Classifying annual daily hydrographs in Western North America using t‐distributed stochastic neighbour embedding

• Published in: Hydrological processes Vol. 36; no. 1
• Main Authors: Tang, Weigang, Carey, Sean K.
• Format: Journal Article
• Language: English
• Published: Hoboken, USA John Wiley & Sons, Inc 01.01.2022; Wiley Subscription Services, Inc
• Subjects: Biological activity; Catchment area; Catchments; Classification; Coders; Daily hydrographs; Deep learning; Embedding; hydrograph similarity; Hydrographs; Hydrology; Machine learning; Physical properties; Reduction; Redundancy; Rivers; Stochasticity; Stream discharge; Stream flow; streamflow classification; Two dimensional flow; t‐SNE map; North America
• ISSN: 0885-6087; 1099-1085
• DOI: 10.1002/hyp.14473

Flow regimes are critical for determining physical and biological processes in rivers and their classification and regionalization traditionally seeks to link patterns of flow to physiographic, climate and other information. There are many approaches to, and rationales for, catchment classification, with those focused on streamflow often seeking to relate a particular response characteristic to a physical property or climatic driver. Rationales include such topics as prediction in ungauged basins (PUB), and providing guidance for model selection in poorly understood hydrological systems. The annual daily hydrograph (ADH) is a first‐order easily visualized integrated expression of catchment function, and over many years the average ADH is a distinct hydrological signature that differentiate catchments from each other. In this study, we use t‐SNE, a state‐of‐the‐art technique of dimensionality reduction, to classify 17 110 ADHs for 304 reference catchments in mountainous Western North America. t‐SNE is chosen over other conventional methods of dimensionality reduction (e.g., PCA) as it presents greater separability of ADHs, which are projected on a 2D map where the similarities are evaluated according to their map distance. We then utilize a Deep Learning encoder to upgrade the non‐parametric t‐SNE to a parametric approach, enhancing its capability to address ‘unseen’ samples. Results showed that t‐SNE successfully clustered ADHs of similar flow regimes on the 2D map and allowed more accurate classification with KNN. In addition, many compact clusters on the 2D map in the coastal Pacific Northwest suggest information redundancy in the local streamflow network. The t‐SNE map provides an intuitive way to visualize the similarity of high‐dimensional data of ADHs, groups catchments with like characteristics, and avoids the reliance on subjective hydrometric indicators. t‐SNE converts Annual Daily Hydrographs (ADHs) into a 2D map, where ADHs from different western North American flow regimes show distinct clusters. This provides an intuitive way to visualize and compare the similarity among a large number of hydrographs. Incorporating a Deep Learning encoder, we provide a convenient tool to project new hydrographs on to the map, identify similar counterparts and establish the flow regime for any given ADH.