The study area lies between 6∘20′27′′ and 7∘51′17′′ N and between 39∘16′58′′ and 40∘04′40′′ E (approximately 5730 km2), located 400 km south-east of Addis Ababa in the Oromia region of Ethiopia (Fig. 1). The area lies within the Bale Mountains and covers 25 % of the afro-alpine ecosystems of Africa (de Deus Vidal and Clark, 2020; Carbutt, 2020; Gehrke and Linder, 2014). The Bale Mountains were formed by lava outpourings of the Trappean series, which created a vast lava plateau considerably eroded and flattened by later glaciations (Berhe et al., 1987; Groos et al., 2021; Williams, 2017). The soils in the area tend to be shallow, gravelly, and fertile silty loams of reddish-brown to black colour, as the parent rock is mainly basaltic and trachytic (Miehe and Miehe, 1994). The study area provides a water source for over 30 million people in Ethiopia, Somalia, and Kenya and contributes to five major perennial rivers (Fig. 1).

The climate in the study area varies from north to south, owing mainly to differences in elevation, aspect, and the influence of lowland hot air masses, which also influence the annual migration of the Inter-Tropical Convergence Zone and the Indian Ocean Monsoon (Miehe and Miehe, 1994). The mean monthly minimum and maximum temperatures are 5.6 and 21.4 ∘C, respectively, for an altitudinal range of 2700 to 4377 m above sea level (a.s.l.). The mean annual ground temperature ranges from 7 to 11 ∘C (Groos et al., 2022). The lowest and highest temperatures recorded on the Sanetti Plateau were -15 and 26 ∘C, respectively (Hillman, 1988), where frosts occur during all clear nights throughout the year (Groos et al., 2022). Rainfall is variable throughout the study area and ranges from 800 to 1500 mm annually (Woldu et al., 1989).

The study area is known for plenty of vascular plant species ranging from tropical rainforests to the afro-alpine ecosystem. It is classified into five vegetation zones: (1) Gaysay Valley grasslands covered by a small grassland situated within a broad flat valley at an altitude of 2600–3000 m a.s.l. in the northern part of the study area; (2) dry evergreen montane forest ranging from 2600 to 3200 m a.s.l. dominated by Juniperus procera and Hagenia abyssinica; (3) afro-alpine in the central part of the study area, including the Sanetti Plateau between about 3200 and 4377 m a.s.l. characterized by short, herbaceous plant communities dominated by Alchemilla, Helichrysum, and Artemisia shrubs; (4) an Ericaceous belt spanning most of the escarpment areas between 3100 and 3800 m a.s.l. and forming a belt between the high-altitude afro-alpine plateau and the lower Harenna Forest (the vegetation in this belt is dominated by Erica arborea and Erica trimera species scattered up to 4276 m a.s.l.; and (5) the Harenna moist montane forest: this forest is one of the most extensive and largely natural forests remaining in Ethiopia. It covers the southern slope of the mountains, extending between 1450 and 3200 m a.s.l. (Asefa et al., 2020; Kidane et al., 2019; Miehe and Miehe, 1994; Fig. 1). Topographically, the study area has a maximum slope of 58.63∘. Almost 50 % of the area is exposed to east-, south-east- and south-facing slopes.

We used and processed historical aerial photographs (HAPs) taken in 1967 and 1984. They have a forward overlap of approximately 60 % and a side overlap of approximately 30 % and were captured using an aerial 23×23 cm film frame camera, which scanned at a resolution of 1200 dots per inch. The qualities of all the scanned HAPs were excellent enough for processing with an image quality value above 0.5 units. Worldview-1 satellite imagery (DigitalGlobe Foundation) was used for the extraction of on-image selected GCPs and applied for pre-processing (Groos et al., 2021). A subset of the GCPs was used for internal accuracy assessment. To assess the external accuracy of our generated DEMs and orthomosaics, we used GCPs collected in the field and additionally from five sources of data: (1) Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM) data with a spatial resolution of 15 m, (2) data from the Shuttle Radar Topography Mission (SRTM) with a spatial resolution of 30 m, (3) Advanced Land Observing Satellite Phased Array Type L band Synthetic Aperture Radar data (ALOS PALSAR) with a spatial resolution of 12.5 m, (4) TanDEM-X with a spatial resolution of 90 m from Rizzoli et al. (2017), and (5) the Copernicus DEM (COP-DEM) with a spatial resolution of 30 m (European Space Agency and Airbus, 2022). Each of the five downloaded DEMs was projected onto the Adindan UTM coordinate system zone 37 N (EPSG: 20137) using ArcGIS Desktop (Version 10.8.2) (ESRI Inc., 2021).

