Nearest neighbor based one-class classification of remote sensing imagery

Abstract

The task of one-class classification is to recognize one specific land-cover class of interest in the remote sensing image. To extract the specific class, the feature space is partitioned into two classes, the class of interest and the other class, with the nearest neighbor classifier. This reduces the effort of training sample selection in the classification. The training samples are selected for the class of interest firstly. Then, training samples of the other class are collected near the samples of the class of interest. As spatial proximity of a sample pair is often correlated to spectral similarity, the spatially adjacent samples of the two classes should create margins to distinguish the specific class of interest from the other class. Using the two kinds of samples, the specific class of interest is classified with nearest neighbor rule. The good performance of one-class classification is validated in the experiment of remote sensing classification.