Cloud Classi cation Using Error-Correcting Output Codes

Novel arti cial intelligence methods are used to classify 16x16 pixel regions (obtained from Advanced Very High Resolution Radiometer (AVHRR) images) in terms of cloud type (e.g., stratus, cumulus, etc.). We previously reported that intelligent feature selection methods, combined with nearest neighbor classi ers, can dramatically improve classi cation accuracy on this task. Our subsequent analyses of the confusion matrices revealed that a small number of confusable classes (e.g., cirrus and cirrostratus) dominated the classi cation errors. We conjectured that, if the class labels in the data were re-represented so that these cloud classes are more easily distinguished, then additional accuracy gains might result. We explored this hypothesis by replacing each class label with a set of error-correcting output codes, a general technique applicable to any classi cation algorithm for tasks with at least three classes. Our initial results are promising; error correcting codes signi cantly increased classi cation accuracy compared with using standard representations for class labels. To our knowledge, this is the rst successful integration of k-nearest neighbor classi ers and error-correcting output codes (i.e., where k is, e ectively, small). One conclusion is that environmental scientists should always select, for their classi cation tasks, a classi er that reduces both variance and learning bias.