Premium
A Comparison of Time‐Frequency Signal Processing Methods for Identifying Non‐Perennial Streamflow Events From Streambed Surface Temperature Time Series
Author(s) -
Partington D.,
Shanafield M.,
Turnadge C.
Publication year - 2021
Publication title -
water resources research
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.863
H-Index - 217
eISSN - 1944-7973
pISSN - 0043-1397
DOI - 10.1029/2020wr028670
Subject(s) - hilbert–huang transform , flow (mathematics) , computer science , time series , streamflow , wavelet transform , set (abstract data type) , data set , time–frequency analysis , series (stratigraphy) , field (mathematics) , data mining , environmental science , wavelet , mathematics , artificial intelligence , machine learning , geography , geology , radar , telecommunications , drainage basin , paleontology , geometry , cartography , white noise , pure mathematics , programming language
The determination of flow state remains an important challenge in non‐perennial stream catchments. To identify periods of flow and no‐flow, previous studies deployed temperature sensors on streambed surfaces and interpreted the resulting time series data using a moving standard deviation approach. However, this technique requires the specification of multiple, subjective constraints. To identify suitable alternative approaches, we tested six time‐frequency analysis methods from three categories: (a) Fourier transform, (b) wavelet transform, and (c) empirical mode decomposition. We compared each of the methods abilities to discern periods of flow from synthetic and field data of streambed temperature time series data. When tested using a synthetically generated data set, the efficacy of methods ranged from moderate to high, with 86%–99% accuracy. When applied to a field data set, greater variability in performance was observed, with 66%–90% accuracy. This accuracy represents a sound ability to determine the percentage of time for which a stream flows and does not flow. However, in the presence of a noisy signal, determining the number of specific flow events as well as correctly identifying timing of activation and cessation remains a challenge that most methods struggled with; this has implications for understanding eco‐hydrological functioning. Differences observed between methods included variations in the ease of implementation and evaluation of results, as well as computational requirements and the ability to handle discontinuous time series data. Based on these results, we suggest five areas for future research to improve the general understanding of time‐frequency analysis techniques amongst practicing hydrologists.