THULab/nexrad_storm_catalog

NEXRAD Storm Catalog (TsFile)

Apache TsFile version of Silurian/nexrad-storm-catalog.

Overview

A metadata index pairing storm events from the NOAA Storm Events Database with nearby NEXRAD WSR-88D weather-radar stations. Each row is one (event, station) pair where the station was within 150 km of the event location, scoped to 2015–2024.

This catalog is metadata only. It does not contain reflectivity, velocity, or any other radar payloads — it points at radar windows so a downstream pipeline can fetch the matching Level-2 volumes from the Unidata NEXRAD Level-2 archive on AWS Open Data (s3://unidata-nexrad-level2/).

• Scale: 583,674 (event, station) rows; 154 radar stations with at least one event (of 158 in the station lookup); 17 storm types grouped into 6 categories; events from 2015-01-01 to 2025-01-01 (UTC).
• Built by: joining NOAA Storm Events records (2015–2024) to WSR-88D station coordinates, mapping each event_type into the 6-way category.

Schema (TsFile structure)

• Time (INT64, milliseconds) — event start time, converted from the source begin_datetime_utc (UTC epoch milliseconds).
• station_id (TAG) — WSR-88D station identifier (e.g. KTLX, KABR); query one radar site with WHERE station_id='KTLX'.
• event_type (TAG) — one of the 17 storm types; combine with the station tag, e.g. WHERE station_id='KTLX' AND event_type='Tornado'.
• event_id (FIELD, INT64) — NOAA Storm Events identifier.
• ef_scale (FIELD, STRING) — EF rating for tornado events (null/empty for other events).
• end_datetime_utc (FIELD, STRING) — event end time, preserved as an ISO UTC string (the source timestamp column other than the Time column).
• begin_lat (FIELD, DOUBLE), begin_lon (FIELD, DOUBLE) — event latitude/longitude in degrees.
• distance_km (FIELD, DOUBLE) — great-circle distance from the station to the event location.
• magnitude (FIELD, DOUBLE) — event magnitude (units depend on magnitude_type).
• magnitude_type (FIELD, STRING) — magnitude units / scale (e.g. EG, MG).
• state (FIELD, STRING) — US state / territory name.
• episode_narrative (FIELD, STRING) — free-text NWS narrative for the surrounding episode (up to ~26 KB per row).
• category (FIELD, STRING) — coarse 6-way grouping: severe_convective, heavy_rain_flood, winter_weather, tropical, other_severe, general.

Type mapping: source int64 stays INT64, source double stays DOUBLE, and all text columns stay STRING (no value truncation). station_id and event_type are used as the two device TAGs, so the table is partitioned into one time series per (station_id, event_type) radar-site/storm-type combination (1,769 combinations with data). Rows sharing a (station_id, event_type, Time) key are separated by the minimum 1 ms offset so every device timeline is strictly increasing; no rows are dropped.

Usage

Install the Apache TsFile Python SDK (pip install tsfile) and read a converted file:

from pathlib import Path
from tsfile import TsFileReader

path = Path("nexrad_event_catalog.tsfile")
with TsFileReader(str(path)) as reader:
    schemas = reader.get_all_table_schemas()
    print("tables:", list(schemas))
    table_name = next(iter(schemas))
    table = schemas[table_name]
    columns = [column.get_column_name() for column in table.get_columns()]
    print("columns:", columns)
    field_names = [
        column.get_column_name()
        for column in table.get_columns()
        if column.get_column_name() not in {"Time", "time"}
    ]
    if field_names:
        with reader.query_table(table_name, field_names[:3], batch_size=1024) as result:
            batch = result.read_arrow_batch()
            if batch is not None:
                print(batch.to_pandas().head())

Source & license

• Original dataset: https://huggingface.co/datasets/Silurian/nexrad-storm-catalog
• Author / publisher: Silurian AI (catalog assembly, categorisation, packaging)
• Upstream data: NOAA Storm Events Database (US-government public domain) and NOAA National Weather Service WSR-88D station metadata
• Paper: none
• License: Apache-2.0 (the value-add join/categorisation/packaging layer). Upstream NOAA records are public domain.

Conversion notes

• nexrad_stations.csv is not converted. It is a static WSR-88D station lookup (station id, name, lat/lon, elevation, state, RDA type) with no time column and no temporal ordering — it is not a time series. It remains available in the original dataset for joining on station_id. Station metadata is therefore not duplicated into this TsFile.
• Episodes. The same episode_narrative is shared by all events of an NWS episode, so the text is highly repetitive across rows; this is faithful to the source and was not deduplicated.
• Missing values. ef_scale (562,123 nulls), magnitude (222,091) and magnitude_type (333,660) are genuinely absent for the event types they do not apply to; they are preserved as null (rendered as empty STRING for the text fields) and were not imputed or dropped.
• No rows dropped. All 583,674 source rows are present; the TsFile chunk metadata reports 583,674 rows.