Assessing the Reliability of Model-Based Estimates of High- and Low-Frequency Variability
Abstract
Model-based estimates of natural internal variability are an integral component of climate change detection and attribution (D&A) studies. Some analysts have claimed that current climate models systematically underestimate the amplitude of observed internal variability on decadal timescales. If this were the case, estimates of the "detection time" - the time required to discriminate a slowly-evolving anthropogenically-forced signal from the background noise of unforced variability - would be biased towards earlier signal detection. A further consequence of such variability errors would be spurious inflation of the ratio between the amplitude of signal and noise trends. It is therefore important to assess the ability of current models to reproduce the amplitude of observed internal variability, particularly on the multi-decadal timescales critical for anthropogenic signal detection. We provide such an assessment using CMIP-3 simulations of forced and unforced climate change. Our specific focus is on global-scale changes in SSTs and the temperature of the lower troposphere (TLT). We perform band- and high-pass filtering of modeled and observed SST and TLT data. The band-pass filtering focuses on variability at timescales of 10 years (with half-power points at 5 and 20 years), while the high-pass filtering has a half-power point at two years, and excludes all variability on timescales > 5 years. For both SST and TLT, we find no evidence that the CMIP-3 models systematically underestimate the amplitude of the variability of observed band-pass filtered data. On average, the CMIP-3 models overestimate observed variability in the 5 to 20-year band by roughly 15 to 20%. This suggests that the S/N ratios and detection times obtained in previous D&A studies with SST and TLT are likely to be conservative estimates. The CMIP-3 models have a weak functional relationship between the amplitude of high- and low-frequency. We conclude from this that model errors in well-observed interannual variability may not provide reliable information on the size and direction of model errors in low-frequency variability. The lesson for model evaluation studies is that model skill in replicating observed temperature variability should be assessed as a function of frequency band, and not with a single metric influenced by variability on multiple timescales.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2011
- Bibcode:
- 2011AGUFMGC14B..01S
- Keywords:
-
- 1610 GLOBAL CHANGE / Atmosphere;
- 1616 GLOBAL CHANGE / Climate variability;
- 1626 GLOBAL CHANGE / Global climate models