A novel generative adversarial network framework for super-resolution reconstruction of remote sensing

Introduction Remote sensing super-resolution (RS-SR) plays a crucial role in the analysis of remote sensing images, aiming to improve the spatial resolution of images with lower resolutions. Recent advancements in RS-SR research have been largely driven by the integration of deep learning techniques, especially through the application of Generative Adversarial Networks (GANs), which have shown significant effectiveness in advancing this field. While GAN has achieved notable advancements in this field, its tendency toward pattern collapse often introduces artifacts and distorts textures in the reconstructed images. Methods This study introduces a novel RS-SR model, termed the Diffusion Enhanced Generative Adversarial Network (DEGAN), designed to improve the quality of RS-SR images through the incorporation of a diffusion model. At the heart of DEGAN lies an innovative GAN architecture that fuses the adversarial mechanisms of both the generator and discriminator with an integrated diffusion module. This additional component utilizes the noise reduction capabilities of the diffusion process to refine the intermediate stages of image generation, ultimately improving the clarity of the final output and enhancing the performance of remote sensing super-resolution. Results In the test dataset, the peak signal-to-noise ratio (PSNR) increased by 0.345 dB at 2× scaling and 0.671 dB at 4× scaling, while the structural similarity index (SSIM) was improved by 0.0087 and 0.0166, respectively, compared to the current state-of-the-art (SOTA) approach. Discussion These results indicate that DEGAN significantly improves the super-resolution reconstruction performance of remote sensing images. The introduction of the diffusion module and attention mechanism effectively reduces noise and enhances image clarity, addressing common issues of texture distortion and artifacts in remote sensing image super-resolution reconstruction.