A comparison of modeling techniques to predict hydrological indices in ungauged rivers

Abstract

Predicting the natural flow regime in ungauged rivers is an important challenge in water resource management and ecological research. We developed models to predict 16 hydrological indices in a river network covering the northern third of the Iberian Peninsula. Multiple Linear Regression (MLR), Generalized Additive Models (GAMs), Random Forest (RF) and Adaptive Neuro Fuzzy Inference System (ANFIS) were used and compared according to their prediction accuracy. The results showed that predictive performance varied greatly depending on the modeled hydrological attribute. The magnitude and frequency indices were predicted with excellent accuracy. In contrast, no technique was capable of developing precise models for hydrological indices of timing, duration and rate of change. This is mainly related to the lack of proper environmental databases on the scales on which these flow regime patterns are influenced. In addition, complex modeling techniques did not always outperform linear models and no single approach was optimal for all indices. ANFIS and GAMs provided the best results; however, other issues such as computational cost and the level of knowledge required to apply the method and interpret the results should be taken into account.