Abstract
<jats:p> Machine learning (ML) methods have emerged as a promising approach for dissolved oxygen (DO) mapping. However, the uncertainties arising from model configurations such as hyperparameter setting and random initialization remain insufficiently characterized. Here we quantify this uncertainty through an efficient ensemble intercomparison. ML-based DO mapping uncertainty spatially peaks in the oxycline and temporally increases during the periods of historical observation scarcity. These patterns are primarily driven by intrinsic DO variability and extrinsic observing strategy. Our results further show that with the deployment of the Argo-O <jats:sub>2</jats:sub> floats, global uncertainty has been reduced by 37.8% over the last two decades. However, uncertainty reductions are regionally uneven: substantial decreases occur mainly where sampling is spatially uniform, including the Atlantic, Indian, and Southern Oceans. These findings emphasize that enhancing spatial uniformity is equally important as increasing the number of observations, positioning uncertainty analysis as a bridge between understanding DO variability and designing future observing-systems. </jats:p>