The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
higher potential for
data
longitudinal and
transverse
coverage
Flight constraints
Less impact of
time, daylight,
night, season
clouds
Daylight flying,
clean atmosphere
necessary
Production range
May be automated,
thus a greater
production
Higher need of
editing control
Budget
25%-33% of
budget:
photogrammetric
compilation
Production
Software: depends
on qualified
commercial &
technical people
Software for the
end-user: slow
process of
identification and
manual extraction.
Not reliable if
automated, implies
editing, especially
at large scales
Data acquisition
limited by a largely
contrasted area
Data can be
acquired.
Successfully used
in coastal
cartographic
production
Difficult and
expensive
Processing
Groups
Correlation
Feature extraction
Definition of zones
or areas
Edge limits 2-D
Results
Edges or limits
3-D
Edges and zones
3-D
Table 1. Comparison of photogrammetry vs punctual LIDAR
From the tests and trials carried out in the Photogrammetry
Laboratory of the Technical School of Surveying, Geodesy and
Cartography of the Technical University of Madrid (Spain) with
both techniques and information, provided by the National
Geographic Institute (IGN) of Spain, from the same zone
(Segovia) flown over with Vexcel Ultracam D digital camera
and LIDAR sensor, processed with DIGI3D/MDTop
photogrammetric software for LIDAR GTBiberica (Inpho
DTMaster) information, we have come to the following
conclusions: •
• The classical photogrammetric technique is very dependable
and accurate but the production process is time-consuming,
very specialized, and thus costly.
• If we consider semi-automation processes, i.e. automatic
correlation with breakline drawing by operator, production
times are going to be improved and even the HR will be less
specialized, however the software and even the quality of
the images make this methodology slow since it involves
much editing of the correlated points. On the other hand, if
we want to decrease the number of correlated points (a
greater interval) we should increase the number of
breaklines, so we would go back to the above-mentioned
case regarding production times.
• Full automation implies a large amount of correlated points.
The automatic extraction of the breaklines provided by the
current software is still the cause of many inaccuracies
which force a revision, and adding that information in the
classical way by restitution. Therefore, the time spared is
lost in a stereoscopic revision.
• Maybe the aspect that stands out is the actual information
source, the image, a metric document, of high geometric
resolution and increasingly better radiometric resolution.
Besides, sensors allow ever-increasing information in the
spectral range.
• Regarding the LIDAR techniques, they appear satisfactory
as far as the number of points (density/sq meter) and
precision (it is necessary to eliminate systematic errors,
calibration, etc.) However we have to use very specific
software and at times the filtering and classification
processes are long and intricate. In any case, a metric
verification of the information supplied appears to be
convenient.
The use of one technique or another will depend on parameters
such as cost, time and quality, independently, two at a time or
all together, as we shall see below (decision triangle).
3. CARRYING OUT THE INTEGRATION OF
INFORMATION
To the extent in which the resolution of the LIDAR sensors has
been progressively increasing with time and in view of the
possibility that these sensors could collect spectral information
in other bands and not only the information provided by the
LIDAR, researchers have been developing procedures and
techniques of classification based on the fusion of information
provided by the LIDAR with the information provided by the
regular photogrammetric cameras. Haala and Brenner (1997)
reconstructed 3-D models of cities and vegetation using a
combination of LIDAR data with another data source.
The ML (Maximum Likelihood) classifier is appropriate in our
case to solve the problem by using several bands of spectral
information and other attributes simultaneously [TSO and
Mather (2001)]. The initial hypothesis of this method is that the
classes we want to obtain are distributed with the same
likelihood in the image considered; although this is not always
the case, the method is improved and extended in the well
known Bayesian decision method which assigns a different
likelihood of occurrence to each class [Swain and Davis (1978),
Strahler (1980), Hutchinson (1982), Mather (1985), Maselli et
al (1995)].
In our case we carry out the classification of the LIDAR scatter
plot by turning the RGB image into HIS, so as to get the colour
information through the H and S attributes encoded to 8 bits. In
addition we have the information of the panchromatic camera
that provides us the channel I with an encoding of 16 bits.
To this spectral information we add the R component of the
infrared camera, encoded to 8 bits, linearly independent of the
previous attributes, the LIDAR intensity level encoded to 8 bits,
and finally the Z increment of the point, i.e. the Z difference
between the first and the last pulse. In all, we carry out the
classification of the LIDAR points with 6 independent attributes
in order to improve the results of the classification that would
be obtained by using only part of the attributes.