Antarctic Time Machine

Abstract

This thesis presents a fully automated framework for transforming mid- 20th-century aerial photographs from the U.S. Trimetrogon Aerial (TMA) archive into geospatial datasets reconstructing past glacier elevations. Although the TMA imagery provides a unique and extensive record of Antarctic glacier conditions, it has remained largely underused due to its analogue format, degraded image quality, and the high manual effort typically required for processing. Previous efforts have relied heavily on manual digitisation and expert intervention. This work introduces a modular, end-to-end pipeline that automates the entire reconstruction process, combining semantic segmentation, metadata extraction, georeferencing, and Structure-from-Motion (SfM) photogrammetry.

The workflow begins with semantic segmentation using a custom-trained U-Net model, which classifies pixels in degraded grayscale aerial images into six categories: snow, ice, water, rock, clouds, and sky. Despite limited training data (100 manually labelled images) and challenges such as low contrast and artefacts, the model achieves an overall accuracy of 73% and an F1-score of 71%. By masking out unusable regions such as sky and ocean, this step significantly improves the reliability of the photogrammetric reconstruction.

An automated metadata extraction module complements the segmentation by retrieving key parameters, including focal length, altitude, and fiducial marker positions, directly from the images. Using a combination of optical character recognition and computer vision techniques, it recovers essential information and estimates missing values by exploiting redundancy across flight series. This reduces the need for manual transcription and converts handwritten image annotations into structured digital formats.

The geo-referencing component establishes a spatial link between historical images and modern coordinate systems. It uses LightGlue, a recent deep-learning-based matching algorithm, along with a progressive tiling strategy adapted to the characteristics of historical imagery. By matching tie points between the TMA scans and Sentinel-2 satellite imagery, the system automatically generates ground control points (GCPs) with positional accuracies of just a few meters, therefore dramatically improving upon the original, often kilometre-scale geolocation estimates.

In the final stage, the segmented images, extracted metadata, and GCPs are automatically passed to Agisoft Metashape, which is integrated into the processing pipeline via its Python API. This stage performs Structurefrom- Motion photogrammetry to generate dense point clouds, orthophotos, and digital elevation models (DEMs) without user interaction. Applied across the Antarctic Peninsula, the pipeline successfully reconstructed 3D glacier surfaces for 49 glacier systems. Validation against the high-resolution Reference Elevation Model of Antarctica (REMA) shows median elevation differences of approximately 90 meters across full glacier extents and 76 meters in topographically stable areas.

While the outputs do not yet match the accuracy of fully manual processing, the developed system enables large-scale, repeatable reconstruction of historical glacier surfaces at a scale previously unattainable. By combining all components into a modular, end-to-end framework, this work makes the TMA archive broadly accessible for contemporary cryospheric research and extends observational baselines by over half a century. All code and workflows are openly available on GitHub, and the resulting data products, including semantic masks, metadata tables, geo-referenced image positions, and 3D glacier models, are publicly released to support further scientific use.