The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
476
The main reason is that these high-rise segments of study area
are covered densely by vegetations so that no laser penetration
is possible there. This causes them to be classified in both
“High-rise” and “Slightly variable” regions, and consequently to
be misclassified as “Building”. In order to eliminate these
erroneous pixels from “Building” class, an area threshold equal
to 70 square meters is defined and applied to the resulting image.
The final results are shown in Fig. 6(b). As can be seen all the
remaining polygons are parts of the roofs of buildings.
Figure 6 - (a) detected building pixels -including high-rise,
dense vegetation- (b) Building detection final results
At the middle of some building roofs in Fig. 6(b) there are some
pixels classified as non-building while they also should belong
to the building class. This happens wherever a high-rise feature
exists on the roof. In the case of the large building at the east of
the scene, the height difference on the roof changes so abruptly
that our system fails to detect the whole building. As a result,
our system has detected two individual buildings there.
4. ACCURACY ASSESSMENT
The final output of our system is a raster image with pixels
classified into two classes; “Building” and “Non-building”
pixels. To assess the accuracy of the detected buildings, we
have compared our results of building detection with the
reference map provided by the data provider. Due to some
unknown reasons one of the buildings of the scene is not
defined in the reference map. So we excluded the corresponding
detected pixels from our results. Fig. 7 shows our results and the
reference map.
Figure 7 - Building detection results versus Ground truth
The comparison of the resulting image with the reference map is
done by the calculation of the Confusion Matrix using the RSI
ENVI4.2 software. Table (1) shows the computed confusion
matrix.
Confusion
matrix
Reference Map
Our Results
Building
Non-Building
Total
Building
82185 (TP)
1380 (FP)
83565
Non-Building
9596 (FN)
133783 (TN)
143379
Total
91781
135163
226944
Table (1) The pixel-by-pixel comparison results for our system.
In Table (1), TP (i.e. True Positive) shows the number of pixels
which have a “Building” label in both datasets. Similarly TN
(i.e. True negative) equals to the number of pixels having “Non
building” labels in both compared datasets. The definition of FN
and FP numbers is straightforward.
The evaluated data in Table (1) are the results of a pixel-by
pixel comparison between our results and the reference map.
The two following objective metrics [Lee et. al 2003] are
employed by some authors (Sohn and Dowman 2007) in order
to provide a quantitative assessment of our building detection
system.
Completeness = 100 * (TP / TP + FN)
Correctness = 100 * (TP / TP + FP)
The evaluated amount of Completeness metric for our results
equals to 89.5% which shows the building detection percentage
[Sohn and Dowman 2007]. And the amount of Correctness
metric for our results equals to 98.3% which shows this
percentage of the “Building” detected pixels belong to buildings
indeed.
The commission and omission errors are also evaluated in
percentages and listed in Table (2). Errors of commission
represent pixels that belong to another class that are labeled as
belonging to the class of interest. Errors of omission represent
pixels that belong to the ground truth class but the classification
technique has failed to classify them into the proper class.
For instance, the amount of 1.65 for Commission error of
“Building” class states that 1.65% of pixels in “Building” class
do not really belong to building roofs. The Omission error of
“Building” class in Table (2) shows that 10.46% of building-
roof pixels have been misclassified by our system as “Non
building”. It’s obvious that Commission and Omission errors
for “Building” class are complementary amounts to the
aforementioned Correctness and Completeness metrics.
Class
Error
Commission
Omission
Building
1.65
10.46
Non-building
6.69
1.02
Table (2) The evaluated Commission and Omission errors for
our results in percentages
The Overall accuracy is another metric which evaluates the
accuracy of any classification process. This metric can be
evaluated using the following formula: