Browsing by Subject "random projection"

Sort by: Order: Results:

Now showing items 1-2 of 2
  • Seitola, Teija (Finnish Meteorological Institute, 16-1)
    Finnish Meteorological Institute Contributions 126
    The ability of climate models to simulate the climate variability is of great importance when considering the reliability of, for instance, multiannual or longerterm predictions. The aim of this thesis is to study the 20th century lowfrequency variability patterns in the Earth system and how these patterns are represented by the current modelling systems. Another, equally important objective is to enable efficient spatiotemporal analysis of highdimensional climate data sets. Decadal scale variability and predictability, from the point of view of the Nordic region, is also summarised in this study. The work is based on the nearsurface temperature of two 20th century reanalyses, obtained from the NOAA/OAR/ESRL and ECMWF, and historical climate model simulations from the coupled model intercomparison project 5 (CMIP5) data archive. In addition, a millennial Earth system model simulation is analysed. The analysis relies on a powerful dimensionality reduction method, called random projection (RP), which is introduced as a preprocessing for highdimensional climate data sets to enhance or enable the analysis. The spectral decomposition of the data sets is based on randomised multichannel singular spectrum analysis (RMSSA), which is one of the main achievements of this thesis. It is shown that dimensionality reduction obtained by RP preserves the main spatial and temporal patterns with high accuracy. In addition, RMSSA is shown to provide an efficient tool for identifying different variability modes in highdimensional climate data sets. This study shows that the 20th century variability patterns in the two reanalysis data sets are very similar. It is also shown that none of the studied climate models can closely reproduce all the variability modes identified in the reanalyses, although many aspects are simulated well. Taking into account the rapidly accumulating amount of data and increasing dimensionality of data sets, RP is a promising method for dimensionality reduction. The results of the model evaluation can be useful in model development due better understanding of the deficiencies in representing the lowfrequency modes. In addition to nearsurface temperature, it would be a natural extension to include more variables in the analysis, especially because RP allows efficient data compression.
  • Seitola, Teija; Silen, Johan; Järvinen, Heikki (2015)
    In this article, we introduce a new algorithm called randomised multichannel singular spectrum analysis (RMSSA), which is a generalisation of the traditional multichannel singular spectrum analysis (MSSA) into problems of arbitrarily large dimension. RMSSA consists of (1) a dimension reduction of the original data via random projections, (2) the standard MSSA step and (3) a recovery of the MSSA eigenmodes from the reduced space back to the original space. The RMSSA algorithm is presented in detail and additionally we show how to integrate it with a significance test based on a red noise null-hypothesis by Monte-Carlo simulation. Finally, RMSSA is applied to decompose the 20th century global monthly mean near-surface temperature variability into its low-frequency components. The decomposition of a reanalysis data set and two climate model simulations reveals, for instance, that the 2-6 yr variability centred in the Pacific Ocean is captured by all the data sets with some differences in statistical significance and spatial patterns.