Parallel Connected Generative Adversarial Network with Quadratic Operation for SAR Image Generation and Application...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep Convolutional Neural Networks (CNNs) excel in computer vision but struggle with Synthetic Aperture Radar (SAR) image classification due to limited labeled data and differing imaging mechanisms.
• Existing methods face challenges in accurately classifying SAR images compared to optical imagery.

Purpose of the Study:

• To enhance SAR image classification by generating additional labeled SAR data using specialized Generative Adversarial Networks (GANs).
• To address the limitations of insufficient labeled SAR data and the unique characteristics of SAR imaging for improved classification performance.

Main Methods:

• Proposed novel GAN architectures, PWGAN and CNN-PGAN, designed specifically for SAR image generation.
• Incorporated quadratic operations and statistical characteristics of SAR images into the GAN framework.
• Developed parallel GAN structures with multiple generators and discriminators tailored to target categories.

Main Results:

• Experimental validation on the TerraSAR-X single polarization dataset demonstrated the effectiveness of the proposed GAN models.
• The generated SAR data significantly improved the performance of SAR image classification tasks.
• The specialized GANs showed superior ability in representing SAR data characteristics.

Conclusions:

• The proposed GAN-based approach effectively addresses the challenge of limited labeled SAR data for classification.
• The novel GAN architectures, PWGAN and CNN-PGAN, offer a promising solution for enhancing SAR image analysis.
• This work contributes to advancing machine learning applications in remote sensing with improved SAR image classification.