The degradation of image quality in space-based remote sensing is a critical challenge due to atmospheric disturbances. This paper proposes a new model to simulate image blur effects caused by turbulence and aerosol scattering. It also analyzes a distortion vector field to simulate the distortion effects from atmospheric turbulence. Using this time-varying physical model, we present a generative adversarial network called MSFFA-GAN. It uses a multi-scale feature fusion and attention mechanism to analyze and apply optimal constraints on deep neural networks for atmospheric impact parameters. This helps our network handle complex atmospheric conditions that cause image degradation. Experimental results show that MSFFA-GAN improves PSNR by 5.05 dB and SSIM by 4.43%. It effectively restores degraded images and enhances the image quality of remote sensing systems.