Machine Learning Reconstruction of Glider Dissolved Oxygen: Accuracy Across Missions, Depths, and Years

Dissolved oxygen (DO) supports marine life and biogeochemical processes, but many glider missions lack oxygen measurements. To reconstruct missing DO, we compared multilayer perceptron (MLP), adapted Self-Attention-based Imputation for Time Series (SAITS), random forest (RF), and histogram gradient boosting (HGB) using 41 northern Gulf of Mexico glider missions (2013–2025), including 12 missions with measured DO, plus shipboard and towed observations. Reconstruction errors for missions excluded from training were about 1.6 times those from training on each mission's first two-thirds and testing its final third. RF and HGB had the lowest errors. Hiding optical inputs during MLP training reduced errors when those inputs were missing during prediction. Final-third errors defined depth ranges for annual analysis. Within these ranges, median errors in annual oxygen anomalies were 0.147–0.284 mL L⁻¹, compared with measured interannual variability of 0.130–0.143 mL L⁻¹. At 150–400 dbar, however, tree-model errors were smaller than measured interannual variability.