The archive consists of approximately 2000 vertical aerial images, acquired during a photogrammetric survey by the former Institut für Angewandte Geodäsie (IfAG), Frankfurt am Main, Germany (today Federal Agency for Cartography and Geodesy (Bundesamt für Kartographie und Geodäsie - BKG)). The survey covered the areas on the western Antarctic Peninsula near Adelaide Island and the Grandidier Channel between 6 and 20 February 1989. Aerial photographs were taken using a Zeiss RMK A 8.5/23 camera with Agfa Aviphot Pan 200 aerial film at multiple image scales (1:70000, 1:30000, 1:15000, 1:10000, 1:5000) at different locations. Approximately 61 % of the images were acquired at an average flight elevation of 5895 m, yielding a nominal photoscale of 1:70000 with forward overlap of about 60 %. The film positives used here are preserved in the Archive for German Polar Research (Archive für deutsche Polarforschung – AdP) at the Alfred Wegener Institute (AWI) in Bremerhaven, Germany. There, they form part of the aerial image archive of BKG, which is part of the Federal Ministry of the Interior. This collection comprises a total of more than 20 000 images and was donated to the AdP/AWI in 2017 by the BKG together with the necessary rights of use. The films were digitized using a Leica DSW 700 scanner by GTA Geoservice GmbH, Neubrandenburg, Germany. Image positives that are 23 cm × 23 cm are scanned at an average scanning resolution of 12.5 µm and provided in 8-bit radiometric resolution, resulting in digital images of size 20 232 × 18 829 (380.948 × 10⁶) pixels (Fig. ). The average ground sampling distance (GSD) of these images corresponds to 0.875 m. To our knowledge, no scientific publication has extensively used this unique archive to study glacier elevation changes to date.

For this study, we selected images from the IfAG dataset that were acquired at a uniform scale of 1:70000. These images were selected due to their consistent coverage and suitability for generating historical DEMs across the study area. Higher-resolution images targeting specific glacier front locations (taken at image scales of 1:30000, 1:15000, 1:10000, and 1:5000) are excluded from the analysis, as identifying stable areas for co-registering historical DEMs to modern reference datasets proved challenging at the individual glacier scale.

We used the Reference Elevation Model of Antarctica (REMA, version 2) mosaic as a reference DEM to extract stable (or static) ground elevation for co-registering our historical DEMs derived from the aerial imagery . The REMA mosaic was downloaded from the OpenTopography portal (https://portal.opentopography.org/, last access: 17 December 2025). It is compiled from multiple REMA strips that are generated using very high resolution (0.32 to 0.5 m) WorldView-1,2,3 and GeoEye-1 satellite imagery through Surface Extraction from TIN-based Searchspace Minimization (SETSM) software (Howat et al., 2019). The mosaic is created to provide a more consistent and complete DEM product with blending and feathering of strip DEMs to avoid edge artefacts. REMA mosaic tiles are co-registered to ICESat-2 and Tandem-X 90 m PolarDEM. Given that REMA is the only high-resolution DEM available with high accuracy, it can be used as an appropriate reference DEM for processing the IfAG data. We used ICESat-2 ATL06 L3A Land Ice Height data from the summer months of 2020 and 2021 to evaluate the vertical accuracy of the derived IfAG DEMs on ice-free areas. Glacier outlines, ice-free areas, and rock outcrops are taken from the and Antarctic Digital Database (High resolution vector polygons of Antarctic rock outcrop v7.3, ).

To determine the planimetric information (longitude and latitude) of image centers, we manually estimated their locations using the survey index map from the Institut für Angewandte Geodäsie (IfAG) at a scale of 1:500000, archived at AWI. The map was digitized and georeferenced to establish the initial camera locations. Planimetric coordinates of the image centers are extracted with respect to the WGS84 datum. The approximate elevation of the camera flown at a scale of 1:70000 is 5985 m above ground. We sample the terrain elevation relative to the WGS84 ellipsoid at each initial horizontal location of the camera positions from the REMA mosaic. Therefore, an initial flying height above the ellipsoid (geodetic height) is incorporated as a third dimension into the initial 3D coordinates of the cameras. We input these 3D coordinates with an accuracy estimate of 1000 m to Agisoft Metashape, along with initial Yaw, Pitch, Roll of 0, 0, 0°, with accuracies of 180, 10, 10°, respectively.

We estimated Shannon entropy for each image as an indicator of texture. This metric is used to assess whether variations in DEM coverage were related to image texture and to filter out low-texture images prior to SfM processing. Shannon entropy measures the variability of the data based on the probability of occurrence and is described in Eq. (): 1H(X)=-∑i=1np(xi)log⁡2p(xi) where H(X) is the entropy of the random variable X and represents the average level of information or uncertainty inherent in the possible outcomes, p(xi) is the probability of the ith outcome xi, log⁡2 denotes the logarithm base 2, and n is the total number of possible outcomes. The Shannon entropy value ranges from 0 to log_x n, where n is the number of bins (for an 8-bit image is 256) and x is the base. We here used a base of 2, therefore, the range is 0–8 for all of our images.

