Streamflow prediction in ungauged watersheds remains a critical challenge in hydrological science due to the absence of in situ measurements, particularly in remote, data-scarce, and developing regions. This review synthesizes recent advancements in artificial intelligence (AI) for streamflow modeling, focusing on machine learning (ML), deep learning (DL), and hybrid modeling frameworks. Three core methodological domains are examined: regionalization techniques that transfer models from gauged to ungauged basins using physiographic similarity and transfer learning; synthetic data generation through proxy variables such as NDVI, soil moisture, and digital elevation models; and model performance evaluation using both deterministic and probabilistic metrics. Findings from recent literature consistently demonstrate that AI-based models, especially Long Short-Term Memory (LSTM) networks and hybrid attention-based architectures, outperform traditional conceptual and physically based models in capturing nonlinear hydrological responses across diverse climatic and physiographic settings. The integration of AI with remote sensing enhances generalizability, particularly in ungauged and human-impacted basins. This review also addresses several persistent research gaps, including inconsistencies in model evaluation protocols, limited transferability across heterogeneous regions, a lack of reproducibility and open-source tools, and insufficient integration of physical hydrological knowledge into AI models. To bridge these gaps, future research should prioritize the development of physics-informed AI frameworks, standardized benchmarking datasets, uncertainty quantification methods, and interpretable modeling tools to support robust, scalable, and operational streamflow forecasting in ungauged watersheds.
Gacu et al. (Sun,) studied this question.