Dimensionality reduction for large spatio-temporal datasets based on SVD

Many models for spatio-temporal measurements Z(s; t) can be written as a sum of a systematic component and a residual component: Z = M + E. The approach presented here incorporates two Singular Value Decompositions (SVD). The first SVD is applied to the space-time data matrix Z with cross-validation to choose the number ofsmoothed singular vectors to use as temporal basis functions for modelling spatially varying temporal trend in the matrix M. The second SVD is applied to the spatio-temporal matrix E of residuals from the trend models fitted at each site; it represents spatially correlatedshort time scale temporal processes. The remaining stochastic structure is explained by simple autoregressive models fit to the final residuals. The procedure is applied to 30 years of daily temperature data from Sicily.