Description of Tropical SST Predictions
Data details:
The current version of LIM SST forecasts combines SST data from the Comprehensive Ocean Atmosphere Data Set ( COADS: Woodruff et al. 1993) with data from the NCEP real time surface marine. COADS data is currently available for dates later than the year 2000. However, it is not available in real time. The SST data set prepared in a manner closest to COADS is the NCEP Surface Marine data set, which is provided by NCEP and summarized into COADS-compatible monthly statistics at PSD. The Surface Marine data set begins in 1991. Although the Surface Marine SST data set is extremely similar to COADS in the tropics, the data sets are not identical. Nevertheless, SST statistics in the two data sets are well within the uncertainties of those statistics over their common period, so we developed the LIM prediction model by training it on a combined data set: COADS (1951-1991) and NCEP Surface Marine (1991-2000).
Model details:
Predictions of Global Tropical Sea Surface Temperature Anomalies (SSTAs) are made using linear inverse modeling procedure discussed in Penland and Magorian 1993: J. Climate, 6, 1067-1076). Anomalies were calculated relative to 1951-2000 COADS-NCEP climatology, smoothed by three months running mean procedure and projected onto the leading EOFs containing about 2/3 of the variance. Predictions of the global tropical strip use 20 EOFs; predictions of the tropical IndoPacific alone use 17 EOFs.
Description of tropical SST index predictions:
The prediction model was validated using a jackknifing procedure using six five-year verification periods in the interval between 1970 and 2001. We present the prediction (blue lines) and verification (red lines) time series corresponding to these validation periods for various tropical SST indices (see map). Gaps in the prediction time series are due to the lags incurred between verification periods in the jackknifing procedure. Correlation skill scores are between predictions and verifications as shown. Dotted lines are confidence intervals corresponding to one standard deviation of the expected prediction error ( Penland and Matrosova 2001). Root-mean-square errors (bottom of the page) are normalized to this standard deviation and presented for each index. If the errors are Gaussian, 95% of those errors should lie between the limits (light black lines) shown.
The projection onto the optimal structure ( Penland and Sardeshmukh 1995) is presented in a manner similar to those of the indices. The maps displayed on the optimal structure page are an average of those estimated for each of the jackknifing periods; we verified that no two of these maps were correlated at less than 90%. The optimal structure index at any time is the projection of the SST map onto the optimal structure. We present the Nino 3.4 SST anomaly (blue line) together with the optimal structure index 8 months earlier (red line) and the correlation between them. Also shown is a scatter plot of these time series. The red star compares the current value of Nino 3.4 SST anomaly with the value of the optimal structure index 8 months earlier; the red arrow shows the current optimal structure index.
Interpretation of SST maps:
SSTs are consolidated onto a 4 degrees (latitude) by 10 degrees (longitude) grid. Values of SST and SST anomalies represent the average value within a grid box. Numerical tables of SST predictions are presented in tabular form, with latitudes representing the center of the grid box along the left side, and longitudes representing the center of the grid box along the top. Grid boxes are left blank on the map when the observational record between 1950 and 2000 at that location was considered too sparse to provide useable information in training the statistical prediction model.