Unsupervised Band Selection Based on Covariance Matrix for Hyperspectral Image Classification

Abstract

<title>Abstract</title> Spectral-spatial classification in hyperspectral imagery has garnered significant interest due to the potential to leverage detailed spectral and spatial information for improved accuracy. However, noisy, or closely spaced spectral bands in original hyperspectral images can detrimentally affect classification, often leading to their exclusion based on expert knowledge. In this paper, we propose an unsupervised band selection method utilizing a covariance matrix for hyperspectral image classification. First, spectral band analysis is performed as a preprocessing step for the proposed method. Second, the absolute difference between the spectral wavelets for the different classes is calculated to select the minimal number of bands based on the maximum absolute difference. After that, the covariance matrix is extracted from the selected band to select the most important bands. Third, the global feature was extracted from the original hyperspectral image (HSI) using selected bands determined by a deep convolutional neural network (CNN) model. Finally, SoftMax and support vector machine (SVM) algorithms were employed for classifying the different classes within the HSI images. To evaluate the effectiveness of our approach, it was compared against contemporary methods using two widely used public HSI datasets: Indian Pines and Salinas-A. The experimental results demonstrate convincingly that our band selection technique outperforms alternative methods in terms of classification accuracy.