- 1Universität Zürich, RSL, Remote Sensing of Environmental Changes, Switzerland
- 2Universität Zürich, Data Science for Sciences, Switzerland
- 3Universität Zürich, Glaciology and Geomorphodynamics World Glacier Monitoring Service, Switzerland
Trend determination for earth surface processes requires long and continuous and certain measurements, but long-term records of landscape change are often limited in temporal and spatial extent. Scanned historical aerial imagery serve as a valuable resource to derive data products like digital elevation models (DEMs) to document the historical state of the Earth's surface and to calculate trends for different processes e.g. glacier dynamics.
Classic Structure-from-Motion (SfM) photogrammetry workflows have demonstrated the capability to automatically generate DEMs and orthoimage mosaics from such historical images, as highlighted in a few studies. These workflows typically consist of the following steps: (a) pre-processing, (b) tie-point extraction, (c) matching, (d) bundle adjustment, (e) dense reconstruction, (f) co-registration, and (g) orthoimage mosaic generation. However, classic methods struggle with the challenges historical imagery coming with. For example: inconsistent image quality, limited metadata documentation, image distortions and distinct viewpoint geometries.
Recently, advances in robotics and computer vision have introduced learned models for tasks such as tie-point identification, matching, dense reconstruction as well as part of the co-registration stage (e.g. SuperPoint, ALIKE, SuperGlue, LoFTR and more). These networks have shown promising results in different stereo-matching scenarios by outperforming classic SfM methods. However, since they were primarily developed for modern robotics and computer vision tasks, their performance on scanned historical aerial imagery remains uncertain. As historical imagery exhibits the properties described above, these networks were not optimised with them during training.
We boost existing pipelines in tie-point extraction and matching with these models and compare the quality of resulting DEMs from different model combinations together. We also highlight issues encountered when applying these learned models to historical aerial imagery and proposes solutions to address them. We demonstrate our findings using scanned historical images from the Southern Patagonian Ice Field (Chile) recorded in 1980, particularly for the Grey & Dickson Glacier area, as well the south-west flank of Cordon Mariano Moreno Mountain and adjacent fjords. These two sites providing different acquisition geometries and overlaps. The results evaluate the average RMS reprojection error following the bundle adjustment, to determine the quality of different extractors and matchers as well as the median distance between closest points to evaluate the co-registration.
How to cite: Kugler, L., Ioli, F., Wegner, J. D., Dussaillant, I., Rada, C., and Piermattei, L.: Advances in Historical Aerial Image Analysis: Boosting SfM Pipelines with Learned Models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17688, https://doi.org/10.5194/egusphere-egu25-17688, 2025.