Machine Learning Modelling for Future Hydroclimatic Extremes Under Climate Change: A Review

Hydroclimatic extremes, such as droughts, floods, and extreme rainfall have been increasing worldwide leading to severe impacts on society and ecosystems. For that reason, hydrological modelling research has advanced to improve flood and rainfall prediction and control.  This estimation has been traditionally carried out using physical and process-based hydroclimatic models, however, they have limitations due to their physical-based nature. They often require a large amount of different hydro-geomorphological monitoring datasets, as well as in-depth knowledge and expertise regarding hydrological parameters, which must be correctly selected, calibrated, and further interpreted to ensure the reliability of the model. In recent years, data-driven hydrological modelling, such as Machine Learning (ML), Artificial Intelligence (AI), and Deep Learning (DL) methods have demonstrated a great deal of promise for enhancing the forecasting of hydroclimatic extremes. In data-driven modelling, the models use a generalized relationship between input and output disregarding the physical mechanism behind the process, built based on historical data. ML methods have some advantages over physical-based models, such as not requiring an understanding of internal specific mechanisms, which can be highly complex to reproduce, as well as having a higher calculation efficiency which may provide a quicker response to extreme events of high-intensity and short duration such as urban flash floods. Although there have been significant advances from the scientific community toward understanding and testing different ML and AI models for various hydrological applications, there are still limitations in their applications. A huge challenge that remains in ML modelling for future extreme floods, is its ‘black-box’ nature where the interactions among various components are unknown, which hinders its further use in supporting important decision-making. Along with that, other challenges in the current hydroclimatic modelling approaches presented by the hydrological community are data availability and assimilation, uncertainty analysis, and model generalisation. Some studies have addressed these issues, showing satisfactory results, especially for hybrid models between ML and traditional process-based approaches and ensembles of multiple methods. However, in light of so many new methodologies and algorithms, we must address their benefits and drawbacks, through an interdisciplinary effort. Understanding the best way to select appropriate methodologies for different settings of data availability, climate variability, and uncertainty, generating rapid and interpretable responses to urgent hydrologic hazards.