Hybrid and Integrative Evolutionary Machine Learning in Hydrology: A Systematic Review and Meta-analysis

It has been claimed throughout the last two decades that hydrological machine learning (ML) models may produce more accurate and resilient simulations than previous approaches. However, one of the key obstacles to applying these approaches in the field of hydrology is the degradation of estimation error in ML models. The application of integrative MLs (a.k.a. hybrid MLs)—by embedding meta-heuristic algorithms (MHAs) with the base learner-would increase the accuracy of simple MLs. This review analyzes the cons and pros of the last decade's publications on the application of evolutionary-embedded ML models in three distinguished realms of hydrological systems, including (i) modeling of surface hydrology (river flow, river water quality, and sediment transport), (ii) hydrometeorology (drought modeling, precipitation, evaporation, and evapotranspiration), and (iii) groundwater hydrology. The overall conclusion of this study reveals that the use of evolutionary MHAs has the ability to improve the accuracy of normal MLs by up to 31.27% on average. Neural networks, neuro-fuzzy inference systems, and support vector machines are the most commonly employed base learners that have been fine-tuned using evolutionary MHAs. It should be highlighted that one of the main disadvantages of using MHAs is their high computing cost in comparison to typical optimization techniques. Among the several MHAs investigated, the genetic algorithm (GA) is the most widely utilized evolutionary algorithm for tuning and optimizing ML models. Nonetheless, particle swarm optimization, which is a swarm-based algorithm, produced significantly more accurate findings than the GA and is recommended as a preliminary proposal for hydrological application.