| 2004 
> low. 
oil. 
; were 
verage 
most 
99. in 
wo of 
one in 
ing on 
res for 
andard 
> most 
class 
ilarity, 
e most 
ication 
ind the 
ximum 
hat the 
ervised 
| cover 
e seen 
Lower 
to bad 
  
  
s was to 
for each 
ing data 
Inmixing 
y. Fuzzy 
Jculating 
mbership 
; method 
Bayesian 
Training 
ore than 
«els. were 
es. The 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
a'posteriori probabilities produced by BML were used to 
represent the proportions of different land cover classes within 
each MODIS-pixel. 
The a'posteriori probabilities were compared to training 
proportions of different classes and measures like bias, RMSE 
and correlation computed. The best classes were Agricultural 
land (bias —15.7, RMSE 40.2, correlation 0.61), Open bog (-2.8, 
17.4, 0.55) and Water (-3.3, 27.1, 0.43). The worst classes were 
forest classes, then the correlations were very low (0.20-0.04). 
4.3 Classification of featuresets 
The classification errors of land cover classes are represented in 
figure 3. For each featureset, classifications were performed 
with varying k for both training and test sets. The a'priori 
probabilities of classes were equal. The final overall accuracy 
was estimated as the mean overall accuracy of the training and 
test sets with highest test set overall accuracy. Overall 
accuracies of different featuresets are rather low, the feature 
extraction using principal component analysis and texture 
increase the accuracy only a little. The use of MODIS-features 
increases the overall accuracy more, from 43% to about 50%. 
Classification accuracy of land cover classes 
100; 1 T f T 
  
  
   
  
S 
8t 70| 
> 
o 
S 60; eve ee RX 
am rm eT im Nu Wiss RC Mein me 
o ——— m M 
© 
c 5 
S... | 
5 | 
Si | 
ve i 
9s 
q 30| 
Oo | 
20} 
10} 
0* n 4. Y ed À. t 
0 1 2 3 4 5 6 7 
Featureset 
Figure 3. Solid line represents the overall accuracy of land 
cover classes, dashed line the mean of producer's accuracies and 
dash-dot line the mean of user's accuracies of classes. Lines 
with "x" perpresent the cases when the a'prior probabilities 
have been estimated from MODIS images. Lines with "o" 
represent the cases when classes have been merged to four. 
The best class is water, its classwise accuracies are about 90% 
and the variation between featuresets is quite low. The classwise 
accuracies of forest classes are low, varying from 2096 to 4594. 
Forest classes are mainly mixed with each other but also to 
agricultural land with featuresets 1-3. The class agricultural land 
is classified reasonably well with featuresets 4-6, then the 
classwise accuracies are more than 80%. Agricultural land 
benefits from the use of MODIS-features. Classwise accuracies 
are lower with featuresets 1-3, producer's accuracies are less 
than 30% and user's accuracies about 80%. This means that if 
image pixel is classified as agricultural land, then it quite likely 
that it is agricultural land in the field. Agricultural land is 
mainly mixed with open land and pine dominated forest. The 
accuracies of class open bog are moderate with featuresets | and 
2, between 60-70%. The use of texture and MODIS- 
classification decrease the classwise accuracies. Open bog is 
mainly mixed with pine and spruce dominated forest. The class 
open land is classified rather badly, classwise accuracies are 
aroung 40%. It is mainly mixed with forest classes and 
93] 
agricultural land. The mixing of different classes happen a quite 
similar way with different featuresets. 
The overall accuracies of tree species vs. development class 
classification varied from 20% (featuresets 1 and 2) to 30% 
(featureset 5). When the classes are merged to three tree species 
classes, then the overall accuracies vary from 38% to 45%. 
When the classes are merged to four development classes, the 
overall accuracies vary from 49% to 57%. This means that the 
features are a bit more sensitive to the size of the trees than 
species. 
The overall accuracies of tree species vs. soil type classification 
varied from 30% (featuresets 1 and 2) to 38% (featureset 5). 
When the classes are merged to three tree species classes, the 
overall accuracies vary from 51% to 57%. When the classes are 
merged to three soil type classes, the overall accuracies vary 
from 47% to 53%. This means that the features are a bit more 
sensitive to the tree species than soil type. 
4.3.1 A’priori probabilities from MODIS: The use of the 
a'posteriori probabilities of MODIS NDVI-classification as 
a'priori probabilities of Bayes classification increases the overall 
accuracies of featuresets 1-3 from about 43% to more than 50% 
(figure 3). The increase is smaller for featuresets 4-6, which 
already use information from MODIS in one way or another. 
The increase of overall accuracy is about 1-3 %-units. The 
increase of classwise accuracies are very small for class water, 
otherwise the increase can be even 25 %-units with the 
featuresets 1-3. Increase is much smaller for featuresets 4-6. 
When the mixing of classes is studied, it can be seen that forest 
classes are more mixed between each other than using equal 
a'priori probabilities. 
In the cases of tree species vs. development class or soil type 
the a’prior probabilities of classes were acquired from the 
a'posteriori probabilities of classes pine, spruce and deciduous 
tree of MODIS-classification. These did not increase the overall 
4 
accuracies much, only 1-3 %-units depending on the featuresets 
or classes. 
4.3.2 Merging of classes: The merging of classes to four classes 
(water, forest, agricultural and open land, and open bog) 
increases the overall accuracies to about 65%-75%, depending 
on the featureset (figure 3). The accuracies are lower with 
featere sets 1-3 when equal a'priori probabilities are used, but 
increase as MODIS-features are used. When a'priori 
probabilities have been acquired from MODIS-classification, 
the overall accuracies do not vary much between different 
featuresets, they are always around 75%. 
4.3.3 Comparison of pixelwise and standwise classification: 
Tree species classes of land cover classification system were 
used to compare pixelwise and standwise classification. Data 
for standwise classification was acquired by computing the 
stand means of the featuresets. Stand means were divided to 
training and test sets and classified using Bayes rule for 
minimum error. Figure 4 illustrates the overall accuracies as 
well as the mean of user's and producer's accuracies as the 
function of featuresets. In general, standwise classificaton 
increases the overall classification accuracy, the gain is 10%- 
units or more. But the drawback is that as the classwise 
accuracies of pine increase, those accuracies for deciduous tree 
decrease. This is most likely due to that the pine stands are 
larger than deciduous tree stands, so mean values are more 
reliable for pine and the amount of stands for deciduous tree is 
rather small, especially compared to pine.