dwefer/radar-dino-fieldtoken-v1

Radar-DINO field-token v1

This artifact contains the inference-only teacher backbone and reference analysis data for the field-token Radar-DINO model trained on 8,416 KHTX GridNC scans.

Expected input

Each NetCDF scan must contain a 301 by 301 (or larger) 1 km horizontal grid and the following fields at 2,000 m:

1. reflectivity
2. specific_differential_phase (KDP and kdp are accepted aliases)
3. differential_reflectivity
4. cross_correlation_ratio
5. spectrum_width

The package center-crops to 300 by 300 grid cells, applies the fixed normalization stored in manifest.json, and preserves missing values with the training fill value.

Installation

python -m pip install \
  "radar-dino[netcdf,hub,analysis,plot] @ git+https://github.com/DanielWefer/radar_dino.git@main"

Usage

from radar_dino import RadarDINO

dino = RadarDINO.from_pretrained(
    "dwefer/radar-dino-fieldtoken-v1",
    device="auto",
)
result = dino.analyze("/path/to/KHTX_scan.nc")
pngs = dino.save_plots(result, "/path/to/output")

result.feature                 # normalized 384-D vector
result.attention               # heads x fields x 300 x 300
result.umap                    # 2-D UMAP transform
result.tsne                    # approximate display-only t-SNE position
result.cluster                 # fitted PCA/HDBSCAN cluster, -1 means noise
result.cluster_probability
result.neighbors               # five closest reference scans by cosine similarity
pngs                           # attention, UMAP, and t-SNE PNG files

The five attention PNGs show each normalized radar field beside its mean-head attention. UMAP and t-SNE PNGs show the complete reference population colored by HDBSCAN cluster, with the input scan marked by a red star.

Reference analysis

The artifact includes normalized 384-D reference features, sanitized scan metadata, a fitted 30-component PCA model, an HDBSCAN model fit on the first 20 principal components, and a cosine UMAP model fit on the original normalized features. The fixed reference t-SNE is included for visualization.

Scikit-learn t-SNE has no out-of-sample transform. A new scan's result.tsne coordinate is therefore a nearest-neighbor interpolation for display only. Clusters are assigned with the fitted PCA/HDBSCAN pipeline, not in UMAP or t-SNE space.

Limitations

This model was trained on KHTX scans and has not been established as a meteorological classifier or severe-weather decision product. Cluster IDs are unsupervised groups, and label -1 denotes HDBSCAN noise rather than a physical radar category. Similarity can reflect coverage, missingness, and data quality as well as meteorological structure.

Provenance and integrity

• Training source commit: 0f8ffd8942691287e392442bcd33ba134297cdb0
• Final epoch-99 checkpoint SHA-256: 64d21ce2ef344fa5a133d225d997a9abfc563c7dcebb099534aa66cc2a51f9b6
• Exported model.safetensors SHA-256: c62cf017d98d84fef63a096ac4236a69a102c2675e2129de033c7f66566ceca0

The model weights use safetensors. The fitted PCA, HDBSCAN, and UMAP estimators use joblib serialization and must only be loaded from a trusted model repository.