A Deep Stochastic Adaptive Fourier Decomposition Network for Hyperspectral Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Deep learning excels in hyperspectral image (HSI) classification but requires numerous labeled images and has many parameters.
• Acquiring and labeling HSI data is challenging, exacerbating deep learning's limitations in this domain.

Purpose of the Study:

• To propose a novel deep network architecture, SAFDNet, for HSI classification that addresses the limitations of current deep learning methods.
• To leverage unsupervised feature extraction to reduce the reliance on large annotated datasets and minimize model complexity.

Main Methods:

• Developed SAFDNet, a deep network architecture utilizing stochastic adaptive Fourier decomposition (SAFD) for unsupervised feature extraction.
• Employed fewer convolution kernels within the network to significantly decrease the number of learnable parameters.
• Utilized SAFD, a signal processing tool with a strong mathematical foundation, to build the unsupervised feature extraction mechanism.

Main Results:

• SAFDNet demonstrated strong performance on three popular HSI classification datasets.
• The proposed method requires only a small number of annotated images for effective classifier training.
• SAFDNet significantly reduced the number of parameters compared to other deep learning approaches.

Conclusions:

• SAFDNet offers an effective solution for hyperspectral image classification by overcoming the data and parameter challenges of deep learning.
• The integration of SAFD provides powerful unsupervised feature extraction, enabling high performance with limited labeled data.
• SAFDNet represents a significant advancement in deep learning for HSI analysis, outperforming state-of-the-art methods.