1156
'5*- - v f 4 r r
■ -
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
In order to test the accuracy of the classifications in a change-
detection process, the classified buildings were compared with
the building objects in the Ordnance Survey topographic
database (used to generate the OS MasterMap® product). In
the test area, all the significant building objects which were
either greater than 50 m 2 in area, or had a postal address (and
hence were residential buildings) were determined from the
topographic database. In total, this revealed 965 buildings to
test, of which 34 had been demolished. In addition to these,
there were 17 significant new buildings on the image, making a
total of 51 Category A changes.
4.2 Detecting demolitions and new buildings
Both the decision tree and object based classifications were
tested using the same method of post-classification change
detection. Demolitions and new buildings were considered
independently.
Demolitions were identified by intersecting the areas classified
as buildings with the known OS MasterMap® building
polygons. For each building polygon, if at least 50% of its area
was classified as a building, that building polygon was
considered to be verified by the classification. If less that 50%
of its area was classified as a building, then it was considered to
be a change (i.e. the building was considered to have been
demolished).
To identify new buildings, the first step was to mask out all the
regions in the test site where constructions would be unlikely to
have occurred. These consisted of all roads, rail or water
bodies present in the OS MasterMap® data. All existing
buildings in the data were also masked out, together with a 3 m
buffer around each building, to help eliminate any remnant-
objects produced by misalignment between the image and the
topographic data or by the draping effect of the DSM. The
remaining area, consisting of vegetation, farmland and man
made surfaces, was then searched for any objects classified as
buildings. Objects smaller than a given size threshold were
filtered out, to leave a set of potential new buildings.
4.3 Results of post classification change detection
Table 2 shows the results for the decision tree classification,
and Table 3 shows the results for the object-based classification.
It can be seen in both cases that, of the 51 Category A changes
on the image, 49 were successfully identified, with only one
actual change not flagged as a change (false negative) each for
demolitions and for new buildings. These errors were caused
by a single feature - a residential building that had been
demolished and rebuilt with a similar footprint. Such rebuilds
are inevitably difficult to detect when the footprint in the map
database of the recently demolished building is similar to the
footprint in the image of the building constructed in its place.
Demolitions
New
Total
Actual changes
34
17
51
Objects classified as
changes
288
161
449
Actual changes correctly
classified (True Positives)
33
16** '
49
Non-changed objects
classified as change
(False Positives)
255
150
405
Actual changes not
classified as changes
(False Negatives)
1
1
2
% Classified as changes
that were actual changes
11%
10%**
11%
% of actual changes
classified as changes
97%
94%
96%
Table 2. Results of change detection from decision tree
classification for the 2 km 2 Heathrow test site.
In the decision tree results (Table 2) there are a large number of
false positives, in which objects which haven’t actually changed
are falsely identified as changes. These errors are caused by a
variety of factors. One factor was that the change detection
works on classified features, rather than individual pixels. In
order to do this, the groups of pixels classified by the decision
tree had to be grouped into contiguous areas and converted to
vector form. The process of grouping and vectorising
inevitably leads to a slight degradation in the quality of the
results. A second factor is the misclassification of an area of
active construction as a building. In the construction site,
earthworks for a new road were of a similar spectral nature to
the buildings, and were elevated above the average ground
surface height, making them similar in height to the building
objects nearby. Other false alarms were caused by the presence
of large vehicles and shipping containers - features which were
much in evidence in this area of active construction work.
Demol
New
Total
Actual Changes
34
17
51
Objects classified as changes
79
96
175
Actual changes correctly
classified (True Positives)
33
16"*
49
Non-changed objects classified
as change (False Positives)
46
84
130
Actual changes not classified as
changes (False Negatives)
1
1
2
% Classified as changes that
were actual changes
42%
17%
28%
% of actual changes classified
as changes
97%
94%
96%
Table 3. Results of change detection from object-based
classification for the 2 km 2 Heathrow test site.
Of the 161 objects flagged as new builds, 11 contained actual
changes, but this included 16 individual new buildings as
many were close together and so were merged in the
classification
* Of the 88 objects flagged as new builds, 12 contained actual
changes, but this included 16 individual new buildings as
many were close together and so were merged in the
classification