Task-Oriented GAN for PolSAR Image Classification and Clustering.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Computer Vision

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) image interpretation faces challenges due to data complexity and limited labeled samples.
• Generative Adversarial Networks (GANs) offer data modeling without assumptions but lack task-specific orientation.
• Existing methods struggle with the small sample problem in PolSAR data analysis.

Purpose of the Study:

• To propose a novel Task-Oriented GAN for enhanced PolSAR image interpretation.
• To address the challenges of PolSAR data analysis and the small sample problem.
• To improve PolSAR image classification and clustering using a task-specific GAN framework.

Main Methods:

• A Task-Oriented GAN is developed, incorporating a generator (G-Net), discriminator (D-Net), and a task-specific network (TaskNet or T-Net).
• T-Net functions as a classifier or clusterer, guiding G-Net to generate task-beneficial synthetic data.
• The method integrates specific PolSAR information into the GAN architecture for improved data mining.

Main Results:

• Task-Oriented GAN successfully generates task-specific fake data to augment limited training sets, mitigating overfitting.
• The proposed method demonstrates strong performance in PolSAR image classification and clustering tasks, even with small amounts of labeled data.
• Visualized comparisons confirm the effectiveness of T-Net in generating relevant fake data compared to standard GANs.

Conclusions:

• Task-Oriented GAN overcomes the task-free limitation of traditional GANs, offering a powerful tool for PolSAR image analysis.
• The approach effectively addresses the small sample problem in PolSAR interpretation by leveraging task-specific synthetic data generation.
• The method enables researchers to mine inherent PolSAR data information without prior data hypotheses, advancing the field.