(2)

A subset of ADHs were manually labelled with flow regime types. The determination of flow regime type was based on visual assessment to the shape of ADHs, which is subjective to some degree. 2480 ADHs were labelled that represent seven distinctive flow regimes in the WNA domain (Fig. 2). ADHs selected for the same type typically come from streams that are geographically clustered. Class 1 is characterized by frequent rain events during winter and low flows in summer, with streams mostly located in coastal Pacific North West (PNW). Class 2 is similar to Class 1, but has a distinct snowmelt-driven spring freshet, and are mainly located in the interior PNW. Class 3 has extremely low winter flows, large spring freshet and many summer rain events. All of Class 3 gauges are located in Alaska. Class 4 exhibits a large spring freshet followed by summer storm events with gauges located in the Northwest Territories and northern Alberta. Class 5 is similar to Class 4, with later summer rain events and gauges are located in Yukon and northern British Columbia. Class 6 is characterized by dominant snowmelt freshet in spring that accounts for more than 70% of the annual discharge with gauges primarily located Canadian Rockies between British Columbia and Alberta. Class 7 has a spring freshet along with occasional winter events and gauges are most often located in Idaho. The labelled ADHs were used to evaluate the performance of t-SNE, with details provided in later sections.

Figure 2: Normalized ADHs for the seven flow regimes. The number in the parenthesis indicates the number of ADHs in that class.

t-distribution Stochastic Neighbor Embedding (t-SNE, van der Maaten and Hinton (2008)) is a state-of-theart technique for dimensionality reduction and high-dimensional data visualization. It is a variant of the SNE that was originally proposed by Hinton and Roweis (2002). SNE represents similarity between datapoints using conditional probabilities that are converted from pairwise Euclidean distances (Eq. 3).