Estimating natural streamflow using a combined extension and routing approach

Abstract

Extension-based streamflow naturalization methods struggle with identifying mutations and selecting key features, while routing methods overlook the contribution of interstation runoff. This study proposes a Combined Extension and Routing (CER) approach to address these issues. The CER approach employs multiple change detection techniques to identify the earliest significant mutation and a multiple linear factors reconstruction method to select key features influencing natural flow. The CER models, implemented using extreme gradient boosting, Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Convolutional Neural Networks, and multiple linear regression, were evaluated in two snow-dominated catchments in the Yellow River, China. Results show that CER models effectively captured both peak and low flow events, achieving Nash–Sutcliffe efficiency of about 0.9 when comparing the estimation results from a water balance model. This study highlights the importance of stable land conditions for the CER approach's effectiveness, providing a reliable framework for natural streamflow estimation.