Distance 
0.7 | 
  
  
  
  
  
  
  
0.5} 2 3 
4 5| 6| 7| 8 9 10 1 Level I 
  
  
  
  
  
  
  
M "f 
  
  
  
  
  
  
02} Ed 
0.1} 
  
  
  
  
  
  
  
  
  
  
  
Level II 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
3 410 9412532 6 82926 5201244 45 35 19 39 46 42 14 28 23 40 30 48 16 47 18 17 49 36 13 52 24 34 11 15 31 33 7 27 43 37 38 51 50 22 21 2 
  
  
Figure 3 The dendrogram of the last 51 mergers in the aggregating process. The final vegetation regions are compiled from the 
clusters created at level I. 
Modification 
Some extreme cases such as islands and barriers have to 
be removed from regions. Islands are small units located 
within large regions. They are very different from their 
surroundings (e.g., water bodies, urban areas) so that they 
cannot be merged into the regions nearby. These islands 
were forced to join one of their neighboring regions. 
Barriers are another type of extreme units. In most cases 
they are small elongated units cutting through two very 
similar units. These barrier units were also identified and 
removed, resulting in the merge of similar areas which 
might otherwise be arbitrarily separated. The ten final 
regions are shown in Figure 4. 
Evaluation 
The fundamental difference between the region partition 
method developed in this research and the traditional 
hierarchical clustering method is that the former employs 
the polygon neighborhood relationship to maintain the 
1004 
contiguity of regions. Each pair of units merged were the 
most similar neighbors but not necessarily the most 
similar units in the study area. It is necessary to evaluate 
whether natural differences exist among the regions 
generated. One method used was comparison of within- 
and between-region variations. When constructing 
regions, the attempt is to maximize homogeneity within 
a region and maximize heterogeneity between regions 
(Amedeo and Golledge, 1975). Low variances are 
indication of homogeneity of regions while high variances 
are indication of heterogeneity. The multivariate F-test 
was used to compare variances and it was found that 
between-region variance was significantly larger than 
within-region variation at œ=0.1 level. The second 
evaluation method used was comparison of the means. 
For each NDVI composite, the means of the ten regions 
were compared using Least Square Means method. No 
pairs of regions were found equal in more than two 
layers. These two types of evaluations prove that the 
regions generated by the model differ in NDVI of the 
growing season. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996