MULTISCALE KURTOSIS FOR EFFICIENT HYPERSPECTRAL IMAGE CLASSIFICATION

Feature extraction is employed in hyperspectral imaging to reduce spectral redundancy while retaining informative contentand providing discriminative representations of hyperspectral signatures. Efficient and low-cost feature extraction plays a crucial role especially in on-board applications. This paper proposes and investigates a classification approach that employs the kurtosis of wavelet coefficients as a compact and discriminative scale-dependent feature. By inheriting the wavelet transform’s ability to characterize signal singularities through their decay across scales, this feature is evaluated in terms of its discriminative power when applied solely to the spectral component and used as input to machine learning classifiers. Preliminary results suggest that the proposed multiscale features can achieve classification accuracies comparable to those obtained with joint spatial–spectral descriptors, indicating their potential discriminative effectiveness. An additional empirical analysis investigates optimal scale selection for kurtosis computation, further reducing feature dimensionality and computational complexity.