Using Self-Organizing Maps to evaluate the NASA GISS AR5 SCM at the ARM SGP site

Description

Cluster analyses have gained popularity in recent years to establish cloud regimes using satellite and radar cloud data. These regimes can then be used to evaluate climate models or to determine what large-scale or subgrid processes are responsible for cloud formation. An alternative approach is to first classify the meteorological regimes (i.e. synoptic pattern and forcing) and then determine what cloud scenes occur. In this study, a competitive neural network known as the Self-Organizing Map (SOM) is used to classify synoptic patterns over the Southern Great Plains (SGP) region to evaluate simulated clouds from the AR5 version of the NASA GISS ModelE single-column model (SCM). In detail, 54-class SOMs have been developed using North American Regional Reanalysis (NARR) variables averaged to 2 x 2.5 degree latitude-longitude grid boxes for a region of 7 x 7 grid boxes centered on the ARM SGP site. Variables input into the SOM include mean sea-level pressure and the horizontal wind components, relative humidity, and geopotential height at the several standard pressure levels. These SOMs are produced for the winter (DJF), spring (MAM), summer (JJA), and fall (SON) seasons during the period 1999–2009. This synoptic typing will be associated with observed cloud fractions and forcing properties from the ARM SGP site and then used to evaluate simulated clouds from the SCM. SOMs provide a visually intuitive way to understand their classifications because classes are related to each other in a two-dimensional space. An example of a feature map for meteorological features at 925 hPa is provided in Figure 1. ARM ARSCL and GISS SCM mid-level (2–6 km) cloud fractions are given in Figure 2. It is clearly seen that the majority of the mid-level clouds occur in association with low-pressure systems. Despite producing clouds during the right conditions, the SCM produces too few mid-level clouds during cyclonic activity.