IEEE Geoscience and Remote Sensing Magazine - June 2022 - 129

Moreover, we take advantage of this approach for optimizing
the coregistraion of canvases from different campaigns.
Indeed, processing archived aerial images usually
leads to bad convergence during camera calibration, which
can be improved by setting the sensor properties (the focal
length and boresight when initializing calibration). However,
since the different canvases Cn
are all related by new
GCPs (tie points between campaigns), we can propagate the
orientation information from C0
through C .n Since the canvas
of reference needs to be the most recent campaign, the
orientation is propagated backward in time from one canvas
to another, with no reevaluation of the reference (i.e, C0
C ,1 C1 to ,C2 and so on or C0 to ,C1 C0 to ,C2 and so forth).
to
In practice, this is done by simultaneously performing the
external orientation and bundle adjustment with hundreds
to thousands of tie points for each image. The spatial distribution
of the tie points (between every campaign and its
reference) is not statistically biased, so the process converges
rapidly and leads to coregistered canvases after steps 1 to 3 in
the " From Scanned Aerial Images to Digital Surface Models
and the Associated Pitfalls " section. In a similar manner, we
can back-propagate the aerotriangulation to older canvases.
When the time between two campaigns is too long, one can
chose a given Cn
as a new reference for a " very old " canvas.
We call this approach Backward Time, Nearest-Neighbor Propagation
(BTN2P). Finally, since it relies on a well-established
bundle adjustment algorithm, the accuracy of the canvas
coalignment is easy to assess.
It is worth noting that a multiblock (or multicanvas) approach
has been proposed for painting restoration using
multiple cameras [26]. However our methodology varies
in many regards. First, the scene depth for landscapes varies
through time, while in paintings, changes are mainly in
the radiometry (i.e., restoration does not affect the depth
of a scene). Second, historical surveys come with a badly
constrained orientation or no flight path, while for painting
restoration, the locations of the cameras are accurately
known and fixed. Finally, and most critically, the scanned
sensors are badly calibrated, as discussed. Because of these
considerations, in our method, internal and external calibrations
are done by BTN2P as opposed to iteratively adding
canvases [26], which would lead to the nonconvergence
of archived scanned images.
OVERVIEW OF THE WORKFLOW
In the following, we summarize our approach for the production
of time series DSMs from archived aerial images:
1) Download scanned aerial images from online catalogs
or manually scan achieved films of every campaign in
the times series.
2) Preprocess the images from the campaigns with FiducialLib
(Algorithm 1) to transform the scanned pictures
into photos taken by digital cameras.
3) Designate the campaign of reference and select the
GCPs for it. In our strategy, archived images closest in
time to the data set (e.g., maps and GPS measurements)
JUNE 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
used to select the GCPs are the campaign of reference.
Select high-quality GCPs on the images of the campaign
of reference from base maps from national surveys/institutions,
orthoimages, and GPS measurements.
4) Process the campaign of reference and evaluate the relative
and absolute orientation of its cameras (steps 2 and
3 of the classic photogrametry workflow). The result is
the canvas of reference, C0
[ " function ImgTo3DCloud
()C0 " in Algorithm 2].
5) Process the other campaigns using BTN2P. The tie points
between canvases Cn
and C0 are used for evaluating the
internal and absolute orientation of the cameras [function
ImgFromC0To3Dcloud(, )CCm0
in Algorithm 2].
Note that the evaluation can be performed from an orthoimage
as far as common tie points can be found with
C .m
6) Dense pixel matching results in a 3D cloud for every resulting
canvas Cn
rithm of the user's choice.
7) (Optional) While this does not concern the data discussed
here, in case of poor tie point matching between
canvases and unbalanced spatial dispersion of the tie
points between canvases (which may happen when radiometry
varies between canvases and if the time span is
too long in regard to actual changes in a landscape), 3D
homothetic compensation can be done by assessing the
Euclidean isometry that minimizes the point-to-point
distance between a 3D cloud and a reference, considering
a residual threshold of the user's choice [either the 3D
cloud derived from C0
or from an external source, such
Dpts Dpts
33 m
as " function CloudCorr(, ,Threshold) " in
Algorithm 2]. The computing transformation matrix is
applied to the external orientation of the cameras, so the
3D cloud and derived orthoimages can be corrected with
Euclidean isometry to preserve the relative shape of the
cloud (as opposed to polynomial correction), enabling
mass balance calculations of time series. We provide the
workflow in Algorithm 2.
APPLICATION TO MOUNTAINOUS TERRAIN
GEOMORPHOLOGICAL SETTINGS
To demonstrate the benefits of the workflow presented in this
study, the analysis of the rate of geomorphic change, we applied
our methodology to two steep-slope environments in
two geological contexts. First, we chose a region in the tropical
environment of Réunion Island, in the Indian Ocean.
The area of interest is the Remparts and Langevin rivers [2],
which incised the flank of the Piton de la Fournaise volcano
during the past 60 Ka [4], leading to deep canyons between
the remaining stable parts of the volcano (called planèzes). In
these canyons, the difference in elevation between the planèzes
and valley bottom ranges from 1,500 to 90 m, with
slopes up to 60º. The second area of reference is the Bossons
glacier in the French Alps, beside the city of Chamonix. Here,
the geomorphological context is different since glacial valleys
are shallower and wider than river canyons. However,
129
by using shapes from a motion algo
IEEE Geoscience and Remote Sensing Magazine - June 2022

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