Alpert S. Methods for hyperspectral satellite image supervised classification under training samples of limited quality

The aim of the present dissertation is to improve an accuracy of supervised classification hyperspectral satellite images (HSI) under training samples of limited quality, particularly when the samples are contaminated and include only small number of constituent elements. The-state of-the art of approaches to satellite image classification was considered and it was shown, that modern algorithms for HSI supervised classification are very sensitive to the characteristics of training samples. So formulated the aim to develop and justify new methods for HSI supervised classification under the aforementioned conditions. The mathematical platform of new methods for HSI supervised classification is the evidence theory of Dempster-Schafer. The new method for HSI supervised classification under contaminated training samples is proposed. This method differs from the known solutions of same problems by introducing a procedure of assessing classification value of the spectral bands by a special empirical function and by specific approach to partition of the spectral feature space, that allows to use only the most informative spectral bands. A new effective method for HSI supervised classification under simultaneously contaminated training samples and restrictions on their volume was offered in the dissertation also. Both proposed methods for HSI supervised classification are implemented algorithmically and on the software. Assessing the effectiveness of the developed methods was conducted using real satellite images obtained by the Hyperion/EO-1 on-board sensor. Results show that the developed methods are more accurate than known method Support Vector Machine and algorithm for classifying objects based on spectral topological characteristics. The proposed methods can be used for computer analysis of hyperspectral and multispectral satellite imagery in remote searching for minerals and hydrocarbons, solving environmental problems, etc.