Methods for Starting the Detection of Undocumented Multiple Changepoints

• Published in: Journal of climate Vol. 21; no. 18; pp. 4887 - 4899
• Main Authors: Gérard-Marchant, Pierre G. F., Stooksbury, David E., Seymour, Lynne
• Format: Journal Article
• Language: English
• Published: Boston, MA American Meteorological Society 15.09.2008
• Subjects: Algorithms; Changes; Climate change; Climatology. Bioclimatology. Climate change; Datasets; Detection; Drought; Earth, ocean, space; Economic models; Estimates; Estimation methods; Exact sciences and technology; External geophysics; Geophysics. Techniques, methods, instrumentation and models; Hurricanes; Hydrologic data; Least squares; Linear regression; Metadata; Meteorology; Methods; NOTES AND CORRESPONDENCE; Precipitation; Regression analysis; Robustness (mathematics); Statistical variance; Statistics; Stream discharge; Stream flow; Streamflow data; Studies; Temperature; Time series; Trends; United States; Georgia; Sweden, Malmoehus, Lund; USA, Georgia; USA, Georgia, Flint R; time series analysis; algorithms; Climatology; Data analysis; hydrology; Change detection; Regression model; Change point; case studies; North America; streamflow
• ISSN: 0894-8755; 1520-0442
• DOI: 10.1175/2008JCLI1956.1

Four algorithms are given, as a first step toward the practical detection of undocumented multiple changepoints. These algorithms are based on the two-phase regression method of Lund and Reeves, as well as the robust method of Lazante. The result of each method is a set that contains statistically detectable changepoints; each candidate is then either independently validated as a changepoint or discarded. This is demonstrated and the methods are compared on artificial data, and then the methods are implemented on streamflow data from the Flint River in southwest Georgia. Most notably, the method based on two-phase regression was able to detect a well-known yet undocumented drop in streamflow from a local drought that no other methods have so far been able to detect.