Hiperspektral görüntüler için spektral-uzamsal en yakın altuzay sınıflandırıcıları

Abstract

Spectral-spatial nearest subspace classifiers for hyperspectral images

The nearest subspace classifiers that make effective use of spectral and spatial information for hyperspectral image classification are proposed within the scope of this thesis. The methods are proposed, inspired by the geometric interpretation of the nearest subspace classifier method. Nearest Subspace Classifier (NSC) is a simple classifier that works on the assumption that samples from the same class approximately lie on the same subspace. Based on this assumption, NSC method analyzes the closeness between the test sample and the space spanned by the within-class training samples. Then, the label of test sample is assigned to the closest class. However, NSC only considers the spectral information and neglects the spatial information. In addition to the assumption in the NSC method, we add another assumption that spatially adjacent pixels quite likely belong to the same class. By combining these two assumptions, we conclude that spatially adjacent pixels must approximately lie on the same subspace as well. Based on these assumptions, we propose the spectral-spatial nearest subspace classification (SSNSC) approach as a generic classification framework that performs classification by analyzing the closeness between the subspace spanned by the samples used for spatial information and subspace spanned by the within-class training samples. In this thesis, Canonical Correlation Analysis is used to measure the closeness between these two subspaces.

In this thesis, we developed four different methods that are specialized versions of the SSNSC approach. They differ from each other based on how they utilize spatial information. The first one uses spatial window as spatial information, and other three use superpixels. The methods are tested on the mucilage datasets as well as the datasets commonly used in the literature.

The strength of the proposed classification approach is that i) the geometric interpretation of the proposed approach is meaningful and easy to understand, ii) it does not need long training times, (iii) it has a closed-form solution, (iv) it is simple to test and does not require laborious parameter-tuning processes, (v) it outperforms the existing solutions when the number of labelled training samples is scarce, vi) it is robust to noisy and outlier pixels.