ul 2004 
  
  
  
   
wLIRA[jG 
Figure 5: Standard per-pixel classification result using the 
maximum likelihood classifier. 
3.2 Classification performance (KHAT & Z-statistic) 
Figure 6 shows the LCM classification results expressed in 
KHAT values for the fourteen different MA values. The figure 
clearly shows that varying MA values have more effect on the 
more fragmented p1996 image than on the p1990 image. 
However, the sensitivity of both images towards different MA 
values is similar. Overall, KHAT values decrease for increasing 
MA. Speafically, the largest decrease is for MA values 
between 5.5 ha and 100 ha, whereas MA shows a small 
decrease between MA |, and MA, and a minor decrease 
between MA 259 and MA. 
The Z statistics showed that the majority of MA settings are 
significantly different at the 0.05 probability level. This means 
that almost all LCM-classification results are different. 
Specifically, the more the MA values are dissimilar, the higher 
the Z values. However, no significant difference occurs 
between MA ,4, and MA,;, MA y, and MA,,,, and MA, and 
MA3sp for the p1990 image. For the p1996 image a similar 
'pattern' occurs, but with a small shift of plus 50 ha, that is, no 
significant difference occurs between MA,s;, and MA 5s, MA» 
until MA;4,, MA ,4, and MA, It seems that both images react 
similarly on different values for MA, although the p1996 image 
is more fragmented than the p1990 image. 
-5- P1996 TM 
1,000 
0,950 
0,900 
0,850 
0,800 
0,750 
  
KHAT 
0,700 
0,650 
0,600 
0,550 
0,500 
  
Minimum Area (ha) 
Figure 6: KHAT values for different values of minimum-area 
(MA) in LCM classification. 
  
3.3 Variability and arrangement of forest cover (PLAND & 
NP) 
Figure 7 shows the LCM classification results as measured by 
the two class-related metrics PLAND (percentage of landscape) 
and NP (number of patches) for the two LCM forest classes 
mainly logged forest (mLF) and mainly heavily logged forest 
(mHLF), and the intermediate LCM class mainly shrub (mSH). 
For all MA values up to 400 ha, PLAND is almost constant 
(less than 2% cover difference) for all three LCM classes in 
both images, except for the LCM class mainly shrub of the 
pl996 image. This LCM class increases between MA ,, and 
MA |, from 34% to 39%. For MA values higher than 400 ha. 
PLAND changes for all three classes about 5% (cover). The 
figure shows also clearly that NP for p1996 is higher than NP 
for p1990 for the entire MA range. For the class mainly heavily 
logged forest NP of p1990 is higher, because of the enormous 
reduction of this class in 1996. 
  
-—e-mLF p1990 ||—«—iuLF 121996 
-U-mHLF p1990| |-5—iuHI F pian 
—k-mSH p1990 || mSli 2124 
50% 7 = — 
45% = 
40% in 1 | 
35% 
Percentage of Landscape 
  
  
  
  
  
  
3 
  
30% 
25% 
PLAND 
20% 
15% 
  
  
10% 
  
5% 
  
  
0% T T T T T T T T T T T T T 
+ € u s i > Ur > a ^ x 
oT mS SN o No o PDF 
^ 
Minimum Area (Ha) 
  
—O- mLF p 1990 ——rüLF iiit 
7U- mHLF p 1990 || —- niHI F 1886 
—- mSH p1990 ||-#rsil LYLE 
Number of Patches 
  
  
  
  
  
350 
  
   
| 
300 
250 
200 
NP 
  
3 S iu j 2 [s] z M D 
SS 0. m GR qv aM 5 SUO 
Minimum Area (Ha) 
Figure 7: LCM classification results as measured by the two 
landscape pattern metrics PLAND (percentage of landscape) 
and NP (number of patches) for forest cover. 
Being a fragmentation measure of the patch type. higher NP 
values for the p1996 image means that indeed this image is 
more fragmented than the p1990 image for all three vegetation 
classes. Generally, NP tends to decrease when increasing MA 
for all LCM classes and for both images. Specifically, the 
influence of MA on NP is largest for MA values smaller than 
100 ha; above MA oy NP stays constant for all classes. This