Control of Storminess over the Pacific and North America by Circulation Regimes

Abstract

Circulation regimes are identified from a cluster analysis applied to 5-day means of the combined anomalies of 500 hPa geopotential height and the 250 hPa zonal wind over the extended Pacific-North America region obtained from ERA-Interim reanalyses. Five regimes are identified: the Arctic Low (AL), Pacific wavetrain (WT), Alaskan Ridge (AR) and Pacific Trough (PT) and Arctic High (AH). Regime dependent anomalies of the frequency of atmospheric rivers and measures of the storm tracks (850 hPa covariance between high pass filtered meridional wind and temperature and 300 hPa variance of high pass filtered meridional wind) are shown to be significant. The circulation regimes provide statistically significant modulation of observed precipitation (obtained from the Multi-Source Weighted-Ensemble data set). We examine regime-specific anomalies, normalized anomalies and the occurrence of extreme values of precipitation. We identify regions in which the ratio of frequency of occurrence of extremes (5th and 95th percentile) to the climatological frequency in each regime is significantly high. Precipitation is decreased over the Pacific Northwest in the AL regime, but is increased over the central and Eastern US, with a large increase in the probability of extreme precipitation, enhanced atmospheric river occurrence and stronger storm tracks. In the WT regime, Pacific atmospheric rivers are shifted towards Alaska, accompanied by a decrease in storm track magnitude over the Eastern North Pacific. The AR regime is characterized by an increase in moisture flux into the Southwest, increases in extreme precipitation and atmospheric rivers and storm track related heat transport. Increases in precipitation, and atmospheric rivers along the Southwest coast are noted in the PT regime, along with increased precipitation over Texas and Florida. The Northeastern region experiences increases in extremely dry conditions. Significant drought conditions occur over the Midwest US. Significant increases in precipitation over a broad region of the Northwest occur during the AH regime.

Appendix: Clustering details

1.1 Algorithm

Let S be the sum of squared distances between each state (expressed in PC space) and the centroid of the cluster to which it is assigned. Let the cluster separation $\varDelta$ be the sum of squared distances of each cluster centroid about the origin, weighted by the cluster population. Since one can easily show that $(N-1) \sigma ^{2} = S + \varDelta$, where N is the number of states and $\sigma ^{2}$ the variance of all N points (which is fixed for the given data set), maximizing the ratio $R = \varDelta /S$ is equivalent to minimizing S.

The iterative algorithm to seek the minimum of S is initiated by randomly identifying k initial points called seeds, whose coordinates form the first estimate of the cluster centroids. Each of the N points in the data set is then assigned to the seed to which it is closest, so that k groups (clusters) are formed. Each of the k centroids (average coordinates) are then recomputed based on all members in that cluster, and with this new definition of the centroids, the cluster assignments are recomputed. This iterative process of re-defining clusters proceeds until S has converged to what is assumed to be a minimum within the set of all possible choices of k clusters. Following a method suggested by Molteni (personal communication) one can ensure more rapid convergence by choosing the k seed points so that they are truly representative of the entire set of points. In PC space, this is accomplished by ensuring that each seed has a norm less than a maximum value in the full PC space, that each pair of seeds has a minimum distance from each other, and that each pair of seeds belongs to a different sector in the plane of the leading two PCs.

1.2 Synthetic time series for PCs

The temporal correlation structure of any time series can be captured in a synthetic time series by computing the Fourier harmonic coefficients of the original time series for all frequencies, retaining the original amplitudes, but randomizing the phases for each frequency. This random-phase approach (Christiansen 2007) preserves the spectrum. Since (for an infinite time series) the spectrum is the Fourier transform of the temporal autocovariance function (Jenkins and Watts 1968), the synthetic time series should retain the original temporal correlation information.This approach needs to be modified when applied to daily data for many winters, thus to discontinuous segments: the random-phase calculation is applied to the time series of seasonal means, and separately to the deviations of the daily data from those seasonal means. Finally the two synthetic time series are added (Straus 2010). Figure 16 shows the auto-covariance function of the original leading PC (in black curves) from the reanalysis and the results from the random-phase method (blue curve).

Fig. 16

Lagged Autocovariance of leading Principle component of the reanalysis data (black lines) and average over 100 simulations of the synthetic PC stochastic process (blue line). The red lines indicate plus and minus one standard deviation limits from the stochstic process