The utility of historical image repositories is often limited due to the lack of geo-referencing. A good example is the TriMetrogon Aerial (TMA) archive, a collection of historical aerial images of Antarctica between 1940 and 2000. These images are difficult to use, as their geolocation is only approximately, with location errors in the order of tens of km. This study addresses this challenge by developing an automated geo-referencing workflow that leverages recent advancements in machine-learning-based tie-point matching. We use the algorithm LightGlue, to establish tie-points between geo-referenced Sentinel-2 satellite imagery and historical non-geo-referenced aerial images. To aid the process, we use already known approximate positions of the historical images. Due to the sub-optimal and inhomogeneous quality of the aerial images, only a portion of the images can be geo-referenced directly by matching. For the remaining images, we employed alternative means of geo-referencing, again based on tie-point matching. Out of a subset of 4,459 images located inside the research area, 3,393 images could be geo-referenced, a percentage of 76%. Reasons for the geo-referencing failing are insufficient contrast or an approximate position too far away from the real position. The workflow can easily be applied to historical images from other archives, to enhance the usability of historical image repositories for scientific research.

The TriMetrogon Aerial (TMA) archive is an archive of historical images of Antarctica taken by the US Navy between 1940 and 2000 with analogue cameras. The analysis of such historic data can give a view of Antarctica's glaciers predating modern satellite imagery and provide unique insights into the long-term impact of changing climate conditions with essential validation data for climate modelling. However, the lack of semantic information for these images presents a challenge for large-scale computer-driven analysis. Such information can be added to the data using semantic segmentation, but traditional algorithms fail on these scanned historical grayscale images, due to varying image quality, lack of colour information and artefacts in the images. To address this, we present a deep-learning-based U-net workflow. Our approach includes creating training data by pre-processing and labelling the raw images. Furthermore, different versions of the U-net are trained to optimize its hyperparameters and augmentation methods. With the optimal hyper-parameters and augmentation methods, a final model has been trained for a use-case to segment 118 images covering Adelaide Island. We tested our approach by segmenting challenging historical images using a U-net model with just 80 training images, achieving an accuracy of 73% for 20 validation images. While no test data is available for our use case, a visual examination of the segmented images shows that our method performs effectively. The comparison of the hyper-parameters and augmentation methods provides directions for training other U-net-based models so that the presented workflow can be used to segment other archives with historical imagery. Additionally, the labelled training data and the segmented images of the test are publicly available at https://github.com/fdahle/antarctic_segmentation.

A huge archive of historical images of the Antarctica taken by the US Navy between 1940 and 2000 is publicly available. These images have not yet been used for large-scale computer-driven analysis as they were captured with analog cameras. They were only later digitized and contain no semantic information. Most modern deep-learning based semantic segmentation algorithms are trained on modern images and fail on these scanned historical images, due to varying image quality, lack of color information, and most crucially, due to artifacts in both imaging as well as scanning (e.g. Newton's rings). The analysis of such historic data can give a view on Antarctica's glaciers predating modern satellite imagery and provide a unique insight into the long-term impact of changing climate conditions with essential validation data for climate modelling. An important first step for analysis of such data is the extraction and localization of semantic information, e.g. where in the image is water, rocks, or snow. In this work we present the first deep-learning based method to perform semantic segmentation on historical imagery archives of the Antarctic Peninsula. Our results show that our method can handle very challenging images even after being trained with only a low number of training data and catch the general semantic meaning of a scene. For eight test images we achieve an accuracy of 74%, where the majority of errors can be explained by the classification of ice as snow.

In many countries digital maps are generally created and provided by Cadastre, Land Registry or National Mapping Agencies. These maps must be accurate and well maintained. However, in most cases, the update process of these maps is still done by hand, often using satellite or aerial imagery. Supporting this process via automatic change detection based on traditional classification algorithms is difficult due to the high level of noise in the data, such as introduced by temporary changes (e.g. cars being parked). This paper describes a method to detect changes between two time steps using 2.5D data and to transfer these insights to a digital map. For every polygon in the map, several attributes are collected from the input data, which are used to train a machine-learning model based on gradient boosting. A case study in Haarlem, in the Netherlands, was conducted to test the performance of this proposed approach. Results show that this methodology can recognize a substantial amount of changes and can support - and speed up - the manual updating process.