Global-scale reconstruction of daily river dissolved oxygen using transformer

Dissolved oxygen (DO) concentration, a critical indicator of aquatic ecosystem health and metabolic activity, exhibits substantial spatiotemporal variability. Despite its ecological significance, large-scale prediction of DO dynamics and identification of its drivers have proven challenging, primarily constrained by scarcity and spatiotemporal inconsistency in water quality data. Meanwhile, extensive hydrometeorological datasets and watershed characteristics have become increasingly accessible. In this study, we developed a transformer model to reconstruct daily DO concentrations by integrating hydrometeorological and DO observations with basin attributes, leveraging data from 403 globally distributed monitoring stations. Our findings reveal that the transformer model achieves a Nash-Sutcliffe Efficiency (NSE) of 0.72 in reconstructing daily DO levels on a global scale. Thermal variables, particularly the 30-day cumulative air temperature, emerge as the predominant drivers of daily DO dynamics across global river basins. Agricultural basins exhibited the most rapid warming rates and severe oxygen depletion. This study highlights the potential of transformer models for imputing missing values in long time series data, thereby facilitating large-scale systematic analyses of DO patterns and their driving factors.