Robust classification of remote sensing images

The performance of a supervised classifier is restricted by the “quality” of the training data since outliers will deteriorate the classification accuracy. Unfortunately, their existence is unknown a priori. A robust estimator is capable of minimizing their influence in estimating the sample statistics but the estimate will become less accurate in case no outlier presents. In this sense, the profit of utilizing a robust estimator is by itself a random process. It has been found in the authors´ experiments of land cover classification, however, that significant improvement can be achieved in most cases and no significant degradation has occurred. This is a strong indication that outliers do exist in the data sampled from remote sensing images. And it is highly recommendable that the robust statistics be used in supervised land cover classification