In recent decades, the use of machine learning has significantly increased, raising concerns about its application and the uncertainties it introduces across various fields-including hydrology. Explainable Artificial Intelligence (XAI) offers a framework for understanding machine and deep learning models, enabling clear interpretation of models, their parameters, and specific behaviours under simulated conditions.
Machine learning approaches have gained substantial ground in the realm of hydrological modelling. Studies have demonstrated that incorporating multiple physical input parameters can enhance or diminish models' interpretability and increase or decrease their predictive performance. Therefore, it is crucial to determine what is physically plausible within the model, identify patterns that extend beyond established physics, and ascertain whether these patterns provide a foundation for future research or reveal key mistakes in the modelling process.
This work presents a series of examples of how hydrological machine learning models can provide basic interpretability, but also to discuss the model interpretability, utilizing methods ranging from feature engineering to model-agnostic techniques-including applications of SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations). This presentation aims to demonstrate how explainability can serve as a basis for model improvement and also highlight instances where physical processes are underrepresented, leading to increased model uncertainty.
Speaker: Gerald Corzo Perez (IHE Delft, Netherlands)