Juris Sinica-Sinavskis defended his PhD thesis “Approaches to an equivalent reduction of multidimensional image spectral bands for object classification.”
Supervisor: Dr.sc.comp., senior researcher Ints Mednieks
The thesis focuses on multidimensional image processing problems associated with the selection of their spectral bands and pixel classification. New, relatively simple, and generic non-supervised EMCR, EXCR, ESCR, ECBG, and supervised XECT band selection procedures for identifying informative spectral band sets in hyperspectral images are proposed. Bands with a higher value of entropy are selected to ensure the informativity of spectral band subsets. The correlation of band images is restricted to avoid redundancy of the subsets.
The methods developed in the thesis have been tested on several hyperspectral images, of which Indian Pines, University of Pavia, Salinas are often used in remote sensing to compare the obtained results. The informativity of the band subsets is characterized by the performance of Bayesian, k-NN, SVM, ELM, and other classifiers. The ECBG procedure provided better results than the alternatives using the SVM classifier.
In the thesis also the solutions for the fusion of two multidimensional images and tree species classification are proposed based on a Bayesian-type classification principle. A simplified method for identification of skin lesions (melanoms and common nevi) is developed using a small number of spectral bands from multispectral images.