Climate change significantly impacts water quality and quantity, intensifying extreme weather events, such as floods, droughts, and heat waves. Rising temperatures can increase humidity and dryness, disrupt the water cycle, cause saltwater intrusion into upstream lakes due to sea-level rise, and reduce dissolved oxygen in rivers, thereby deteriorating freshwater quality. Thus, accurate prediction of key climate variables, such as precipitation and temperature, is essential for mitigating detrimental impacts. This study evaluates three modeling approaches, including Process-Based (PB) models, Deep Learning (DL) models, and Process-Based Deep Learning (PBDL) models, to highlight their strengths and limitations.
Our assessment shows that PB models, which are based on physical laws and account for complex interactions between the atmosphere, land, and water bodies, require high parameterization and computational simplifications, which can lead to inaccurate results. DL models can uncover complex relationships from large datasets. They are effective in co-predicting variables, simulating General Circulation Model (GCM) outputs, optimizing PB models, and filling spatiotemporal data gaps. However, their performance depends on the availability of extensive temporal-spatial data, particularly for extreme events. The other group, PBDL models, known as physics-informed or hybrid models, can integrate the strengths of PB and DL approaches. Indeed, these models consider physical laws, such as mass balance and energy conservation, while leveraging DL's pattern recognition capabilities. Even with limited data, these models achieve superior predictions by combining pre-trained PB model outputs, which reduces computational demands.
Although these methods are used to evaluate (actual) evapotranspiration, snowmelt rate, soil permeability, hydraulic conductivity, and the effect of a warming climate on water temperature and streamflow, the interconnected influences on water systems, especially water quality indicators such as dissolved oxygen, heavy metals, nutrients, and water clarity, remain underexplored, presenting a critical research gap. Findings confirm that incorporating simultaneous predictions from DL models with proper variable selection and hyperparameter tuning can further enhance model robustness. Advancing PBDL models through integrating well-calibrated hydrological models, expanding spatiotemporal data coverage, and improving measurement accuracy yields more reliable climate change predictions and bolsters sustainable water resource management strategies.
To identify promising solutions, researchers are encouraged to address the non-stationary behavior of natural systems, considering not only meteorological factors (e.g., wind speed and solar radiation) but also the compound impacts of anthropogenic climate change on water resources. Additionally, selecting appropriate models and coupling them can improve an overall understanding of climate and water system interactions.
