Thë International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
2.4. Decision process
The statistical tools of least squares adjustments provide the
mathematical foundation necessary to come with a valid
analysis of the obtained results and automatically decide
whether change occurred or not. During the execution of the
matching loop, several criteria are considered such as Shift dX.
A threshold of 0.5 pixels is set to classify a matching as
successful. During the application of the least square matching
process, a significant amount of visually successful matches
was not correctly identified, because the value of the shift was
not close to our threshold. The distinction between a successful
match and an unsuccessful one is the range of dX. If the
variation is less than 0.5 pixels, then the match is considered
successful, otherwise rejected.
Figure 4: GDs in the case of an edge represented by three levels
3. RESULT AND DISCUSSION
When multiple edges exist, they can result from the actual shape
of the edge or random condition such as noise, shadow,
different projection, etc. We compensate for theses effects by
introducing a scaling factor describing the expected geometry of
the edge. Two new GDs are created on the both side of the
vector edge. The means of these GDs are assigned the values of
//* ± dj depend on which side they are (Fig. 5). The standard
deviation of the expected geometry GDs expresses the
uncertainty of this information and it is defined as a constant
based on the distance of each edge to vector edge i/, and d 2 ■
The value of standard deviation corresponds to value of d i 13 ,
for a threshold of 99%. The experiments showed that the
standard deviation should be limited to 40% at the distance of
standard deviation to compensate random condition. The
mathematical expression of scaling factor is:
(x-/j.-d t ) 2 (x-tu+d 2 f
Kj(x) = Max{e ~ 2a ' ,e ~ 2cT ‘ }, (21)
(■x-M,) 2
G,(x) = e ~ 2a ‘
The weight distribution P(x) on the axes perpendicular to the
vector edge is given by the following formula:
P(x) = Max{K, (x) x G, (x), G 2 (x), K 3 x G 3 (jc)} (22)
Figure 5: Gaussian distribution in the case of an edge
represented by four levels
The developed algorithm in this paper is implemented using
Microsoft's visual C# Net < Visual Basic 6 and Esri's
MapObjects ActiveX component. The datasets and assessment
process is described as follow. The available datasets to test
change detection algorithm performance is described in table 1.
datasets
properties
Digital image of
city
• IslamAbad
• 1/10000
• 1988
• TIFF format
Digital image of
city
• The same
area(IslamAb
ad)
• 1/10000
• 2002
• TIFF format
Vector
information
• 1988
• Esri's
shapefile
format
Table 1: datasets and their characteristics to test proposed
methodology
3.1 Evaluation of implemented system
3.1.1 Accuracy assessment of developed least square
matching algorithm
The accuracy of the measurement refers to how close the
measured value is to the true or accepted value. So, we have
considered 50 buildings with known information in the test
area, and compared the implemented system results with
expected ones, in order to evaluate the accuracy of developed
building change detection algorithm. The result of this
evaluation has been shown in table 2.
Comparison between expected and
computed result
Percentage
(%)
Buildings that algorithm can
detect their changes correctly
Upper 70%
Buildings have changed but the
algorithm can not detect their changes
Under 10%
Buildings have not changed but the
algorithm detect changes;
Approximately
20%
Table 2: the accuracy assessment of implemented system