taziksh/itrdb-tree-rings

International Tree-Ring Data Bank (ITRDB), tidy Parquet

The International Tree-Ring Data Bank is NOAA NCEI's public archive of tree-ring data contributed by dendrochronologists since the 1970s: about 7,000 sites on six continents, some series going back thousands of years. Researchers use it to reconstruct past temperature and drought.

NCEI publishes the archive as roughly 27,000 fixed-width "Tucson format" text files. The format has many quirks, so this dataset parses those files into long, typed Parquet tables. It also joins site metadata from the NCEI study API and keeps a per-file parse report.

Glossary.

• A series is one measured core or radius of one tree, for example LMO01A.
• A .rwl file holds a site's raw measurements: one value per series per year.
• A chronology (.crn file) is a site-level mean of detrended series. Detrending is called standardization. A chronology's values are a dimensionless ring-width index (RWI) that averages about 1.0. Sample depth is the number of series behind each year's value.
• Chronology types:
  • standard: the detrended mean.
  • residual: the same with autocorrelation removed.
  • ARSTAN: autocorrelation modelled and added back.

Configs

config	one row is	rows
ring_widths (default)	one measurement of one series in one year: total, earlywood or latewood width, in mm	61,943,261
densities	one wood-density measurement of one series in one year, in g/cm³	8,979,752
chronologies	one chronology value in one year, with sample depth	4,840,762
sites	one NCEI study × site: location, species, investigators, DOI, citation	6,983
files	one in-scope source file, with its parse status, warnings and header lines	26,901

from datasets import load_dataset
rw = load_dataset("taziksh/itrdb-tree-rings", "ring_widths", split="train")
sites = load_dataset("taziksh/itrdb-tree-rings", "sites", split="train")

-- duckdb: mean total ring width per year for one site, plus its coordinates
SELECT r.year, avg(r.value) AS mean_mm, any_value(s.latitude) AS lat, any_value(s.longitude) AS lon
FROM 'hf://datasets/taziksh/itrdb-tree-rings/data/ring_widths/*.parquet' r
JOIN 'hf://datasets/taziksh/itrdb-tree-rings/data/sites/*.parquet' s USING (study_id)
WHERE r.site_code = 'alge001' AND r.variable = 'total ring width'
GROUP BY r.year ORDER BY r.year;

ring_widths and densities

column	meaning
study_id	NCEI study ID. Joins to sites.study_id; landing page at https://www.ncei.noaa.gov/access/paleo-search/study/<id>. Null for 22 files no study claims
site_code	ITRDB site code taken from the file name, e.g. alge001
source_file	path under https://www.ncei.noaa.gov/pub/data/paleo/treering/, e.g. measurements/africa/alge001.rwl
variable	total ring width, earlywood width or latewood width. In densities: maximum density, minimum density, earlywood density, latewood density, total ring density or density
series_id	series ID as written in the file. In 6 files a repeated ID with overlapping years gets an X suffix, so no measurement is lost (dplR convention)
series_id_raw	the ID before that renaming
year	calendar year as written. Negative means BCE, using the file's own numbering; the files are not consistent about a year 0
value	value_raw × precision: mm for widths, g/cm³ for densities
value_raw	the integer stored in the file
precision	unit of value_raw: 0.01 or 0.001 (0.0001 for some density files)
precision_source	stop_marker (widths) or magnitude (densities). See the quirks section
stop_marker	the end-of-series marker found: 999 (0.01 mm units), -9999 (0.001 mm units), or null

Rows are unique on (source_file, series_id, year). Each source file's rows are contiguous, in file order (series by series, years ascending). Files are ordered by (first_seen, source_file); see "How it is kept up to date".

chronologies

column	meaning
study_id, site_code, source_file	as above
variable	what the chronology was built from: total ring width, earlywood width, maximum density, percent latewood, …
chronology_type	standard, residual or ARSTAN, from the NCEI API or else the file suffix (none / r / a)
chronology_index	0-based number of the chronology within the file (exact duplicate sections removed)
chronology_id	ID in columns 1–6 of the data lines
year, index	year and index value (index_raw / 1000)
index_raw	integer as stored (index × 1000)
sample_depth	number of series contributing that year (null when blank)

Rows are unique on (source_file, chronology_index, year), and use the same file ordering.

sites

One row per (study, site) that has at least one parsed file. Columns:

