A robust forward-chain approach for ex-post and ex-ante seasonal river water-quality forecasting using remote sensing and hydroclimate forecast datasets

Understanding pollution dynamics in urban riverine ecosystems is essential for effective water-quality management and ecological sustainability. However, reliable seasonal forecasting remains challenging due to sparse observations and uncertainty in hydrometeorological drivers. To fill this gap, this study presents a novel forward-chain (FC) forecasting framework for ex-post and ex-ante seasonal river water-quality prediction, explicitly designed to bridge the gap between retrospective model evaluation and operational forecast deployment. The framework uses a data-driven, long short-term memory (LSTM) architecture that integrates Sentinel-2 remote sensing, European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5), and Copernicus Climate Change Service (C3S) seasonal forecast datasets. Model training was conducted exclusively using ERA5 data, while forecast skill was evaluated under both ex-post ERA5-driven and ex-ante C3S-driven conditions. The framework was applied to the lower Ganges River (Hooghly River, India), a large, complex, and highly urbanized river system, and evaluated for dissolved oxygen (DO), electrical conductivity (EC), turbidity, and total suspended solids (TSS) across 1–6 month forecast horizons using Kling–Gupta Efficiency (KGE) and normalized mean absolute error (MAE %). Unlike conventional retrospective studies, the framework explicitly compares ex-post ERA5-driven forecasts with ex-ante C3S-driven forecasts to evaluate operational forecasting skill under realistic conditions. The LSTM model was benchmarked against an autoregressive exogenous-input (ARX) model, while the forward-chain (FC) strategy was compared with a fixed-window (FW) training scheme. Results demonstrated consistent forecast skill across all parameters and lead times, although accuracy generally declined as the forecast horizon increased. MAE increased by approximately 15%–30% under ERA5 forcing and 25%–45% under C3S forcing, reflecting the propagation of hydroclimatic forecast uncertainty. Compared with ARX, LSTM showed stronger advantages for optically responsive and event-driven parameters, particularly turbidity and TSS. The FC strategy consistently outperformed FW training, reducing MAE by 10%–25% with slower error growth across lead times. Overall, the proposed framework offers a potential approach for monthly-to-seasonal river water-quality forecasting, supporting the prioritization of monitoring and water-quality management.