Developing a Historical Precipitation Record - Satellite-based Applications on Climate Change - page 96

variations. This suggests that the modern gauge sampling is adequate to resolve
most variations spanned by the set of EOF modes used. Therefore, a blended EOF-
based reconstruction can be formed using REOF(GPCP) to update the historical
analysis. For the historical analysis, the EOF-based reconstruction from CRU gauge
data was used, gradually shifting from the historical analysis to REOF(GPCP) data
over 1979 to 1988. We call this reconstruction the REOF(blend).
7.4 An Improved CCA-Based Reconstruction
An indirect reconstruction method based on canonical correlation analysis (CCA)
was developed using analyses of SST and sea-level pressure (SLP). The CCA
(Barnett and Preisendorfer 1987 ) was adapted for reconstructing precipitation
(Smith et al. 2009a ). The method, referred to as RCCA, is summarized here. The
RCCA specifies precipitation anomalies from SST and SLP anomalies, which are
both much better sampled over the oceans. Annual averages are specified using
annual predictors. For annual average anomalies, both SST and SLP are related to
large-scale precipitation variations. Our goal is to use the superior oceanic sampling
of the predictors to more reliably determine oceanic precipitation multi-decadal
variations. The RCCA of annual-average precipitation anomalies is computed
75 S-75 N over both land and ocean regions, although we are most concerned
with improving the oceanic multi-decadal signal. In polar latitudes, both predictor
data and satellite precipitation training data are not reliable enough for meaningful
reconstructions.
As with the EOF-based reconstruction, GPCP anomalies are used to form base
statistics. The predictors include a SST analysis (Smith et al. 2008a ) and a SLP
analysis (Allan and Ansell 2006 ). These predictor analyses allow the reconstruction
to extend back to 1900. The SLP analysis is through 2004, with updates afterwards.
Because the SLP updates are computed differently from their historical analysis, the
update variance is larger than in the historical period. Therefore, the RCCA base
period used is 1979-2004 to keep that artificial variance change out of its statistics.
The predictor and GPCP anomalies are averaged annually before computing the
reconstruction statistics. The annual predictor fields are then used to reconstruct the
annual anomalies beginning 1900.
This indirect RCCA can resolve large-scale precipitation variations, including
multi-decadal variations. Reliability is shown by comparison over land regions
where there are independent data for validation. Gauge data are included in the
GPCP base data, so for 1979-2004 the validation is not independent. Outside that
period, no gauge data are used in the RCCA, so the gauges can be used for
independent validation before 1979. To avoid sampling differences, averages of
the annual RCCA and GHCN gauge analysis are computed over only regions with
gauge data (Fig. 7.1 ). The GPCP base data averaged over gauge regions are also
included for comparison. The strong correlation between the RCCA and the GHCN
in the independent period shows the ability of this method to resolve both