We processed 548 images from 12 flightlines photogrammetrically in Agisoft Metashape version 2.1.1 in 8 different projects (subsets; Fig. ). The subsets are selected to optimize computational efficiency while encompassing well-distributed stable areas for co-registration, ensuring high absolute accuracy. The Structure from Motion (SfM) workflow is run on a computer with an NVIDIA RTX 500 Ada Generation GPU (32 251 MB, 100 compute units, 2550 MHz) and a CPU AMD Ryzen Threadripper PRO 5955WX 16-Cores 128 GB RAM. All point cloud generation parameters are detailed in Table .

Preprocessed images were imported into Metashape, and the tie points were generated at the native resolution of the imagery, facilitating precise and independent alignment of each subset. We set thresholds on quality parameters to reduce the number of incorrect tie points . Initially, we removed all tie points that have a reconstruction uncertainty of more than 10, which is equivalent to a camera base to height ratio of 2.3 (parallax angle of 23°). This removes all points that have a poor viewing angle, which may lead to weak 3D reconstruction and increased uncertainty in depth estimation.

We filtered out tie points with low projection accuracy caused by their poor localization. Tie points are poorly localized when the features they represent are large or less distinct, making it harder to locate their exact position in the images. To remove these points, we applied a projection accuracy threshold of 5, which was measured as the average image scale of the feature across overlapping images. Then, we reduced the reprojection errors of all subsets to less than 0.5 pixels. The filtered tie points were used to estimate the intrinsic and extrinsic camera parameters, which were then applied to generate a dense point cloud for each subset at medium quality, i.e., at a scale sixteen times lower than the original image scale. We selected a medium accuracy in “depth map generation” as a compromise between the required accuracy and coverage (Table ). To increase the robustness of the generated DEMs, we further excluded the points that were found in fewer than three scenes. We then generated a DEM by gridding the filtered point cloud at 10 m posting in the Antarctic Polar Stereographic (EPSG code 3031) coordinate system. This resolution corresponds to approximately three times the effective GSD of the input images processed at medium quality (originally ∼ 3.5 m), thereby minimizing interpolation artefacts and aligning with the resolution of the 10 m REMA DEM to avoid additional resampling.

We used the Ames Stereo Pipeline (ASP v3.4)'s pc_align.py tool that is embedded in HSfM for multistage co-registration of the generated raw DEMs, based on the Iterative Closest Point (ICP) algorithm . At each stage of ICP co-registration, a 12-parameter transformation matrix is calculated according to the expected offset and differences between the reference and raw DEM. In the first step, the rigid body transformation of the point-to-plane algorithm with translation and rotation is applied to the whole raw DEM with respect to the reference DEM (REMA mosaic). In the second step, in addition to translation and rotation, scaling is also added to further converge the offsets to the reference DEM. It is achieved by a similarity-point-to-plane algorithm. The point-to-plane algorithm is more robust to outliers than the point-to-point algorithm and thus converges faster . In the last stage of ICP co-registration, the alignment is refined by applying translation, rotation, and scaling corrections to stable areas only. Ice-free areas are taken from and the Antarctic Digital Database (ADD) rock outcrop mask. Stable areas for co-registration are determined by: (1) Filtering these areas to exclude slopes greater than 30° and minimize steep terrain-induced errors. (2) Manually removing blunders in areas where feature matching failed by cross-referencing with the Landsat Image Mosaic of Antarctica (LIMA) and orthoimages (Fig. ). We set the default expected offset values at each stage of the three-step ICP co-registration procedure to 2500, 500, and 100 m, based on visual inspection of multiple IfAG DEMs. After ICP co-registration, we used algorithm from the demcoreg package for subpixel co-registration over stable areas, which has a higher accuracy compared to ICP, as demonstrated by a reduction of NMAD after co-registration . This method estimates and corrects systematic offsets by relating elevation differences to terrain slope and aspect.

Finally, we mosaicked the 6 subsets that belong to the mainland Antarctic Peninsula into one product using the average value of the overlapping area to create a mosaic DEM. We provide DEMs that belong to two islands, i.e., Adelaide Island, PPI, as two separate files, along with the pre-coregistration point clouds in eight LAS files. We set the minimum elevation value to 0 and reclassified DEMs to eliminate negative values. We also provide a “bad” pixel mask raster for the mainland DEM, where our DEM is affected due to clouds or artefacts in the reference DEM. We generated this mask by excluding outlier elevation differences with respect to REMA mosaic (values exceeding 4 standard deviations from the mean) within glacier-covered areas, calculated across a binned elevation raster. We suggest using the DEM data where the bad pixel mask raster value is 1.

The IfAG camera extrinsics are updated by applying the transformation matrix obtained from ICP co-registration and the 3D shift vector from co-registration of IfAG DEMs, and orthomosaics are generated using Metashape at the original resolution. We use Metashape’s void-filling option of the DEMs to generate orthoimages without voids. We provide orthomosaics of 8 subsets in 8 TIFF files.