• IDs and names: study_id, site_id, study_code (e.g. ALGE001), study_name, site_name, location (an NCEI hierarchy such as Continent>Africa>Northern Africa>Algeria).
• Location:
  • latitude, longitude: the centre of the site's bounding box.
  • The raw bounding box: northernmost_latitude, southernmost_latitude, easternmost_longitude, westernmost_longitude, plus geometry_type (POINT or POLYGON).
  • Elevation: elevation_min_m, elevation_max_m.
• Years covered: earliest_year_ce, most_recent_year_ce.
• Species: species_codes (ITRDB 4-letter codes such as PISY), species_scientific_names, species_common_names.
• People and dates: investigators, contribution_date, study_version.
• Citation: dataset_doi, landing_page, dataset_citation (NCEI's "Cite as" text; fill in the [date] you accessed the data), publication_citations, publication_dois.
• data_license, data_license_url: set for only 38 sites (24 CC BY 4.0, 14 CC0).
• source_files: the parsed files belonging to this site.

files

One row for every in-scope file in the NCEI listing, including skipped and failed files. Columns:

• status: ok, skipped or failed, with error giving the reason.
• Labels: variable, chronology_type, label_source (api or file-name suffix) and study_link (api_file_list, or study_code for the fallback join).
• size_bytes and source_mtime from the NCEI directory index.
• first_seen (the build date the file first appeared) and shard (which Parquet shard of its table holds its rows). Together with size and mtime, this table is the sync state.
• n_series, n_rows, first_year, last_year.
• header: the Tucson header lines as written, which carry site name, species, elevation and coordinates in the contributor's own words.
• events: a list of parser warnings, as {event, series, count}. These are what this card's quirk counts come from. Example: SELECT * FROM files, unnest(events) t(e) WHERE e.event = 'COLUMN_LAYOUT'.

Source, snapshot, version

• Source: NOAA NCEI World Data Service for Paleoclimatology, https://www.ncei.noaa.gov/pub/data/paleo/treering/.
  • Data files come from measurements/<region>/ and chronologies/<region>/. The regions are africa, asia, atlantic, australia, centralamerica, europe, northamerica/{canada,mexico,usa} and southamerica.
  • Site metadata comes from the NCEI paleo study search API (https://www.ncei.noaa.gov/access/paleo-search/study/search.json?dataPublisher=NOAA&dataTypeId=18).
• Snapshot: directory listing retrieved 2026-10-28 19:49 UTC (12:49 PT); study metadata the same day; parsed 2026-10-29. Build time and per-run changes are in data/build_summary.json.
• Version: the ITRDB has no release versions any more. updates/version.txt was last updated in 2022 (v7.22). Files are added and edited continuously, so treat the snapshot date as the version.
• Mirror: Zenodo record 14816488 ("Virtually all of the ITRDB", CC0) is an unofficial rescue copy from 5 Feb 2025. The uploader says they are not affiliated with NOAA. It was not used here: NCEI is online, more recent, and authoritative.

Reconciliation with the source listing

	.rwl	.crn
files in the NCEI region folders (excluding *noaa* template files)	10,411	16,490
parsed OK	10,071	14,837
skipped by design (see below)	339	1,647
failed to parse	1	6

Every row of the data tables traces back to an ok file, and the per-file row counts in files sum to the table sizes.

Failed files (7):

• chronologies/asia/russ053w{,a,r}.crn: 0 bytes on the NCEI server. The same data is inside russ053w_crns.crn.
• chronologies/northamerica/usa/az558{,a,r}.crn: written in .rwl layout (6-character values, no sample depth), so the scale cannot be determined.
• measurements/northamerica/usa/ut528.rwl: header only, no data lines.

Skipped by design:

• 1,647 *_crns.crn files. Each is a Schweingruber-collection composite of the same site's standard, residual and ARSTAN chronologies, which exist as separate files, plus a "RAW" mean that can be derived from the measurements.
• 339 .rwl files measuring something other than width or density:
  • percent latewood: 184
  • blue intensity and delta blue intensity: 97
  • tracheid or cell anatomy: 35
  • basal area increment: 4
  • unlabelled: 19

What is not included

These parts of /paleo/treering/ are excluded:

