What question did this study set out to answer?

The aim is to quantify a century of urban growth using deep learning techniques on historical and contemporary imagery.

May 13, 2026Open Access

Reconstructing a Century of Urban Growth Through Deep Learning-Based Colorization and Segmentation of Historical Aerial and Satellite Imagery: Les Sables-d’Olonne, France (1920–2024)

Key Points

The aim is to quantify a century of urban growth using deep learning techniques on historical and contemporary imagery.
Developed a deep learning pipeline utilizing a Pix2Pix colorization stage and U-Net++ segmentation.
Analyzed historical aerial imagery from 1920 to 1971 and satellite data up to 2024.
Employed few-shot adaptation to enhance model accuracy for historical data.
The colorization stage achieved a PSNR of 35.21 dB and an SSIM of 0.9762.
Segmentation model displayed strong performance on modern imagery with mIoU of 0.9789.
Few-shot adaptation for historical data resulted in reliable building footprints with mIoU between 0.53 and 0.65.

Abstract

Coastal urbanization is increasingly constrained by legacy land-use patterns and escalating climate risks, yet long-term morphological trajectories remain poorly quantified due to the absence of multispectral data in pre-satellite archives. This study introduces a scalable deep learning pipeline that bridges a century-scale domain gap through an attention-enhanced Pix2Pix colorization stage and a few-shot U-Net++ segmentation stage, enabling automated reconstruction of urban expansion from panchromatic historical aerial imagery (1920–1971) and digital aerial photographs (1997) to contemporary very-high-resolution satellite data (2024) in Les Sables-d'Olonne, France. The novelty of the approach lies in coupling generative colorization with epoch-specific fine-tuning to overcome radiometric and annotation bottlenecks that have historically prevented quantitative urban reconstruction from pre-satellite archives. The colorization stage achieved high spectral fidelity (PSNR 35.21 dB, SSIM 0.9762), and segmentation performed strongly on modern imagery (mIoU 0.9789). While the segmentation model performed strongly on modern imagery, direct transfer to historical data exhibited substantial domain shift due to radiometric discrepancies. Few-shot adaptation on year-specific calibration sets recovered reliable building footprints (mIoU 0.53–0.65) across the full timeline. Multi-scalar analysis of the reconstructed footprints revealed constrained anisotropic expansion: early saturation of the coastal historic core, followed by rapid inland peri-urbanization post-1971 driven by geographic barriers. This spatiotemporal shift has entrenched spatial lock-in, placing recent development in retro-littoral zones that are vulnerable to submersion and characterized by severe vegetation loss. The framework unlocks previously inaccessible historical archives for quantitative urban monitoring, providing critical insights into legacy effects of unconstrained growth and informing resilient coastal planning under climate change.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper