A machine-learning pipeline for merging and georeferencing very large archives of historical aerial photographs

We introduce SPLAT (Stitching Pipeline for Large Aerial Tilesets), a new method for the automated and simultaneous stitching and geolocation of large archives of aerial historical photographs. Millions of physical prints and films of aerial images are archived around the world. The growing share of these images being digitally scanned motivates the development of a robust, scalable, and efficient method to process them and make them suitable for quantitative analysis. Existing solutions can stitch datasets comprising hundreds of historical images, either using time-consuming manual geolocation and tie point collection or, alternatively, automated techniques in those cases where extensive and detailed metadata is available.To our knowledge, SPLAT is the first pipeline to be able to process millions of images. It can do so by using a simplified model for the reconstruction of the image positions, making use of available metadata, and optimizing the amount of manual labor involved in the process.We apply SPLAT to a dataset of 1.67 million images from the National Collection of Aerial Photography (NCAP) in Scotland, which includes coverage of numerous countries, including former British colonies, between the 1940s and 1980s. We report the results of SPLAT on a selected group of African countries. We show we are able to geolocate and retrieve footprints for more than 50,000 images. An analysis of the goereferencing error on a set of evaluation ground control points yields a median (mean) error of 76 (197) meters on the prediciton of their location.