SOMs provide an unsupervised learning-based approach to the analysis of complex, possibly nonlinear geophysical datasets. Initial SOM analysis summarizes data variability into a two-dimensional, spatially organized grid of distinct, generalized patterns, or modes, that collectively provide a discrete, nonlinear summary of the input data space. For example, when applied to geopotential height data, the resulting (discrete) patterns provide a nonlinear classification of the continuum of circulation patterns that can often be directly interpreted. Once defined, patterns are also used to classify the data to study, for example, pattern frequencies at different times, pattern transitions, and trajectories through the grid space.
Applying SOMs to the satellite composite archive is a natural extension of its prior usage in the climate field. SOMs have been applied successfully to a wide variety of climate questions in recent years, e.g., characterizing North Atlantic sea level pressure, evaluating IPCC climate models in the polar regions, and studying Antarctic sea-ice extent. SOMs have not been as widely used for image analysis, a task which brings unique problems not found with, e.g., ERA-40 gridded atmospheric data. As such, we are also exploring whether SOMs are in fact suited to this particular dataset. One finding is the importance of preprocessing to improve the data signal in the imagery. For example, we are exploring a number of cloud masking algorithms (e.g., simple thresholding) to better identify the areas of the image we want the SOM to consider.
Early results for June 2004 data based on a 512x512 pixel domain covering the Amundsen-Bellingshausen Sea region (an area previously shown to feature cloud streaming into West Antarctcia) show the viability of the SOM-based approach. The extracted SOM patterns have clear structure with, in many cases, unique and interpretable cloud patterns. Classifying the data by SOM pattern shows that the SOM successfully groups images from different periods that share similar cloud patterns. Likewise, the SOM readily captures persistence, i.e., similar cloud patterns in serial timesteps map to the same SOM pattern.
We anticipate being able to provide significant, new and unique cloud-based climate insights as we expand the analysis time period and improve the analysis methodology.
![](images/flag.png)
![](images/blue_ribbon.png)