IEEE Geoscience and Remote Sensing Magazine - June 2022 - 122

recent images), the use of scanned archived photos requires
important preparation.
In that regard, the use of a large number of images for detailed
time series, large geographical areas, and granular geomorphic
analysis requires the automation of all the steps that
can be mechanized for the production of DSM from archived
aerial photos. In this article, after highlighting the limitations
of current methodologies for DSM production using archived,
scanned images, we introduce a new workflow with
automatic steps and show its application and results on the
production of DSMs for tropical and alpine contexts, namely,
the Remparts and Langevin rivers and Bossons glacier.
FROM SCANNED AERIAL IMAGES TO DIGITAL
SURFACE MODELS AND THE ASSOCIATED PITFALLS
Automatically processing archived images for photogrammetry
is a challenging task because numerous steps are
hampered by flaws that can make the process difficult and
propagate errors throughout the workflow. Digital photogrammetry
applied to overlapping aerial scanned images is
based on four main steps. The first consists of recovering,
from raw images and external sources (e.g., camera calibration
reports for sensor and optical systems), the technical
information necessary to use scanned photos as though
they were acquired with a digital camera (e.g., digital single-lens
reflex cameras and UAVs). The second step involves
extracting the relative orientation between the images, using
tie points (determined manually or automatically, depending
on the software that is used). The third requires
placing this orientation in a georeferenced (i.e., absolute)
coordinate system using ground control points (GCPs) acquired
in the field or from a reference base map. Finally, in
the fourth, pixel correlation is employed to derive a final
DSM for an area. Each step requires particular caution and
is associated with limitations that we highlight in the following
(being as agnostic as possible about software).
STEP ONE AND POTENTIAL ISSUES REGARDING THE
ASSESSMENT OF AUTOCOLLIMATION
Before they were accessible as downloadable images from
online catalogs, photos were acquired with a so-called pinhole
camera, and after being physically archived, they were
digitized using a scanner. Therefore, an important task is to
process pictures to obtain images as though they had been
taken by a digital camera and to obtain related metadata (this
what is called preprocessing in the following sections). First, it
is necessary to find the position (in a scanned photo) of fiducial
marks to calculate the exact internal orientation of the
pixels. This task may be easy for the human eye, but to make
it automatic for processing tens of photos or more, it is necessary
to consider several aspects. For example, the pattern as
well as the number of marks can change from one camera
manufacturer to another, and when it comes to multidecennial
archives, many patterns must be taken into account. In
addition, as shown in Figure 1, photos can be in gray scale or
color [Figure 1(a) and (b)]. In some cases, fiducial marks may
122
be barely detectable in low-contrast areas [e.g., Figure 1(c)].
Others can be small [Figure 1(d) and (e)]. Finally, fiducial
marks may be cut or missing due to film misalignment in a
scanner during manual digitization [e.g., Figure 1(f)].
Once they are identified and located in each picture, fiducial
marks are used to reproject and crop the images so that
they are mapped as though they had been digitally acquired.
For this, an affine transformation is generally considered to
overcome potential misalignment due to manual scanning.
After remapping, it is necessary to define the intrinsic geometry
of the sensor (the height and base of the field of view
pyramid that corresponds to the photos) by using the specifications
of the camera, such as 1) the focal length F, defined
as
Ff ,ijpij
= #;; with fij;
the focal length in the i,j-axis and
00
p |ij the pixel size in the i,j-axis; 2) the main point ([ ,]),ij
called the boresight; and 3) the skew coefficient of the axis, s,
that causes shear distortion, and the aperture or sensor/pixel
size. These parameters are used to derive the intrinsic matrix
(i.e., the internal calibration) of the camera K as
K=
R
T
S
S
S
f
s
i
i00
f
j
00
1
j
V
X
W
W
W
.
(1)
These parameters and camera specifications are often provided
with a calibration certificate, but they can be reestimated,
although it usually requires additional calculations
for good convergence.
Cropping/remapping the pictures and deriving the internal
calibration of the camera is critical. With every photo
in a survey being taken with the same camera, the matrix K
is constant, and to be relevant, pictures must be accurately
cropped along the fiducial marks. Depending on the software
used for autocollimation, the steps described previously
can be purely manual or partially automatic. With
manual procedures and large data sets, which are time
consuming, pixel misalignment between images may result
due to inaccurate fiducial mark pointing and imprecise
cropping. Such inaccuracy has a direct impact on the quality
of subsequent operations and their products. In the case
of semiautomatic procedures, when they exist (i.e., to the
authors' knowledge, there is no purely automatic method),
we observe (in the MicMac suite, e.g., [22] and [23]) up to
10% misalignment due to erroneous fiducial mark detection
on hundreds of images in one survey. Autocollimation
results in the internal calibration of the camera, which will
be used to estimate the relative orientation of the device
that took the pictures under consideration.
STEP TWO: THE RELATIVE ORIENTATION
Relative orientation consists of inferring the relative position
of the sensor for every image, accounting for 1) the
homologous points (tie points) between photos and 2)
the internal calibration of the camera that was obtained
during the previous step. Considering the type of implementation
(which depends on the software package), this
step may also be used to derive internal parameters, such
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE JUNE 2022
IEEE Geoscience and Remote Sensing Magazine - June 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2022
