Climate field reconstruction of North Atlantic sea surface temperatures

<p>Pseudo-proxy experiments test the skill and sensitivity of two extended climate field reconstruction (CFR) methodologies in reconstructing northern North Atlantic Summer sea surface temperatures (SSTs). The Summer target data originate from one millennium-long simulation of the CESM LME (Otto-Bliesner et al. 2016) .  The experiments test the reconstruction skill systematically for input data mimicking SST marine proxies and instrumental observations, including characteristics such as sparse distribution in space, varying signal-to noise ratio and age uncertainties.</p><p>The Bayesian hierarchical model BARCAST assumes implicitly that the target variable is described as an AR(1) process in time (Tingley & Huybers 2010), while the proxy surrogate reconstruction (PSR) method makes no such assumption (Graham et al. 2007). The PSR selects climate analogues from the simulated instrumental period in our study.</p><p>Results show that both methodologies generate skillful reconstructions for perfectly dated input data, and the PSR is superior when realistic noise levels are chosen for the input data. When the input is perturbed with age-uncertainties, the methodologies are unable to generate acceptable skillful reconstructions. Facilitating in form of data clustering is tested for both methodologies in the attempt of improving reconstruction skill. This proves successful for the PSR methodology, with the best skill obtained using n=3 clusters over the reconstruction region.</p><p>Additionally, and addressed as a topic for discussion, we detect weak temporal persistence in the input data and the BARCAST reconstructions. The lack of SST persistence is found to be partly due to the input data sampling frequency: Summer means (June, July, August) averaged for every year. Analyses show that the simulated SST data exhibit weaker memory from one Summer to the next, compared to year-to-year variability based on annual means. Similar results are also found for instrumental observations. This finding stands in contrast to results of previous studies on terrestrial reconstruction, where climate reconstructions and individual proxy records exhibit strong persistence properties, also targeting the Summer season (Werner et al. 2018, Nilsen et al. 2018)<sup>,</sup>.</p><p> </p><p>References:</p><p>Graham, N.E. et al. (2007), Clim. Change, 83, 241-285, doi: 10.1007/s10584-007-9239-2</p><p>Otto-Bliesner, B.L. et al (2016), Bull. Amer. Meteor. Soc., 97, 735-754, doi: 10.1175/BAMS-D-14-00233.1</p><p> Nilsen, T. et al. (2018), Clim. Past, 14, 947-967, doi: 10.5194/cp-14-947-2018</p><p>Tingley, M. P. and P. Huybers (2010a), J. Clim., 23, 2759–2781, doi: 10.1175/2009JCLI3015.1</p><p>Werner, J. P. et al. (2018), Clim. Past, 14, 527-557, doi: 10.5194/cp-14-527-2018</p>

Assessing the performance of the BARCAST climate field reconstruction technique for a climate with long-range memory

The skill of the state-of-the-art climate field reconstruction technique BARCAST (Bayesian Algorithm for Reconstructing Climate Anomalies in Space and Time) to reconstruct temperature with pronounced long-range memory (LRM) characteristics is tested. A novel technique for generating fields of target data has been developed and is used to provide ensembles of LRM stochastic processes with a prescribed spatial covariance structure. Based on different parameter setups, hypothesis testing in the spectral domain is used to investigate if the field and spatial mean reconstructions are consistent with either the fractional Gaussian noise (fGn) process null hypothesis used for generating the target data, or the autoregressive model of order 1 (AR(1)) process null hypothesis which is the assumed temporal evolution model for the reconstruction technique. The study reveals that the resulting field and spatial mean reconstructions are consistent with the fGn process hypothesis for some of the tested parameter configurations, while others are in better agreement with the AR(1) model. There are local differences in reconstruction skill and reconstructed scaling characteristics between individual grid cells, and the agreement with the fGn model is generally better for the spatial mean reconstruction than at individual locations. Our results demonstrate that the use of target data with a different spatiotemporal covariance structure than the BARCAST model assumption can lead to a potentially biased climate field reconstruction (CFR) and associated confidence intervals.

How wrong are climate field reconstruction techniques in reconstructing a climate with long-range memory?

