Quantifying the Effect of Additional Training Data When Using Machine Learning to Predict Streamflow in Ungauged Basins

One of the greatest challenges in hydrology is in the accurate prediction of streamflow in ungauged basins. In recent years, machine learning models have made great strides in the ability to accurately predict these flows, but these models are extremely data intensive. However, resources for streamflow monitoring are often quite limited so it is important to gather data as efficiently as possible. In this study, we establish a relationship between the amount of data used in training and the quality of predictions in ungauged basins for Long-Short Term Memory-based neural network models. First, using the CAMELS dataset, we create numerous training sets with a different number of basins in each and corresponding testing sets that use basins outside of the training set to simulate ungauged basins. We then quantify how changing the size of the training sets and the length of the training data affects the quality of predictions in the testing sets.