PCA-GAN: A Novel Framework for Cloud Removal in Remote Sensing Images

Cloud cover increasingly disrupts Earth observation via optical remote sensing, compromising our ability to monitor critical resources. To address these challenges, we present PCA-GAN, integrating principal component analysis with generative adversarial networks to remove cloud interference from satellite imagery. By reducing data dimensionality before adversarial training, our method achieves state-of-the-art cloud removal while significantly accelerating computation. Compared to current benchmarks, PCA-GAN improves thin-cloud removal accuracy by 0.62 dB (PSNR) while maintaining 97.4% structural similarity (SSIM), and enhances thick-cloud removal by 4.21 dB PSNR and 0.01% SSIM. Notably, these improvements come with 31-40% reduced training time, demonstrating both superior performance and computational efficiency. This advancement enables more reliable Earth observation under increasingly challenging atmospheric conditions.