Enhancing long-lead rainfall forecasting in data-scarce large watersheds using multi-model fusion

The Karkheh River Basin in western Iran is a large, topographically complex watershed with limited precipitation observations and plays a crucial role in regional water supply and agriculture. This paper proposes a comprehensive framework to enhance long-term monthly precipitation forecasts by integrating outputs from four numerical weather prediction (NWP) models provided by the Copernicus Climate Change Service (C3S). Systematic biases in raw forecasts with 1–6-month lead times were corrected using quantile mapping (QM). The bias-corrected outputs were then fused using seven machine learning algorithms: multilayer perceptron (MLP), support vector regression (SVR), Bayesian model averaging (BMA), random forest (RF), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). Model performance was optimized through automated hyperparameter tuning with a tree-structured Parzen estimator (TPE). The trained fusion models produced deterministic forecasts as the data fusion ensemble mean (DFEM) and probabilistic forecasts as the data fusion ensemble members (DFEMs). The results demonstrated that hyperparameter-tuned fusion models consistently outperformed individual NWP models, with MLP, AdaBoost, and LightGBM best capturing precipitation variability. Compared to the best raw NWP model, the integrated bias correction and fusion approach reduced the root mean square error (RMSE) by 9.6 mm and improved the Nash–Sutcliffe efficiency (NSE) and Kling–Gupta efficiency (KGE) by 0.24 and 0.18, respectively. Probabilistic forecasts effectively quantified uncertainty, with prediction intervals reliably encompassing observed precipitation. The proposed framework offers a robust solution for improving forecast accuracy and reliability in data-scarce, complex regions. • Multi-model fusion improves long-term rainfall forecasts in data-scarce watersheds. • Quantile mapping and fusion techniques boost Nash-Sutcliffe efficiency by 0.24. • Hyperparameter-tuned MLP and ensemble methods best capture precipitation variability. • Probabilistic forecasts with uncertainty quantification yield reliable precipitation intervals. • Addressing rainfall forecasting gaps in topographically complex regions with scarce data.