Semisupervised hyperspectral image classification based on generative adversarial networks and spectral angle...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification faces challenges due to the difficulty and cost of collecting ground truth labels, leading to insufficient training samples.
• Existing generative adversarial network (GAN) methods often use cost functions suitable for 2D natural images, which may not be optimal for the unique spectral characteristics of HSIs.

Purpose of the Study:

• To address the limitations of current methods in hyperspectral image classification with limited labeled data.
• To propose a novel semisupervised algorithm that leverages the spectral features of HSIs for improved classification accuracy.

Main Methods:

• Developed a novel semisupervised generative adversarial network (GAN) algorithm for HSI classification.
• Introduced spectral angle distance (SAD) as a loss function to enhance GAN convergence and spectral fidelity.
• Employed multilayer feature fusion and utilized the discriminator for extracting multiscale features from labeled and unlabeled samples.

Main Results:

• The proposed GAN-based method demonstrated improved convergence speed and generated more realistic spectral samples.
• Multilayer features extracted by the discriminator effectively trained a classifier using limited labeled samples.
• Experimental validation on four diverse hyperspectral datasets (Pavia University, Indiana Pines, Salinas, Tianshan) showed highly promising results compared to state-of-the-art methods.

Conclusions:

• The proposed semisupervised GAN approach with SAD loss is effective for hyperspectral image classification, especially when training data is scarce.
• The method successfully extracts discriminative multiscale spectral features for accurate classification.
• This work offers a promising direction for advancing HSI classification techniques in remote sensing and related fields.