Rainfall-Runoff Modelling: A Review of Empirical, Physically-Based and Machine Learning Approaches and Climate Change Applications

Abstract

Rainfall-runoff (R-R) models are the foundational computational tools of applied hydrology, underpinning flood design, drought monitoring, reservoir operation, irrigation management, and climate change impact assessment. This paper provides a comprehensive global review of R-R modelling that integrates, for the first time in a single synthesis, model classification across four functional categories, satellite-based input data sources, calibration and validation frameworks, CMIP6 climate projection integration, and emerging technological frontiers. Four model families are examined in detail: empirical (Rational Formula, SCS-CN), conceptual (HBV, GR4J, Sacramento), physically-based (SWAT, VIC, HEC-HMS, MIKE SHE), and data-driven (ANN, LSTM, Random Forest, hybrid physics-ML). Remote sensing products including GPM-IMERG, CHIRPS, MSWEP, SoilGrids250m, and Copernicus DEM GLO-30, all accessible through Google Earth Engine, have substantially resolved input data scarcity barriers in ungauged basins. The Nash-Sutcliffe efficiency (NSE) and Kling-Gupta efficiency (KGE) metrics carry fundamental limitations documented in recent literature that require multi-metric evaluation with confidence intervals. CMIP6 multi-model ensembles driven by Shared Socioeconomic Pathways (SSPs) and corrected through quantile mapping represent the current standard for climate-driven streamflow projections. Published CMIP6-driven studies project divergent regional trajectories: increasing runoff in monsoonal South and Southeast Asia, substantial streamflow decreases in Mediterranean and semi-arid regions, and non-linear cryospheric transitions in Himalayan and Andean basins. LSTM networks consistently outperform calibrated process-based models across large-sample multi-basin benchmarks. Synthesis across eight global hydroclimatic regions reveals that LSTM consistently outperforms process-based models in arid catchments, while SWAT and VIC remain indispensable for CMIP6-driven multi-sector impact assessments. Persistent challenges include hydroclimatic non-stationarity, parameter equifinality, structural model uncertainty, and inadequate human water use representation. Physics-informed neural networks, digital twin frameworks, and IoT-integrated cloud decision support systems are identified as the most promising emerging technologies for next-generation operational hydrology.