{"success":true,"data":{"author":"deepguess","bucketAccessType":"public","bucketName":"deepguess/tornet-temporal","bucketPrefix":"","category":"Climate & Weather","cloudProvider":"huggingface","createdAt":"2026-11-21T20:46:01.348Z","description":"Storm-centered sequences of NEXRAD WSR-88D weather radar data for tornado detection and short-term tornado prediction, extending the TorNet benchmark with time. It covers 24,862 storm events from NEXRAD Level-II archives between 2013 and 2022, at 130 radar sites across the contiguous US, all after the dual-polarization upgrade. Each event has 8 to 22 consecutive volume scans (about 4 to 5 minute cadence, median 13) on a 128 x 128 grid at 1 km, with 24 channels: reflectivity, radial velocity, spectrum width, differential reflectivity (ZDR), correlation coefficient (CC) and specific differential phase (KDP) at four elevation angles (0.5 to 1.8 degrees). Events fall into three classes, TOR (confirmed tornado, with EF rating), WRN (tornado warning issued, no tornado) and NUL (no severe weather). Each event folder holds sequence.npz with the radar array (T, 24, 128, 128), scan times, storm center, radar site, EF magnitude and labels; catalog.csv gives TorNet-compatible splits, and year-based splits (train 2013-2021, test 2022) are recommended. Load with numpy. Processing and labels are CC BY 4.0 by deepguess; the raw NOAA/NWS radar is public domain.","displayTitle":"TorNet-Temporal: Dual-Polarization NEXRAD WSR-88D Radar Sequences for Tornado Detection (24,862 Storm Events, 2013-2022)","huggingFaceUrl":"https://huggingface.co/datasets/deepguess/tornet-temporal","id":"14efd13d-8a3c-5737-a5e9-bca9f2bc3390","license":"CC-BY-4.0","licenseEvidenceUrl":"https://huggingface.co/datasets/deepguess/tornet-temporal","objectCountEstimate":24863,"probeStatus":"ok","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"deepguess/tornet-temporal\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"deepguess/tornet-temporal\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=deepguess, license:cc-by-4.0, size_categories:10K<n<100K, downloads=347"],"format":"mixed","generatedAt":"2026-11-21T20:48:53.375Z","layout":{"example":"tornet_1000043_NUL/sequence.npz","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"tornet_1000043_NUL/sequence.npz","truncatedAt":512}],"sampleObjects":["tornet_1000043_NUL/sequence.npz","tornet_1000048_NUL/sequence.npz","tornet_1000051_NUL/sequence.npz","tornet_1000053_NUL/sequence.npz","tornet_1000151_NUL/sequence.npz"],"sidecars":[{"bytes":"---\r\nlicense: cc-by-4.0\r\ntask_categories:\r\n  - image-classification\r\n  - video-classification\r\ntags:\r\n  - weather\r\n  - radar\r\n  - tornado\r\n  - nexrad\r\n  - meteorology\r\n  - severe-weather\r\n  - dual-pol\r\n  - earth-science\r\n  - climate\r\nsize_categories:\r\n  - 10K<n<100K\r\npretty_name: \"TorNet-Temporal: Temporal Dual-Pol NEXRAD Radar for Tornado Detection\"\r\n---\r\n\r\n# TorNet-Temporal: Temporal Dual-Pol NEXRAD Radar for Tornado Detection\r\n\r\nA 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**.