Efficiency-driven 3D CNN architectures for hyperspectral classification

Hyperspectral imaging enables the simultaneous capture of spatial and spectral information across multiple wavelengths, yielding high-dimensional data suitable for a wide range of applications. 3D Convolutional Neural Networks (CNNs) can completely exploit the hyperspectral data structure through 3D convolutional filters, which jointly extract spatial and spectral features. This process improves classification performance by increasing intraclass variation and reducing interclass variation [1]. On the other side, the high computational cost of deep CNN architectures — both in terms of resource consumption and training time — when processing such high-dimensional data necessitates optimization techniques. These can be approached through dimensionality reduction or more efficient network architectures [2]. The former reduces the input dimensionality by transforming the data into a lower-dimensional yet representative form, while the latter focuses on streamlining the network architectures. Two distinct approaches for enhancing hyperspectral classification efficiency using 3D CNNs are proposed. The first method employs feature extraction, projecting the data in a proper domain and automatically selecting relevant components in the transformed space based on the entropic normalized information distance. This approach is an adaptive and automatic method where the number of features to be selected is not pre-defined but automatically given [3]. The second methodology focuses on determining the filters size setting of convolutional layers in a 3D CNN, guided by Heisenberg’s uncertainty principle. This principle inspires a rule for relating the spatial and spectral dimensions of convolutional filters as the network depth increases, enabling the network to learn discriminative features that capture both fine spatial resolution and broad spectral characteristics [4]. The effectiveness of CNNs in the proposed approaches is assessed using both raw and transformed input data. Both the features selected by the entropybased method and the architectures with Heisenberg-based cascaded filter setting demonstrate a significant reduction in training time while preserving high classification accuracy. These strategies provide solutions for processing hyperspectral data, aimed at enhancing operational efficiency.