Clustering
lModified ISODATA clustering
–Allows for merging and splitting of clusters
lCluster pixels within each cover type (corn, soybeans, rice ...)
–Exclude bad fields
–Principal component clipping to remove outliers
–Merge & split clusters within cover types
lCombine all the cover type clusters back into one signature file
•Known pixels are separated by cover type and clustered, within cover type, using a modified ISODATA clustering algorithm, which allows for both merging and splitting of clusters.
•Clustering is done separately for each cover type (or specified combination of cover types, such as all small grains). The clustered cover types are then assembled together into one signature file.
•Clustering is based on the LARSYS (Purdue University) ISODATA algorithm.  It performs an iterative process to divide pixels into groups based on minimum variance.  The pairs of clusters in close proximity (based on Swain-Fu distance) are merged under a NASS modification.  High variance clusters can be split into two clusters (variance of first principal component is used as a measure).
•The output of any clustering program is a statistics file: stores mean vectors and covariance matrices of final set of clusters for all cover types.
•Three criteria for program termination (i.e. end of cluster merging and splitting):
•No further merges or splits possible.
•A minimum number of clusters (user specified) is reached.
•An infinite loop (self-detecting) is entered.
•An Analysis District can contain one or more scenes, the clustering is performed on all candidate segments found in the scene.  Clustering is an iterative process as it may take a few runs before a clean data set is successfully run through the clustering routine.