Global monthly ocean dissolved oxygen (1960–2023) reconstructed to 5902 m with BLENDR, a Bayesian-optimized ensemble learning framework

Abstract

Oceanic oxygen levels, crucial for marine ecosystems and biogeochemical cycles, have declined significantly over the past few decades due to climate change, posing severe environmental risks. However, historical dissolved oxygen (DO) measurements, especially below 2000 m, remain sparse, limiting comprehensive annual and seasonal analyses. Here, we introduce the BLENDR framework (Bayesian-optimized Learning and ENsemble modeling for Data Reconstruction), a Bayesian-optimized ensemble of six machine-learning models (Random Forest, XGBoost, LightGBM, CatBoost, Extremely Randomized Trees and Histogram-based Gradient Boosting) fused via a spatially coherent dynamic weighting scheme, to reconstruct global monthly DO distributions at a 1° × 1° resolution from the surface to 5902 m from 1960 to 2023. Validation against an independent dataset demonstrated that BLENDR achieves better performance than any individual model, with an R2 of 0.968. Our dataset captures depth-dependent deoxygenation, with the most pronounced decline occurring between 150 and 200 m at approximately -0.12 µmol kg−1 yr−1, and shows severely accelerated oxygen loss in the Arctic Ocean and North Atlantic over the past decade. This work provides a long-term, global monthly DO product from the ocean surface to 5902 m. The bathypelagic DO data provided in this work are a significant contribution to deep ocean oxygen dynamics and global biogeochemical cycles. The data product is publicly accessible at https://doi.org/10.5281/zenodo.19705526 (Han and Zhou, 2026).