TorNet

TorNet is a benchmark dataset of storm-centered NEXRAD WSR-88D dual-polarisation radar samples for tornado detection, released by MIT Lincoln Laboratory (Veillette et al., 2024). This is the complete v1.1 dataset, all ten years and both partitions, repackaged as Parquet so it loads with datasets by year and partition without downloading or extracting per-year archives. Nothing is dropped or altered: every field of every netCDF sample is here, the radar values in their physical units.

• 203,133 samples, 2013 to 2022: 171,666 train, 31,467 test (the authors' partitions).
• Four scans per sample, about five minutes apart, the newest last.
• Six products per scan on a polar grid of 120 azimuths × 240 range gates × 2 tilts (0.5° and 0.9°): reflectivity DBZ (dBZ), radial velocity VEL (m/s), specific differential phase KDP (°/km), correlation coefficient RHOHV, differential reflectivity ZDR (dB), spectrum width WIDTH (m/s). Empty gates are NaN.
• Three categories: TOR (confirmed tornado, with EF rating), WRN (tornado-warned storm with no tornado), NUL (random non-tornadic storm).
• Per-scan labels (frame_labels): 1 if a tornado is present in that scan, so a sample's four labels read like 0011 for a tornado that appears in the third scan.

Layout

data/train/<year>/part-NNNNN-of-MMMMM.parquet    256 samples per shard, zstd
data/test/<year>/part-NNNNN-of-MMMMM.parquet
catalog.parquet                                  one row per sample, every scalar field

Loading

from datasets import load_dataset

ds = load_dataset("jakeryderv/tornet", "2015")            # one year, train and test splits
ds = load_dataset("jakeryderv/tornet")                     # all ten years
ds = load_dataset("jakeryderv/tornet", data_files={"train": ["data/train/2013/*.parquet",
                                                             "data/train/2015/*.parquet"],
                                                   "test": "data/test/2014/*.parquet"})  # a fold
catalog = load_dataset("jakeryderv/tornet", "catalog", split="train").to_pandas()  # 1 s, no arrays

sample = ds["train"].with_format("numpy")[0]
sample["VEL"].shape      # (4, 120, 240, 2): scan, azimuth, range, tilt
sample["frame_labels"]   # array([0, 0, 1, 1])

Radar arrays are stored as float16 (the netCDF values are exactly representable; the round trip is lossless) and come back as float32 under with_format("numpy"). The range-folded mask is uint8. Selecting a year or partition downloads only those shards.

Fields

field	type	meaning
DBZ, VEL, KDP, RHOHV, ZDR, WIDTH	float16 (4, 120, 240, 2)	the products, physical units, NaN where empty
range_folded_mask	uint8 (4, 120, 240, 2)	1 where VEL / WIDTH are range-folded
filename	string	the netCDF name in the original release, e.g. train/2015/TOR_150425_...nc
partition, year	string, int16	the authors' split and the sample's year
category	string	TOR, WRN or NUL
ef_number	int8	EF rating for TOR, −1 otherwise
event_id, episode_id	int64	NWS storm event and episode ids
site	string	NEXRAD site, e.g. KTLX
lat, lon	float32	storm location
label	int8	the newest scan's label
frame_labels	int8[4]	tornado present in each scan
time	int64[4]	each scan's time, Unix seconds
az_lower, az_upper	float32	azimuth limits of the chip, degrees
rng_lower, rng_upper	float32	range limits of the chip, metres
tornado_start_time, tornado_end_time	int64	the tornado's interval, Unix seconds (v1.1); 0 when none

The CoordConv coordinates used by the published detector are derived from the azimuth and range limits; they are not stored.

Source and license

Original release: Veillette, M. S., et al., TorNet: A benchmark dataset for tornado detection and prediction using full-resolution polarimetric weather radar data, Artificial Intelligence for the Earth Systems, 2024, https://doi.org/10.1175/AIES-D-24-0006.1. Data on Zenodo, https://doi.org/10.5281/zenodo.12636522 (per-year records), code at https://github.com/mit-ll/tornet, pretrained detector at https://huggingface.co/tornet-ml/tornado_detector_baseline_v1.

The dataset is distributed under CC BY 4.0, as released. This repackaging changes the container format only; cite the paper and the Zenodo record when you use it.

@article{veillette2024tornet,
  title   = {TorNet: A Benchmark Dataset for Tornado Detection and Prediction Using Full-Resolution Polarimetric Weather Radar Data},
  author  = {Veillette, Mark S. and Kurdzo, James M. and Stepanian, Phillip M. and McDonald, Joseph and Samsi, Siddharth and Cho, John Y. N.},
  journal = {Artificial Intelligence for the Earth Systems},
  year    = {2024},
  doi     = {10.1175/AIES-D-24-0006.1}
}