| Abstract: |
Rainfall-runoff (RR) modeling forms the analytical backbone of sustainable water resources management, underpinning flood forecasting, reservoir operation, drought planning, and watershed-scale decision making. This review and meta-analysis synthesizes findings from more than ninety peer-reviewed studies published between 2000 and 2024, spanning empirical, conceptual, physically-based distributed, and data-driven artificial-intelligence approaches to RR modeling. The paper traces the historical evolution of modeling paradigms, from the unit hydrograph and Soil Conservation Service Curve Number methods to contemporary machine-learning architectures such as artificial neural networks, support vector machines, and long short-term memory networks, and examines how each paradigm addresses the persistent challenges of parameter uncertainty, spatial heterogeneity, and non-stationarity induced by climate and land-use change. A structured methodology, informed by PRISMA-style screening, was used to identify, filter, and synthesize the literature quantitatively and qualitatively. The meta-analysis reveals that hybrid physically-informed machine-learning models consistently outperform purely empirical or purely data-driven counterparts in terms of Nash-Sutcliffe efficiency, particularly in data-scarce and ungauged basins. The critical analysis further highlights recurring limitations across the literature, including overreliance on single-basin validation, inadequate uncertainty quantification, and limited transferability of trained models across climatic regimes. The discussion situates these findings within the broader goals of sustainable water governance, arguing for greater standardization of benchmarking datasets and increased integration of remote-sensing inputs. The review concludes that future RR modeling research must prioritize explainability, transfer learning, and coupling with decision-support frameworks to translate modeling advances into actionable water resour |
