1157
tu
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
In the object-based classification, the true positives were the
same as for the other method. The difference in the object-
based method is in the number of false positives. Although
there are still many features incorrectly flagged as changes,
there are significantly fewer than in the decision tree method.
Of the 965 buildings present in the topographic database, 886
(92%) were detected by the method, leaving 79 predicted to be
demolished. Of these, 46 were false alarms, caused mainly by
low-rise buildings such as industrial sheds and small residential
properties that were in the topographic database but were below
the height threshold required to be classified as buildings. For
new buildings, 16 of the 17 actual changes were identified, but
84 other objects were flagged as new buildings but turned out to
be false alarms. These errors were mainly from objects such as
caravans, lorries and shipping containers being mis-identified as
new buildings and from garages and that fall outside the
Category A specification.
5. CHANGE DETECTION PRODUCTION TRIAL
The results of the change detection led us to choose the object-
based method for further development. To test the method in a
more realistic environment, a prototype production system was
developed, using a combination of software already used within
the production area and software required for the classification
and change detection process. It was decided that the process
would be tested on a live production job and directly compared
with the manual process of change detection currently
employed.
Two test sites in Sunderland in NE England were chosen for the
production trial, one area (site A) of 23 km 2 , the other (site B)
of 25 km 2 The DMC imagery was processed to provide the
inputs required by the object based classifier. These inputs
consisted of an orthorectified 4-band image mosaic, a DSM and
a slope map derived from the DSM. The newly-acquired
NGATE module of SOCET-SET was used to produce the DSM
fully automatically (i.e. without any seamline editing or other
manual processing). The same rule-set used on the first test
dataset was also used in the trial, to test whether the rules could
successfully be applied to different areas.
The object-based change detection method was applied to the
Sunderland image data and OS MasterMap topographic data.
This resulted in a set of polygons representing potential
demolitions and new buildings. These were presented to the
image interpreters using a similar user-interface to the one they
would normally use for manual change detection. The interface
was modified slightly to automatically direct the user to each
potential change in turn, and zoom in to the image at that point.
The user then compared the image to the topographic data in
the area of potential change and clicked a button to either
accept or reject the change. Once a button was clicked, the
result was recorded and the user was immediately presented
with the next potential change on the list.
5.1 Results of the trial
In site A, 142 potential changes were detected. Of these, 35
were accepted as real Category A changes (representing 25%
completeness) and just one real change was missed. In site B,
427 potential changes were detected, of which 77 were accepted
as genuine (18%). There were 14 real changes that were missed
in site B, which are discussed below. In terms of time taken to
identify the changes in the two sites, it was found that the
application of the automatic change detection process reduces
the overall time by 50% (compared with the manual process).
This is a significant improvement on the manual process, and
would be improved further by some small changes to the user
interface, which were requested by the operators.
5.2 Changes that were missed
The results show that 14 of the changed features were not
detected by the automatic process, even though the initial test
had very good completeness statistics. The nature of the
omissions in site B were:
• 5 demolitions
• 4 new buildings
• 5 minor modifications to school buildings
Of the 5 demolitions, 3 were small buildings which did not have
an address in the database, and therefore were assumed to be
insignificant, non-residential buildings. In the manual change
detection process, these were recorded as changes, even though
they did not meet the exact criteria to be considered as Category
A change. This is a situation in which the human operator will
err on the side of caution, while the automated process simply
filters these features out. A second reason for the omissions
also involves an interpretation of the specification, this time
relating to minor changes to school buildings. The automated
process ignored minor changes to existing buildings, while the
human operator treated school buildings as a special case and
marked up any changes in such features. This could be
overcome by identifying schools in the topographic data and
treating them as special cases in the automatic process. A third
reason for the discrepancies was the time lapse between the
imagery and the topographic data. In several cases, there were
genuine changes between the map data and the image, but the
map data was more up-to-date, having been updated via field
survey since the imagery had been flown. The human operator
can readily verify this, while the automated process cannot.
5.3 Opinions of the operator
When asked for his opinion of the automated process, the
operator considered that it would halve the time taken to collect
the change data, and that it was “a very useful tool for change
intelligence”. To determine whether this evaluation of the
process is widely accepted, further tests of the system are
planned.
6. CONCLUSIONS AND FURTHER WORK
An automated tool for detecting changes to buildings using an
object-based classifier has proved to be sufficiently successful
in a research environment to be taken up in a production trial.
In this trial, the tool was well received and was shown to
significantly reduce the amount of time taken to identify the
changes over a 48 km 2 area. At the time of writing, a large
production trial is planned, which will use recently-captured
imagery in an area which has changed significantly since the
previous update to the topographic data. This will allow us to
test the method in another area, and to determine its
effectiveness and efficiency in an area with a large number of
changes.