Datasets:

huilinsigehigh
/

thunder_storm_aug353

Modalities:

Text

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 1 rows

strategy_summary dict	files list	total_files int64
{ "strategy": "S2 (frac >= 30 dBZ) Top-320 from thunder-all + all 73 storm-lite", "total_days": 353, "storm_days": 73, "picked_days": 320, "duplicates_removed": 40, "pooled_pixels": 23721099947, "mean": 3.2191830122595695, "std": 6.866652014996538, "p90": 11.99375, "p95": 19.53125, "p99": 31.11875...	[ { "source": "thunder-all", "rank_in_S2": 263, "name": "day_20210316.npy", "path": "D:\\datasets\\thunder-all\\2021\\day_20210316.npy", "frac_ge_30_score": 0.004859885114143632, "mean": 1.6266800165176392, "std": 4.908071041107178, "max": 59.625, "p95": 11.31875, "p99": 26.393...	351

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Thunder Storm Augmented (351 days)

A curated radar reflectivity dataset combining storm-focused days with distribution-similar augmentation days from the broader thunder-all archive.

Note: Originally planned as 353 days; 2 days were excluded due to upload issues (see below).

Composition

Total: 351 daily files (.npy)
Source breakdown:
- 73 days from the storm-only seed dataset (thunder_single_144f)
- 280 additional days selected from thunder_all (2021–2023) using strategy S2: rank by per-day fraction of pixels ≥ 30 dBZ, take Top-320, dedupe against seed.

Per-file Format

Identical to upstream thunder_all:

File: day_YYYYMMDD.npy
Shape: (144, 499, 840) — 144 timesteps × 499 lat × 840 lon
dtype: float32
Units: dBZ (radar reflectivity), observed range ≈ [0, 80]
Time resolution: 10 minutes (144 = 24 h × 6)

Pooled Statistics

Metric	storm-lite (73)	this dataset (353)	thunder_all full (652)
pixels	4.4 B	23.7 B	39.4 B
mean dBZ	3.54	3.22	2.40
std	7.16	6.87	5.64
p95	20.66	19.53	14.47
p99	31.91	31.12	27.52
%≥20 dBZ	5.37%	4.81%	2.73%
%≥30 dBZ	1.45%	1.26%	0.65%
%≥40 dBZ	0.152%	0.133%	0.069%
JS-div vs storm-lite	0	0.00050	0.00455

The augmented set preserves storm-lite's reflectivity distribution very closely (JS = 5e-4) while offering ~5× more days for training.

Selection Methodology

For each day in thunder_all we computed per-day features (mean, std, max, percentiles, threshold-fraction) on GPU. Strategy S2 ranks days by the fraction of pixels at or above 30 dBZ — a proxy for convective coverage. The Top-320 days were combined with the original 73 storm days; 40 overlapping days were deduplicated.

Other strategies were also evaluated (max-dBZ, p99, JS-divergence, Wasserstein, multi-feature L2). Results are summarized in the upstream thunder_all repository.

File Manifest

The full file list (chronological) is in manifest.json.

Excluded / Bad Files

Corrupt files (data damaged, permanently excluded):

2021/day_20210422.npy — truncated (13.4 MB instead of 230 MB)
2023/day_20230927.npy — truncated (18.5 MB instead of 230 MB)

Upload-pending files (data intact, excluded only due to upload stall, may be added later):

2023/day_20230730.npy — upload blocked at 99% during HF server re-validation
2023/day_20230804.npy — upload blocked at 82% during HF server re-validation

License

CC-BY-4.0