Generative adversarial network-based super-resolution reconstruction of remote sensing images.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Satellite imaging is vital for resource management and urban planning.
• Existing super-resolution methods face challenges with image defects, computational inefficiency, and limited generalization.

Purpose of the Study:

• To develop a lightweight framework for enhancing remote sensing image quality.
• To address limitations of current super-resolution techniques in handling image defects and computational demands.

Main Methods:

• Proposed SDGAN (Super-Densely Connected Generative Adversarial Network) framework.
• Integration of a super-dense residual module with 2D convolution kernels.
• Utilized a dual-discriminator system with attention U-Net for improved efficiency.

Main Results:

• Achieved state-of-the-art performance on UCMerced-LandUse and WHU-RS19 datasets.
• Demonstrated significant runtime reductions: 21.72% and 14.47% over baselines.
• Effectively balanced reconstruction quality with computational efficiency.

Conclusions:

• SDGAN offers an efficient and effective solution for super-resolution in remote sensing.
• The framework minimizes artifacts and enhances local feature representation.
• SDGAN shows promise for real-world applications in urban planning and resource management.