In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010,1 APRS, Vol. XXXVIII, Part 7B
279
separated by the textural and the spectral features, the structural
features used in the classification are still not efficiency enough,
and additional features need to be investigated in the future. The
classification errors shown in the confusion matrix also cause
some of the completeness and correctness values in Table 3 to
be very low.
class
completeness
correctness
tilled cropland
78,9%
97,8%
unfilled cropland
57,1%
100%
cropland (untilled + tilled)
77,1%
97,9%
grassland
96,6%
61,0%
Table 3: Completeness and correctness of the classification.
Figure 4. Training and test objects super-imposed to the
Ikonos scene. Blue: training objects; green: correct
test objects; red: errors.
Figure 5 shows an example for a grassland object misclassified
as ‘untilled cropland’. The reason for this is the fact that the
bare soil is visible in a large part of the object.
Figure 5. Classification errors, 'grassland' object classified as
'untilled cropland'
As mentioned before, the main focus of our approach is the
verification of the GIS objects. It is embedded in a semi
automatic system that uses the automatic tool to focus the
attention of the human operator to possible errors in the GIS.
Thus, work is saved largely due to the fact that the operator
needs no longer to check any object that was accepted by the
automatic module. Under these circumstances, and given the
fact that quality control is essentially carried out to remove
errors in the data base, classification errors that cause errors in
the GIS to remain undetected, i.e. the erroneous acceptance of a
wrong object, are to be avoided by all means. As a
consequence, the acceptance of objects has to be very reliable.
On the other hand, the erroneous rejection of a correct GIS
object may reduce the economical effectiveness of the system,
but it will not result in an error remaining in the data base. The
confusion matrices for the verification process carried out on
the basis of the classification results described above are
presented in
Table 4 and 5 for cropland and grassland objects, respectively.
Note that these numbers also contain objects that were rejected
based on the texture-based classification described in (Busch et
al„ 2004).
^~~~~--~~-^automatic
reference
accepted
rejected
correct
162 (66.4%)
57 (23.4%)
false
1 (0.4%)
7 (2.9%)
Table 4. Confusion matrix for the verification of cropland.
^~~~--^_automatic
reference ^
accepted
Rejected
correct
20 (22.5%)
61 (56.6%)
false
0 (0.0%)
25 (28.1%)
Table 5: Confusion matrix for the verification of grassland.
The confusion matrix in Table 4 shows that our approach does a
reasonably good job in verifying cropland objects. Only one of
eight errors in cropland objects (Table 1) remains undetected,
and the number of wrong cropland objects in the GIS is thus
reduced by 87.5%. The economical efficiency is at 66.4%, i.e.
66.4% of the cropland objects need not to be inspected by the
human operator because these objects were accepted
automatically. Of the 23.4% of the objects that are erroneously
rejected by the system, 3.3% were rejected by the texture-based
classification described in (Busch et al., 2004). Unfortunately,
the verification of grassland objects is far less successful. On
one hand, all errors contained in the GIS could be detected, but
on the other hand, the efficiency of the system is only at 22.5%.
Of the objects rejected erroneously by the system, 5.6% can be
attributed to texture-based classification. Unlike with cropland,
the texture-based classification rejected one object correctly. It
is clear that the classification of grassland and cropland objects
still needs to be improved. In particular, the structural features
used for classification seem to require a revision.
5. CONCLUSION AND OUTLOOK
The method used to separate cropland from grassland objects
described in this paper achieved reasonable results when
applied to the verification of cropland objects, but the results
for grassland objects are still unsatisfactory. In the future we
will revise the structural features used for classification, which
apparently fail to separate grassland from cropland objects
properly in the current version of the approach. For instance,
rather than focusing on the orientation of lines alone, we could
also consider the distance between lines by designing features
that highlight periodical patterns corresponding to parallel lines.
Structural features based on other types of analysis, e.g.
variograms, could be added to the classification process. In
addition, we could try to use training for determining the
parameter y of the Gaussian Kernel in the SVM classification.
Finally, we need to analyse which features are the most relevant
ones and have the biggest impact on the classification result.
The main goal of our approach is its application for the
verification of ATKIS grassland and cropland objects. In
ATKIS one agricultural object may consist of different
management units. For instance, a cropland object may consist
of fields covered by different crops. It has been stated above