Fast Multi-dimensional Ensemble Empirical Mode Decomposition for the analysis of Big Spatiotemporal Data Sets

Abstract:

In this big data era, it is more urgent than ever to solve two major issues: (1) fast data transmission method that can facilitate access to data from non-local sources, and (2) fast and efficient data analysis methods that can reveal the key information from the available data for particular purposes. Although approaches in different fields to address these two questions may differ significantly, the common part must involve data compression techniques and fast algorithm.

In this paper, we introduce the recently developed adaptive and spatiotemporally local analysis method, namely the fast multi-dimensional ensemble empirical mode decomposition (MEEMD), for the analysis of large spatiotemporal dataset. The original MEEMD uses ensemble empirical mode decomposition (EEMD) to decompose time series at each spatial grid and then pieces together the temporal-spatial evolution of climate variability and change on naturally separated timescales, which is computationally expensive. By taking the advantage of the high efficiency of the principle component analysis/empirical orthogonal function (PCA/EOF) expression for spatiotemporally coherent data, we design a lossy compression method for climate data to facilitate its non-local transmission. In addition to that, we also explain the basic principles behind the fast MEEMD through decomposing PCs instead of original grid-wise time series to speedup computation of MEEMD. Using a typical climate dataset as an example, we demonstrate that our newly designed methods can (1) compress data with a compression rate of one to two orders; (2) speed up the MEEMD algorithm by one to two orders.