A Review of Self-Organizing Map Applications in Meteorology and Oceanography

Coupled ocean-atmosphere science steadily advances with increasing information obtained from long-records of in situ observations, multiple-year archives of remotely sensed satellite images, and long time series of numerical model outputs. However, the percentage of data actually used tends to be low, in part because of a lack of efficient and effective analysis tools. For instance, it is estimated that less than 5% of all remotely sensed images are ever viewed by human eyes or actually used (Petrou, 2004). Also, accurately extracting key features and characteristic patterns of variability from a large data set is vital to correctly understanding the interested ocean and atmospheric processes (e.g., Liu & Weisberg, 2005). With the increasing quantity and type of data available in meteorological and oceanographic research there is a need for effective feature extraction methods. The Self-Organizing Map (SOM), also known as Kohonen Map or Self-Organizing Feature Map, is an unsupervised neural network based on competitive learning (Kohonen, 1988, 2001; Vesanto & Alhoniemi, 2000). It projects high-dimensional input data onto a low dimensional (usually two-dimensional) space. Because it preserves the neighborhood relations of the input data, the SOM is a topology-preserving technique. The machine learning is accomplished by first choosing an output neuron that most closely matches the presented input pattern, then determining a neighborhood of excited neurons around the winner, and finally, updating all of the excited neurons. This process iterates and fine tunes, and it is called self-organizing. The outcome weight vectors of the SOM nodes are reshaped back to have characteristic data patterns. This learning procedure leads to a topologically ordered mapping of the input data. Similar patterns are mapped onto neighboring regions on the map, while dissimilar patterns are located further apart. An illustration of the work flow of an SOM application is given in Fig. 1. The SOM is widely used as a data mining and visualization method for complex data sets. Thousands of SOM applications were found among various disciplines according to an early survey (Kaski et al., 1998). The rapidly increasing trend of SOM applications was reported in Oja et al. (2002). Nowadays, the SOM is often used as a statistical tool for multivariate analysis, because it is both a projection method that maps high dimensional data to lowdimensional space, and a clustering and classification method that order similar data patterns onto neighboring SOM units. SOM applications are becoming increasingly useful in geosciences (e.g., Liu and Weisberg, 2005), because it has been demonstrated to be an