final_program_abstracts[1]

11 IMSC Session Program
Quantifying uncertainties associated with calibrating proxy
data against instrumental records – an ensemble approach
Monday - Parallel Session 3
Ailie Gallant 1 , Joelle Gergis1 and Karl Braganza 2
1 School of Earth Sciences, University of Melbourne, Australia
2 Australian Bureau of Meteorology, Australia
One of the goals of rigorous scientific analysis is to quantify the uncertainty
associated with the data. In palaeoclimate reconstructions, uncertainty can arise from
a variety of sources including in the measurement of the climate proxy and the
assumptions of stationarity and linearity between the palaeodata and the climate
variable. It is important to address these uncertainties when aiming to produce reliable
estimates of the range of past climate variability. When developing palaeoclimate
reconstructions, proxy data is often tested and/or calibrated using an instrumental
record. Present practice normally identifies a single calibration period against which
the record is ‘tuned’ and subsequently tested against an independent verification
period. If the reconstruction shows skill in representing the instrumental data during
this verification period, it is often recalibrated using the entire period of overlap with
the instrumental data.
We argue that there is a trade off between improving the performance of the proxy
transfer function and potentially over-fitting reconstructions to the instrumental
record. Therefore, we propose a methodology whereby the uncertainty associated with
calibrating the proxy data to the instrumental record is included in the final
reconstruction by using a simple bootstrapping technique to create an ensemble of
reconstructions based on a suite of randomly selected calibration periods.
The technique is demonstrated using a newly developed multi proxy rainfall
reconstruction for southeast Australia (1796–1989) with 80 overlapping years of
instrumental data. Using a Monte-Carlo approach, n years within this overlap period
were randomly sampled to provide a series of years for calibration, while the
remaining years were used for validation. This was performed 10,000 times to
produce an ensemble of reconstructions. The final rainfall reconstruction was
calculated as the mean of the 10,000-member ensemble, with the 95th and 5th
percentiles of the distribution used to provide robust uncertainty estimates.
By using multiple, randomly selected calibration periods, we argue that the potential
of over-fitting the reconstruction to one period from the instrumental record is
reduced. As rainfall displays more stochastic variability than temperature, we suggest
that this approach helps minimise calibration biases associated with the assumption of
linearity. Using this technique, we found that the suite of reconstructions showed
variations of over one standard deviation from the ensemble mean, proving that the
contribution of the differences associated with the choice of the calibration period is
not insignificant.
Abstracts 76