LightSpecNet: Asymmetric Knowledge Distillation for Hyperspectral Image Classification

Hyperspectral image (HSI) classification plays a pivotal role in various fields such as precision agriculture, environmental monitoring, and mineral identification. However, existing methods are often hindered by the high computational complexity and challenges in deploying on edge devices. To address these issues, we introduce LightSpecNet, an asymmetric dual-branch distillation method that balances spectral fidelity with computational efficiency. LightSpecNet processes raw 3D hyperspectral cubes through a 3D convolutional teacher branch to retain fine-grained spectral details, while a lightweight 2 D convolutional student branch operates on PCA-reduced 2D inputs, significantly reducing computational costs by approximately $80 \%$ without substantial information loss. A simplified feature alignment loss, based on mean squared error (MSE), facilitates the direct transfer of spectral knowledge from the teacher to the student, eliminating the need for complex probability distribution matching. This efficient knowledge distillation mechanism not only accelerates the training process but also maintains high classification accuracy. Our extensive experiments demonstrate that LightSpecNet is highly suitable for real-time HSI classification on edge devices, offering a competitive performance while enhancing computational efficiency.