Search Results

You are looking at 1 - 10 of 12 items for

  • Author or Editor: Alexey Kaplan x
  • All content x
Clear All Modify Search
Elizabeth C. Kent and Alexey Kaplan

Abstract

A method is developed to quantify systematic errors in two types of sea surface temperature (SST) observations: bucket and engine-intake measurements. A simple linear model is proposed where the SST measured using a bucket is cooled or warmed by a fraction of the air–sea temperature difference and the SST measured using an engine intake has a constant bias. The model is applied to collocated nighttime observations made at moderate wind speeds, allowing the effects of solar radiation and strong vertical gradients in the upper ocean to be neglected. The analysis is complicated by large random errors in all of the variables used. To estimate coefficients in this model, a novel type of linear regression, where errors in two variables are correlated with each other, is introduced. Because of the uncertainty in a priori estimates of the error covariance matrix, a Bayesian analysis of the regression problem is developed, and maximum likelihood approximations to the posterior distributions of the model parameters are obtained.

Results show that the temperature change in bucket SST resulting from the air–sea temperature difference can be detected. The analysis suggests that bucket SST may be in error by a fraction from 0.12° ± 0.02° to 0.16° ± 0.02°C of the air–sea temperature difference. When this temperature change of the bucket SST is accounted for, a warm bias in engine-intake SST in the mid- to late 1970s and the 1980s was found to be smaller than that suggested by previous studies, ranging between 0.09° ± 0.06° and 0.18° ± 0.05°C. For the early 1990s the model suggests that the engine-intake SSTs may have a cold bias of −0.13° ± 0.07°C.

Full access
Jason E. Smerdon, Alexey Kaplan, and Diana Chang

Abstract

The regularized expectation maximization (RegEM) method has been used in recent studies to derive climate field reconstructions of Northern Hemisphere temperatures during the last millennium. Original pseudoproxy experiments that tested RegEM [with ridge regression regularization (RegEM-Ridge)] standardized the input data in a way that improved the performance of the reconstruction method, but included data from the reconstruction interval for estimates of the mean and standard deviation of the climate field—information that is not available in real-world reconstruction problems. When standardizations are confined to the calibration interval only, pseudoproxy reconstructions performed with RegEM-Ridge suffer from warm biases and variance losses. Only cursory explanations of this so-called standardization sensitivity of RegEM-Ridge have been published, but they have suggested that the selection of the regularization (ridge) parameter by means of minimizing the generalized cross validation (GCV) function is the source of the effect. The origin of the standardization sensitivity is more thoroughly investigated herein and is shown not to be associated with the selection of the ridge parameter; sets of derived reconstructions reveal that GCV-selected ridge parameters are minimally different for reconstructions standardized either over both the reconstruction and calibration interval or over the calibration interval only. While GCV may select ridge parameters that are different from those that precisely minimize the error in pseudoproxy reconstructions, RegEM reconstructions performed with truly optimized ridge parameters are not significantly different from those that use GCV-selected ridge parameters. The true source of the standardization sensitivity is attributable to the inclusion or exclusion of additional information provided by the reconstruction interval, namely, the mean and standard deviation fields computed for the complete modeled dataset. These fields are significantly different from those for the calibration period alone because of the violation of a standard EM assumption that missing values are missing at random in typical paleoreconstruction problems; climate data are predominantly missing in the preinstrumental period when the mean climate was significantly colder than the mean of the instrumental period. The origin of the standardization sensitivity therefore is not associated specifically with RegEM-Ridge, and more recent attempts to regularize the EM algorithm using truncated total least squares could theoretically also be susceptible to the problems affecting RegEM-Ridge. Nevertheless, the principal failure of RegEM-Ridge arises because of a poor initial estimate of the mean field, and therefore leaves open the possibility that alternative methods may perform better.

Full access
Alexey Kaplan, Yochanan Kushnir, and Mark A. Cane

Abstract

Near-global 4° × 4° gridded analysis of marine sea level pressure (SLP) from the Comprehensive Ocean–Atmosphere Data Set for monthly averages from 1854 to 1992 was produced along with its estimated error using a reduced space optimal interpolation method. A novel procedure of covariance adjustment brought the results of the analysis to the consistency with the a priori assumptions on the signal covariance structure. Comparisons with the National Centers for Environmental Prediction–National Center for Atmospheric Research global atmosphere reanalysis, with the National Center for Atmospheric Research historical analysis of the Northern Hemisphere SLP, and with the global historical analysis of the U.K. Meteorological Office show encouraging skill of the present product and identifies noninclusion of the land data as its main limitation. Marine SLP pressure proxies are produced for the land stations used in the definitions of the Southern Oscillation and North Atlantic Oscillation (NAO) indices. Surprisingly, they prove to be competitive in quality with the land station records. Global singular value decomposition analysis of the SLP fields versus sea surface temperature identified three major patterns of their joint large-scale and long-term variability as “trend,” Pacific decadal oscillation, and NAO.

