Machine learning and deep learning in river-basin modeling: A comprehensive review

This study highlights a key question about the capacity of Machine Learning (ML) and Deep Learning (DL) models as standalone, futuristic approaches for river basin modeling. To answer the main research question, a PRISMA-based systematic review framework was used. This study evaluates 133 peer-reviewed studies that used standalone ML/DL models or compared them with conventional hydrological and process-based models to predict various hydrological variables and basin dynamics. Findings synthesis indicates that even though the ML/DL models are better in terms of pattern recognition, data integration, and predictive capacity, especially in non-stationary and non-heterogeneous hydroclimatic conditions, their application alone is not yet as physiologically interpretable and generalized across basins. Overall, the reviewed studies show that hybrid and physics-informed ML/DL models are more effective than traditional empirical models and data-driven models. Nevertheless, there are still difficulties, such as the lack of data, the inequality of computations in different regions, and the uneven assessment of extreme and low-flow states. This study highlights the necessity of hybrid modeling approaches, explainable and physics-guided AI, open data ecosystems, and standardized benchmarking frameworks to enable reliable, transparent, and fair river basin management in changing climatic conditions.