Nonparametric Classification of Satellite Images

Abstract

The task of classifying the objects on a satellite image into predefined categories is the topic of the article. The problems arising while designing a practicable classifier are discussed. The general conditions for robustness of a classifier are provided. To solve the problems mentioned, a robust classification approach is proposed aiming at completely nonparametric unsupervised clustering with consequent association of the clusters with target categories using multiple sources of the testing and training data. The nonparametric clustering used is primarily based on ranking and grouping. Completely nonparametric cluster union and cleaning procedures are presented; theoretical basics for other parts of the approach are provided. The software implementation and complexity of the methodology are discussed. The approach aims at getting the highest possible classification accuracy under real conditions for images with more than 100 million pixels.