Interpretability of machine learning with hydrological applications

This doctoral thesis focuses on the interpretability of the machine learning (ML) considering two specific topics to achieve a better interpretation of machine findings: feature importance and feature effects. Feature importance helps to identify features that drive the ML model response, while feature effects provide a visualization of the partial behavior of the ML model as a function of a subset of features. Exploiting one of the most powerful visualization tool, Accumulative Local Effect (ALE) plot, I develop new approaches to obtain insights on feature importance. Moreover, I employ these new techniques in combination with other promising ML methods in hydrological applications. First, I aim to understand a catchment hydrological response by investigating how sub-basins of a selected natural watershed contribute to its stormflow response. Second, I prove that using ML tools and feature importance measures helps to enhance an early warming system based on monitored discharges in specific watershed cross-sections.