Full access
Chie Ihara, Yochanan Kushnir, Mark A. Cane, and Alexey Kaplan

Abstract

The relationship between all-India summer monsoon rainfall (ISMR) and the timing of (El Niño–Southern Oscillation) ENSO-related warming/cooling is investigated, using observational data during the period from 1881 to 1998. The analysis of the evolutions of Indo-Pacific sea surface temperature (SST) anomalies suggests that when ISMR is not below normal despite the co-occurrence of an El Niño event, warming over the eastern equatorial Pacific starts from boreal winter and evolves early so that the western-central Pacific and Indian Ocean are warmer than normal during the summer monsoon season. In contrast, when the more usual El Niño–dry ISMR relationship holds, the eastern equatorial Pacific starts warming rapidly only about a season before the reference summer so that the western-central Pacific and Indian Oceans remain cold during the monsoon season.

Full access
Jason E. Smerdon, Alexey Kaplan, and Daniel E. Amrhein

Abstract

The commenters confirm the errors identified and discussed in Smerdon et al., which either invalidated or required the reinterpretation of quantitative results from pseudoproxy experiments presented or used in several earlier papers. These errors have a strong influence on the spatial skill assessments of climate field reconstructions (CFRs), despite their small impacts on skill statistics averaged over the Northern Hemisphere. On the basis of spatial performance and contrary to the claim by the commenters, the Regularized Expectation Maximization method using truncated total least squares (RegEM-TTLS) cannot be considered a preferred CFR technique. Moreover, distinctions between CFR methods in the context of the discussion in the original paper are immaterial. Continued investigations using accurately described and faithfully executed pseudoproxy experiments are critical for further evaluation and improvement of CFR methods.

Full access
Jason E. Smerdon, Alexey Kaplan, and Daniel E. Amrhein

Abstract

Pseudoproxy experiments evaluate statistical methods used to reconstruct climate fields from paleoclimatic proxies during the Common Era. These experiments typically employ output from millennial simulations by general circulation models (GCMs). It is demonstrated that multiple published pseudoproxy studies have used erroneously processed GCM surface temperature fields: the NCAR Community Climate System Model (CCSM), version 1.4, field was incorrectly oriented geographically and the GKSS ECHO-g FOR1 field was corrupted by a hemispheric-scale smoothing in the Western Hemisphere. These problems are not associated with the original model simulations; they instead arose because of incorrect processing of the model data for the pseudoproxy experiments. The consequences of these problems are evaluated for the studies in which the incorrect fields were used. Some quantitative results are invalidated by the findings: these include all experiments that used the corrupted ECHO-g field and those aspects of previous CCSM experiments that focused on Niño-3 reconstructions. Other results derived from the CCSM field can be reinterpreted based on the information provided herein and their qualitative characteristics remain similar.

Full access
Alicia R. Karspeck, Alexey Kaplan, and Mark A. Cane

Abstract

The seasonal and interannual predictability of ENSO variability in a version of the Zebiak–Cane coupled model is examined in a perturbation experiment. Instead of assuming that the model is “perfect,” it is assumed that a set of optimal initial conditions exists for the model. These states, obtained through a nonlinear minimization of the misfit between model trajectories and the observations, initiate model forecasts that correlate well with the observations. Realistic estimates of the observational error magnitudes and covariance structures of sea surface temperatures, zonal wind stress, and thermocline depth are used to generate ensembles of perturbations around these optimal initial states, and the error growth is examined. The error growth in response to subseasonal stochastic wind forcing is presented for comparison.

In general, from 1975 to 2002, the large-scale uncertainty in initial conditions leads to larger error growth than continuous stochastic forcing of the zonal wind stress fields. Forecast ensemble spread is shown to depend most on the calendar month at the end of the forecast rather than the initialization month, with the seasons of greatest spread corresponding to the seasons of greatest anomaly variance. It is also demonstrated that during years with negative (and rapidly decaying) Niño-3 SST anomalies (such as the time period following an El Niño event), there is a suppression of error growth. In years with large warm ENSO events, the ensemble spread is no larger than in moderately warm years. As a result, periods with high ENSO variance have greater potential prediction utility.

In the realistic range of observational error, the ensemble spread has more sensitivity to the initial error in the thermocline depth than to the sea surface temperature or wind stress errors. The thermocline depth uncertainty is the principal reason why initial condition uncertainties are more important than wind noise for ensemble spread.

Full access
Jason E. Smerdon, Alexey Kaplan, Diana Chang, and Michael N. Evans

Abstract

Canonical correlation analysis (CCA) is evaluated for paleoclimate field reconstructions in the context of pseudoproxy experiments assembled from the millennial integration (850–1999 c.e.) of the National Center for Atmospheric Research Community Climate System Model, version 1.4. A parsimonious method for selecting the order of the CCA model is presented. Results suggest that the method is capable of resolving multiple (3–13) climatic patterns given the estimated proxy observational network and the amount of observational uncertainty. CCA reconstructions are compared to those derived from the regularized expectation maximization method using ridge regression regularization (RegEM-Ridge). CCA and RegEM-Ridge yield similar skill patterns that are characterized by high correlation regions collocated with dense pseudoproxy sampling areas in North America and Europe. Both methods also produce reconstructions characterized by spatially variable warm biases and variance losses, particularly at high pseudoproxy noise levels. RegEM-Ridge in particular is subject to significantly larger variance losses than CCA, even though the spatial correlation patterns of the two methods are comparable. Results collectively indicate the importance of evaluating the field performance of methods that target spatial climate patterns during the last several millennia and indicate that the results of currently available climate field reconstructions should be interpreted carefully.

Full access
Jason E. Smerdon, Alexey Kaplan, Diana Chang, and Michael N. Evans
Full access