DeepTornado: A Tornado Radar Dataset over China

This paper describes DeepTornado, a publicly available tornado radar dataset over China, developed to facilitate radar-based tornado research and applications. DeepTornado is constructed based on confirmed tornado cases from 2017–2024 and operational S-band radar data from the China New Generation Weather Radar (CINRAD) network, providing consistently processed and quality-controlled radar samples curated from 143 cases. A four-category empirical labeling scheme is used to jointly consider tornado reports and radar signatures: TOR (confirmed supercell tornadoes with compact and intense velocity couplets), WRN (warning tornado-like supercell mesocyclones without confirmed reports), WEK (confirmed tornadoes with weak or ambiguous signatures), and NUL (non-tornadic background scenes). The dataset contains 2,064 labeled samples (739 TOR, 1,189 WRN, 136 WEK) and 5,314 NUL samples. Each sample provides a cropped data patch, including reflectivity, radial velocity, and spectrum width at the three lowest elevation angles, interpolated to a Cartesian grid with 250-m spacing. Expert training, multi-labeler reconciliation, and final review ensure consistent labeling and quality control.