Multi-Day Salinity Forecasting in the Vietnamese Mekong Delta: Horizon-Dependent Performance and Interpretation of XGBoost, LSTM, and a Weighted Hybrid

Short-range salinity forecasts can support water management in tide-influenced deltas, but model performance may change with lead time. This study compared extreme gradient boosting (XGBoost), a long short-term memory network (LSTM), and a weighted hybrid for forecasting daily mean salinity one to five days ahead at Vam Kenh station in the Vietnamese Mekong Delta. Using 721 daily observations from six dry seasons (February–May, 2020–2025), we trained and tuned the models on 2020–2024 through seasonal time-series cross-validation and tested them independently on the full 2025 season. We fitted a direct model for each horizon, and derived hybrid weights solely from development-period out-of-fold predictions. XGBoost was the stronger constituent model at one to three days (RMSE 1.158 g/L; NSE 0.867 at one day), whereas LSTM outperformed XGBoost at five days (RMSE 1.657 versus 1.890 g/L; NSE 0.735 versus 0.656). The hybrid gave the best or equal-best estimates at every horizon (five-day RMSE 1.611 g/L; NSE 0.750), although its improvement over the best constituent model was modest (0–3.2% RMSE). Empirical 90% prediction intervals calibrated only on development-period residuals achieved 0.97–1.00 coverage in the test season, indicating conservative width. SHAP and permutation analyses identified recent salinity, dry-season timing, and upstream hydrological conditions as the main sources of predictive information, varying by architecture and horizon. The results demonstrate horizon-dependent model complementarity at one station and one independently tested season, rather than spatial transferability or operational readiness.