Nagendra, Harini 
classification (p<<0.01). The degree of bias in patch metrics due to unsupervised classification can be related to the 
Mann-Whitney index — the higher this index, the greater the bias. This bias in patch metrics estimated from 
unsupervised classification does not seem directly related to the degree of misclassification (difference between 
supervised and unsupervised classification accuracy). 
  
1 2 3 4 5 6 7 8 9 10 11 12 13 
SIZE | 41.5 1844 1516 135.6 |392 |690 [61.1 | 45.6 | 518 | 435 66.9 | 45.3 | 54.8 
SHP |403 1825 150.0 1344 [379 [626 [59.6 [44.1 | 50.0 | 424 65.0 | 43.6 | 55.6 
NND 1376 | 82.5 | 445 [309 | 368 |63.8 | 54.6 | 412 | 45.3 | 39.8 59.3 | 40.2 | 50.8 
SUN 13754 1321 1297 [462 1152 |349 1507 1416 1329 92 |500 | 394 | 32.0 
Table 1. Differences between estimates of patch size (SIZE), patch shape (SHP) and nearest neighbor distance (NND) 
derived from supervised and unsupervised classification. In the first three rows, values represent outcomes of a one- 
tailed Mann-Whitney U test assessing the significance of these differences, for patch metrics denoted by row headings 
and landscapes denoted by column headings. The cutoff value for significance at p=0.01 is 2.3. Values in the fourth row 
(S-UN) depict the difference in percentage accuracy of supervised and unsupervised classifications, for comparison. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Landscape metrics of mean patch size, mean patch shape, mean nearest neighbor distance and Shannon landscape 
diversity determined from unsupervised classification were also significantly greater than those determined from the 
supervised classification (p<0.01), based on a one-tailed Wilcoxon’s signed-ranks test. However, no significant 
differences could be detected for indices of interspersion-juxtaposition and contagion assessed from supervised and 
unsupervised classifications. 
Metrics of landscape structure quantify spatial as well as non-spatial pattern. The bias in the non-spatial landscape index 
of diversity is easily explained. The above-mentioned algorithm for unsupervised classification assigns pixels to 
spectral classes based on minimum Euclidean distance, and not based on a user-prescribed number of standard 
deviations. In supervised classification however, the spectral classes are user-defined, and may have greatly different 
variances. Spectral classes defined by unsupervised classification will therefore tend to have a similar variance when 
compared to supervised classification. As a result, spectral classes defined by unsupervised classification will also have 
a more even distribution of pixels. Estimates of diversity calculated from unsupervised classification will therefore tend 
to be higher. However, indices of interspersion-juxtaposition and contagion, both spatial metrics describing levels of 
patch and pixel type adjacencies, are not significantly biased by unsupervised classification. Indeed, there was no 
reason this might have been expected. 
Unsupervised classification however creates significantly larger patches, of more irregular shapes, at greater distances 
from neighboring patches of the same cover type. This is an unexpected result. As space is not explicit in the clustering 
algorithm, it is not apparent how a bias in spatial pattern can result from unsupervised classification. Differential 
within-class spectral variances combined with spatial auto-correlation may provide a clue to this behavior. 
In images with random noise, local variance tends to be similar to overall variance in the image. Real-life landscapes 
however tend to demonstrate high levels of spatial auto-correlation. Large features (patches) within the image contain 
clusters of pixels with very similar spectral values. The local variance tends therefore to be much lower compared to 
overall variance (Warner and Shank, 1997). Spectral image classification is an effort at providing spatial differentiation, 
to make clear this pattern of spatial heterogeneity. If spatial auto-correlation exists, the spatial characteristics of features 
delineated by the classifier will depend greatly on the within-class variance of class signatures. An increase in the 
within-class variance of spectral values should lead to the definition of larger features, i.e. greater patch size. 
Information on class signatures derived from supervised and unsupervised classifications was used to compare the 
within-class spectral variability of classes defined by supervised and unsupervised classifications. The intent was to 
determine whether classes identified by unsupervised classifications had significantly greater within-class variance in 
spectral values, compared to classes identified by supervised classification. For twelve landscapes, the median of 
within-class variances was calculated for signatures derived using unsupervised and supervised classifications, for all 
four bands of imagery. As can be seen from Table 2, the median within-class variance for unsupervised classification 
was higher than for supervised in all landscapes for bands 1, 2 and 3 — with the single exception of landscape 1 (band 
3). The median within-class variance in the infrared band for unsupervised classification is however higher than that for 
supervised in only 7 cases, and is in fact lower in 5 landscapes. A Wilcoxon’s signed-ranks test declared within-class 
variance for signatures derived using unsupervised classification to be significantly higher than for signatures derived 
from supervised classification for bands 1, 2 and 3, though not for band 4 (p<0.01). 
The results of this statistical test, though not conclusive, indicate that unsupervised classification tends to delineate 
classes with higher within-class spectral variance, compared to supervised classification. Given spatial auto-correlation, 
a higher within-class variance can result in the inclusion of more pixels, hence larger patches, while delineating patch 
boundaries. Due to the nature of the shape index, by chance alone, larger patches prove more likely to have irregular 
shapes (McGarigal and Marks, 1994). Also, given a bounded landscape, the creation of fewer, larger patches would 
imply that these patches are farther apart. Thus, nearest neighbor distances would increase. The difference in within- 
  
958 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 
class 
near 
| = | 
B/ 
| el 
B/ 
  
4 CC 
In re 
(Fon 
main 
our k 
Lanc 
patch 
and. 
1995 
land: 
Unst 
it cai 
estin 
et al. 
neigl 
Indic 
Furtl 
class 
purp 
class 
Inve: 
Asse 
isar 
coun 
there 
cogn 
impli 
ACK 
I thai 
N.V. 
and t 
trave 
Bask 
Rese: