Predicting South African
Seasonal Rainfall Using
A Combination of MOS and
Perfect Prognosis
contributed by Willem A.
Landman1 and Lisa Goddard2
1South African Weather Service
Numerical model global predictions from
the ECHAM4.5 AGCM are downscaled statistically to produce 3-month seasonal
rainfall probabilistic forecasts for 1028 South African rainfall stations.
Method:
GCMs
typically overestimate rainfall amounts and often spatially distort patterns of
rainfall variability. Such systematic biases suggest the need to downscale GCM
simulations. Successful recalibration to regional rainfall over southern Africa
has been achieved using a perfect prognosis approach (Landman et al. 2001) and
a model output statistics approach (MOS) (Landman and Goddard 2002). Here, a
method of empirical downscaling is presented where MOS equations are developed
using 24-member ensemble GCM simulation rainfall data (the ensemble was forced
with simultaneous observed SSTs for each of the 3-month seasons considered) and
then 24-member ensemble rainfall real-time forecast fields at different
lead-times from the same GCM are subsequently used in these MOS equations. It
is therefore assumed that the skill with which the GCM can produce forecast at
lead-times is as good as skill obtained from simulation data, reminiscent to
the assumption of a perfect prognosis approach where “perfect” forecasts are
assumed. The ECHAM4.5 predictions are generated for DJF 2003/04, JFM 2004 and FMA
2004, by persisting observed October 2003 SST anomalies on top of the monthly
varying annual cycle of climatological SSTs. At initialization, ensemble
members differ from each other by one model day integration for both the
simulation and forecast data.
Canonical
correlation analysis (CCA) is the mathematical technique used to set up the MOS
equations (Landman and Goddard 2002) and also to do the “perfect prognosis”
downscaled forecasts with. The predictor (GCM rainfall) and predictand (1028
rainfall stations) are first prefiltered using EOF analysis.
Ranked
probability skill scores (RPSS) of about 0.2 obtained from doing a 33-year
cross-validated MOS with simulation GCM rainfall data during austral summer
months, are higher than the RPSS obtained from “raw” GCM data. Although the
perfect prognosis component of the scheme has not yet been verified, real-time
forecast skill over southern Africa at short lead-times of only a few months
should not be significantly different to scores obtained from using simulation
data (Landman and Goddard 2002).
The Forecast:
Figures
1 to 3 show the probabilistic forecasts produced by the statistical downscaling
system. Forecasts at the 1028 rainfall stations are interpolated to a 2.5° x
2.5° grid using the Cressman analysis method. For DJF 2003/04 (Figure 1), close to equal probabilities are forecast for
each of the three categories of below-normal, near-normal and above normal, except
over the far western areas where high probabilities of above-normal rainfall
are forecast. However, forecasts made for JFM and FMA 2004 (Figures 2 and 3) over the austral
summer rainfall regions show progressively larger probabilities assigned to the
below-normal category. High probabilities (> 60%) of below-normal rainfall
are forecast over the north-eastern areas of the forecast region.
References:
Landman,
W. A., and L. Goddard, 2002: Statistical recalibration of GCM forecasts over
southern Africa using model output statistics. Journal of Climate, 15,
2038-2055.
Landman,
W. A., S. J. Mason, P. D. Tyson, and W. J. Tennant, 2001: Retro-active skill of
multi-tiered forecasts of summer rainfall over southern Africa. International Journal of Climatology, 21, 1-19.
Captions:
Figure 1: Probabilistic forecasts (%) for DJF 2003/04.
Figure 2: Probabilistic forecasts (%) for JFM 2004.
Figure 3: Probabilistic forecasts (%) for FMA 2004.