Possible causes of data model discrepancy in the temperature history of the last Millennium

Abstract

Model simulations and proxy-based reconstructions are the main tools for quantifying pre-instrumental climate variations. For some metrics such as Northern Hemisphere mean temperatures, there is remarkable agreement between models and reconstructions. For other diagnostics, such as the regional response to volcanic eruptions, or hemispheric temperature differences, substantial disagreements between data and models have been reported. Here, we assess the potential sources of these discrepancies by comparing 1000-year hemispheric temperature reconstructions based on real-world paleoclimate proxies with climate-model-based pseudoproxies. These pseudoproxy experiments (PPE) indicate that noise inherent in proxy records and the unequal spatial distribution of proxy data are the key factors in explaining the data-model differences. For example, lower inter-hemispheric correlations in reconstructions can be fully accounted for by these factors in the PPE. Noise and data sampling also partly explain the reduced amplitude of the response to external forcing in reconstructions compared to models. For other metrics, such as inter-hemispheric differences, some, although reduced, discrepancy remains. Our results suggest that improving proxy data quality and spatial coverage is the key factor to increase the quality of future climate reconstructions, while the total number of proxy records and reconstruction methodology play a smaller role.

Figure 1

Comparison of hemispheric model temperatures, pseudoproxies and real proxy reconstructions over the past Millennium. Real proxy temperature reconstructions (black, grey shaded area indicating: 90% ensemble range), model truth (red) and pseudoproxy reconstructions using synthetic proxies based on local (blue) and hemispheric mean (yellow) correlations for the CESM1-CAM5 (member 1) simulation over the past Millennium. (a) SH, (b) NH and (c) standardized NH-SH difference. A 30-year loess filter was applied to all curves. Data in (a) and (b) are centered on the climatological base period of 1961–1990.

Figure 2

Same as Fig. 1 but for the HadCM3 last millennium simulation. Plots for all other simulations in the SM (Supplementary Figs 44–55).

Figure 3

Reconstruction skill and low frequency amplitude. (a) Temporal mean Reduction of Error (RE) skill, in the SH (left) and NH (right). Higher values indicate higher reconstruction skill. (b) Low-frequency amplitude defined as the difference between average temperatures over 1950–1999 (present-day) and 1600–1649 (LIA). Boxplots are across all model simulations and reconstruction ensemble members. Dashed black (red) horizontal lines are the median values of the real proxy experiments (model truth). Bold lines are medians, boxes represent the interquartile range and whiskers the 90% range. Additional skill metrics and PPE as well as the results from each simulation are shown in the SM.

Figure 4

Inter-hemispheric correlations and differences. (a) Correlations between the NH and the SH over the period 1400–1999. Distributions are shown across all model simulations and ensemble members. (b) same but for NH-SH differences. Boxplots are defined as in Fig. 3.

Figure 5

Superposed Epoch Analysis of the temperature response to volcanic eruptions. SH (left) and NH (right) temperature anomalies relative to 5-year pre-eruption means. Lines (shading) represent ensemble medians (9–95% range). Red: model truth, black: real proxies, other colors: PPE. Cyan shading is the 5–95% range of monte-carlo sampled years during non-volcanic periods. Distributions are across all model simulations and ensemble members. Model truth and real proxy data are shown in all panels for better comparison.

Figure 6

Detection and Attribution scaling factors. Amplitude (‘scaling factor’) of the response to all-forcing fingerprints in real proxies and pseudoproxies for the SH (left) and NH (right) over the period 1400–1900. Circles (triangles): HadCM3 (CESM-LME) ensemble members. Symbols represent the ensemble median, vertical lines the 90% range.