\r\n\r\n## Dataset Summary\r\n\r\n- **24,862 storm events** from NEXRAD Level-II radar archives (2013-2022)\r\n- **8-22 consecutive radar scans** per event (~4-5 min cadence, ~45-90 min total; median 13 frames)\r\n- **24 channels**: 6 dual-pol radar products x 4 elevation angles\r\n- **128x128 spatial grid** at 1km resolution, storm-centered\r\n- **3 categories**: TOR (tornado), WRN (tornado-warned but no tornado), NUL (null/no severe)\r\n- **130 unique NEXRAD radar sites** across CONUS\r\n\r\n## Data Format\r\n\r\nEach event is stored as a compressed NumPy archive (`sequence.npz`) containing:\r\n\r\n| Key | Shape | Type | Description |\r\n|-----|-------|------|-------------|\r\n| `data` | (T, 24, 128, 128) | float32 | Radar volume sequence (T varies 8-22, median 13) |\r\n| `scan_times` | (T,) | str | UTC timestamps per frame |\r\n| `center_time` | scalar | str | Event center time |\r\n| `lat` | scalar | float64 | Storm center latitude |\r\n| `lon` | scalar | float64 | Storm center longitude |\r\n| `radar` | scalar | str | NEXRAD site ICAO (e.g., KTLX) |\r\n| `mag` | scalar | int64 | EF-scale magnitude (-1 if no tornado) |\r\n| `label` | scalar | int64 | Binary label (1=tornado, 0=no tornado) |\r\n| `category` | scalar | str | TOR, WRN, or NUL |\r\n\r\n### Channel Layout (24 channels)\r\n\r\nThe 24 channels represent 6 dual-polarimetric radar products at 4 elevation angles (0.5, 0.9, 1.3, 1.8 degrees):\r\n\r\n| Channel Index | Product | Elevation |\r\n|--------------|---------|-----------|\r\n| 0-3 | Reflectivity (REF) | 0.5, 0.9, 1.3, 1.8 |\r\n| 4-7 | Velocity (VEL) | 0.5, 0.9, 1.3, 1.8 |\r\n| 8-11 | Spectrum Width (SW) | 0.5, 0.9, 1.3, 1.8 |\r\n| 12-15 | Differential Reflectivity (ZDR) | 0.5, 0.9, 1.3, 1.8 |\r\n| 16-19 | Correlation Coefficient (CC) | 0.5, 0.9, 1.3, 1.8 |\r\n| 20-23 | Specific Differential Phase (KDP) | 0.5, 0.9, 1.3, 1.8 |\r\n\r\n### Category Definitions\r\n\r\n- **TOR**: Storm produced a confirmed tornado (EF0+). Label=1.\r\n- **WRN**: Storm had a tornado warning issued but no confirmed tornado. Label=0.\r\n- **NUL**: No severe weather. Randomly sampled from tornado-day radar scans at non-tornadic locations. Label=0.\r\n\r\n## Recommended Splits\r\n\r\nWe 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.\r\n\r\n| Split | Years | Approx. Events | Purpose |\r\n|-------|-------|----------------|---------|\r\n| Train | 2013-2021 | ~21,800 | Model training |\r\n| Test | 2022 | ~3,040 | Final evaluation |\r\n\r\nFor train/validation splitting, we recommend holding out one earlier year (e.g., 2021) as a validation set.\r\n\r\nA `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.\r\n\r\n**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.\r\n\r\n## Usage\r\n\r\n```python\r\nimport numpy as np\r\nimport torch\r\nfrom pathlib import Path\r\n\r\n# Load a single event\r\nevent = np.load(\"tornet_1000855_TOR/sequence.npz\")\r\ndata = event[\"data\"]          # (T, 24, 128, 128), T varies per event\r\nlabel = int(event[\"label\"])   # 1 for tornado\r\ncategory = str(event[\"category\"])  # \"TOR\"\r\nprint(f\"Frames: {data.shape[0]}, Label: {label}, Category: {category}\")\r\n\r\n# Select 8 consecutive fr","encoding":"utf8","key":"README.md","name":"README.md","size":6710,"truncated":true}],"sourceObjectKey":"tornet_1000043_NUL/sequence.npz","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"deepguess/tornet-temporal\", filename=\"tornet_1000043_NUL/sequence.npz\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"tornet_1000043_NUL/sequence.npz","totalObjects":24863},"summary":"Storm-centered sequences of NEXRAD WSR-88D weather radar data for tornado detection and short-term tornado prediction, extending the TorNet benchmark with time.","title":"TorNet-Temporal: Dual-Polarization NEXRAD WSR-88D Radar Sequences for Tornado Detection (24,862 Storm Events, 2013-2022)","updatedAt":"2026-11-21T22:08:34.004Z","visibility":"public","websiteUrl":"https://huggingface.co/datasets/deepguess/tornet-temporal"}}