The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
298
(up-to-date) building databases were edited manually, by field
surveying. Out-of-date databases were derived by simulating
changes to the reference databases by inserting new and
deleting some of the existing buildings. In Marseille, 107
changes were simulated (89 new and 18 demolished buildings)
and 40 changes (23 new and 17 demolished buildings) were
simulated in Toulouse. The out-of-date databases were
converted to binary image files (building vs. no building)
having the same GSD as the input data. These binary building
masks were distributed to the participants along with the CIR
orthophotos and the DSMs.
2.2 Evaluation Procedure
Each group participating in the test was asked to deliver a
change map in which each building is labelled either as
unchanged, demolished or new. However, both the
representation of the results of change detection and the output
formats varied considerably between the individual algorithms.
In addition, the definitions of the classes that are discerned in
the change detection algorithms are not identical. Whereas the
algorithm by Champion (2007) exactly matches the test
requirements, this is not the case for the other two algorithms
used in this study. In these two cases it was thus decided to use
a building label image representing the updated building map
for the evaluation. The evaluation consists of a comparison of
the outcomes of each algorithm to ground truth (i.e. the initial
reference database). Two quality measures are computed for the
evaluation: the completeness, i.e. the percentage of the actual
changes that are detected by an algorithm, and the correctness,
i.e. the percentage of the changes detected by an algorithm that
correspond to real changes (Heipke et al., 1997):
Completeness = - e 0.
TP + FN 1
TP
Correctness = e |0,ll
TP + FP L J
In Equation 1, TP, FP, and FN are the numbers of true positives,
false positives, and false negatives, respectively. They refer to
changes in the change map compared to actual changes in the
reference. Thus, a TP is an entity reported as changed
{demolished or new) that is actually changed in the reference. A
FP is an entity reported as changed by an algorithm that has not
changed in the reference. A FN is an entity that was reported as
unchanged by an algorithm, but is changed in the reference.
Finally, an entity reported as unchanged by an algorithm and
also being unchanged in the reference is a true negative (77V). In
this context, the entities to compare can be buildings, which
results in per-building quality measures, or pixels in a rasterised
version of the change map, which results in per-pixel quality
measures. In the cases where it was decided to use a building
label image representing the updated map for the evaluation, the
rules for classifying an entity as a TP, a FP, a FN, or a 77V had
to be defined in a slightly different way. Any existing (i.e.
unchanged) building in the reference database is considered a
TN if a predefined percentage (T h ) of its area is covered with
buildings in the new label image. Otherwise, it is considered a
FP, because it does not have a substantial correspondence in the
new label image, which thus indicates a change. A demolished
building in the reference database is considered a TP if the
percentage of its area covered by any building in the new label
image is smaller than T h . Otherwise, it is considered to be a FN,
because the fact that it corresponds to buildings in the new label
image indicates that the change detection algorithm has not
found this building to have been demolished. A new building in
the reference database is considered a TP if the cover
percentage is greater than T h . Otherwise, it is considered a FN.
The remaining large areas in the new label image that do not
match any of the previous cases correspond to objects wrongly
alerted as new by the algorithm and thus constitute FPs.
3. CHANGE DETECTION APPROACHES
3.1 Method 1 - (Matikainen et al., 2007)
The building detection method of the Finnish Geodetic Institute
(FGI) was originally developed to use laser scanning data as
primary data. In this study, it is directly applied to input
correlation DSM and orthophotos. The method includes the
following stages:
1. Pre-classification of DSM height points to separate ground
points from above-ground points, using (Terrasolid, 2008).
2. The region-based segmentation of the DSM into
homogeneous regions and calculation of various attributes
for each segment, using (Definiens, 2008)
3. The classification of the segments into ground and above
ground classes by using the pre-classification.
4. The definition of training segments for buildings and trees on
the basis of training data sets.
5. The construction of a classification tree by using the
attributes of the training segments (Breiman et al., 1984).
6. The classification of above-ground segments into buildings
and trees on the basis of their attributes and the classification
tree.
7. A post-processing to correct small, misclassified areas by
investigating the size and neighbourhood of the areas.
The output of the method also consists of a building label image
representing the new state of the database, which is used for
evaluation in this study.
3.2 Method 2 - (Rottensteiner, 2007)
The input data of this method consist of a DSM obtained by
LIDAR or stereo-matching techniques. A geocoded NDVI
image, the initial building data base, and a Digital Terrain
Model (DTM) can optionally be used. If no DTM is available, it
is derived from the DSM by hierarchic morphologic filtering. If
the initial database is available, it can be used to introduce a
bias that favours a classification consistent with the initial data
base, because in most scenes only a small percentage of
buildings will actually have changed. The workflow of the
method consists of three stages. First, a Dempster-Shafer fusion
process is carried out on a per-pixel basis and results in a
classification of the input data into one of four predefined
classes: buildings, trees, grass land, and bare soil. Connected
components of building pixels are grouped to constitute initial
building regions. A second Dempster-Shafer fusion process is
then carried out on a per-region basis to eliminate regions
corresponding to trees. The third stage of the work flow is the
actual change detection process, in which the detected buildings
are compared to the existing map. A very detailed change map
is generated in this process. The output of the method consists
of a building label image representing the new state of the data
base and in change maps describing the change status both on a
per-pixel and a per-building level (Refer to (Rottensteiner, 2007)