Fig. 2.

SSA. The plots show the different eigenvalues obtained by SSA for each of the four time series ranked by order of importance according to the variance they explain. The dominant eigenvalue in three of the four sites (c, Gikongoro; d, Muhanga; and b, Kabale) corresponds to the trend and is followed by a pair of eigenvalues associated with the seasonal cycle. In Kericho (a), this order is reversed, and the trend corresponds to the subdominant eigenvalue. The trend components account for 10%, 20.5%, 14.7%, and 18% of the total variance, respectively, for Kericho (a), Gikongoro (c), Muhanga (d), and Kabale (b). The respective reconstructed components are shown in Fig. 1 (those for the pair of subdominant eigenvalues in b–d and the dominant pair in a are plotted in Fig. 5). The bars specify the 95% confidence intervals generated with Monte Carlo simulations of red noise. Specifically, the error bars computed for each empirical orthogonal function represent 95% of the range of variance found in the state-space direction defined by that empirical orthogonal function in an ensemble of 999 red-noise realizations. Thus, the bars represent the interval between the 0.5% and 99.5% percentiles, and eigenvalues lying outside this range are significantly different (at the 5% level) from those generated by the red-noise process against which they are tested. The dominant eigenvalues for Gikongoro (c), Kabale (b), and Muhanga (d) and the subdominant eigenvalue for Kericho (a) lie outside this interval and are significantly different from noise. The application of SSA in combination with this red-noise test is known as “Monte Carlo SSA” (38). The ssa toolkit freeware software from www.atmos.ucla.edu/tcd/ssa was used for the analysis. A large order was selected to force a better separation of constituent components mainly at the lower frequencies (22) (a, Kericho; b, Kabale; c, Gikongoro; d, Muhanga).