Skip to Main Content
This paper presents a new semiautomated soft classification method that is a hybrid between supervised and unsupervised classification algorithms for the classification of remote sensing data. Continuous iterative guided spectral class rejection (IGSCR) (CIGSCR) is based on the IGSCR classification method, a crisp classification method that automatically locates spectral classes within information class training data using clustering. This paper outlines the model and algorithm changes necessary to convert IGSCR to use soft clustering to produce soft classification in CIGSCR. This new algorithm addresses specific challenges presented by remote sensing data including large data sets (millions of samples), relatively small training data sets, and difficulty in identifying spectral classes. CIGSCR has many advantages over IGSCR, such as the ability to produce soft classification, less sensitivity to certain input parameters, potential to correctly classify regions that are not amply represented in training data, and a better ability to locate clusters associated with all classes. Furthermore, evidence is presented that the semisupervised clustering in CIGSCR produces more accurate classifications than classification based on clustering without supervision.