The study of Hydrological processes is of paramount importance in understanding and effectively managing our water resources. These processes are inherently complex and non-linear, posing challenges for hydrological modelling. Traditional approaches, such as physics-based and lumped rainfall-runoff models, are widely used but often face limitations. Their computational demands can be substantial or their predictive accuracy is constrained by inherent simplifications. Neural hydrology offers an alternative, leveraging machine learning to learn directly from data. Deep learning methods, inspired by rainfall-runoff model structures, can estimate fluxes at the watershed outlet and between lumped storage units, offering insights into intermediate processes like soil moisture and snowpack dynamics. This approach enables neural hydrology to evaluate the capability of machine learning models to replicate key hydrological functions effectively.
Despite these advancements in machine learning applications, several challenges remain unaddressed: (i) the interpretability of dynamic processes is still incomplete, ii) limited ability of users to insert and re-use the existing domain knowledge.
We are convinced that neural hydrology can be significantly advanced by adopting a Bayesian framework. This approach not only enables meaningful uncertainty quantification and probabilistic predictions but also facilitates the integration of soft and hard knowledge about hydrological processes through physics-informed learning. Moreover, it aligns the learned models with human-level understanding, enhancing their interpretability and applicability. Our research goal is to develop Bayesian Neural Hydrology as new cutting-edge approach that holds the potential to revolutionize our understanding of hydrological systems.
In order to assess the performance of Bayesian Neural Hydrology as an innovative approach that combines accurate modelling capabilities with improved physical representation of dynamic systems, we will integrate prior hydrological knowledge within a Bayesian framework. This integration aims to enhance forecasting accuracy, out-of-sample generalization, and data efficiency of neural hydrology. Furthermore, we will incorporate mass balance equations into the Bayesian neural hydrology framework in a controlled and interpretable manner. Lastly, we will leverage the variable memory capabilities of LSTM-based neural hydrology within the Bayesian neural hydrology network to yield a tool for watershed modelling in accordance with physical principles.
