Application of Machine Learning Models to Predict Maximum Event Water Fractions in Streamflow

Abstract

Estimating the maximum event water fraction, at which the event water contribution to streamflow reaches its peak value during a precipitation event, gives insight into runoff generation mechanisms and hydrological response characteristics of a catchment. Stable isotopes of water are ideal tracers for accurate estimation of maximum event water fractions using isotopic hydrograph separation techniques. However, sampling and measuring of stable isotopes of water is laborious, cost intensive, and often not conceivable under difficult spatiotemporal conditions. Therefore, there is a need for a proper predictive model to predict maximum event water fractions even at times when no direct sampling and measurements of stable isotopes of water are available. The behavior of maximum event water fraction at the event scale is highly dynamic and its relationships with the catchment drivers are complex and non-linear. In last two decades, machine learning algorithms have become increasingly popular in the various branches of hydrology due to their ability to represent complex and non-linear systems without any a priori assumption about the structure of the data and knowledge about the underlying physical processes. Despite advantages of machine learning, its potential in the field of isotope hydrology has rarely been investigated. Present study investigates the applicability of Artificial Neural Network (ANN) and Support Vector Machine (SVM) algorithms to predict maximum event water fractions in streamflow using precipitation, soil moisture, and air temperature as a set of explanatory input features that are more straightforward and less expensive to measure compared to stable isotopes of water, in the Schwingbach Environmental Observatory (SEO), Germany. The influence of hyperparameter configurations on the model performance and the comparison of prediction performance between optimized ANN and optimized SVM are further investigated in this study. The performances of the models are evaluated using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and Nash-Sutcliffe Efficiency (NSE). For the ANN, the results showed that an appropriate number of hidden nodes and a proper activation function enhanced the model performance, whereas changes of the learning rate did not have a major impact on the model performance. For the SVM, Polynomial kernel achieved the best performance, whereas Linear yielded the weakest performance among the kernel functions. The result showed that maximum event water fraction could be successfully predicted using only precipitation, soil moisture, and air temperature. The optimized ANN showed a satisfactory prediction performance with MAE of 10.27%, RMSE of 12.91%, R2 of 0.70, and NSE of 0.63. The optimized SVM was superior to that of ANN with MAE of 7.89%, RMSE of 9.43%, R2 of 0.83, and NSE of 0.78. SVM could better capture the dynamics of maximum event water fractions across the events and the predictions were generally closer to the corresponding observed values. ANN tended to underestimate the events with high maximum event water fractions and to overestimate the events with low maximum event water fractions. Machine learning can prove to be a promising approach to predict variables that are not always possible to be estimated due to the lack of routine measurements.