A Deep Learning-based Ocean Mesoscale Eddies Dataset for the South China Sea Spanning 1993 to 2021

This dataset provides a daily-resolution mesoscale eddy product for the South China Sea spanning 1 January 1993 to 31 December 2021, derived from multi-mission satellite altimeter sea level anomaly (SLA) data with a spatial resolution of 0.25° × 0.25°. Mesoscale eddies are identified using a deep-learning-based object detection framework, namely the YOLOF (You Only Look One-level Feature) model. The dataset consists of two main components: Eddy identification resultsThe eddy detection outputs are stored in NetCDF format under the directory “identificated_eddy”, organized by year (1993–2021). For each day, two separate files are provided: Cyclonic.nc and Anticyclonic.nc, corresponding to cyclonic and anticyclonic eddies, respectively. Each file records the properties of all detected eddies on that day, including time, latitude, longitude, amplitude, eddy kinetic energy (EKE), and radius. Training and evaluation sample setThe directory “sample_set” contains the datasets used for training, validation, and testing the YOLOF model, organized into train, validation, and test subsets. Each subset includes images and label folders, which store visualization images generated from SLA and geostrophic velocity fields, along with the corresponding bounding-box annotation files. This component enables reproducibility of the eddy detection experiments and supports further development of machine-learning-based eddy identification methods. This dataset offers a comprehensive and consistent characterization of mesoscale eddy activity in the South China Sea over nearly three decades. It is suitable for studies on eddy statistics, spatiotemporal variability, ocean dynamics, model validation, and data-driven oceanographic applications.