INVESTIGATION OF HYPERSPECTRAL FEATURE SYSTEMS FOR CLASSIFICATION OF NATURAL AND ANTHROPOGENIC OBJECTS
S.M. Borzov, E.S. Nezhevenko, S.I. Orlov, O.I. Potaturkin, S.B. Uzilov
Institute of Automation and Electrometry, Siberian Branch, Russian Academy of Sciences, Novosibirsk, Russia
Keywords: object detection, reflection spectra, spectral channels, vegetation indices, classification methods and algorithms, image processing
Abstract
A method for processing a set of hyperspectral data in order to form a representative system of features is proposed, and object classification is carried out using 7 different systems of features in narrow spectral intervals (30 nm) of the visible and near IR spectral regions based on measured spectral brightness coefficients (SBC). It is shown that it is advisable to use systems of three features for classification of 12 types of vegetation and camouflage coverings. At the same time, traditional vegetation indices often used for plant research provide insufficiently high accuracy of classification of objects of selected types. Simultaneous use of two difference indices is more effective in comparison with them. However, the best classification accuracy is provided by systems of three features, which are integrated values of the SBC in specially selected spectral ranges. Note that the classification accuracy is equal to or close to 100% in almost all cases when classifying objects into two classes.
|