• NOAA template files (*-noaa.rwl, *-rwl-noaa.txt, *-crn-noaa.txt). They repeat the Tucson data in tab-separated form; their metadata comes in through the study API instead.
• measurements/correlation-stats/: COFECHA crossdating reports.
• Non-standard folders:
  • isotope/
  • reconstructions/
  • the per-paper folders (zhao2018/, miramont2025/, vothrurup2026/, breitenmoser2014/, …)
  • signalfree/
  • updates/ (older Schweingruber/WSL copies, which overlap the region folders)
  • the 2017 itrdb-v713-*-crn.zip snapshots

Licence and citation

NCEI attaches no explicit licence to most ITRDB studies:

• Access constraints are "None".
• Only 38 of the 6,983 sites carry a licence in the API: 24 CC BY 4.0 and 14 CC0. See the data_license column.
• The data are contributed by individual investigators, so they are not simply a US-government work.
• NCEI's ISO metadata adds a no-warranty, no-liability disclaimer.

What NCEI asks for (verbatim):

Please cite original publication, online resource, dataset and publication DOIs (where available), and date accessed when using downloaded data. If there is no publication information, please cite investigator, title, online resource, and date accessed.

The ITRDB readme files add: "PLEASE CITE ORIGINAL CONTRIBUTOR WHEN USING THIS DATA!"

For each site you use, cite sites.dataset_citation (NCEI's "Cite as" format, e.g. Touchan, R. (2010-10-20): NOAA/WDS Paleoclimatology - Touchan - Leid Mohamad Ouali - CDAT - ITRDB ALGE001. NOAA National Centers for Environmental Information. https://doi.org/10.25921/sah9-fk68. Accessed [date].) and the entries in publication_citations. For the archive as a whole, cite the International Tree-Ring Data Bank, NOAA NCEI World Data Service for Paleoclimatology, with the snapshot date above.

Parsing and known quirks

The .rwl reader is a Python port of the decisions in dplR's read.tucson(). That reader was rewritten in 2026, after a sweep of the full ITRDB, to report rather than hide problems. The .crn reader follows dplR's read.crn() and adds layout checks. Every guess is recorded in files.events.

Validation. 150 random .rwl and 150 random .crn files were checked against NCEI's own tab-separated template versions:

• .rwl: 1,085,611 shared values, 0 disagreements.
• .crn: 93,557 shared values. The only disagreements (2 files) come from templates that do not scale by 1000.

The only systematic difference is negative codes; see "Negative values" below.

Units and stop markers. Each .rwl series ends with a stop marker:

• 999 means the values are in units of 0.01 mm.
• -9999 means units of 0.001 mm.

Series:

• 199,891 series use 999.
• 125,185 series use -9999.
• 2 series have no marker and are assumed to be 0.01 mm (event NO_STOP_MARKER).

Density units are inferred, not read. Density files store integers in 0.01, 0.001 or 0.0001 g/cm³, and their stop marker does not track this. For example, most Schweingruber earlywood and latewood density files end in 999 but hold values like 671, which means 0.671 g/cm³.

• The scale is therefore chosen per file, as the one that puts the file's median nearest a typical density for that variable (precision_source = magnitude, event DENSITY_SCALE_INFERRED).
• In 659 files this disagrees with the marker.
• In 11 files it is still ambiguous (DENSITY_SCALE_AMBIGUOUS).
• value_raw is always kept, so you can rescale.

Negative values. 228,435 raw values in 2,933 files are negative codes other than the terminator: -999, -888, -8, -7 and others. Following dplR (2026), these are treated as missing and produce no row (event NEGATIVE_AS_MISSING). NCEI's template files and older readers record them as 0, i.e. as a locally absent ring. Zero values that are actually written in the file (217,255 of them) are kept as 0.

Column-layout problems (.rwl):

• 17 files contain lines where the fixed 6-character columns and a whitespace split disagree. Those lines are dropped (COLUMN_LAYOUT).
• Tab characters are expanded to 8-column stops.
• Over-wide IDs that push the year out of columns 9–12 are recovered (YEAR_MISPLACED).
• Exact duplicate lines are removed (DUPLICATE_LINE).

Chronology files:

• 3,354 .crn files contain their whole chronology twice, header and all. This looks like an NCEI reformatting artefact; the duplicate sections are removed (DUPLICATE_SECTION).
• 32 files use a non-standard ID width: 5, 8 or 3 characters instead of 6 (LAYOUT_ID5, LAYOUT_ID8, LAYOUT_ID3). They are parsed only if every line then gives consecutive decades.
• 93 chronologies have a mean far from 1000, the usual ×1000 index scale (INDEX_SCALE_UNUSUAL). Some of these hold raw mean widths (finl001, the Kauri newz1xx files) or density anomalies centred on 0 (co517x). They are kept, but check before you average across sites.
• **** overflow cells are treated as missing (OVERFLOW_VALUE).

