• Multimodal deep learning predicts optical and non-optical water quality. • 14-days forecasts reveal cumulative nutrient dynamics missed at short horizons. • Fine-tuning with 1% local data outperforms 50% retraining in sparse regions. • Scenario analysis prioritizes spatially-targeted interventions for management. Accurately forecasting river-water quality and assessing the impacts of prospective land-use or climate interventions remain central challenges for basin managers, particularly where monitoring networks are sparse. We present a multimodal deep-learning framework that fuses static and dynamic drivers, land-use composition, population density, Köppen climate class, meteorological and hydrological variables, and air-quality indicators, within a shared representation learned by a multimodal encoder–decoder architecture that combines LSTM, Transformer, and MLP encoders with a Transformer-based decoder. Trained on > 4 million samples from 1,801 monitoring stations in China, the model achieves Nash–Sutcliffe efficiencies of 0.72–0.96 for key parameters at 3-day lead time and maintains, or improves, predictive performance at 14-day, unveiling latent trends in slowly responding pollutants (e.g., total nitrogen, COD). A modular fine-tuning approach that freezes the globally pretrained encoders uses only 1% local data yet outperforms a baseline retrained with a 50/50 train–test split of the local data, underscoring strong cross-site transferability. SHapley Additive Explanations (SHAP) analyses reveal that static land-use and socioeconomic features dominate baseline water quality, while air-quality and hydro-meteorological drivers govern short-term variability through cumulative, non-linear processes. Scenario simulations further indicate that replacing built or bare land with tree cover can reduce total nitrogen by up to 60% in densely populated, high-rainfall regions. Together, these findings demonstrate a pathway from reactive monitoring to proactive basin governance: sparse observations are translated into reliable forecasts; “what-if” scenarios become quantitative pollutant-reduction estimates; and managers gain an interpretable, scalable tool to align urban growth, ecological restoration, and climate-adaptation goals.
Wang et al. (Wed,) studied this question.