International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
° measurement on the ground of at least 50 GCPs with GPS
or topography;
° observation of the landmarks on all the photograms;
. observation of two tie points in each of the nine canonical
positions, in each model.
The units measured and recorded the image coordinates of all
the observed points, so that it would be possible to manage a
wide range of different elaborations.
The units also calculated a manual aerial triangulation for the
blocks, in order to know the best possible results attainable with
the used data. In order to obtain reliable results, the control
points were subdivided into two disjoint sets: the proper GCPs,
used for aerial triangulation calculation, and the CKPs, used
exclusively for independent validation of results.
To summarize, three EO datasets will be considered; they are
indicated with the following acronyms: APP - produced with
direct sensor orientation; 447 - calculated with light integrated
sensor orientation; A7 - derived from traditional, manual aerial
triangulation.
4. DATA ANALYSIS
The pilot unit of the University of Pavia prepared a library of
objects and procedures for data analysis, working in the Matlab
environment. First, all the collimated points are stereoplotted in
each model where they are visible; stereoplotting happens on
single models, as this is the way map compilation is usually
performed. For each stereoplotted point, much data is stored:
the point's name, the images constituting the model, the plotted
coordinates, the residual parallaxes estimation and many others.
For CKPs, the differences between the plotted coordinates and
the true ones are calculated and recorded. Residual parallaxes
are estimated in the object space, measuring the skewness of the
homologous lines; they are also re-projected onto the image
space, via the local image scale factor (Casella, Franzini, 2003).
Once the stereoplotting has been performed, it is possible to
analyze and plot the data in many ways. There is the possibility
of analyzing residual parallaxes for all the tie points and of
estimating accuracies for CKPs, for instance. Moreover, it is
possible to consider the whole set of the plotted points but also
to filter them in a very flexible way, maintaining only those
coming from along-track models or those coming from across-
track models, for example. Many other combinations are
possible.
Within our test, nine datasets were considered, with respect to
the nature of the involved models:
. Whole: the whole set of the plotted points;
. Along: the whole set of the points measured within along-
track models;
* Along 60: points measured within along-track models
whose base corresponds to the ordinary 6096 overlap (as in
the Along dataset models having longer bases are also
considered);
. Across: the whole set of the points measured within
across-track models;
. Strip 1, Strip2, Strip3: single strips;
* Across 1-2, Across 2-3: points measured within across
models belonging to strips 1 and 2 only or 2 and 3 only.
882
5. SELECTED RESULTS
To summarize, an in-depth analysis was performed for each of
the six blocks for which the measurements are finished (Pavia,
Perugia, Como, Palermo, Parma, Vercelli), for each of the nine
above listed datasets, and for each of the three sets of exterior
orientations. Due to space limitations, it is not possible to show
all the results: we will only focus on three datasets which are
Whole, Along60 and Across.
Pavia's block will be considered first (Table 1). Regarding
accuracy, RMSEs are around 15 cm for £ and N and around 50
cm for Z, for the Whole dataset and the APP EO; results are
similar for the subsets 4long60 and Across. Moving to the AAT
EO, planimetric accuracy gets worse, while the altimetric one
doesn't significantly change; the subsets 4/ong60 and Across
show the same behaviour as the Whole one. If the A7 EO is
considered, planimetric and altimetric accuracies are respec-
tively around 15 and 30 cm for all the datasets. Of great interest
is the fact that planimetric accuracy is the same for APP (direct
sensor orientation) and for A7 (traditional, manual aerial
triangulation). For altimetry APP is significantly worse and this
is mainly due to the existence of a systematic error which
probably could have been eliminated with a proper calibration:
this is confirmed by the presence of similar systematic errors in
all the blocks. Again it is noticeable that the Z standard devia-
tions of APP and AT are substantially the same: a proper direct
sensor calibration could give similar results as AT.
If residual parallaxes are taken into consideration (Table 2),
RMSE is around 21 microns for the APP EO, which is quite a
high value: for the A/ong60 subset figures are better, while they
are rather worse for the Across one. For AAT EO results are
worse for the Whole set, but they are better if we focus on the
Along60 set. It must be considered that the A/ong60 dataset
represents the usual map compilation environment, therefore
AAT guarantees good stereovision quality. The 47 EO globally
gives results which are comparable to APP, and this is astonish-
ing; further analysis shows that results are very good for
Along60, as the RMSE is around 5 microns, and are worse than
APP for Across: the aerial triangulation adjustment tied along-
track much more than across-track. Residual parallaxes and
accuracies are surprisingly quite uncorrelated: for the Along60
dataset, the passage from APP to AAT halves the parallaxes, but
accuracies don't change or get worse.
As there is insufficient space to analyze the other blocks in the
same manner, some remarks will be raised, in comparison with
Pavia. Aerial triangulation gives substantially the same results
for all the blocks: 13-15 cm in planimetry and around 30 cm in
altimetry, RMSE, and these represent the attainable quality of
the images examined. Focusing on the Along60 dataset, Figure
2 shows that the application of light integrated sensor orienta-
tion, that is the passage from APP to AAT, always gives a
benefit in terms of accuracy. The same happens for parallaxes,
as the central diagram of Figure 4 highlights.
For some blocks the APP EO itself gives good results, while in
other blocks results are not very good for APP, even if they
become rather better after the application of the smoothing light
integrated sensor orientation. The AAT EO, which doesn't
require any ground measurements, guarantees planimetric
RMSEs below 33 cm and results are much better for some
blocks; altimetric RMSEs are not greater than 61 cm and are
much better for two blocks.
Intern
The r
georef
intere:
and ir
height
Furthe
effect:
small :
Casell
Cavall
GPS/I
Parma
0,7
0.61
0.4
0.3
RMSE (meters)
0,21
[AL
0.0(
0,7(
0,6t
0,5(
0.4(
0,3(
RMSE (meters)
0,20
0.)
0.0(
0,7
0,66
0,50
0.4(
0,30
RMSE (meters)
0,20
0.16
0,00
