In an Empirical Orthogonal Function (EOF) decomposition the normalisation of the EOF maps and PC time series is arbitarry. Consider the definition

E(x,t) = ∑_{i=1^N Ei(x) pi(t)}

One can multiply the EOF E_i(x) by an arbitrary factor A as long as you divide the time series p_i(t) by the same factor A.

You can choose between four normalisation schemes, in addition to applying no normalisation to the output products. These are a completely separate set of normalisations from the input normalisations, and are applied to the calculated data. For details on input data normalisation by std dev, see here.

Spatial normalisation options are applied to the EOF maps before calculating the PC time series. Time Series normalisations are computed after the PC time series. The PC time series are divided by the scaling factor, and the EOF maps multiplied by it.

Maximum methods normalise by the maximum absolute value in the given data (EOF map for spatial, PC for time series).

Variance methods normalise by the standard deviation of the data (EOF map for spatial, PC for time series). More details on each of the options and their usage:

• Spatial maximum. This implies that all EOF plots will run from -1 to +1, which makes them easy to plot. The PC time series will have the units of the original field and smaller amplitudes for higher values of EOF number i.
• Spatial variance. This resembles the previous option, but makes the EOFs and PCs more comparable if there are few isolated areas with large amplitude in the EOF.
• Time series maximum. Not very useful, included for symmtery.
• Time series variance. This option implies that the EOF maps have the units of the original field. They will have a smaller amplitude for larger EOF number i.