Stretching time series further in the past with the best possible accuracy is essential to the understanding of climate change impacts and geomorphological processes evolving on decadal-scale time spans. In the first half of the twentieth century, large parts of the polar regions were still unmapped or only superficially so. To create cartographic data, a number of historic photogrammetric campaigns were conducted using oblique imagery, which is easier to work with in unmapped environments as collocating images is an easier task for the human eye given a more familiar viewing angle and a larger field of view. Even if the data obtained from such campaigns gave a good baseline for the mapping of the area, the precision and accuracy are to be considered with caution. Exploiting the possibilities arising from modern image processing tools and reprocessing the archives to obtain better data is therefore a task worth the effort. The oblique angle of view of the data is offering a challenge to classical photogrammetric tools, but the use of modern structure-from-motion (SfM) photogrammetry offers an efficient and quantitative way to process these data into terrain models. In this paper, we propose a good practice method for processing historical oblique imagery using free and open source software (MicMac and Python) and illustrate the process using images of the Svalbard archipelago acquired in 1936 by the Norwegian Polar Institute. On these data, our workflow provides 5 m resolution, high-quality elevation data (SD 2 m for moderate terrain) as well as orthoimages that allow for the reliable quantification of terrain change when compared to more modern data.
The processing, or reprocessing, of historical imagery into digital
elevation models (DEMs) and orthoimages enables the extension of time series for
long-term geomorphological change analysis
Processing nadir imagery over mostly stable land areas where good-quality
modern data are available for reference has proven easily accomplished and
effective The scanning usually does not conserve a consistent internal geometry
(the fiducial marks are not consistently found at the same pixel
coordinates), and the images therefore need their internal orientation to be
standardized. No metadata are present in the files. If not always strictly necessary,
it is often very helpful to have access to (1) calibration information (i.e.,
calibrated focal and position of fiducial marks) and (2) location information
in the form of ground control points (GCPs) or navigation data. Stereo overlap is low (typically 60 % along-track and 20 %
cross-track) compared to modern standards (typically around 80 %
along-track and 60 % cross-track).
Challenging scenarios are, however, common when metadata are impossible to find
or image quality is particularly bad. For instance, large datasets over
Antarctica include regions with very little off-glacier stable areas to
provide sufficient GCPs, as well as large areas of
low-contrast snow fields yielding no tie points and bad or no correlation
Another common type of imagery is oblique imagery. It was acquired primarily
in areas previously unmapped or only superficially so. Such imagery has
the advantage of providing a point of view that is easier to comprehend for a
human user, and identifiable points can be seen from long distances
(even if the image resolution decreases with distance from the camera
position). However, for both analytical and computer-based processing,
oblique data add a layer of challenges compared to nadir imagery.
The footprint of each image is not very well defined as it extends to
infinity in a cone shape and excludes part of the ground because of terrain
occlusion (e.g., behind ridges and peaks). Therefore, the term
The resolution of the imagery is dependent on (1) the distance to the
camera (the further away the camera, the lower the resolution) in a much
bigger way than nadir imagery (here, the variation in the camera-to-ground
distance can vary several fold!) and (2) the slope of the terrain (slopes
facing the camera will have a very high number of pixels per horizontal unit
of surface).
In this paper, we present and discuss the challenges and solutions related to the processing of oblique (e.g., non-nadir) historical imagery using the 1936 survey of the Svalbard archipelago as an example. We then provide some concise comments on the observed change in elevation over glaciers and rock glaciers.
In the summers of 1936 and 1938, Norwegian expeditions lead by Docent Adolf
Hoel went to Svalbard to photograph the area from the air. Bernhard Luncke
was in charge of the planning and acquisition of aerial photographs. Their
goal was to collect images suitable for creating topographic maps of Svalbard
in
In 1936, good weather conditions allowed for a total effective flight time of
89 h. 3300 images covering an area of 40 000 km
The goal of the 1938 expedition was to photograph the areas that were not
covered in 1936. Bad weather conditions and technical problems prevented the
completion of the planned acquisition, but 48 h of flight time resulted in
2178 photographs covering a total area of about 25 000 km
The photographs are owned and managed by the Norwegian Polar Institute
(NPI).
In this paper, a subset of 72 pictures taken in the summer of 1936 in two
lines along the western (43 pictures) and eastern (29 pictures) coasts of the
Prins Karls Forland, eastern Svalbard, was used as an example (see
Fig.
Historical oblique imagery can be processed through a nearly unmodified
typical structure-from-motion photogrammetric pipeline, as described in this
section and in Fig.
Workflow of the structure-from-motion photogrammetric pipeline used in this paper. Data are represented in gray and processes in green.
