Resnet3d 18 Tornet

ResNet3D-18 for Tornado Detection

A 3D convolutional neural network trained on temporal dual-polarimetric NEXRAD radar sequences for tornado detection and prediction.

Model Description

This model uses a ResNet3D-18 backbone (3D adaptation of ResNet-18) with a dual-head architecture:

• Detection head: Classifies whether a tornado is currently occurring in the radar sequence
• Prediction head: Classifies whether a tornado will occur, using only pre-tornado frames

The model processes 8 consecutive radar volume scans (~40 minutes of data) with 24 dual-polarimetric channels across a 128x128 km storm-centered grid.

Architecture Details

Parameter	Value
Backbone	ResNet3D-18 (BasicBlock, layers=[2,2,2,2])
Parameters	33.3M
Input shape	(B, 24, 8, 128, 128) -- channels, time, height, width
Output shape	(B, 4) -- [det_class0, det_class1, pred_class0, pred_class1]
Channels	24 (6 dual-pol products x 4 elevation angles)
Temporal frames	8
Spatial resolution	1 km/pixel, 128x128 grid

Channel Layout

Channels	Product	Description
0-3	REF	Reflectivity at 0.5, 0.9, 1.3, 1.8 deg
4-7	VEL	Radial velocity
8-11	SW	Spectrum width
12-15	ZDR	Differential reflectivity
16-19	CC	Correlation coefficient
20-23	KDP	Specific differential phase

Performance

Validation Set (2022, 2,117 events)

Head	AUC	CSI	F1
Detection	0.927	0.652	0.789
Prediction	0.993	0.883	0.938
Combined	0.960	--	--

Test Set (3,685 events)

Head	AUC	CSI	F1	Precision	Recall
Detection	0.896	0.540	0.702	0.602	0.841
Prediction	0.988	0.856	0.922	0.958	0.889
Combined	0.942	--	--	--	--

Per-category prediction accuracy: TOR 88.9%, WRN 96.8%, NUL 99.1%

Comparison with Literature

Model	Year	Frames	Channels	AUC
TorNet (MIT Lincoln Lab)	2024	1	12	~0.88
This model	2026	8	24	0.942

Training Details

Parameter	Value
Dataset	tornet-temporal (24,862 events)
Train split	2013-2021 (~19,061 events)
Optimizer	AdamW (lr=1e-3, weight_decay=0.01)
Scheduler	Cosine annealing with 3-epoch linear warmup
Batch size	256
Epochs	20
Mixed precision	FP16 (AMP) with FP32 classification heads
GPU	NVIDIA H100 NVL (96GB)
Training time	~4 hours

Usage

import torch
import numpy as np
from model_resnet3d import DualHeadResNet3D  # from this repo

# Load model
model = DualHeadResNet3D(in_channels=24, arch="resnet18")
state = torch.load("best.pt", map_location="cpu")
model.load_state_dict(state["model_state_dict"])
model.eval()

# Load a radar sequence
event = np.load("tornet_EVENTID_TOR/sequence.npz")
data = event["data"][:8]  # first 8 frames, shape (8, 24, 128, 128)
x = torch.from_numpy(data).float().unsqueeze(0)  # (1, 8, 24, 128, 128)
x = x.permute(0, 2, 1, 3, 4)  # (1, 24, 8, 128, 128) -- channels first

with torch.no_grad():
    out = model(x)  # (1, 4)
    det_prob = torch.softmax(out[:, :2], dim=1)[:, 1]  # detection probability
    pred_prob = torch.softmax(out[:, 2:], dim=1)[:, 1]  # prediction probability

print(f"Detection probability: {det_prob.item():.3f}")
print(f"Prediction probability: {pred_prob.item():.3f}")

Real-World Deployment

How to use this in an operational setting

1. Data ingestion: Ingest real-time NEXRAD Level-II data from the Unidata IDD or AWS NEXRAD archive. Extract the 6 dual-pol products (REF, VEL, SW, ZDR, CC, KDP) at 4 elevation angles (0.5, 0.9, 1.3, 1.8 deg).

2. Storm tracking: Use an existing storm tracker (e.g., TINT, SCIT, or a simple reflectivity centroid tracker) to identify storm cells and extract 128x128 km storm-centered patches.

3. Temporal buffering: Maintain a rolling buffer of the last 8 radar volume scans (~40 minutes) for each tracked storm cell. New scans arrive every ~4-5 minutes.

4. Inference: Run the model on each storm cell's 8-frame sequence. The prediction head output is most useful operationally -- it answers "will this storm produce a tornado?"

5. Thresholding: Use a probability threshold of ~0.66 (optimized for CSI on the test set) to generate tornado warnings. At this threshold:

   • Precision 95.8%: When the model warns, it's almost always right
   • Recall 88.9%: Catches ~89% of actual tornadoes
   • Lead time: The prediction head uses pre-tornado frames, providing inherent lead time

6. Integration: Feed predictions into NWS warning decision support systems or automated alerting pipelines. The model runs in <50ms on a modern GPU, well within real-time requirements.

Limitations

• Trained on CONUS NEXRAD data only (2013-2022). May not generalize to other radar networks or non-US storm environments.
• Requires all 24 dual-pol channels. Single-pol radars are not supported (see 12-channel ablation for degraded single-pol performance).
• Storm-centered input assumes a working storm tracker upstream.
• The detection head (AUC 0.896) is weaker than the prediction head (AUC 0.988). Active tornado signatures may be more variable than pre-tornado mesocyclone patterns.
• NUL (null) training samples are drawn from tornado-day radar scans only. The model has not been tested on truly quiescent weather.

Citation

@model{resnet3d-18-tornet,
  title={ResNet3D-18 for Temporal Radar Tornado Detection},
  author={DeepGuess},
  year={2026},
  url={https://huggingface.co/deepguess/resnet3d-18-tornet},
}