A Hyperspectral Image Classification Method Based on Multi-Discriminator Generative Adversarial Networks.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral remote sensing images (HSIs) offer rich spectral information for various applications.
• Deep learning, particularly Generative Adversarial Networks (GANs), shows promise for HSI classification.
• Existing GANs face challenges like spectral ambiguity, gradient vanishing, and limited sample quality.

Purpose of the Study:

• To address limitations in GAN-based HSI classification, specifically pattern collapse and diversity deficiency.
• To propose and evaluate a Multi-Discriminator Generative Adversarial Network (MDGANs) for improved HSI classification.
• To investigate the impact of the number of discriminators on classification performance across different HSI datasets.

Main Methods:

• Implemented a semi-supervised classification framework using MDGANs on three hyperspectral datasets.
• Introduced a multi-discriminator collaboration mechanism to enhance the rigor and accuracy of sample generation.
• Analyzed the influence of varying numbers of discriminators on the model's judgment ability and sample quality.

Main Results:

• MDGANs demonstrated improved judgment capabilities and more accurate spectral sample generation compared to single-discriminator GANs.
• The proposed method effectively mitigated noise in generated spectral samples, enhancing overall classification performance.
• The number of discriminators had a variable impact on classification results, depending on the specific hyperspectral dataset.

Conclusions:

• Multi-discriminator collaboration in GANs is a viable strategy for improving hyperspectral image classification.
• MDGANs offer a robust solution to common GAN limitations, leading to more accurate and reliable HSI analysis.
• Further research can explore optimal discriminator configurations for diverse HSI classification tasks.