Extended Data Fig. 4: An overview of the image pre-processing and structure from motion (SfM) pipelines.

To improve feature selection during the SfM reconstruction, we enhance the digitized images by increasing contrast through histogram stretching and sharpening features using the Dehaze Tool in Adobe Lightroom. This radiometric enhancement step improves photogrammetric reconstructions over ice and snow, which tend to be lower contrast than the surrounding land. Finally, since scanning does not preserve the internal geometry (images can be rotated, translated, and warped), we locate the four fiducial marks on the edges of each image and apply a projective transformation that maps the images to a standardized internal geometry. Owing to the large number of images in the dataset, we use an automated pipeline for fiducial mark identification. We convolve the edges of the image with an idealized fiducial template to identify target regions. Next, inside the target regions, we convolve the image with a Laplacian of Gaussian filter to locate the fiducial spot. We process the aerial photographs in a standard photogrammetric workflow in Agisoft Metashape 1.6.0. In brief, we first extract up to 40,000 keypoints from each image. Keypoint matching across all the images provides the constraints to solve for the unknown parameters, including the relative camera locations/orientations and the camera distortion parameters. Adding ground control points (GCPs), with specified (x,y,z) positions, enables the absolute georeferencing of the model. Finally, a multi- view stereo (MVS) reconstruction converts the sparse 3D model to a dense 3D point cloud. We perform the MVS reconstruction with a Dense Quality of Medium (meaning that depth maps are generated at 1/4 the image resolution) and Dense Filtering at Moderate to Aggressive.