Evaluating Depthwise Separable Filters in Heisenberg-Guided 3D CNN for Efficient Hyperspectral Classification

Multidimensional data, such as hyperspectral signatures, provide highly informative content for a wide range of applications. However, their high dimensionality imposes significant computational and memory demands, particularly in real-time or on-board processing scenarios [1]. Consequently, there is a growing need for low-cost, computationally efficient methodologies. Deep learning approaches, especially three-dimensional Convolutional Neural Networks (3D-CNNs), have shown strong performance in hyperspectral image classification by jointly exploiting spatial and spectral information. Nevertheless, their high complexity and extensive training requirements limit deployment in resource-constrained environments. To overcome these challenges, two complementary strategies can be adopted: dimensionality reduction or lightweight architectures that efficiently extract relevant features. This study focuses on the latter, highlighting architectural optimization as a key factor for edge-deployable deep models. Previous research has demonstrated that network design critically affects performance; in particular, 3D-CNNs employing cascaded convolutional filters sized according to the Heisenberg principle can substantially reduce training time while preserving accuracy, particularly when combined with wavelet-based preprocessing [2]. Building on these results, this work focuses on the introduction of separable convolutional filters within Heisenberg-guided 3D-CNN architectures, aiming to further enhance computational efficiency and model compactness [3]. Relationships among filter size, network depth, and data characteristics are also analysed to provide a principled basis for lightweight architectures design. Preliminary experiments confirm that the proposed Heisenberg-guided 3D-CNN employing separable filters can achieve substantial reductions in computational cost without compromising accuracy, thus advancing the development of lightweight and real-time hyperspectral image analysis models. References [1] N. Ghasemi, J. A. Justo, M. Celesti, L. Despoisse, and J. Nieke, Onboard Processing of Hyperspectral Imagery: Deep Learning Advancements, Methodologies, Challenges, and Emerging Trends, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 4780–4790, 2025. [2] V. Bruni, G. Monteverde, and D. Vitulano, Heisenberg Principle-Inspired Filters Size Setting in 3D CNN for Hyperspectral Data Classification, IEEE Proceedings of 14th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing, pp. 1-5, 2024. [3] Z. Zhao, X. Xu, J. Li, S. Li, and A. Plaza, Gabor-Modulated Grouped Separable Convolutional Network for Hyperspectral Image Classification, IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–17, 2023.