Dimensionality reduction in hyperspectral imaging using standard deviation-based band selection for efficient classification

Abstract

Abstract Hyperspectral imaging generates vast amounts of data containing spatial and spectral information. Dimensionality reduction methods can reduce data size while preserving essential spectral features and are grouped into feature extraction or band selection methods. This study demonstrates the efficiency of the standard deviation as a band selection approach combined with a straightforward convolutional neural network for classifying organ tissues with high spectral similarity. To evaluate the classification performance, the method was applied to eleven groups of different organ samples, consisting of 100 datasets per group. Using the standard deviation is an effective method for dimensionality reduction while maintaining the characteristic spectral features and effectively decreasing data size by up to 97.3%, achieving a classification accuracy of 97.21% compared to 99.30% without any processing. Even in comparison with mutual information– and Shannon entropy–based band selection methods, the standard deviation exhibited superior stability and efficiency while maintaining equally high classification accuracy. The results highlight the potential of dimensionality reduction for hyperspectral imaging classification tasks that require large datasets and fast processing speed without sacrificing accuracy.