A study on weighing a semi-labelled sample using spectral and spatial context information

Abstract

ABSTRACT Classification of spectrally similar objects is a hard task, mainly when using moderate resolution data. Even though hyperspectral data are a useful source of information, the Hughes phenomenon is highlighted when limited number of training samples are used. For data classification and to mitigate this drawback, the number of training samples needs to be increased in the methodology. In this study, we investigated the estimation of the weights of semi-labelled samples using spectral and spatial context information by relaxation process in a two-steps methodology. The weights of semi-labelled samples in parametric classifier were estimated iteratively in the first step using spectral information only. In the second step, addition of spatial information was done by a relaxation process. This study investigated a more refined approach, improved by the inclusion of spatial context information in the relaxation process. The aim of this work was to mitigate the Hughes phenomenon and improve the separation of similar classes. The proposed methodology was tested using the data (hyperspectral image) from a study area, where the land cover classes are spectrally similar and the accurate separation of these classes was a hard task. Even though several experiments were performed, only a selected number of representative experiments are presented in this work. The results showed that the inclusion of context information can be used for the successful mitigation of the Hughes phenomenon allowing almost twice the number of bands used and increase the classification overall accuracy by up to 8%.