Multi-task Evidential Neural Networks for Stage-Discharge Rating Curves with Single-Stage Inference: Comparison Against Gaussian Processes, Deep Ensembles and the Published Laraque Curves on the Congo River at Brazzaville-Kinshasa

Reliable stage-discharge rating curves are essential for river monitoring, yet their development remains challenging when based on sparse data measurements and when uncertainty quantification is required. This study evaluates an evidential neural network (ENN) framework for rating-curve construction at the BZV-KIN station on the Congo River and compares its performance with Gaussian Process (GP) and Deep Ensemble Bayesian Neural Network (DE-BNN) approaches. The methodology is assessed using both single-task and multi-task formulations incorporating additional hydraulic and environmental variables, and its transferability is tested across eight independent gauging stations spanning the Congo, Orinoco, Niger, and Mekong basins. Results show that all methods produce highly consistent central rating curves within the operational stage range, with out-of-sample performance exceeding R^2=0.995, NSE =0.995, and KGE =0.993 at all transfer stations. The multi-task formulation substantially improves the training-set likelihood but yields only marginal changes in predictive accuracy and uncertainty width. The ENN framework provides the additional advantage of distinguishing aleatoric and epistemic uncertainty through belief and plausibility bounds, while maintaining predictive performance comparable to competing approaches. A major finding is the identification of a persistent discharge offset of approximately 3000 to 5000 m³/s between Acoustic Doppler Current Profiler (ADCP) based rating curves and the historical data collected using current meters. Temporal cross-validation and an independent validation using 255 GRDC observations from the Kinshasa station indicate that this discrepancy is most likely attributable to differences between measurement technologies rather than to model structure or sample limitations. These results demonstrate that uncertainty-aware machine-learning models can generate robust and transferable rating curves from limited ADCP datasets while highlighting the need for explicit record harmonization when integrating modern ADCP measurements with historical discharge archives.