Ocean deoxygenation is an important consequence of climate change, but its long-term evolution in the Mediterranean Sea remains difficult to quantify because dissolved oxygen observations are unevenly distributed across space, depth, and time. Here, we develop a monthly three-dimensional reconstruction of dissolved oxygen across the Mediterranean Sea from 1976 to 2025 using a multilayer perceptron neural network trained on quality-controlled observations from CARIMED and CTD-O2WMED. Observations from MOOSE cruises are withheld from model training and used as an independent benchmark for model evaluation and validation.

The reconstruction is designed to represent long-term temporal variability while avoiding the imposition of a recurrent seasonal signal in deep waters, where oxygen seasonality is expected to be weak. Model performance is assessed using withheld observations, six independent research cruises not included in the training dataset, depth-stratified error diagnostics, and comparisons with oxygen changes associated with documented ventilation events. Across the independent cruises, the reconstruction achieves a mean absolute percentage error of approximately 3.1%, although larger deviations occur in coastal areas and in deep, data-sparse regions. Preliminary diagnostics further indicate that the selected temporal formulation suppresses non-physical seasonal variability at depth while preserving the main spatial and vertical structures of dissolved oxygen.

The reconstructed fields are used to investigate oxygen variability in intermediate waters at 300–400 m and in deep waters below 2000 m, with particular attention to changes associated with deep convection and water-mass renewal. The analysis evaluates whether recent Mediterranean oxygen evolution is characterized by continued deoxygenation or includes periods of slowdown, stabilization, or partial recovery associated with changes in ventilation.

By integrating heterogeneous observations into a spatially and temporally consistent product, this study provides a framework for examining long-term oxygen changes in relation to hydrographic variability and for investigating interactions among circulation, ventilation, and marine biogeochemistry across multiple timescales.