Impossible years. Some years are later than the build year. 35 chronology files (5,854 rows) have them, for example swed022t (header 1127–1987, data to 2049) and the floating chronology cana166 (years 6255+). The cause is either floating, relatively dated chronologies or errors in the source; NCEI's own template files show the same years. They are kept and flagged (FUTURE_YEAR). Filter with year <= 2026 if you need calendar years only.

Duplicates across files. The ITRDB contains re-contributions and site collections. 12,335 series appear with identical values in two different .rwl files (306 files involved, e.g. ak095/ak096, cana314/cana322). Deduplicate on (series_id, year, value_raw) if you pool sites.

Labels. Variable and chronology type come from the NCEI API's per-file variable list where it exists (24,545 of 26,901 files). Otherwise they come from the file-name suffix:

suffix	meaning
none / w	total width
e	earlywood width
l	latewood width
x	maximum density
n	minimum density
i	earlywood density
t	latewood density
p	percent latewood

On chronologies, a trailing r means residual and a means ARSTAN.

The suffix is not reliable on its own: w sometimes means tracheid diameter and b blue intensity. That is why the API label wins when both exist.

Study links.

• 24,356 files are linked to a study through the API's file list.
• 530 more are linked by matching site_code to the study code.
• 22 files have no study.
• 44 files are listed by two studies; the first one is used.

Other:

• Some outlier widths look like data-entry errors in the source (11,395 values above 20 mm; the maximum is 272 mm). They are kept as written.
• Header metadata inside the Tucson files sometimes disagrees with the data, for example in the declared year range. The sites table uses the NCEI API, not the file headers.

How it is kept up to date

The ITRDB is not static:

• 629 studies were contributed in 2025 and 104 so far in 2026.
• Region folders were modified as recently as 2026-10-14.
• Existing files are edited too: e.g. brit.rwl was modified 2026-08-19.

So the dataset is designed for a monthly sync.

build.py is a self-contained uv script:

uv run build.py                 # full build: crawl, download every file, studies, parse
uv run build.py --update        # incremental (see below)
uv run build.py --update --push # same, then upload data/ + README.md + build.py (needs HF_TOKEN)

• Full build: about 30–60 min. Most of that is downloading about 27k small files with at most 4 connections. Parsing takes under a minute on 8 cores.

• Update (no raw files needed):

  1. If data/ is missing, as on a fresh CI runner, it downloads the previous data/ from the Hub (~250 MB). data/files is the sync state: size, mtime, first_seen and shard for every source file.
  2. It re-crawls the NCEI index and refreshes study metadata from the API.
  3. It downloads and parses only files that are new, whose size or mtime changed, or whose labels or study link changed.
  4. It rewrites only the shards that contain one of those files or a deleted file. All other shards stay byte-identical.

• Storage-friendly layout:

  • New files get first_seen = build date and are appended to the last (small) shard of their table. A file keeps its shard and position for life, so an edit changes one contiguous block.
  • Row groups are cut only at source-file boundaries (≥ 250k rows each), and pages use Parquet content-defined chunking (256 KiB–1 MiB). Unchanged data therefore produces identical bytes, which the Hub's Xet storage deduplicates.

• Measured growth: HF keeps every uploaded version, and Xet stores only chunks it has not seen. I simulated a typical month: 10 newly contributed studies (10 files, 39k rows), 3 files with corrected values and 2 with an added series. Running tools/dedup_estimate.py on every rewritten file gave these numbers of new bytes per month:

  layout	new MB / month
  previous design (alphabetical full rebuild, shards cut by row count)	83–85 (with or without CDC)
  stable order, file-aligned 1M-row groups, no CDC	8.9
  + CDC 256 KiB–1 MiB	7.3 (64–256 KiB pages: 7.0, but files 2% larger)
  + stable sites/files order, no access date in citations	6.9
  + 250k-row groups (current)	4.7

  That is about 0.06 GB a year. About 1.4 MB of each month is the small sites and files tables, which change throughout. Each edited ring-width file costs about 0.25 MB, so a month with heavy edits scales with that number.

• CI: .github/workflows/sync.yml runs --update --push on the 3rd of each month. It needs no cache.