jakeryderv/us-tornado-data-2010-2025

US Tornado Data, 2010–2025

v2.2.1 combines 20,164 recorded U.S. tornadoes, linked damage footprints and county context with radar-derived indicators, NWS warning histories and land-cover summaries. It provides 17 linked source/supporting Parquet tables under analysis/ and two event-level modeling tables under ml/.

This is an independent research compilation for recorded EF-rating analysis, not an official government product or a complete record of every tornado. All original tornado rows and unknown EF ratings are retained. The nine v2.1.0 backbone tables remain unchanged; this release adds radar, warnings, NLCD and reproducible feature views. ERA5 and ACS/TIGER tract enrichment are deferred.

Start with one modeling table

ml/events_onset.parquet has 20,164 rows and 30 columns (about 1.05 MB). ml/events_retrospective.parquet has 20,164 rows and 53 columns (about 2.82 MB). Download either table and ml/feature_dictionary.json without fetching the full source collection. Python needs pandas and pyarrow; geometry tables additionally use geopandas.

Hugging Face

Install huggingface_hub, pandas and pyarrow, then:

from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    "jakeryderv/us-tornado-data-2010-2025",
    repo_type="dataset", revision="v2.2.1",
    allow_patterns=["ml/*"],
))

Use allow_patterns=["analysis/*", "ml/*"] for all 19 tables. For strict reproducibility, replace the tag with the immutable HF commit recorded in the GitHub release receipt.

Select features explicitly

The following works after either download example:

import json
import pandas as pd

spec = json.loads((root / "ml/feature_dictionary.json").read_text())
events = pd.read_parquet(root / "ml/events_onset.parquet")
labelled = events.loc[events.target_known & events.target_ef_rating.notna()]
X = labelled[spec["onset_predictor_columns"]]
y = labelled[spec["target"]]
groups = labelled[spec["group"]]

Use grouped/temporal evaluation and fit imputers/scalers only on training data. Do not use every numeric column as a predictor: targets, metadata and retrospective fields have distinct roles. No fixed train/test split or fitted model is supplied. See ml/feature_dictionary.json and ENRICHMENT.md for the feature contract.

Tables and sources

All filenames below are Parquet. Tables retain their different units of observation. tornado_id links event-level tables; bridge tables connect source-specific IDs.

File under analysis/	Source	Rows	Contents
tornadoes.parquet	SPC + linkage summaries	20,164	Stable tornado ID, final EF label, timing, endpoints, dimensions, casualties and county/link summaries
storm_events.parquet	NCEI Storm Events	23,189	Tornado county segments, narratives, damage, casualties, timing and locations
storm_fatalities.parquet	NCEI Storm Events	1,352	Related fatality records linked by NCEI event ID
storm_locations.parquet	NCEI Storm Events	35,920	Additional reported event locations
tornado_footprints.parquet	NOAA Event Footprint Catalog	24,858	DAT/Storm Events-derived damage regions and associated track attributes
county_context.parquet	Census Population Estimates Program	50,294	Annual county population and housing estimates
county_boundaries.parquet	Census cartographic boundaries	3,234	Generalized 2020 county polygons and identifiers
source_crosswalk.parquet	Derived time/geometry matching	95,222	Accepted, ambiguous and unmatched source links with decision evidence
tornado_counties.parquet	NCEI links + Census	20,842	Tornado-to-county associations and county/year context
radar_detections.parquet	NCEI SWDI / NEXRAD Level III-derived products	1,805,706	TVS, mesocyclone and storm-structure detections, locations/times and available radar indicators
tornado_radar.parquet	Derived proximity/time matching	2,104,931	Tornado–detection links, distances and assumed availability times
warning_updates.parquet	NWS warnings archived by Iowa Environmental Mesonet	846,391	TO/SV warning polygons, issuance times, native fields and resolved VTEC/text fields where linked
tornado_warnings.parquet	Derived warning association	109,133	Tornado–warning links, start-point coverage and issue lead times
event_areas.parquet	Footprint Catalog + SPC	20,164	Accepted footprint union or documented buffered-track sampling geometry
nlcd_samples.parquet	USGS Annual NLCD	40,294	Two prior-year products per covered event: land-cover pixel counts and impervious percentage
source_coverage.parquet	Collection bookkeeping	60,492	Per-event/source status, reason, job ID and extraction definition
record_provenance.parquet	Collection bookkeeping	4,610,010	Record-to-source-asset references

analysis/annual_summary.csv contains 16 yearly backbone/coverage summaries. The backbone inventory is analysis/manifest.json; the eight added tables are inventoried by enrichment/manifest.json, which points to ../analysis.

File under ml/	Rows	Contents
events_onset.parquet	20,164	EF target, normalized onset/location, pre-cutoff radar aggregates, warning counts/lead time and missingness
events_retrospective.parquet	20,164	Onset fields plus final track/impact data, radar through onset +60 minutes and land-cover summaries

Coverage and timing limitations

All 60,492 requested event/source jobs were processed: radar and warnings completed for 20,164 events each; NLCD completed for 20,147, with 17 outside selected CONUS coverage explicitly marked unavailable. There are no failed jobs. Completion does not guarantee a detection, warning, usable raster pixels or a complete historical observing record. Inspect source statuses and feature missingness.

