State‐space discrimination and clustering of atmospheric time series data based on Kullback information measures

Abstract

SUMMARYStatistical problems in atmospheric science are frequently characterized by large spatio-temporal data sets and pose difficult challenges in classification and pattern recognition. Here, we consider the problem of identifying geographically homogeneous regions based on similarities in the temporal dynamics of weather patterns. Two disparity measures are proposed and applied to cluster time series of observed monthly temperatures from locations across Colorado, U.S.A. The two disparity measures are based on state-space models, where the monthly temperature anomaly dynamics and seasonal variation are represented by latent processes. Our disparity measures produce clusters consistent with known atmospheric flow structures. In particular, the temporal anomaly pattern is related to the topography of Colorado, where, separated by the Continental Divide, the flow structures in the western and eastern parts of the state have different dynamics. The results further suggest that seasonal variation may be affected by locally changing solar radiation levels primarily associated with elevation variations across the Rocky Mountains. The general methodology is outlined and developed in the Appendix. We conclude with a discussion of extensions to time varying and non-stationary systems.

Climate Zone Delineation: Evaluating Approaches for Use in Natural Resource Management

“…2, Map). In recent years, new clustering methods have addressed challenges posed by the seasonal structure and autocorrelation inherent in atmospheric time series data (Bengtsson and Cavanaugh 2008;Lund and Li 2009). However, because the cluster analyses here do not rely on time series data, instead using only a matrix of monthly averages (normals), we used more traditional climate clustering methods (Fovell and Fovell 1993;Unal and others 2003) and instead employed bootstrapping (described below) to assess the sensitivity of our results to data artifacts.…”

Time series clustering by a robust autoregressive metric with application to air pollution

“…In this framework, cluster analysis is used to classify fields obtained from observed data to identify "prototype" of spatial patterns. By considering a functional representation, Bengtsson and Cavanaugh [54] modeled the observed time series in a state space setup and classified the sites via hierarchical clustering methods relying on disparity measures based on Kullback information. Kim et al [55] employed k-means clustering for classifying sites based on the temporal fluctuation of PM2.5.…”

Looking for similar patterns among monitoring stations. Venice Lagoon application

“…Recent literature provides some new classification methods based on Functional Data Analysis (FDA, Ramsey and Silverman (2005); Ferraty and Vieu (2006)). In the context of the analysis of atmospheric time series, Bengtsson and Cavanaugh (2008) model the observed time series in a state space setup and classify the sites via hierarchical methods relying on disparity measures based on Kullback information. Ignaccolo et al (2008) apply a partition around medoids method to site classification in the air quality network of Piemonte (Italy).…”