HKNAS: Classification of Hyperspectral Imagery Based on Hyper Kernel Neural Architecture Search

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Neural Architecture Search (NAS) has advanced hyperspectral image (HSI) classification.
• Current NAS methods often optimize architectures independently from network weights, leading to increased search time and performance limitations.

Purpose of the Study:

• To address the limitations of existing NAS methods in HSI classification.
• To develop a more efficient and effective NAS approach for HSI classification tasks.

Main Methods:

• Propose a novel hyper kernel design to directly generate structural parameters, converting dual optimization into single-tier optimization.
• Develop a hierarchical multimodule search space using only convolutional operations integrated into unified kernels.
• Combine hyper kernel search with 3-D convolution decomposition for flexible architecture generation.

Main Results:

• Achieved state-of-the-art results on six public HSI datasets.
• Demonstrated significant reductions in search costs compared to previous NAS methods.
• Generated diverse networks for both pixel-level and image-level HSI classification using 1-D and 3-D convolutions.

Conclusions:

• The proposed hyper kernel-based NAS approach offers a more efficient and flexible solution for HSI classification.
• This method overcomes the limitations of independent architecture and weight optimization in NAS.
• The developed techniques achieve superior performance and flexibility in HSI classification tasks.