The 1967 HAPs were acquired by the United States Air Force under project number AF 58-3 using a four-engine RC-130A aircraft flying at an altitude of 9448.8 m a.s.l. equipped with a KC-1(B) PLANIGON camera. For detailed information about the camera and lens number, calibrated focal length, magazine serial number, roll number, exact acquisition dates, and flight index map, refer to Supplement Fig. S1 and Table S1. The shutter speed and aperture were 1/150 s and F-8, respectively, resulting in a scale of approximately 1:62 000.

The 1984 HAPs were obtained by the SWEDSURVEY company under project number ET 1:5 using a Wild RC 10 camera (Wild Universal Aviogon II lens type, camera serial number 3045) with a calibrated focal length of 152.822 mm and a maximum aperture of F/4.0. The flight campaign had an average altitude of 7600 m a.s.l. and was conducted from 15 to 17 January 1984. For more details on the roll number, strip number, photo number, acquisition date, and flight index map, see Supplement Fig. S2 and Table S2.

The data preparation step included extraction of the calibrated focal length, camera principal point coordinates, camera exposure stations, fiducial marks, and ground control points. All the data preparation steps were performed with Agisoft Metashape Professional (Version 1.8.0.13794) (AgiSoft LLC, 2021), Google Earth, and ArcGIS Desktop (Version 10.8.2) (ESRI Inc., 2021).

Traditional digital photogrammetry requires the 3D position and camera poses or the 3D position of a series of control points to be known to reconstruct a 3D structure from 2D overlapping photographs. SfM is the process of estimating the 3D structure of a scene from a series of 2D images with the absence or presence of camera exposure locations and GCPs. SfM uses the scale-invariant feature transform algorithm (Lowe, 2004) and Hessian-based detectors (Lindeberg, 1998) to address the problem of identification of features in the individual images used for image correspondence, especially for the reconstruction of DEMs and orthomosaics using scanned historical aerial photographs. In this step camera pose and scene geometry are reconstructed in the automatic identification of matching features in multiple images. In our workflow SfM is a specific step that provides camera parameters and a sparse point cloud (Daniotti et al., 2020). Then, dense image-matching algorithms, such as MVS, are used in a subsequent step to densify the point cloud. MVS will capture more angles in between instead of taking two photos from two different viewpoints to increase durability, for example for image noise or surface texture (Wu et al., 2018). The MVS algorithm handles images with more different perspectives, such as sets of images that surround objects, and can also handle very large numbers of images (Fig. 13).

The whole above-mentioned process is referred to as SfM–MVS photogrammetry, with a workflow methodology including multiple different algorithms, computer vision, photogrammetry, and more conventional survey techniques (Eltner and Sofia, 2020). It is applicable for soil erosion, volcanology, glaciology, coastal morphology, mass movements, and fluvial morphology. It was used in at least 65 scientific studies from 2012 to 2015 (Eltner et al., 2016) and also recently (Grottoli et al., 2020; Nyssen et al., 2022; Tomczyk and Ewertowski, 2021). The advantages of SfM–MVS photogrammetry are that it can be applied fully automated (Eltner et al., 2016); is flexible, efficient, and inexpensive; and requires little training (Iglhaut et al., 2019; Westoby et al., 2012). The following steps were sequentially applied for generating DEMs and orthomosaics with SfM–MVS photogrammetry. The whole process is also shown in Fig. 14.

The generated datasets can be used for different applications, and we give some brief examples of five use cases here. Firstly, within our study area, in two locations soil material was excavated and used for constructing and repairing a gravel road (Fig. 4). An example volumetric analysis was done with Agisoft Metashape Professional Edition software (AgiSoft LLC, 2021) for the selected sites. The volumetric change calculated from the 1967 and 1984 DEMs revealed that the removed soil was 14 728.9 m3 for data example 1 (Fig. 4a, b) and 97 352 m3 for data example 2 (Fig. 4e, f).

Secondly, the number of settlements in the afro-alpine ecosystem and central parts of Bale Mountains National Park increased (Fig. 7). There was not much human interference and deforestation until 1967. However, encroachment by humans and degradation of the vegetation were visible from 1984 onwards (Figs. 7 and 8). This is complemented by Reber et al. (2018), who found 870 settlements and 331 rock shelters in only the afro-alpine ecosystem of Bale Mountains National Park.

Thirdly, the resulting orthomosaic datasets could be used as an input to compare the historical distribution of ericaceous species to today's recent distribution in the Bale Mountains and in particular to detect the historical development of the Erica patches scattered across the Sanetti Plateau given the detailed resolution of our datasets. In addition, Mekonnen et al. (2023) examined potential factors responsible for the now fragmented occurrence of Erica on the Sanetti Plateau, which is possibly also due to the topography of the area (Mekonnen et al., 2023). The distribution of the giant root rat over the Sanetti Plateau and the adjacent valleys in the north-western parts of the study area was predicted by Wraase et al. (2023). However, this study could be extended in the future by analysing the historical distribution shift of the giant root rat and further explaining its effect on the landscape and its ability to change landscapes.

