Abhinav Gupta

This study aimed to assess the usefulness of data from donor watersheds to predict streamflow in parent watersheds. For this purpose, Long-Short Memory Network (LSTM) is used as an information extraction algorithm. Data from a total of 434 watersheds were used in this study. Out of these 434 watersheds, 57 watersheds were selected as the parent watersheds. These 57 watersheds were those where streamflow statistical structure changed over the study period (1980-2013 water years). Several LSTM models were developed by using the different number of donor watersheds as training watersheds varying from 1 to 128. It was found that the optimal number of training watershed were much less than 128 for most of the parent watersheds. Increasing the number of donor watersheds beyond this optimal value resulted in a statistically insignificant gain in accuracy. In some cases, the Nash-Sutcliff Efficiency (NSE) decreased, albeit only slightly, when the number of donor watersheds for training increased beyond the optimal value. However, we also found that using data from more watersheds beyond the optimal number of training watersheds results only in a slight decrease in NSE. Therefore, one bears only a small cost by training LSTM against a large number of watersheds compared to the optimal number of watersheds. The results of this study contradict the prevalent idea that LSTM continues to extract hydrologically relevant information as more and more data are fed to the model; this was true only for a few of the 57 parent watersheds used in this study.