeps every version of every file. Its Xet backend stores content-defined chunks of about 64 KiB once each, so a sync only adds the bytes that differ from the previous upload. The Parquet layout is designed to keep that small:

• Rows are in SID order. A SID starts with the genesis year and day, so new storms and their points are appended at the end of both tables.
• Row groups are keyed on the SID year: one per decade before 2000 and one per year from 2000, 43 in each file. The boundaries depend on content rather than row counts. A change to one year, such as provisional revisions, an agency's annual best-track release or a historical reanalysis, therefore rewrites only that year's row group.
• Pages use Parquet content-defined chunking: 16–64 KiB for tracks, pyarrow's defaults for storms.
• Column statistics are written only for the usual filter columns (sid, valid_time, season, basin, lat, lon, wmo_wind, …). This keeps the footer, which changes on every sync, small.

New bytes per sync, estimated with a Xet-style chunker:

scenario	tracks	storms
real NCEI update, 2026-10-27 → 2026-10-29	0.4 MB	0.2 MB
simulated: last 14 days of points missing, recent provisional values revised	0.4 MB	0.2 MB
simulated: one agency adds a season's best track (JMA 2025)	0.3 MB	0.2 MB
simulated: historical reanalysis (NA 1971–75, 25% of points removed)	3.0 MB	0.4 MB

The original layout (128k-row groups, no CDC) cost 2.6 + 1.0 MB per sync in the 14-day simulation. Fixed 8k-row groups matched this layout for recent changes, but cost 15 MB for a reanalysis because every later row group shifts. At 3 syncs a week this layout adds about 0.6 MB per sync, roughly 0.1 GB a year, on top of the initial ~25 MB.

Reconciliation (this snapshot)

• Storms: 13,631 = NetCDF storm dimension = distinct SIDs in the official CSV.
• Track points: 730,348 = Σ numobs = CSV data rows (excluding the header and units rows). (sid, valid_time) is unique.
• All 174 CSV columns × 730,348 rows match the Parquet: 0 mismatches after normalising blanks to null. Row order also matches the CSV.
• Spot checks: Katrina 2005 has wmo_wind_max_kt 150 and wmo_pres_min_mb 902 (HURDAT2). Patricia 2015 has 185 kt and 872 mb. Haiyan 2013 has JMA 125 kt and 895 mb, and JTWC 170 kt.

License and terms

The NCEI terms say: "IBTrACS was developed to support scientific research efforts. The IBTrACS data usage policy follows the World Data Center for Meteorology (WDC), which provides full and open access to the data. We have agreements from the Regional Specialized Meteorological Centers that provide track data to make IBTrACS data open for distribution to contribute to global tropical cyclone research. We also use the World Meteorological Organization's Resolution 40 policy as the guide for commercial use of IBTrACS data." The NetCDF license attribute reads: "These data may be redistributed and used without restriction."

This mirror changes only the format, as described above. The added convenience columns are derived from source values.

Citation

NCEI asks that you cite both the BAMS paper and the dataset in written work that has a bibliography:

Knapp, K. R., M. C. Kruk, D. H. Levinson, H. J. Diamond, and C. J. Neumann, 2010: The International Best Track Archive for Climate Stewardship (IBTrACS): Unifying tropical cyclone best track data. Bulletin of the American Meteorological Society, 91, 363-376. doi:10.1175/2009BAMS2755.1

Gahtan, J., K. R. Knapp, C. J. Schreck, H. J. Diamond, J. P. Kossin, M. C. Kruk, 2024: International Best Track Archive for Climate Stewardship (IBTrACS) Project, Version 4r01. [indicate subset used]. NOAA National Centers for Environmental Information. doi:10.25921/82ty-9e16 [access date].

For this mirror, the subset is "ALL" and the access date is the Retrieved date in the Snapshot table.

Without a bibliography, use: "NOAA's International Best Track Archive for Climate Stewardship (IBTrACS) data, accessed on [Access Date]".

@article{knapp2010ibtracs,
  author  = {Knapp, Kenneth R. and Kruk, Michael C. and Levinson, David H. and Diamond, Howard J. and Neumann, Charles J.},
  title   = {The International Best Track Archive for Climate Stewardship ({IBTrACS}): Unifying Tropical Cyclone Best Track Data},
  journal = {Bulletin of the American Meteorological Society},
  volume  = {91}, number = {3}, pages = {363--376}, year = {2010},
  doi     = {10.1175/2009BAMS2755.1}
}
@misc{gahtan2024ibtracs,
  author    = {Gahtan, J. and Knapp, K. R. and Schreck, C. J. and Diamond, H. J. and Kossin, J. P. and Kruk, M. C.},
  title     = {International Best Track Archive for Climate Stewardship ({IBTrACS}) Project, Version 4r01 (ALL)},
  publisher = {NOAA National Centers for Environmental Information},
  year      = {2024},
  doi       = {10.25921/82ty-9e16}
}

Further reading: Knapp & Kruk (2010), Quantifying interagency differences in tropical cyclone best-track wind speed estimates, MWR 138, doi:10.1175/2009MWR3123.1. Schreck, Knapp & Kossin (2014), The impact of best track discrepancies on global tropical cyclone climatologies using IBTrACS, MWR 142, doi:10.1175/MWR-D-14-00021.1.