Unveiling the Patterns and Drivers of Global Riverine Dissolved Oxygen via a Transformer-Based Deep Learning

Riverine dissolved oxygen (DO) dynamics are critical to aquatic ecosystem health, yet global understanding has long been constrained by sparse, temporally discontinuous, and heterogeneous monitoring data. Conventional statistical and process-based approaches also face difficulties integrating multi-source predictors across chemically ungauged basins. We developed a Transformer-based framework integrating hydroclimatic forcing, basin attributes, land cover, and demographic context to reconstruct daily DO dynamics across 403 river stations from 2001 to 2022. The model achieved a median Nash–Sutcliffe efficiency (NSE) of 0.61 in testing and 0.58 in hold-out validation at unseen stations, but performance varied substantially by region and was lower for some extreme DO values. Integrated Gradients identified 30-day cumulative air temperature as the most influential individual input predictor in the trained Transformer model. The reconstructed records indicate deoxygenation at 95% of the 222 stations selected for trend analysis, with stronger declines in agricultural basins. This work provides a data-driven framework for reconstructing continuous daily DO records from fragmented monitoring observations, with potential to support candidate hotspot identification and inform climate-adaptive freshwater management, subject to regional validation.