BD-VITGAN: a blind dense VITGAN for satellite remote sensing images super-resolution reconstruction

High-resolution (HR) remote sensing images contain valuable information, and drone imagery typically offers higher resolution than satellite images. However, UAVs face challenges in cost and resource consumption when covering large areas. Deep learning techniques have been widely used for super-resolution (SR) reconstruction, but existing algorithms based on bicubic interpolation perform poorly, especially in heterogeneous remote sensing images. Although both generative adversarial networks (GANs) and transformers show potential in image super-resolution, few studies combine these two techniques. To address these challenges, we propose a blind super-resolution framework that integrates the strengths of both Transformer and GAN, aiming to improve model adaptability to multi-source remote sensing data. The key contributions include: (1) a random mixed degradation modeling method, analyzing the impact of real paired data and synthetic degraded data, and revealing the influence of time differences on reconstruction quality; (2) introducing transformer into the generator network to enhance modeling of long-range spatial correlations; (3) designing a color consistency constraint algorithm based on feature space alignment to improve model generalization. Experimental results show that the proposed method achieves significant performance improvements in heterogeneous remote sensing image reconstruction.