Influence of training sampling protocol and of feature space optimization methods on supervised classification results

Land cover map are produced from remote sensing images using per-pixel or, more recently, object-based classifications. Various trainable classifiers and feature space optimization methods can be used to that aim. The choice of both training and control samples is liable to influence the results according to the classification method employed but little is known about the way of choosing an appropriate sampling set. This makes thus the focal point of our study. Using three sampling methods and four discriminative classifiers we compared various classification procedures, some of them including a feature space optimization step. The one that led to the best results was LDA preceded by its feature pre-selection algorithm. Generally, for training samples, class numbers of 40 were necessary to get the best results.