Classification of Hyperspectral Images Using a Combination of Features Extracted From the Weighted Local Kernel Matrix of Spectral and Fractal Features

In recent years, the use of hyperspectral imagery in various fields of Earth science, especially in remote sensing, has significantly increased due to its rich spectral information. However, the classification of these images and the extraction of useful information from them present variues challenges. These challenges include the effective management of high-dimensional data and the achievement of accurate classification when the number of training samples is limited. One of the primary objectives of the remote sensing scientific community has been to improve the accuracy of image classification, thereby facilitating comprehensive investigations of surface phenomena and changes. In recent years, there has been a growing interest in the use of spatial features as a means of improving the classification accuracy of hyperspectral images. Numerous methods have been suggested for the spectral-spatial classification of hyperspectral images. Currently, research is being conducted with the objective of developing simpler yet more accurate methodologies. The existence of intricate relationships between different bands of the hyperspectral image, as evidenced by research in the field of machine vision, has prompted the development of a novel methodology in current research for modelling the complex relationships between spectral and spatial features within a hyperspectral image. The main objective of this article is to present a novel and efficient approach that combines features derived from weighted local kernel matrices of spectral and fractal characteristics for hyperspectral image classification.

 In the present research, hyperspectral images are first subjected to a dimension reduction step. Subsequently, spatial features are generated based on the directional fractal dimension, and these features are further reduced in dimension. In the subsequent stage, the novel features are derived from the weighted local kernel matrices of both the spectral and fractal feature groups. These secondary features consider nonlinear local dependencies between spectral and fractal characteristics, which were not previously considered in other feature generation methods. Ultimately, this stage serves to enhance the accuracy of the classification process. The resulting feature vectors from both groups are then merged, creating a comprehensive vector that is rich in spectral-spatial information for each pixel. Finally, the support vector machine (SVM) algorithm is employed to classify the obtained feature vector and assign labels to each pixel. The experiments conducted as part of this research were carried out on two real hyperspectral benchmark images: one depicting Indian pine and the other the University of Pavia.

The analysis of the outcomes demonstrates the effectiveness of the proposed approach, which incorporates features derived from weighted local kernel matrices of both spectral and fractal characteristics. The classification accuracy of both the Indian Pine and University of Pavia images is enhanced by 20% and 18%, respectively, compared to the exclusive use of spectral features. These findings confirm that incorporating spatial information significantly enhances classification accuracy, particularly in scenarios with limited training samples. Furthermore, the results demonstrate that the proposed method exhibits superior accuracy compared to other studies conducted in this domain.

The enhanced performance of the proposed method in comparison to other competitors can be attributed to the incorporation of local non-linear dependencies between both spectral and fractal features, which have not been considered in previous studies. In the future, further improvements to the proposed approach are anticipated. Firstly, efforts will be made to optimise the efficiency of the proposed method in terms of processing time. Furthermore, the accuracy of the method will be enhanced by considering additional fractal features in subsequent steps. These refinements will be pursued in future research endeavours.