TorNet-Temporal: Temporal Dual-Pol NEXRAD Radar for Tornado Detection

A large-scale dataset of storm-centered NEXRAD WSR-88D radar sequences for tornado detection and prediction, featuring 24-channel dual-polarimetric data across variable-length temporal sequences.

Dataset Summary

• 24,862 storm events from NEXRAD Level-II radar archives (2013-2022)
• 8-22 consecutive radar scans per event (~4-5 min cadence, ~45-90 min total; median 13 frames)
• 24 channels: 6 dual-pol radar products x 4 elevation angles
• 128x128 spatial grid at 1km resolution, storm-centered
• 3 categories: TOR (tornado), WRN (tornado-warned but no tornado), NUL (null/no severe)
• 130 unique NEXRAD radar sites across CONUS

Data Format

Each event is stored as a compressed NumPy archive (sequence.npz) containing:

Key	Shape	Type	Description
data	(T, 24, 128, 128)	float32	Radar volume sequence (T varies 8-22, median 13)
scan_times	(T,)	str	UTC timestamps per frame
center_time	scalar	str	Event center time
lat	scalar	float64	Storm center latitude
lon	scalar	float64	Storm center longitude
radar	scalar	str	NEXRAD site ICAO (e.g., KTLX)
mag	scalar	int64	EF-scale magnitude (-1 if no tornado)
label	scalar	int64	Binary label (1=tornado, 0=no tornado)
category	scalar	str	TOR, WRN, or NUL

Channel Layout (24 channels)

The 24 channels represent 6 dual-polarimetric radar products at 4 elevation angles (0.5, 0.9, 1.3, 1.8 degrees):

Channel Index	Product	Elevation
0-3	Reflectivity (REF)	0.5, 0.9, 1.3, 1.8
4-7	Velocity (VEL)	0.5, 0.9, 1.3, 1.8
8-11	Spectrum Width (SW)	0.5, 0.9, 1.3, 1.8
12-15	Differential Reflectivity (ZDR)	0.5, 0.9, 1.3, 1.8
16-19	Correlation Coefficient (CC)	0.5, 0.9, 1.3, 1.8
20-23	Specific Differential Phase (KDP)	0.5, 0.9, 1.3, 1.8

Category Definitions

• TOR: Storm produced a confirmed tornado (EF0+). Label=1.
• WRN: Storm had a tornado warning issued but no confirmed tornado. Label=0.
• NUL: No severe weather. Randomly sampled from tornado-day radar scans at non-tornadic locations. Label=0.

Recommended Splits

We recommend year-based splitting to prevent data leakage from correlated storm environments. Events from the same storm system can appear within hours of each other at the same radar site, so random splitting risks leaking information across splits.

Split	Years	Approx. Events	Purpose
Train	2013-2021	~21,800	Model training
Test	2022	~3,040	Final evaluation

For train/validation splitting, we recommend holding out one earlier year (e.g., 2021) as a validation set.

A catalog.csv is included with TorNet-compatible train/test assignments if you prefer to match TorNet's splitting methodology. Events not in the catalog can be assigned to either split.

Important: All data is from 2013 onwards (post dual-pol upgrade), so there are no legacy single-pol-only events. All 24 channels contain real data in every event.

Usage

import numpy as np
import torch
from pathlib import Path

# Load a single event
event = np.load("tornet_1000855_TOR/sequence.npz")
data = event["data"]          # (T, 24, 128, 128), T varies per event
label = int(event["label"])   # 1 for tornado
category = str(event["category"])  # "TOR"
print(f"Frames: {data.shape[0]}, Label: {label}, Category: {category}")

# Select 8 consecutive frames for model input (pad/truncate as needed)
T = data.shape[0]
if T >= 8:
    frames = data[:8]                         # take first 8
else:
    frames = np.pad(data, ((0, 8 - T), (0,0), (0,0), (0,0)))  # zero-pad

# Convert to PyTorch tensor: (C, T, H, W) for 3D CNN
tensor = torch.from_numpy(frames).float()  # (8, 24, 128, 128)
tensor = tensor.permute(1, 0, 2, 3)        # (24, 8, 128, 128)

Benchmark Results

Using dual-head architecture (detection + prediction heads) on this dataset:

Architecture	Params	Frames	Val Det AUC	Val Pred AUC	Combined AUC
ResNet3D-18	33M	8	0.919	0.991	0.955
ResNet3D-34	64M	8	0.916	0.985	0.947
Video Swin-T	28M	8	0.888	0.914	0.901
ResNet3D-18 (1-frame ablation)	33M	1	0.870	0.970	0.920

Temporal context (8 frames vs 1 frame) provides a significant boost, especially for detection (+0.049 AUC).

Data Quality Notes

• Variable time steps: Events contain 8-22 frames (not a fixed number). Your DataLoader should pad or truncate to a uniform length.
• KDP fill values: KDP channels (20-23) have a slightly elevated fill rate (~6-7%) compared to other channels (<1%). This is normal -- KDP is a derived product that fails to compute in some conditions.
• 26 corrupt files: A small number of NUL-category events (~0.1%) have corrupted NPZ files. These should be caught by a try/except in your data loader.
• 0.5-degree elevation gaps: ~3% of events have missing VEL/SW data at the lowest elevation (0.5 deg) due to SAILS/MESO radar scan strategies. The model should learn to handle these naturally.
• No dual-pol leakage: All data is post-2013 (after the NEXRAD dual-pol upgrade completed), so dual-pol channel availability cannot serve as a proxy for era or data quality.

Data Source

Raw radar data sourced from the Unidata NEXRAD Level-II Archive on AWS S3 (s3://unidata-nexrad-level2). Tornado reports from the Storm Prediction Center storm reports database.

Citation

If you use this dataset, please cite:

@dataset{tor-ten-v3.2,
  title={TorNet-Temporal: Temporal Dual-Pol NEXRAD Radar for Tornado Detection},
  author={DeepGuess},
  year={2026},
  url={https://huggingface.co/datasets/deepguess/tornet-temporal},
}

License

CC-BY-4.0. Raw radar data is in the public domain (NOAA/NWS). Processing pipeline and labels by DeepGuess.