With the widespread adoption of image acquisition devices and the increasing complexity of application scenarios, motion blur and multi-source blur issues in landscape imagery have become increasingly prominent, making it challenging for traditional methods to meet high-precision restoration demands. To address these challenges, this paper proposes a cross-domain landscape restoration and fusion method based on a hybrid deep learning model. The approach integrates a globally aware generative adversarial network (GAN), a spatial detail enhancement module, a multi-scale feature fusion network, and a weighted spatial pyramid feature fusion network (WSPF-Net), enabling comprehensive utilization of features across different scales and regions. Experimental evaluations conducted on the UrbanScapes, NatureScapes, and HDR_Landscape datasets demonstrate that WSPF-Net achieves PSNR values of 30.87 dB, 31.02 dB, and 29.14 dB, respectively, with significant improvements over comparative models. The SSIM metrics all exceed 0.89, and visual quality scores are consistently above 4.5 points. The system exhibits outstanding performance in image sharpness, detail fidelity, and color reproduction. The proposed method demonstrates superior image restoration capabilities and strong cross-domain adaptability in complex scenarios, offering a novel technical pathway for landscape image processing.