Conclusion
Here, we demonstrate the potential of t-SNE as an approach to compare
similarity among hydrographs that is particularly useful for large data
sets. By converting annual daily hydrographs to 2-D representation,
their degree of similarity is indicated by distance on the map. t-SNE
distance can be used as novel similarity metric, supplementing other
comparative metrics such as Xcorr and Nash-Sutcliffe efficiency. In this
application on reference watersheds in western North America, t-SNE
outperformed PCA analysis for dimensional reduction, suggesting its
potential in classifying and regionalizing streamflows. A deep learning
encoder network was developed and trained to project new data onto
existing maps to identify hydrological counterparts, overcoming previous
challenges of the non-parametric t-SNE approach. While subjectivity in
classification limits deep learning algorithms in some circumstances, we
believe that t-SNE is underutilized in hydrological applications and
that it has considerable potential for extremely large data sets.