DeepTornado: A Tornado Radar Dataset over China

The dataset contains 2,064 labeled samples derived from 143 confirmed tornado cases, along with 5,314 NUL samples representing diverse non-tornadic conditions.DeepTornado is intended to support statistical analyses of tornado-related radar features, data-driven and machine learning–based methodological studies, and related research on severe convective storms under realistic observational constraints. Users are encouraged to consider the inherent limitations of weather radar observations and to construct additional background samples as needed for specific applications.Each sample represents a localized radar data patch and includes three radar variables, namely reflectivity factor, radial velocity, and spectrum width, at the three lowest elevation angles.

The dataset is organized using the directory structure CATEGORY/SAMPLE_ID.npz, where CATEGORY denotes one of “tornado”, “warning”, “weak”, or “null”, corresponding to the TOR, WRN, WEK, and NUL classes, respectively. SAMPLE_ID is a unique integer assigned to each sample in sequential order.Each .npz file contains nine two-dimensional arrays with the keys “ref_0”, “ref_1”, “ref_2”, “vel_0”, “vel_1”, “vel_2”, “sw_0”, “sw_1”, and “sw_2”.

The prefixes “ref”, “vel”, and “sw” correspond to reflectivity factor (dBZ), radial velocity (m s-1), and spectrum width (m s-1), respectively. The suffixes 0, 1, and 2 indicate the three lowest radar elevation angles of 0.5°, 1.5°, and 2.4°. All arrays share a uniform spatial shape of 384 × 384, with a grid spacing of 250 m.

The x axis increases from west to east, whereas the y axis increases from north to south. All arrays are stored as 32-bit floating-point values. Missing values in all arrays are represented by the sentinel value −999.To support analyses across different time periods, samples are coarsely partitioned into three temporal groups (2017–2022, 2023, and 2024), which may be used as non-overlapping subsets in downstream studies.