In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
thoroughly described in (Haralick 1979, Gonzalez and Woods
2002, Zhang 2001). It is obvious how important it is to include
some kind of information that tells us whether the values are
well distributed over the whole matrix, or whether they are
mostly located close to the matrix diagonals (e.g. Difference
Moment or Inverse Difference Moment, Equations 5-10 and 5-
11). In case big values lie close or on the diagonal (Figure 10,
non-urban), the region under investigation is expected to be
homogeneous, whereas if the values are distributed more
homogeneously (Figure 10, urban), the co-occurrence matrix
corresponds to a heterogeneous region.
Figure 10: Horizontal and vertical co-occurrence matrices for 3
different types of terrain.
Usually the quality of the input data (i.e. imagery and DSM) is
responsible for erroneous results. It is very difficult for the
introduced algorithm to produce correct results, if the buildings
to be extracted does not cover a certain number of pixels. For
instance, when dealing with IKONOS imagery (with GSD of
lm) a small house of 8m x 10m will most probably not be
extracted correctly.
The procedures discussed in paragraph 2.2 and 2.3 fall into the
category of feature extraction and make use of the imagery for
deriving the geometric building properties.
The general idea is to use a library where characteristic features
of buildings are stored; by analyzing the image, areas are
searched that correspond to a high degree to the registered
“library buildings”. Characteristic features of a building can be
textural measures (by using the so-called occurrence and co
occurrence descriptors) or similarity measures of the image
grey values.
3. RESULTS
In this chapter an evaluation of the presented methods is given.
It consists of a quantitative and qualitative description, and
moreover shortcomings and weaknesses of the presented
methods are discussed.
Since many subsets are examined that are coming from various
types of line scanning systems, both airborne and spacebome
(ADS40, HRSC-AX, Quickbird, Orbview, IKONOS, SPOT5),
their outcomes will not be listed individually. Errors in the
qualitative evaluation will be given in image space units
(pixels).
The presented outcomes are divided into two groups:
quantitative and qualitative results. Moreover, the three
presented DCM extraction approaches are evaluated
individually. Input data is subdivided into categories depending
on image scale and building density (low urban and urban) of
the investigated areas. Image scale is defined as the scale that
we would expect from an analogue product, e.g. for a 1:10,000
product we expect 1-2 metres accuracy in nature, if the
graphical accuracy and visual perceptivity are 0.1-0.2mm.
Regarding the mentioned image scales the three interpretation
categories are:
1. scale A: 1:1000-1:4000
2. scale B: 1:4000-1:12000
3. scale C:< 1:12000
3.1 Quantitative Assessment of Building Extraction
The aim in the quantitative analysis is to evaluate whether the
presented approaches are practical in sense of completeness of
building detection of the result, i.e. how many buildings were
actually found. It is investigated whether the techniques for
finding potential building candidates are applicable.
Furthermore, an evaluation is carried out to see how many of
these buildings were extracted and to what a degree:
• CFB: Correctly Found Buildings,
• NFB: Not Found Buildings (also includes
insufficiently mapped buildings: building seed point
was determined successfully, but the adaptive region
growing process did not manage to create an area that
covers a reasonable amount of the object),
• WFB: Wrongly Found Buildings, i.e. found objects
were in reality no building exists.
The calculation of CFB (true positive), NFB (false negative)
and WFB (false positive) are briefly explained in the following:
The CFB and NFB percentages are calculated with respect to
the total number of existing buildings in the area under
investigation, whereas the WFB is calculated with respect to the
total number of found buildings (comprising correctly and
wrongly found buildings). Figure 11 shows the way of
computing and a numerical example, respectively.
c .=
3 -c
o =
existing \
buildings
% of correctly found
buildings (CFBt
% of correctly found
buildings
% of not found
buildings (NFB)
existing buildings = 100%
found buildings = 100%
Figure 11 : Illustration for quantitative assessment computation.
We consider that the pre-processing has been carried out
without error, so that the orientation of the imagery and the
derived orthophotos on which we apply the investigated
techniques are correct. We will also not evaluate nDSM
extraction algorithms and their qualities in detail, since this is
not topic of this research.
For the evaluation of the outcomes altogether 13 different
scenes containing 677 buildings were examined.
Table 2 allocates the quantitative analysis.
Concerning the level of detail that can be derived from
individual data sets the Nyquist theorem has to be taken into
consideration (“Sampling rate must be at least twice as high as
