A GAN-enhanced deep learning framework for rooftop detection from historical aerial imagery
Year: 2025
Type: article
Abstract: Precise detection of rooftops from historical aerial imagery is essential for analysing long-term changes in urban development and the evolution of human settlement patterns. Advances in AI-powered computer vision technologies present substantial potential in realizing the value of a large number of historical aerial images archived by many organizations. However, black-and-white analog photographs have shown considerable challenges for common object detection frameworks due to their limited spatial resolution, absence of color information, and archival degradation. To address these limitations, we introduce a two-stage image enhancement pipeline based on Generative Adversarial Networks (GANs). The procedure of data preprocessing includes image colorization through a DeOldify approach and super-resolution enhancement through Real-ESRGAN. The enhanced images were subsequently employed to train and evaluate rooftop detection models, including Faster R-CNN, DETReg, and YOLOv11n. The results demonstrate that the combination of colorization with super-resolution significantly enhances detection performance, with YOLOv11n achieving a mean Average Precision (mAP) exceeding 85%. This signifies an enhancement of approximately 40% over the original black-and-white images and 20% over images enhanced solely through colorization. The proposed method effectively bridges the gap between archival imagery and current deep learning techniques, facilitating more reliable extraction of building footprints from historical aerial photographs. Potential applications of the method have been discussed in the paper. The codes and documentation for reproducing our results are publicly available at github.com/Pengyu-gis/Historical-Aerial-Photos.
Authors: Pengyu Chen, Sicheng Wang, Sicheng Wang, Cuizhen Wang, Senrong Wang, Senrong Wang, Beiao Huang, Lu Huang, Zhe Zang
Language: English
Subfield: Ocean Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4412499637