The scanning of images usually results in images that are not all the exact
same dimension, and fiducial marks are not necessarily positioned exactly at
the same place. To simplify the image assimilation into the photogrammetric
process, the interior orientation needs to be standardized: the images have
to be geometrically standardized to all be in the same geometry so that the
image resolutions and fiducial mark positions are the same, and the
internal calibration of all the images is the same
In order to compute the parameters of the 2-D-transformation (translation,
rotation, and scaling) for each image, the fiducial marks need to be
identified either manually or through an automated process. If fiducial
marks are in the center inside of a clear target such as a cross, it is easy
to find them automatically. However, this is not the case for our data, in
which
the fiducial mark is a single dot in the film (see Fig.
Fiducial mark in an image from 1936 over Svalbard and a close-up of the top one.
To be able to run a structure-from-motion-based bundle adjustment, tie points
need to be extracted from the imagery. The SIFT algorithm
Once the images are in a coherent internal geometry and the tie points
are extracted, the camera can be calibrated and a bundle adjustment process can
be performed. The NPI archives contained a single piece of information about
the camera used for our data: its focal length of 210.71 mm. From this
information, we could initialize the calibration process. The calibration (in
the form of Brown's distortion model,
The absolute orientation requires GCPs, but no such points were recorded at the time of acquisition in 1936. It is therefore necessary to rely on modern data coverage of the area from which to gather GCPs. To do so, prominent points in stable areas are identified both in the historical images and in a modern source. For our case study, at the time of processing the orientation, the 1990 DEM from the Norwegian Polar Institute represented the best data available and was therefore used. The GSD is 20 m and the SD with the 2008 DEM is below 2 m.
This step is very challenging since most of the lowlands, on top of being
generally unrecognizable by a lack of strong structures, cannot be considered
stable over 50 years: ice-cored moraines and permafrost areas, thermokarst
processes, unstable cliffs, or simply erosion renders the lowlands unstable.
Some areas are also challenging to recognize because of the vastly different
ice cover over the terrain. Therefore, only bedrock features can be used and
it is easiest to identify the mountain tops as GCPs, but even these points
are not always very accurate, as the reference DEMs can be of low resolution
and often soften the mountain ridges and peaks in particular. The process is
further complicated by the strong obliqueness of the historic imagery, making
the task of recognizing even very prominent features difficult. The
automation of the process used by
Once GCPs are identified, a similarity transformation (seven parameters: three for 3-D translations, three for rotation, and one for scaling) is computed first, and then the camera calibration is reestimated using the GCPs.
To generate point clouds, images are grouped into triplets of subsequent
images in a line of flight. There is no point using more than three images as
the area imaged in the central image of a triplet (used as the master image)
will not be seen on further images because of the ca. 60 % overlap. Each
triplet goes through a dense correlation process
In order to only compute data on areas of interest (in our case, the strip of
land closest to the camera positions), the master images need to be masked,
as trying to get data out of the entire viewshed area would make the process
slower, add noise to the data, and create heavier files that are harder to
work with. Specifically, one should mask the sea, the sky, clouds, and
background land visible in the image but not in the area of interest. The
masking can be completely manual, but it is then a very time-consuming task,
especially for large datasets. It is possible to automate the detection of
sea, sky, and clouds by running an entropy-based texture analysis, but
differentiating foreground and background is more difficult
Entropy response values for samples of different terrain features.
Once a three-dimensional point cloud has been generated, it can be filtered
to remove noise and then turned into a number of different raster (2-D)
files. The first step in the filtering is to simply remove points that are
outside of the area of interest (defined by a polygon in
Representation of erroneous data points created by mountain ridges and seashore in the approximate direction of the optical rays. Note that the terrain occluded by the foreground mountain, and therefore not included in the viewshed, is not represented in the point cloud.
The clean three-dimensional point cloud is then used as the input for raster
data creation. The resolution and corner coordinates of the raster needs to
be determined first. From the 1936 Svalbard photos, a 5 m GSD product can be
reliably generated in the areas presented in this paper without requiring
extensive void filling. Here, we want to create an orthoimage (a raster of
colors, here gray scales; see Fig.
Rasters extracted from a single point cloud.
Once all rasters are generated for each point cloud, they can be mosaicked
using the distance raster as an input for the creation of a Voronoï
diagram: for each point of the final raster, the color and elevation from the set
of rasters presenting the lowest range value (distance to camera) are selected
and recorded, creating the data shown in Fig.
Raster mosaic.