Fourthly, the very-high-resolution information from the 1984 orthomosaic can be employed for river and riverside altimeter change assessments (Fig. 9). Already visible by using the orthomosaics is that the river named Welmel Tika and its side changed in both width and altitude, thereby reducing previously forested area.

Lastly, our generated datasets are useful for studying changes in surface geomorphology and urban areas. Currently, in the north-western corner of the study area above Adaba, there is an artificial lake dam called Melka Wakena that was built for hydroelectric power generation. Back in 1967 the area was cultivated land and had a river course inside today's dam (Fig. 10). The urban area of Adaba in the north-western part of the study area, for example, also expanded from 53.83 ha in 1967 to 107.09 ha in 1984 and to 475.79 ha in 2011 (Fig. 11).

The quality of our generated DEM dataset was assessed based on external and internal accuracy assessment. For external accuracy assessment, the quality of the resulting DEM dataset was assessed by using external sources, i.e. 509 GPS control points (CPs) collected in the field as reference points as well as ASTER GDEM (30 m), SRTM DEM (30 m), COP-DEM (30 m), TanDEM-X DEM (90 m), and ALOS PALSAR (12.5 m) points for both 1967 and 1984. The values of elevation (in metres) for each CP from each of the DEMs were extracted, and then accuracy assessment was done using linear regression models by comparing the descriptive statistics of the elevation, minimum elevation difference with field control points, first quantile, median, third quantile, maximum elevation difference from CPs, residual standard error, multiple R-squared correlation coefficient, and root mean square error. Comparing the 1967 and 1984 DEMs with external globally freely available DEMs is beneficial and gives additional information regarding the accuracy of DEMs with complex topography like the present in our study area.

Our generated DEMs have higher accuracies than previous DEMs (Fig. 5 and Table 2). As demonstrated by its RMSE value of 3.44, the DEM generated from the 1984 HAPs had the best quality. The 1967 DEM also had a very good quality, as indicated by an RMSE value of 3.55. The multiple R-squared value was 0.9998 (equal to the 1984 DEM), and there was no significant difference between the 1967 and 1984 DEMs (Table 2). Our datasets are accurate for doing an in-depth analysis as the spatial resolution is very high (0.84 and 0.98 m for 1967 and 1984, respectively) compared to hitherto freely available DEMs and satellite images. In addition, our datasets are spatially aligned with other global images and relate to external reference systems too (Fig. 6).

In addition, the resulting DEM dataset was also assessed by using on-image selected GCPs for internal accuracy assessment (12 out of 49 GCPs and 7 out of 27 GCPs for the 1967 and 1984 HAPs, respectively; Table 1). It indicates a high quality of internal image network geometry as huge dense point clouds for historical aerial photographs. This was illustrated by the low sub-pixel values of tie point reprojection errors of 1.49 and 0.68 pixels and the RMSEs of 4.24 and 0.88 cm on the control points for the years 1967 and 1984, respectively.

Among the five readily available DEMs used for comparison, the Copernicus DEM (30 m) had the highest accuracy (RMSE value of 4.9) complemented by the validation maximum RMSE value of 5 (European Space Agency and Airbus, 2022). In similar high mountain areas, carried out to evaluate the performance of DEMs for tectonogeomorphic analysis in the South American Andes, the Copernicus DEM (30 m) also performed best (Del Rosario González-Moradas et al., 2023). Thus, complementing the comparison made for freely available DEMs for our study area, and it showed the best accuracy (Fig. 5 and Table 2). However, the SRTM DEM was the next most accurate (RMSE value of 5.64) in our study area, which has a complex topography and a huge elevation difference of 3800 m. These results agree with Sena et al. (2020). The SRTM DEM showed better precision (Fig. 5 and Table 2). However, others (Shebl and Csámer, 2021; Chowdhuri et al., 2021; Jalal et al., 2020) concluded that ALOS PALSAR is more accurate than the SRTM and ASTER DEMs and obtained different results for our study area. The ALOS PALSAR DEM (Fig. 5 and Table 2) was the third-most accurate DEM, with an RMSE vertical accuracy of 11.38. The ASTER GDEM was the fourth-most accurate in the area as the residual standard error value was 11.42 with an RMSE value of 11.54, indicating that there is a gap in elevation values compared to the field GPS control points (Fig. 5). Thomas et al. (2014) also attested to a higher accuracy of the SRTM compared to the ASTER GDEM. The TanDEM-X 90 m (Rizzoli et al., 2017) DEM was the least accurate, with an RMSE value of 11.9. However, TanDEM-X 90 is more accurate compared to the Multi-Error-Removed Improved-Terrain DEM and SRTM 90 m resolution accuracy assessment checked within six continents at 32 floodplain locations using high-resolution lidar DEMs as a reference (Hawker et al., 2019).