Model Validation Through SOM-based Pattern Analysis: Evaluating HadAM3 GCM Hindcasts Over Southern Africa
Abstract
The decomposition of the time-space structure in atmospheric data has long been a core analysis technique for feature identification and interpretation. Most common of these methods has been EOF analysis and cluster analysis. However, the linear and orthogonal nature of these methods inherently imposes some restrictions on the results. Nonlinear Primary Component Analysis may be used to extend EOF techniques, while Self Organizing Maps (SOMs) may be used to extend clustering techniques. Such technique developments are of particular importance for improving seasonal forecasting by evaluating the shorter time frame of model performance (as opposed to the more common temporal means and anomalies). In this analysis SOMs are applied to time series of spatial fields from the NCEP reanalysis and the HadAM3 GCM forced with observed SSTs. Using 6-hourly data the SOM finds reference vectors that span the high-dimensional input data space, and forms a mapping of this high-dimensional space to a 2-dimensional lattice. In effect the procedure identifies arch-types in the data. The original data fields may then be associated with the reference vectors, in effect performing a clustering of the data against the reference vectors identified. The 2-dimensional mapping may then be used to effectively visualize the ordered relations of the input data. In this manner the relative frequency of occurrence of synoptic events may be determined, as well as the temporal evolution of synoptic events, and the velocity of the synoptic evolution. Using this approach the GCM data for the summer season over southern Africa (DJF) is evaluated against the NCEP reanalysis for the same period. This allows bias and error in the model to be identified in terms of circulation modes and evolution.
- Publication:
-
AGU Spring Meeting Abstracts
- Pub Date:
- May 2001
- Bibcode:
- 2001AGUSM...A41A07H
- Keywords:
-
- 3309 Climatology (1620);
- 3319 General circulation;
- 3337 Numerical modeling and data assimilation;
- 3364 Synoptic-scale meteorology