This simple approach creates visible boundaries between each area, especially in the orthoimage, and one might want to perform some radiometric equalization. The mosaicking could be avoided by first merging all the point clouds and performing the rastering process only once from the full combined cloud using a weighting system based on the point-to-camera distance value. However, given that individual clouds can be composed of over 50 million points, the amount of memory required to load dozens of clouds into software (either for merging or rastering) is an almost impossible task, even for a heavy workstation.
In this section, we first discuss the final product quality and then point out the weaknesses remaining in the data that required manual filtering. Figure 8 shows an overview of the output data for the Prins Karls Forland dataset.
A first approach to determine the data quality is to check the residuals of the orientation. Here, after using GCPs to reevaluate the calibration, the average residual of the tie points was 0.803 pixels, and the GCP reprojection error was of 4.108 pixels with an RMSE of 18.15 m. The relatively high RMSE value is most likely due to the relatively soft DEM used as an input for GCP identification. The quality of the data is better estimated by pulling statistics from various differences of DEMs (DoD) between our data and ground-truth DEMs.
Overview of the data resulting from the processing of the 1936 images over Prins Karls Forland. From left to right: Hillshade, orthoimage, DEM difference with the NPI 2008 DEM.
Data over Millerbreen.
Data over Søre Buchananisen.
In order to determine their accuracy, our 1936 elevation data were compared
to other DEMs. The NPI most modern DEM product (S0 product) over the area is
extracted from the 2008 aerial survey of Svalbard and is also at 5 m GSD. It
has a declared accuracy of 2–5 m
(
These values are significant improvements on the 12.2 m reported by
The key reason to process the 1936 photos using modern technology is that the result can be used as a baseline to assess long-term change related to glacial and periglacial processes, adding important early data points in the time series of DEMs over the area. By comparing these data with modern products, a nearly centennial glacier mass balance could be estimated, for instance. We show here three examples of different processes observed in the differential DEM (dDEM) between our product based on the 1936 imagery and the most recent DEM from the NPI computed from imagery from 2008.
In this example, we show the data gathered for the alpine glacier
Millerbreen. In 1936, it was land terminating, reaching nearly to its
moraine, while in 2008 (and up to today), a large part of the front calves
into a lake that formed inside the barrier formed by the moraine.
Figure
Data over an unnamed rock glacier.
Søre Buchananisen is a large marine-terminating glacier that experienced a
significant retreat (up to 2.0 km) in the 72 years separating the two datasets (27.8 m year
The precision of the data allows for a clear identification of rock glaciers
(or ice-cored moraines, or small debris-covered glaciers) and of their
changes linked to a degradation process. The example shown here for
illustration (Fig.
The value of the information contained in historical photogrammetric surveys for the study of change in the cryosphere is very obvious, as shown by the increasing amount of scientific literature on the topic. With our processing workflow, we achieve an almost fully automated generation of high-quality DEMs even from high oblique photographs. However, such processing still requires some amount of manual work, mostly for GCP identification and for point-cloud cleanup. This cleanup, while crucial to avoid having to deal with areas of vastly erroneous data, does not modify the points that are conserved in the final data and as such does not induce additional noise or noise reduction. Applied on the imagery taken over Svalbard in 1936, our workflow provides high-quality elevation data (STD 2 m for moderate terrain) that allows for the reliable computation of elevation change when compared to more modern data. This may provide an accurate estimate of the retreat and geodetic mass balance of glaciers in the area for the time period since 1936, pushing the beginning of the time series much earlier than previously achievable. The high precision of the data obtained with our workflow also allows for the long-term monitoring of processes presenting elevation changes of smaller magnitude, such as on rock glaciers.
Original imagery from NPI. DEM data available upon request.
LG led the research, wrote the paper, processed the data on
Prins Karls Forland, and supervised the work of NIN and FC. FC did an initial
assessment of the imagery data and of the potential quality of the products
created from them and designed a first version of the workflow
The authors declare that they have no conflict of interest.
The study was funded by the European Research Council under the European Union's Seventh Framework Program (FP/2007-2013)/ERC grant agreement no. 320816 and the ESA projects Glaciers_cci (4000109873/14/I-NB), DUE GlobPermafrost (4000116196/15/IN-B), and CCI_Permafrost (4000123681/18/I-NB). Data scanning and distribution was funded by the Norwegian Polar Institute. Couderette's internship was financed in part through the Erasmus program. We are very grateful to the Norwegian Polar Institute, specifically to Harald F. Aas for providing the raw imagery data for this study and the modern validation data. We also acknowledge the help from Marc Pierrot-Deseilligny and the MicMac team at IGN/ENSG for their help in the development of the software. Edited by: Nicola Masini Reviewed by: Simon Buckley and two anonymous referees