Symmetric All Convolutional Neural-Network-Based Unsupervised Feature Extraction for Hyperspectral Images...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep learning excels at feature extraction (FE) for hyperspectral image (HSI) processing.
• HSI datasets often lack sufficient labeled samples for training deep learning models.

Purpose of the Study:

• To propose a novel unsupervised deep-learning-based FE method for HSI data.
• To address the challenge of limited labeled samples in HSI analysis.

Main Methods:

• An end-to-end trained framework with symmetric encoder and decoder subnetworks.
• Utilizes 3-D convolutional and deconvolutional neural networks for spectral-spatial feature extraction.
• Incorporates a novel cost function with a sparse regular term to enhance feature robustness.

Main Results:

• The proposed method achieves robust and effective feature representation.
• Experimental results demonstrate the method's efficacy on publicly available HSI datasets.
• Successfully overcomes the limitations of parameter-heavy 3-D kernels.

Conclusions:

• The unsupervised deep learning approach provides a viable solution for HSI feature extraction with limited labeled data.
• The developed method yields high-quality features suitable for downstream classification tasks.
• Offers a promising direction for advancing HSI data analysis.