Exploring Catchment Regionalization Through Hydrolstm
Presenter: Luis De la Fuente P41
Co-Author(s): Laura Condon, Hoshin Gupta, Andrew Bennett
Advisor(s): Laura Condon
1Hydrology & Atmospheric Sciences
Hydrologists have been working on the issue of regionalization for decades. It is used when transferring parameters from one calibrated model to another and when searching for similarities between gauged and ungauged catchments. However, a unified method of regionalization that can successfully transfer parameters and identify similarities between different regions, accounting for differences in meteorological forcing, catchment attributes, and hydrological responses, is still an open research question. Machine learning (ML) has shown promising results in the generalization of its results at temporal and spatial scales for streamflow prediction. This indicates that ML models have learned useful regionalization relationships that we would like to extract. In this study, the HydroLSTM representation, which is a modification of the traditional Long Short-Term Memory, is explored to learn meaningful relationships between meteorological forcing and catchment attributes. A promising feature of the HydroLSTM representation is that the learned relationships can generate different hydrological responses across the US. These results indicate the potential to learn more about regionalization just by learning from our ML models.