Developing deep-learning models guided by semantics to colorize historical grayscale aerial imagery for land-cover classification

High-resolution historical aerial imagery is valuable for analyzing long-term land-cover and environmental change, but most deep-learning methods rely on multispectral or color data and cannot fully use grayscale archives. We developed a semantically guided framework that uses OpenEarthMap land-cover labels to guide Pix2pixHD and Hyper-U-Net colorization. Among five models, Hyper-U-Net + Label achieved the best segmentation performance (mean IoU = 0.5296; mean F1 = 0.6738), approaching real-color imagery, while Pix2pixHD + Label produced the most visually realistic outputs in an online survey of 514 participants. Applied to 1980s aerial photographs of Taiwan, colorization increased overall classification accuracy from 0.4599 to 0.6673 and mean IoU from 0.2602 to 0.4274. These results show that semantic guidance improves both visual plausibility and downstream segmentation, helping bridge historical grayscale archives with modern AI-based geospatial analysis and supporting more reliable assessment of long-term land-cover and environmental change.