The Bayesian hierarchical model BARCAST (Bayesian Algorithm for Reconstructing Climate Anomalies in Space and Time) climate field reconstruction (CFR) technique, and idealized input data are used in the pseudoproxy experiments of this study. Ensembles of targets are generated from fields of long-range memory stochastic processes using a novel approach. The range of experiment setups include input data with different levels of persistence and levels of proxy noise, but without any form of external forcing. The input data are thereby a simplistic alternative to standard target data extracted from general circulation model (GCM) simulations. Ensemble-based temperature reconstructions are generated, representing the European landmass for a millennial time period. Hypothesis testing in the spectral domain is then used to investigate if the field and spatial mean reconstructions are consistent with either the fractional Gaussian noise (fGn) null hypothesis used for generating the target data, or the autoregressive model of order one (AR(1)) null hypothesis which is the assumed temperature model for this reconstruction technique. The study reveals that the resulting field and spatial mean reconstructions are consistent with the fGn hypothesis for most of the parameter configurations. There are local differences in reconstructed scaling characteristics between individual grid cells, and a generally better agreement with the fGn model for the spatial mean reconstruction than at individual locations. The discrepancy from an fGn is most evident for the high-frequency part of the reconstructed signal, while the long-range memory is better preserved at frequencies corresponding to decadal time scales and longer. Selected experiment setups were found to give reconstructions consistent with the AR(1) model. Reconstruction skill is measured on an ensemble member basis using selected validation metrics. Despite the mismatch between the BARCAST temporal covariance model and the model of the target, the ensemble mean was in general found to be consistent with the target data, while the estimated confidence intervals are more affected by this discrepancy. Our results show that the use of target data with a different spatiotemporal covariance structure than the BARCAST model assumption can lead to a potentially biased CFR reconstruction and associated confidence intervals, because of the wrong model assumptions.

Pseudo-proxy evaluation of climate field reconstruction methods of North Atlantic climate based on an annually resolved marine proxy network

Two statistical methods are tested to reconstruct the interannual variations in past sea surface temperatures (SSTs) of the North Atlantic (NA) Ocean over the past millennium based on annually resolved and absolutely dated marine proxy records of the bivalve mollusk Arctica islandica. The methods are tested in a pseudo-proxy experiment (PPE) setup using state-of-the-art climate models (CMIP5 Earth system models) and reanalysis data from the COBE2 SST data set. The methods were applied in the virtual reality provided by global climate simulations and reanalysis data to reconstruct the past NA SSTs using pseudo-proxy records that mimic the statistical characteristics and network of Arctica islandica. The multivariate linear regression methods evaluated here are principal component regression and canonical correlation analysis. Differences in the skill of the climate field reconstruction (CFR) are assessed according to different calibration periods and different proxy locations within the NA basin. The choice of the climate model used as a surrogate reality in the PPE has a more profound effect on the CFR skill than the calibration period and the statistical reconstruction method. The differences between the two methods are clearer for the MPI-ESM model due to its higher spatial resolution in the NA basin. The pseudo-proxy results of the CCSM4 model are closer to the pseudo-proxy results based on the reanalysis data set COBE2. Conducting PPEs using noise-contaminated pseudo-proxies instead of noise-free pseudo-proxies is important for the evaluation of the methods, as more spatial differences in the reconstruction skill are revealed. Both methods are appropriate for the reconstruction of the temporal evolution of the NA SSTs, even though they lead to a great loss of variance away from the proxy sites. Under reasonable assumptions about the characteristics of the non-climate noise in the proxy records, our results show that the marine network of Arctica islandica can be used to skillfully reconstruct the spatial patterns of SSTs at the eastern NA basin.

Pseudo-proxy evaluation of Climate Field Reconstruction methods of North Atlantic climate based on an annually resolved marine proxy network

Two statistical methods are tested to reconstruct the inter-annual variations of past sea surface temperatures (SSTs) of the North Atlantic (NA) Ocean over the past millennium, based on annually resolved and absolutely dated marine proxy records of the bivalve mollusk Arctica islandica. The methods are tested in a pseudo-proxy experiment (PPE) set-up using state-of-the-art climate models (CMIP5 Earth System Models) and reanalysis data from the COBE2 SST data set. The methods were applied in the virtual reality provided by global climate simulations and reanalysis data to reconstruct the past NA SSTs, using pseudoproxy records that mimic the statistical characteristics and network of Arctica islandica. The multivariate linear regression methods evaluated here are Principal Component Regression and Canonical Correlation Analysis. Differences in the skill of the Climate Field Reconstruction (CFR) are assessed according to different calibration periods and different proxy locations within the NA basin. The choice of the climate model used as surrogate reality in the PPE has a more profound effect on the CFR skill than the calibration period and the statistical reconstruction method. The differences between the two methods are clearer for the MPI-ESM model, due to its higher spatial resolution in the NA basin. The pseudo-proxy results of the CCSM4 model are closer to the pseudo-proxy results based on the reanalysis data set COBE2. The addition of noise in the pseudo-proxies is important for the evaluation of the methods, as more spatial differences in the reconstruction skill are revealed. More profound differences between methods are obtained when the number of proxy records is smaller than five, making the Principal Component Regression a more appropriate method in this case. Despite the differences, the results show that the marine network of Arctica islandica can be used to skilfully reconstruct the spatial patterns of SSTs at the eastern NA basin.