en
id-
> 1S
m,
led
ery
, at
on.
the
um
ted
the
ved
ass
ean
her
| to
nce
Nagendra, Harini
are then merged. New cluster means are calculated for each band as the average of intensity values in that band for all
pixels present in that cluster. Pixels are then again reassigned to clusters, and this process repeated until the
correspondence between iterations reaches a user-specified level.
The number of clusters was specified as the same as the number of classes in the supervised classification, for
comparability. The minimum distance required between clusters was input as zero, to ensure that the final number of
clusters was the same as that initially specified. The desired correspondence between iterations was specified as 9896.
Once the iterations converged, class signatures (mean in four bands, and the variance-covariance matrix) were assessed.
These signatures were then input into a maximum likelihood classifier, to create a map derived from unsupervised
classification.
Classification accuracy was verified by field visits during the months of August and September 1997 (Nagendra and
Gadgil, 1999 b). Between 70 and 120 points, randomly distributed in each landscape, were used to compare supervised
and unsupervised classifications with the ground data. Each class in the unsupervised classification was assigned to a
unique land cover type in order to maximize correspondence.
2.3 Comparison of Landscape Metrics
Supervised and unsupervised classification maps were used to derive estimates of patch and landscape structure,
through the program FRAGSTATS 2.0 (McGarigal and Marks, 1994). This program computes several metrics of
structure for each patch in the landscape, as well as for the entire landscape. A subset of these was selected, based on
their relevance to species distribution (Forman 1995). Three patch metrics were identified as being concerned with
different aspects of patch structure - patch size, patch shape and the nearest neighbor distance (distance from a patch to
the nearest neighbor of the same type). At the landscape scale, indices of mean patch size, mean patch shape, mean
nearest neighbor distance, Shannon's diversity, interspersion-juxtaposition index and contagion index, were computed.
Definitions of patch and landscape metrics are provided in McGarigal and Marks (1994). They can be summarized as
Patch metrics:
a] Patch size: Area covered by a patch, in hectares.
b] Patch shape index: Complexity of patch shape, compared to that of a square patch of identical area. For a single
patch, the shape index is 1 when square, and increases without limit as the patch becomes more irregular.
c] Nearest neighbor distance: Edge-to-edge distance between the patch and its nearest neighbor patch of the same type.
Landscape metrics:
a] Mean patch size: Average area covered by a patch.
b] Mean patch shape index: Average complexity of patch shape, compared to that of a square patch of identical area.
c] Mean nearest neighbor distance: Average edge-to-edge distance between two nearest-neighbor patches belonging to
the same land cover type.
d] Shannon diversity index: This measure of landscape diversity increases as the number of land cover types increase
and/or as the proportional area distribution of these types becomes more equal.
e] Interspersion-juxtaposition index: The degree to which patches belonging to different land cover types are
interspersed. This value decreases as the distribution of adjacencies among types becomes increasingly uneven.
f] Contagion index of pixel interspersion: The degree to which pixels belonging to different land cover types are
interspersed. This value decreases as the distribution of pixel adjacencies among types becomes increasingly uneven.
For each landscape, a one-tailed Mann-Whitney U Test was used to assess whether patch characteristics estimated from
the unsupervised classification differed significantly from those measured by supervised classification. A one-tailed
Wilcoxon's signed-ranks test was used to assess whether landscape metrics estimated from unsupervised classification
significantly differed from those estimated using supervised classification (Sokal and Rohlf, 1981).
Information on class signatures derived from supervised and unsupervised classifications was compared. For each
landscape, the median of within-class variances was calculated for signatures derived using unsupervised and
supervised classifications, for all four bands of imagery. This information could only be computed for twelve out of
thirteen landscapes, for which class statistics were available. Values taken by the four spectral bands may be correlated
to varying extents. A one-tailed Wilcoxon's signed-ranks test was therefore separately carried out for each spectral
band, to assess whether within-class variance for signatures derived using unsupervised classification is significantly
greater than within-class variance for signatures derived from supervised classification.
3 RESULTS AND DISCUSSION
Between five and nine land cover types were encountered in a landscape. Supervised classification accuracy ranged
from 70-92%, with a mean of 84%. Unsupervised classification accuracy was uniformly worse, ranging from 31-75%
with a mean of 51% (Nagendra and Gadgil, 1999 b).
Table 1 demonstrates that for all thirteen landscapes, patch metrics (patch size, shape and nearest neighbor distance)
estimated using the unsupervised classification were significantly greater than those estimated from the supervised
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 957