Tropical Indian Ocean variability revealed by self-organizing maps

Abstract

The tropical Indian Ocean climate variability is investigated using an artificial neural network analysis called self-organizing map (SOM) for both observational data and coupled model outputs. The SOM successfully captures the dipole sea surface temperature anomaly (SSTA) pattern associated with the Indian Ocean Dipole (IOD) and basin-wide warming/cooling associated with ENSO. The dipole SSTA pattern appears only in boreal summer and fall, whereas the basin-wide warming/cooling appears mostly in boreal winter and spring owing to the phase-locking nature of these phenomena. Their occurrence also undergoes significant decadal variation. Composite diagrams constructed for nodes in the SOM array based on the simulated SSTA reveal interesting features. For the nodes with the basin-wide warming, a strong positive SSTA in the eastern equatorial Pacific, a negative Southern Oscillation, and a negative precipitation anomaly in East Africa are found. The nodes with the positive IOD are associated with a weak positive SSTA in the central equatorial Pacific or positive SSTA in the eastern equatorial Pacific, a positive (negative) sea level pressure anomaly in the eastern (western) tropical Indian Ocean, and a positive precipitation anomaly over East Africa. The warming in the central equatorial Pacific appears to correspond to El Niño Modoki discussed recently. These results suggest usefulness of SOM in studying large-scale ocean–atmosphere coupled phenomena.

Appendix

To examine whether SOM and EOF can capture two climate modes with the same variance, we have generated an artificial SSTA data by

$$ SSTA{\left( {x,y,t} \right)} = pc1{\left( t \right)} \times EOF1(x,y) + pc2{\left( t \right)} \times EOF2(x,y) + noise{\left( {x,y,t} \right)}. $$

Here, pc1(t) and pc2(t) are the first two normalized principal components, and EOF1(x, y) and EOF2(x, y) are the normalized eigenfunctions or spatial patterns of the first two EOF modes of the tropical Indian Ocean shown in Fig. 12a. Note that we made the random noise noise(x,y,t) to contribute to about 10% of the total variance and each of the first two modes explains about 45% of the total variance as they are normalized.

When the SOM method is applied to this data, an SOM array, which is very similar to Fig. 2, is obtained (figure not shown). However, as is clear from Fig. 13, the EOF method failed to capture the original SSTA pattern for the first two modes shown in Fig. 12. The first EOF mode in Fig. 13 has a larger SSTA in the southeastern tropical Indian Ocean, whereas that in Fig. 12a has a larger SSTA in the west. We note that this mixture occurs even when more than three modes have a similar variance and rotated EOF could not provide a remedy for this particular problem (see also Behera et al. 2003). A similar pitfall was also met by Dommenget and Latif (2002).

Fig. 13

First two EOF modes obtained from the artificial SSTA data. Values in parentheses show the variance contribution of each mode