• EF is a post-event damage rating, not a direct measurement of peak tornado wind. Exposure, construction, survey practice and missing damage indicators affect the recorded label. Keep narratives, other EF fields and survey-derived wind estimates out of onset predictors.
• Reported onset is a hindsight anchor. The onset view uses final-catalog time/location. Radar availability is observation time plus an assumed five-minute latency, not verified real-time arrival. This supports conditional onset analysis, not an operational forecasting claim.
• Radar is a nearby association. Queries cover 20 km around the reported start and ±60 minutes; the onset view filters by its availability cutoff. Nearby detections are not confirmed parent-storm identities. Missing detections do not establish an absence of radar coverage. These are derived records, not Level II scans.
• Warnings have archive limits. Relevant TO/SV histories are linked when at least one polygon covers the start point. Background query rows remain in the source table. The IEM polygon export omits standalone cancellation rows, so it is not a complete VTEC timeline and some active-warning counts can be overstated.
• Footprints are damage-derived. DAT/Storm Events regions may be nested or reconstructed. The crosswalk retains uncertainty; it is not independently verified tornado identity. Where no accepted footprint is available, land cover uses a fixed 500 m buffer around the reported endpoint track/point.
• NLCD summaries are retrospective. Prior-year, native 30 m land-cover and impervious grids are sampled within the final event area. No-data and areas containing no pixel centers stay missing. No national raster grids are bundled.
• County context is coarse. Population/housing values describe linked whole counties, not people/buildings struck. Estimates use vintages 2020 and 2025; the fixed 2020 map is not a historical county-boundary series.
• 2010 is a study-scope choice, not the EF transition or a completeness threshold. Rare EF classes, spatial/reporting biases and related records require care.

Coverage/missingness reports in research/ are split by year, state/territory and EF class. Unknown EF affects 1,525 records; only eight are EF5. Land-cover fractions are missing for 317, including 299 areas with no pixel centers. Rural/low-exposure under-rating, changing reporting practices, radar sampling, warning archive gaps and shared source lineage limit interpretation. County totals are not exposure struck, and the saved NLCD WCS metadata does not resolve its collection revision. See RESEARCH_READINESS.md for intended/inappropriate uses, measured coverage, split guidance and the limits on conclusions.

The onset view supports conditional recorded-tornado analysis. It cannot establish live availability or support safety decisions, true-wind estimates, property-level risk assessment or causal claims. Dictionary schema 2 uses eligible_for_conditional_onset; available_by_onset is null for candidates because actual delivery is unverified. post_nlcd_valid_fraction is quality metadata, not a default retrospective predictor. Table values and schemas are unchanged from v2.2.0.

Reproducibility and files

Both hosts share the same release manifest and checksummed payload. Kaggle also provides the complete payload in release.zip.bin; extract it with Python zipfile. The extra suffix preserves upstream compressed source files. HF stores the payload files directly. Core analysis/modeling does not require the full archive.

Original backbone source files remain included. The compact additions retain normalized source records, geometry, extraction settings, URLs, timestamps and hashes. Required warning texts and small schema documents are bundled in enrichment/supporting_assets.zip to avoid tens of thousands of tiny downloads. Working caches, checkpoints and benchmarks are excluded.

Offline ML regeneration uses the retained linked tables. Full raw re-extraction requires upstream downloads; checksums cannot recover subsequently removed upstream files. To run the standalone enrichment verifier on a full downloaded payload, first expand the support bundle with the repository's python -m scripts.release_data restore-support PATH command. Ordinary release verification checks the bundle checksum without requiring extraction.

Use release_manifest.json, SHA256SUMS, schema.json, ANALYSIS.md, ENRICHMENT.md, LINKAGE.md, and ml/feature_dictionary.json. The GitHub release receipt records the shared v2.2.1 label, immutable HF revision and Kaggle numeric version. This is a fixed snapshot with no scheduled refresh; later changes receive new versions, and historical releases remain available.

Sources, reuse and citation

• NOAA/NWS SPC tornado database.
• NOAA NCEI Storm Events.
• NOAA Event Footprint Catalog, incorporating DAT and Storm Events.
• Census Population Estimates Program.
• Census 2020 cartographic boundaries.
• NOAA NCEI SWDI, NEXRAD-derived products.
• Iowa Environmental Mesonet warning archive, archiving NWS warning products.
• USGS Annual NLCD.

The data designation is U.S. Government Works, subject to source-specific notices and third-party exceptions described in DATA_SOURCES.md. Annual NLCD is marked CC0 by USGS, and IEM materials are public domain with attribution requested. MIT covers project code and original documentation, not ownership of government records. Credit NWS/NOAA, Census, USGS and IEM's archive service as applicable. The compilation and transformations are not endorsed by these organizations.

Cite Jake Van Slyke, US Tornado Data, 2010–2025, v2.2.1, the relevant host revision and access date, together with the original sources. CITATION.cff contains machine-readable citation information, release date, data URL and code revision. source_registry.json gives each product’s recommended citation, terms, version/coverage, actual acquisition dates and mapping into derived tables. No DOI is assigned. Code, methodology and release receipts are maintained in the GitHub repository.

Release v2.2.1; 250 shared files; 1,335,942,053 bytes.