The use of small training sets containing mixed pixels for accurate hard image classification: Training on mixed spectral responses for classification by a SVM

The accuracy of a supervised image classification is a function of the training data used in its generation. It is, therefore, critical that the training stage of a supervised classification is designed to provide the necessary information. Guidance on the design of the training stage of a classification typically calls for the use of a large sample of randomly selected pure pixels in order to characterise the classes. Such guidance is generally made without regard to the specific nature of the application in-hand, including the classifier to be used. The design of the training stage should really be based on the classifier to be used since individual training cases can vary in value as can any one training set to a range of classifiers. It is argued here that the training stage can be designed on the basis of the way the classifier operates and with emphasis on the desire to separate the classes rather than describe them. An approach to the training of a support vector machine (SVM) classifier that is the opposite of that generally promoted for training set design is suggested. This approach uses a small sample of mixed spectral responses drawn from purposefully selected locations (geographical boundaries) in training. The approach is based on mixed pixels which are normally masked-out of analyses as undesirable and problematic. A sample of such data should, however, be easier and cheaper to acquire than that suggested by conventional approaches. This new approach to training set design was evaluated against conventional approaches with a set of classifications of agricultural crops from satellite sensor data. The main result was that classifications derived from the use of the mixed spectral responses and the conventional approach did not differ significantly, with the overall accuracy of classifications generally ∼92%.