Georgia Tech Dissolved Oxygen dataset 2026

This is a global dataset including Atlantic, Pacific, Indian, Arctic, Southern Ocean, Japan/East Sea and Mediterranean Sea from surface to 2000m at 1 degree resolution. The time range is from 1965 to 2025. Data source includes bottle, CTD-O2 and Argo-O2 profiles. Gridded dissolved oxygen maps generated from quality controlled profile data are available. (IAP) indicates Institute of Atmospheric Physics QC profile data (Gouretski et al., 2024). (NCEI) indicates NOAA National Centers for Environmental Information QC profile data (a.k.a. World Ocean Database, Mishonov et al., 2023). For the NCEI QC data, tanh bias correction is applied. Correction=1.69 x 1/2 * [1 + tanh{ (z-Ztrans)/Width }] with Ztrans=150m and Width=30. - Optimal Interpolation method (Ito 2021, Geosci Data Journal) Climatological annual means: (1) GTOI_clim_IAP_g1x1z67_65C5_ltm.nc (2) GTOI_clim_NCEI_g1x1z67_65C5_ltm.nc Yearly anomalies: (3) GTOI_anom_IAP_g1x1z67_65C5_yearly.nc (4) GTOI_anom_NCEI_g1x1z67_65C5_yearly.nc - Machine Learning method, modified from (Ito et al, 2024, J. Geophys. Res. Machine Learning and Computation). - The specific modifications are as follows: - PyTorch MLP (original version used Scikit-Learn) with the AdamW optimizer on GPU- IAP T/S data for inference, re-tuning of hyperparameters. - The predictor variables are (T, S, GMT=(global mean temp), sine & cosine (2*pi*month/12), Z (depth), Coordinates (lon/lat for Atl, Pac, and Indian Ocean, and polar coordinate for the Arctic and Southern Ocean). - Train-test split: randomly select 12 years from 61 years of data (~ 20%)- Reproducibility: random seed = 0 is used throughout the code whenever randomization occurs- Temporal K-fold CV is used for performance assessment and hyperparameter tuning- Each ocean basins are trained separately, such that there is no shared information between basins. - The resulting maps are joined with edge smoothing Annual means: (5) GTML_annual_IAP_g1x1z67_65C5.nc (6) GTML_annual_NCEI_g1x1z67_65C5.nc