For the IfAG archive, the Shannon entropy value ranged from 4.3 to 7.5, with an average of 6.76 for all selected images (Fig. ). No scanning-related artefacts are observed in the processed DEMs. Terrain shadows are visible in some places in the South1 subset and may be the cause of 7.18 (high) entropy value in these areas. In Table , we summarize the total proportional area and the percentage area coverage of our DEMs of two land classes in our study area 1. Glacier 2. Ice-free areas. We masked our DEMs using the respective layers from . We estimate area coverage percentages by dividing the number of valid pixels in a class by the total pixels of the class. Overall coverage (Glacier + Ice-free Areas) of all DEMs ranges between 20 % and 45 %, with on-glacier coverages spanning 20 %–42 %. We excluded images from the western part of Adelaide Island due to the lack of stable areas and insufficient image features (low entropy; Fig. 4), and from north of Adelaide Island near the Grandidier Channel, where images predominantly cover water pixels (Fig. 1). Notably, DEMs for PPI and North1, South1 show higher coverage, which corresponds to the relatively higher average entropy of their source images (Figs. , ).

To evaluate the vertical accuracy of IfAG DEMs with independent surface elevation data, we have taken ICESat-2 ATL06 data from the summers of 2020–2021. Around 70 000 points are available in the ice-free areas in the study area. We estimated the uncertainty with respect to filtered ICESat-2 data on 1. Stable areas used for the co-registration, and 2. All ice-free areas. The ICESat-2 validation dataset was reduced from ∼ 70 000 to ∼ 6000 points after outlier filtering, highlighting the validation challenges in a complex terrain (Table , Fig. ). Our conservative outlier filtering (Sect. ) removed unreliable ICESat-2 data caused by clouds but also highlights misalignment issues with ICESat-2, likely at steep cliffs (further slope filtering reduced the no.of observations to ∼ 1000).

All our DEMs have vertical accuracy with respect to ICESat-2 of less than 8 m (maximum NMAD of 7.21 m for Mainland IfAG DEM). A better accuracy of less than 5 m is observed on stable areas used for co-registration (4.16 m for Mainland IfAG DEM). Due to blunders and outliers on slopes greater than 30°, DEMs have biases up to 0.48 m (Mean offset for IfAG mainland mosaic for all slopes). These biases reduce to 0.3 m (Mean offset for IfAG mainland mosaic for filtered, slopes less than 30°), when only lower slopes are considered (Table ). Furthermore, the accuracy of ICESat-2 is known to degrade at higher curvatures ; the uncertainties of our DEMs may therefore be overestimated in such regions.

Few DEMs based on similar historical aerial imagery are available for comparison with our IfAG dataset. recently published an elevation dataset for Larsen B glaciers, derived from 1968 aerial imagery. They reported vertical uncertainties relative to the REMA strip DEMs from 2021 of 15.22 and 19.21 m for Crane and Flask glaciers, respectively. Another dataset covering the Greenland Ice Sheet, based on 1978–1987 aerial imagery, demonstrated varying vertical accuracies according to the year of the campaign of up to 8.8 m for slopes < 20° and 10.3 m for all slopes when validated against Airborne Topographic Mapper (ATM) data from 1994–2014 . estimated detailed elevation and volume changes of 16 individual glaciers in northern Antarctic Peninsula using FIDASE archives from 1956–1957 austral summers. Using least square surface matching with modern DEMs derived from WorldView-2 imagery, they obtained historical DEMs with post-matching biases varying between +1 to -5.9 m and uncertainties between 7.3 to 28.2 m. In contrast, our IfAG DEMs, derived from 1989 imagery and validated against REMA and ICESat-2 data, exhibit lower vertical uncertainties (e.g., Adelaide Island: NMAD 6.39 m with REMA, Mainland: NMAD 7.21 m with ICESat-2, Fig. , Table ), outperforming the Larsen B datasets by 3–5 times and matching or exceeding the accuracy of the Greenland, northern AP datasets on average.

The larger uncertainties in the Larsen B DEMs may be attributed to the lower quality of the 1968 imagery, poor stereo overlap, manual tie point and GCP placement, and the lack of precise camera positioning. Similarly, the elevated uncertainties in the FIDASE dataset can be attributed to the age and condition of the 1956–1957 film negatives, and differences in type of scanners used in the digital archiving, which led to scaling issues in these image-derived products . Although both the Larsen B and IfAG datasets lack camera calibration reports and rely on imprecise initial camera positions, our DEMs benefit from a refined estimated camera model and an iterative co-registration approach using the Iterative Closest Point (ICP) algorithm. These methods effectively reduce vertical errors, even for subsets with large initial geolocation offsets (up to 6500 m on North1). Additionally, our DEMs show minimal vertical biases relative to REMA (e.g., PPI: -1.22 m, Mainland: 0.49 m, Adelaide Island: 0.82 m, Fig. ), compared to higher biases reported for Larsen B (e.g., Crane: 6.4 m vs. REMA, -4.16 m vs. ASTER; Flask: -0.06 m vs. REMA, -9.01 m vs. ASTER) and FIDASE (up to -5.9 m vs. WorldView-2) . While the Greenland dataset benefits from access to camera calibration reports and extensive terrestrial GPS-based ground control, our comparable accuracies were achieved through the co-registration with and validation against spatially well-distributed reference terrain.