In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
477 
Fig. 3. (a) First-last-echo difference model and (b) the derived 
enhanced vegetation mask 
Next, building regions are segmented by inverting the DSM and 
applying a fill sinks procedure (Arge et al., 2001; GRASS 
Development Team, 2010). All high objects are considered as 
sinks and filled up to the minimum elevation in the individual 
region in order to guarantee a hydrologically consistent 
elevation model. This model is subtracted from the original 
DSM and thresholded at a certain minimum height in order to 
remove artefacts, i.e. overestimation of building outlines or the 
influence of low vegetation (Fig. 4). 
Fig. 4. Outline detection of building footprints by fill sinks and 
height constraint 
The remaining building segments are enhanced by applying a 
morphological opening, which further smoothes and removes 
remaining overestimation of the building outlines (Fig. 5). 
Fig. 5. Outline enhancement by morphological opening. 
The segments are classified into buildings and non-buildings 
using a classification tree (Breiman et al., 1993; Maindonald 
and Braun, 2007) derived from a training area. As training 
segments building footprints and non-building segments are 
selected from the derived segments. For those, several statistical 
features such as first order statistic on elevation, object heights, 
first-last-echo difference, standard deviation of slope and aspect 
derived from the DSM, and geometrical object properties such 
as area and shape indices. Table 1 lists all the input features 
which were calculated to build up the classification tree. 
By applying the classification tree (Themeau and Atkinson, 
1997) the sample set is divided into subsets which are tested 
and compared in order to define the optimal splitting rule 
between both classes. In fact, the developed rule base is a box- 
classifier in feature space, which has crisp thresholds at each 
node (rule). The features can occur in multiple hierarchies of 
the classification tree. The levels, i.e. the complexity of the 
classification tree, can be regulated by defining a complexity 
parameter, which is also known as pruning. In general, the 
complexity of a classification tree should be kept minimal in 
order to avoid modelling the data itself instead of describing the 
class specific characteristics. 
Object Feature 
DSM 
FLDM 
Segments 
Stdev object height 
X 
Mean object height 
X 
Max object height 
X 
Min object height 
X 
Mean FLDM 
X 
Area 
X 
Shape (perimeter/area) 
X 
X 
Shape (circumscribing 
X 
X 
circle) 
Stdev slope 
X 
Stdev curvature 
X 
Table 1. Object features calculated as classification input 
4.2 Change detection 
The building change detection procedure is based on the 
automatically extracted building footprints and their attributes 
exclusively. The procedure distinguishes the following cases: 
unchanged building or building part 
new building 
demolished building 
new building part 
demolished building part 
The change detection compares spatially related building 
footprints and their attributes derived from each epoch 
individually. In order to be able to detect also gradual changes 
at buildings such as the construction of a new story, not only 
the appearance of another object polygon is checked but also 
the mean difference of the elevation in the segment part. There 
are several methods how to measure detection success of 
building footprint extraction (Rutzinger et al., 2009) In the 
following the change detection results are evaluated by 
calculating the overall accuracy as 
overall accuracy = TP / (TP+FP+FN) (1) 
with true positives (TP), which are segment parts classified as 
change which are also changes in the reference and the false 
positives (FP), which are segment parts classified as change 
where no changes occur in the reference. False negatives (FN) 
are changes which are in the reference but are not detected by 
the method. 
5. RESULTS 
5.1 Building detection 
The vegetation mask is derived for both input data sets and then 
merged in order to get maximum vegetated area. The building 
segments from both epochs are derived by the fill sinks 
approach (Sect. 4.1) and were further selected by a minimum 
height of 2.5 m and minimum area of 10 sqm. The shape of the