Non-Linear, Multi-Variate Analysis of Slow Climate Variations
Collaborator: William Rossow (Remote Sensing Climate Group, RSCG, NOAA/CREST)
Our ultimate goal is to determine the non-linear, multi-variate relationships that express the processes coupling atmosphere, ocean, and land states and circulations and producing slow modes of climate variation. To do that we need analysis methodologies that can handle the data volume and do not limit the analysis to particular space-time scales or to assumptions of linearity and/or low-dimensionality.
The 66-yr NCEP reanalysis provides an example of what a multi-state climate oriented dataset might look like and it is statistically robust enough (large sample size, uniform sampling) so that it can be used to test methods for dealing with various data problems. Thus, the first step towards our goal is to investigate various non-linear multi-variate analysis methods to develop a practical tool kit for such analyses.
Extracting the large scale structure of climate data records presents a number of challenges due to the large size of the data, its high physical and mathematical dimension, anomalies due to artifacts in the data and space and time sampling, methodological data inhomogeneities, and inhomogeneities in units when considering multiple variables. To overcome these challenges, we adapt “best of breed” techniques from data mining and statistical learning.