Probabilistic prediction of chlorophyll-a in a highly regulated river using a multi-source Bayesian Neural Network

Predicting riverine algal blooms remains a major challenge due to the high stochasticity of aquatic systems and the complex, non-linear interactions among environmental drivers. To address this, the present study establishes a robust probabilistic forecasting framework for the middle and lower Han River, China, by integrating multi-source datasets into a Bayesian Neural Network (BNN), complemented by a hybrid interpretability approach merging Bayesian posterior inference with Generalized Additive Model (GAM). By fusing heterogeneous long-term hydrological, meteorological, water quality, and ecological data, the model effectively captures dynamic environmental interactions and therefore provides reliable probabilistic forecasts of chlorophyll-a (Chl-a) concentrations for 1-to-7-day horizons. The BNN architecture explicitly performs uncertainty quantification to mitigate inherent data noise and model uncertainty by delivering exceedance probability of bloom predictions, which help decision-makers minimize false negatives near critical alert thresholds. Key ecological findings elucidate a dual driving mechanism, whereby short-term forecasts are predominantly governed by algal biological inertia, whereas medium-to-long-term trends are constrained by environmental carrying capacity. Specifically, bloom outbreaks hinge on a multi-factor environmental window featuring water temperature exceeding 23°C, optimal light intensity, and stable hydrological conditions. GAM analysis reveals a nonlinear relationship between total phosphorus (TP) and Chl-a, indicating the limited efficacy of nutrient reduction in high-phosphorus regimes. Methodologically, this study underscores the necessity of combining multi-source data fusion with uncertainty quantification and non-linear attribution to advance deep learning applications in complex ecological systems.Keywords: Bayesian Neural Network; Uncertainty quantification; Algal Bloom Prediction; Middle and Lower Han River; Multi-source datasets; Generalized Additive Model