Laplacians of the 1951 to 1980 MOHSST4 climatology (Bottomley et al., 1990) were calculated for data-void 1° boxes.
SSTs for the data-void 1° boxes were then computed by solving Poisson's equation with the Laplacians as forcing and the values in data-filled boxes as boundary conditions, as has been done by Reynolds (1988) and Bottomley et al. (1990) in constructing their climatology.
The 1° resolution absolute SST data output by the Laplacian interpolation were converted to anomalies, and the anomalies were smoothed 1:2:4:6:4:2:1 east-west, then north-south. Anomalies rather than actuals were smoothed, to avoid smoothing strong SST gradients. After smoothing, the anomalies were converted back to SSTs. SST gradients were thus retained with approximately 1° resolution, whereas the anomalies, which vary much more slowly geographically, were retained on about their original 5° resolution.
Examples of the final result of this process, in the form of anomalies for January 1878 and January 1983, are shown in Figures 1d and 3b (see color well). These fields (and others not shown) suggest that in future versions of GISST, the anomalies should be smoothed with a spatially lower-pass filter. Even in the 1980s, monthly (as opposed to seasonal) in situ SST anomalies are often unreliable on a 5° space scale (Folland et al., 1993).
A sampling experiment was carried out to assess the reliability of the analysis technique when used on sparse data. In it, the complete analysis, including the Laplacian stage, was repeated for each month of the El Niño years 1982 to 1983, but with the basic SST data prior to the ''filling and quality control" stage omitted from 5° boxes where there were no data in the corresponding months of the El Niño years 1877 to 1878. In this experiment, January 1877 corresponds to January 1982, and December 1878 corresponds to December 1983. Also, data for 1981 and 1984 needed for this experiment were reduced to the coverages available in the corresponding months of 1876 and 1879. Figures 4a, b, c, and d, which should be compared with Figures 3a and b, illustrate the results of the experiment for January 1983. Figure 5a shows the two-year mean difference between the reduced-data analysis and GISST 1.0. (All are in the color well.)
Biases are largest in the eastern tropical Pacific, where the analysis of the reduced data base underestimated the strength of the 1982 to 1983 warm El Niño event. This underestimation exceeded 1°C in some places at the peak of the event in late 1982 to early 1983 (compare Figures 3b and 4d). It is likely therefore that GISST 1.0 itself will have underestimated the strength of the 1877 to 1878 El Niño, owing to the sparser data base then. Figure 5b presents the root-mean-square (rms) difference field between the GISST 1.0 and the reduced-data analyses; it is based on 24 monthly values at each location. The rms differences (which include a bias component) are less than 0.5°C over most of the Atlantic and Indian Ocean north of 40°S, but exceed 0.5°C over much of the Pacific and the Southern Ocean, and are over 1°C in the eastern tropical Pacific, where the reduced-data analyses underestimate the El Niño warmth. Cosine-latitude-weighted summary statistics for the globe and for the east tropical Pacific (20°N to 20°S, east of 170°W) are given in Table 2. The global rms difference (which also includes a bias component) is around 0.5°C before and after the El Niño but reaches 0.7°C in early 1983, when the underestimation in the east tropical Pacific averages nearly 0.7°C (compare also Figures 3b and 4d). However, the global field correlations rise from about 0.55 before the event to about 0.7 during the event and maintain this level, possibly because the El Niño intensifies the overall global anomaly pattern signal. During the El Niño the global bias reaches -0.15°C. For the two-year period as a whole, however, the global bias is only -0.03"C, suggesting, in accord with the "frozen grid" analyses in Bottomley et al. (1990) and Folland et al. (1990), that the sparse coverage available in 1877-1878 does not severely prejudice estimation of annual or multi-annual global mean SST anomalies. Similar, but apparently slightly better, results (not shown) were obtained when a corresponding experiment was made using the El Niño years 1972-1973 with the 1877-1878 coverage. The apparent improvement may have resulted from the slightly reduced coverage of observations in 19721973 relative to 1982-1983.
In a further, longer experiment the analysis for 1981 to 1990 was repeated using coverage for 1881 to 1890. Biases (not shown) were within ±0.1°C in most of the Atlantic and the Indian Ocean, and generally exceeded ± 0.3°C only in the Southern Ocean and parts of the North Pacific. Hemispheric and global average biases were within ±0.01°C. Root-mean-square differences generally exceeded 0.5°C only in the latter areas and in parts of the tropical Pacific and the far northern Atlantic. Correlations (Figure 6. again in the color well) based on 120 values at each location, exceeded 0.8 over most of the Atlantic, much of the eastern Pacific, and much of the Indian Ocean. Low correlations over the Southern Ocean and the central and northern Pacific emphasize the need to acquire more historical data for these regions.
These sampling experiments only assess the impact of reduced areal coverage of 5° box SST data. They do not provide any measure of the effects of increased scatter in the individual monthly 5° box values resulting from, for