Virtual restoration and prediction of urban heritage landscape changes based on generative adversarial networks
Abstract
Urban heritage preservation faces significant challenges due to progressive environmental degradation and structural aging, which threaten the longevity and cultural value of historic landmarks. To address these issues, we propose a unified framework that combines generative adversarial networks (GANs) for virtual restoration with a temporal prediction module for forecasting future degradation. Specifically, our approach includes a Restoration Module based on a GAN architecture augmented with residual connections and attention mechanisms, capable of reconstructing damaged or missing architectural elements with high fidelity; and a Prediction Module that leverages sequence modeling to estimate future landscape deterioration driven by pollution, humidity, and urbanization dynamics. The framework is trained on a curated dataset of 308 high-resolution images of heritage sites, enhanced with manual annotations and synthetic augmentations to simulate both restored and degraded states. Extensive experiments demonstrate that our model outperforms baseline methods in terms of PSNR, SSIM, and FID, validating its superiority in both restoration quality and predictive reliability. This research provides a novel dual-stage approach to urban heritage conservation by integrating visual restoration and long-term environmental forecasting into a single pipeline, offering practical tools for sustainable cultural preservation.