Object-based hail nowcasting training dataset supporting Anthony et al. 2026

Abstract

This dataset contains the tabular, machine-learning-ready records used to train and evaluate object-based nowcasts of severe hail for Australia, as described in Antony et al. (2026). Each record describes one 90-minute segment of an objectively tracked thunderstorm: 37 predictors summarising radar-derived storm intensity, lightning activity, storm kinematics, and the near-storm environment over a 30-minute observation window, together with continuous and binary measures of hail severity over three overlapping forecast intervals (0–30, 15–45, and 30–60 minutes). Records span nine years (2015–2023) of observations from twelve S-band Bureau of Meteorology radars covering eastern Australia, Adelaide, and Melbourne, and are derived from the Australian Unified Radar Archive (AURA), the DTN Total Lightning Network, and the BARRA2 regional reanalysis. Severe hail is defined by a 35 mm threshold on the radar-derived maximum estimated size of hail (MESH). The dataset is intended to support the development, benchmarking, and reproduction of storm-scale hail nowcasting models for Australian conditions.

Methods

Storm objects were identified and tracked from column-maximum reflectivity using the AINT algorithm (Brook et al. 2025) applied to AURA level 2 radar data (see related datasets below), with objects defined by reflectivity ≥ 35 dBZ over an area of at least 30 km². Tracks lasting at least 90 minutes were divided into overlapping 90-minute segments, each yielding one record. Full details of the AURA level 2 correction chain, the SHI-to-MESH conversion, the parcel definitions, and each predictor are given in the accompanying paper.

Data structure

Format: comma-separated values (CSV), UTF-8, one header row / describe actual packaging
File: antony_et_al_2026_training_dataset.csv
Records: 49 columns 402357 rows
Size: 90 MB
Missing data: represented by the literal string `nan`

Column names in the file differ from the parameter names used in Table 2 of the accompanying paper; the crosswalk below gives both.

Identifiers and location

• radar_id: Bureau of Meteorology radar site identifier (see site table below)
• parent_cell_id: AINT storm track identifier
• segment_id: Identifier of the 90-minute track segment
• time_unix [s]: Analysis time (t = 0), seconds since 1970-01-01 00:00:00 UTC
• latitude [°N]: Storm centroid latitude at t = 0
• longitude [°E]: Storm centroid longitude at t = 0

Predictors — storm properties (radar-derived)
 
  storm_area          [StormArea / km²]
      Storm object area at t = 0
  storm_area_trend    [StormAreaΔ / km² hr⁻¹]
      Trend in storm area over the observation window
  storm_duration      [StormDuration / s]
      Elapsed storm lifetime at t = 0 (time since first detection in the track)
  storm_motion_x      [StormMotionE / m s⁻¹]
      Eastward component of storm motion over the observation window
  storm_motion_y      [StormMotionN / m s⁻¹]
      Northward component of storm motion over the observation window
  storm_motion_mag    [StormMotion / m s⁻¹]
      Magnitude of storm motion over the observation window
  deviant_motion_x    [DevMotionE / m s⁻¹]
      Eastward component of deviant motion over the observation window
  deviant_motion_y    [DevMotionN / m s⁻¹]
      Northward component of deviant motion over the observation window
  deviant_motion_mag  [DevMotion / m s⁻¹]
      Magnitude of deviant motion over the observation window
 
Predictors — storm intensity (radar-derived)
 
  MESH_max               [MESHmax / mm]
      Maximum MESH within the storm footprint at t = 0
  MESH_max_trend         [MESHmaxΔ / mm hr⁻¹]
      Trend in maximum MESH over the observation window
  MESH_p90               [MESHp90 / mm]
      90th percentile of MESH within the storm footprint at t = 0
  MESH_p90_trend         [MESHp90Δ / mm hr⁻¹]
      Trend in 90th percentile MESH at t = 0
  MESH_area_ge_20        [MESH20Area / km²]
      Area of MESH ≥ 20 mm within the storm footprint at t = 0
  MESH_area_ge_20_trend  [MESH20AreaΔ / km² hr⁻¹]
      Trend in area of MESH ≥ 20 mm over the observation window
  ETH_p90                [ETHp90 / m]
      90th percentile of echo-top height above radar level at t = 0
  ETH_p90_trend          [ETHp90Δ / m hr⁻¹]
      Trend in 90th percentile echo-top height over the observation window
  azshear_p90            [AzShearp90 / 10⁻³ s⁻¹]
      90th percentile of azimuthal shear, 2–5 km ARL
  azshear_p90_trend      [AzShearp90Δ / 10⁻³ s⁻¹ hr⁻¹]
      Trend in 90th percentile azimuthal shear over the observation window
 
