International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
operation was also applied to the AIRSAR land cover map 1
and the ERS-1 land cover map of 05-09-1994. This cross
operation achieved the best results.
Mismatch occurred mainly in areas classified as secondary
forest in the AIRSAR classification and as primary forest in the
ERS-1 classification, which is most likely due to mis-
classification of secondary forests in the ERS-1 classification.
Furthermore, the pastures of the AIRSAR classification have
mismatches with the secondary forest of the ERS-1 land cover
map.
The conformity of the cross maps between the AIRSAR land
cover map | and the selected ERS-1 land cover maps is
presented in Table 4. The recently cut areas were all classified
into other classes. In the ERS-1 images these areas were not
detectable and therefore they could not be classified, but in the
AIRSAR image the recently cut areas were clearly visible and
consequently also classified.
Similar results were achieved for the conformity of the second
AIRSAR land cover map and the ERS-1 land cover maps.
4.2.1 Discussion
To understand the results of the cross operations, the input
classifications need to be considered. The original data, i.e. the
AIRSAR land cover maps and the ERS-1 land cover maps,
show already differences when comparing them visually, but
also concerning the number of classes, the classification
accuracies, the number of bands and polarizations.
The AIRSAR land cover maps show more details. In the
upscaled AIRSAR maps, several small patches of the other
classes can be distinguished within the large area of primary
forest. In the ERS-1 maps these cannot be differentiated. But
also large areas, e.g. parts of the secondary forest and of the
pastures were not classified correctly in the ERS-1 maps. With
regard to the size of the patches or the number of pixels in a
patch, the small patches have a large number of boundary
(mixed) pixels and few interior (pure) pixels, which can lead to
non-detection and misclassification. Large patches have a large
number of interior (pure) pixels relative to the fewer boundary
(mixed) pixels, so thev have a higher chance to be detected and
classified correctly.
Four pure land cover classes can be distinguished clearly in the
AIRSAR land cover maps, while the ERS-1 land cover maps
have three pure classes and one mixed class. The class recently
cut areas could not be detected in any of the ERS-1 land cover
maps. Due to the start-up of the ERS-1 monitoring svstem
developed by Bijker (1997), not every change was detected
immediately with the image of 28-09-1993 or with the image of
05-09-1994. The land cover changes detected with the images
consist of real, recent land cover changes and the learning effect
of the monitoring system, detecting "old" changes not yet
registered. Due to this learning effect, classes were detected
afterwards that were there but could not be detected earlier.
The AIRSAR land cover map has an overall accuracy of 95%,
while the overall accuracy of the ERS-1 land cover maps ranges
from 65% till 70%. In the original ERS-I land cover
classifications the accuracy for the secondary forest was already
low. It reached only 43%. On the other hand, the pastures
showed a relatively high accuracy (8696) in the ERS-1 land
cover maps, since they were relatively well distinguishable
from the other classes.
Furthermore, the AIRSAR sensor has three bands, the C-, L-
and P-band, all fully polarimetric. These bands are
complementary. The combination of the three bands and their
polarizations makes it possible to accurately separate the four
land cover classes: primary forest, secondarv forest, pastures
and recently cut areas. The ERS-1 sensor has only one band, the
C-band, with only VV polarization. Consequently, the upscaled
AIRSAR land cover maps present more information and a
bigger variety of objects with different shapes and sizes than the
ERS-1 land cover maps.
5. CONCLUSIONS AND RECOMMENDATIONS
5.1 Conclusions regarding upscaling
Both upscaling approaches, stepwise and direct, showed similar
results. Nevertheless, the second approach, direct upscaling
from the same basis to the desired levels of spatial resolution,
was selected because it presents a slightly better outcome
concerning the changes in the number and size of patches.
Another advantage of this approach is the shorter processing
time for the implementation of upscaling, since the interim
results do not need to be calculated. The upscaling leads
immediately to the desired output pixel size.
5.2. Conclusions regarding land cover maps
The cross operation between the AIRSAR land cover map 1 and
the ERS-1 land cover map based on the image of 05-09-1994
provided best results, since it resembles best the AIRSAR land
cover map |, despite of the time lag of one year.
General reasons for non-conformity of classes can be errors in
geo-referencing or in upscaling as well as errors within the
input classifications. Apart from failures of detection or mis-
classification in the input land cover maps caused by mixed
pixels along boundaries of patches, the differences in the
number of spectral bands and polarization between the original
AIRSAR image and the ERS-1 images cause differences in the
classification accuracies of the input data. For this study area
with its particular land cover, the effect of the spatial resolution
Is not as determining as expected.
Concerning the order of upscaling and classification, it was
found that both possibilities provide similar results. It has to be
mentioned that upscaling before classification can add un-
certainty to the pixel value. The classification accuracy could
be prejudiced accordingly. Therefore, it is suggested to apply
classification before upscaling.
The combination of low spatial resolution and high spatial
resolution imagery gives better results than only using frequent
low spatial resolution or only infrequent high resolution. With
the help of high spatial resolution data the information from the
lower spatial resolution data can be improved. Locations of
classes can be derived, failures in classification can be corrected
and consequently the accuracy will improve.
5.3 Recommendations
Since the effect of spatial resolution was less determining than
the number of bands and polarizations of the sensor, it can be
suggested to use another radar sensor for the monitoring system
960
Inter
