PhyEddy-Net Global Mesoscale Eddy Detection Dataset (2010–2021)

This dataset provides daily global mesoscale ocean eddy detections at 0.125° spatial resolution for the period 2010–2021. Detections were generated by PhyEddy-Net, a multimodal physics–data dual-driven deep-learning segmentation framework that fuses three daily satellite-derived fields: sea level anomaly (SLA) from the AVISO+ CMEMS DUACS DT2018 product, geostrophic relative vorticity computed from SLA-derived geostrophic velocities, and sea surface temperature anomaly (SSTA) from NOAA OISST v2.1. After patch-level training, full-global daily maps were produced through sliding-window inference and a fixed 3° spherical-radius deduplication step. The resulting catalogue contains approximately 36.7 million eddy detections (about 3.06 million per year on average). Each record contains the following variables: date: detection date (yyyy-mm-dd) type: eddy polarity (C = cyclonic, AC = anticyclonic) center_lat: centroid latitude (°) center_lon: centroid longitude (°), normalized to [-180, 180] area: eddy area expressed as the number of 0.125° grid pixels in the segmented eddy mask radius_km: equivalent-circle radius (km) derived from area prob: model detection probability (softmax confidence) basin: ocean basin identifier (Atlantic, Pacific, Indian) The dataset has been independently validated against Argo profiling floats using a leave-one-out correction protocol. Compared with the META4 altimetry atlas, PhyEddy-Net achieves a higher hit rate (87.3% vs 85.0%) and type agreement (73.1% vs 71.6%) on 4,054 clear Argo profiles. Note on amplitude: The maximum absolute SLA amplitude is used internally for quality checks but is not included in the released catalogue. Users can extract amplitude from the original SLA product using the provided centroid coordinates, date, and segmentation mask if needed.