A multiscale representation method for efficient hyperspectral data classification

The growing demand for real-time hyperspectral data analysis, especially in remote sensing applications, requires efficient algorithms that can be implemented on edge processing systems with limited computational resources [1, 2]. Multiscale representations, particularly those inspired by or derived from the wavelet transform [3, 4], have widely demostrated their efficiency in signal processing due to their ability to capture both local and global features while reducing redundancies with highly computationally efficient implementations. We present a wavelet-based multiscale characterization method designed for fast and compact representation of hyperspectral signatures. The goal is to facilitate the extraction of meaningful features with minimal computational load. To ensure fast and accurate classification, we integrate machine learning classifiers, optimized for speed and efficiency. Experimental results show that the proposed approach reduces processing time while maintaining classification accuracy; it also shows easy adaptability to new acquisitions, making it a good candidate for on-board processing. References [1] V. Bruni, G. Maiello, G. Monteverde, A. Paglialunga, D. Vitulano, A Wavelet-Based Band Selection Method for Hyperspectral Image Classification, Proc. of WHISPERS 2023 [2] G. De Lucia, M. Lapegna, D. Romano, Unlocking the potential of edge computing for hyperspectral image classification: An efficient low-energy strategy, Future Generation Computer Systems, 2023 [3] V. Bruni, M.L. Cardinali, D. Vitulano, An MDL-Based Wavelet Scattering Features Selection for Signal Classification, Axioms, 2022 [4] S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, 2009.