The CoSA algorithm seeks to automatically identify land cover categories in an unsupervised fashion. A -means clustering algorithm is used on undercomplete approximations of image patches (i.e., feature vectors), which in turn were found using the learned dictionaries. The main steps of the CoSA algorithm are summarized in 24, and some initial results on Barrow image data were shown in Refs. 23 and 25. Since we have two types of learned dictionaries, extracted from multispectral and index band data, each with four different spatial resolutions, we have a large number of clustering scenarios to consider. One important step is determining the number of clusters necessary for good classification, from a domain expert point of view. In this work, a range of fixed number of cluster centers, , is considered for each of the four spatial resolutions, and few millions of multispectral patches randomly drawn from the training area are used to learn the cluster centers. The cluster training performance is evaluated as detailed in 31, from a quantitative (minimum distance convergence), as well as a qualitative (meaningful grouping of land cover features) standpoint. The trained cluster centers are used to generate land cover labels at the same spatial pixel resolution of the original satellite image. Specifically, for every pixel in the image, a clustering classification label is given based on its surrounding context (i.e., pixel patch centered on the respective pixel). In other words, we use a step-size of 1 pixel to extract overlapping patches for classification and do not extend image borders.