Predictors — lightning
 
  lightning_flash_rate                [LFR / flashes hr⁻¹]
      Strokes summed over the storm footprint, divided by the volume scan period, at t = 0
  lightning_flash_rate_trend          [LFRΔ / flashes hr⁻¹ hr⁻¹]
      Trend in lightning flash rate over the observation window
  lightning_flash_rate_density        [LFRD / flashes km⁻² hr⁻¹]
      Lightning flash rate divided by storm area at t = 0
  lightning_flash_rate_density_trend  [LFRDΔ / flashes km⁻² hr⁻¹ hr⁻¹]
      Trend in lightning flash rate density over the observation window
 
Predictors — near-storm environment (BARRA-R2) using closest timestamp before r=0
 
  MUCAPE        [CAPE / J kg⁻¹]
      Most-unstable convective available potential energy
  MUCAPEm10m30  [CAPEm10m30 / J kg⁻¹]
      CAPE in the −10 °C to −30 °C layer (hail growth zone)
  MUCIN         [CIN / J kg⁻¹]
      Most-unstable convective inhibition
  MUEL          [EL / m AGL]
      Equilibrium level height
  WBFZL         [WBFZL / m AGL]
      Wet-bulb freezing (melting) level height
  MUVTEm20      [VTEm20 / K]
      Virtual temperature excess of the lifted parcel at −20 °C
  BWD06         [BWD06 / m s⁻¹]
      0–6 km AGL bulk wind difference
  LR36          [LR36 / K km⁻¹]
      3–6 km AGL lapse rate
  PW            [PW / mm]
      Precipitable water, surface to 300 hPa
  RH36mean      [RH36 / %]
      Mean relative humidity, 3–6 km AGL
  SRH03l        [SRH03L / m² s⁻²]
      0–3 km storm-relative helicity, Bunkers left mover
  SRH03r        [SRH03R / m² s⁻²]
      0–3 km storm-relative helicity, Bunkers right mover
  U06mean       [U06 / m s⁻¹]
      0–6 km AGL mean zonal wind
  V06mean       [V06 / m s⁻¹]
      0–6 km AGL mean meridional wind
 
Predictands
 
  MESH_max_next_30   [mm]
      Maximum MESH within the storm footprint, t = 0–30 min (excluding t = 0)
  MESH_max_15_45     [mm]
      Maximum MESH within the storm footprint, t = 15–45 min
  MESH_max_last_30   [mm]
      Maximum MESH within the storm footprint, t = 30–60 min
  MESH_bool_next_30  [0/1]
      Severe hail flag for t = 0–30 min (1 where MESH_max_next_30 ≥ 35 mm)
  MESH_bool_15_45    [0/1]
      Severe hail flag for t = 15–45 min
  MESH_bool_last_30  [0/1]
      Severe hail flag for t = 30–60 min
 
The continuous MESH_max_* fields are provided so that users may apply alternative
severity thresholds or treat the problem as a regression task. Note that last_30 refers
to the last 30 minutes of the 60-minute forecast window, i.e. the 30–60 minute lead-time
interval.
 
Radar sites
 
  radar_id  Location            Latitude  Longitude  Beamwidth (°)
  --------  ------------------  --------  ---------  -------------
  2         Laverton, VIC       −37.85    144.75     1.0
  3         Wollongong, NSW     −34.26    150.87     1.9
  4         Newcastle, NSW      −32.73    152.02     1.9
  8         Gympie, QLD         −25.96    152.58     1.9
  40        Captains Flat, ACT  −35.66    149.51     1.9
  50        Marburg, QLD        −27.60    152.54     1.9
  64        Buckland Park, SA   −34.61    138.47     1.9
  66        Mt Stapylton, QLD   −27.72    153.24     1.0
  69        Namoi, NSW          −31.02    150.19     1.9
  71        Terrey Hills, NSW   −33.70    151.21     1.0
  72        Emerald, QLD        −23.55    148.24     1.9
  75        Mount Isa, QLD      −20.71    139.55     1.9