Combined Uses of Supervised Classification and Normalized Difference Vegetation Index Techniques to Monitor Land Degradation in the Saloum Saline Estuary System - The Land/Ocean Interactions in the Coastal Zone of West and Central Africa - page 55

Table 1 Change matrices generated (%) through overlay analysis between the four single-date classification results, with the classification

method

Water High mangrove Low mangrove ''tan'' Savanna/rain-fed agriculture Forest

1992

1984 Water 97 0 2 1 0 0

High mangrove 2 42 55 0 0 0

Low mangrove 5 4 83 6 1 1

''tan'' 1 0 7 83 9 0

Savanna/rain-fed agriculture 0 0 0 1 98 1

Forest 0 0 1 0 45 54

1999

1992 Water 96 0 2 2 0 0

High mangrove 5 21 72 1 1 1

Low mangrove 28 3 61 7 1 0

''tan'' 11 0 5 78 7 0

Savanna/rain-fed agriculture 0 0 0 2 96 2

Forest 0 0 2 1 26 70

2010

1999 Water 92 1 5 2 0 0

High mangrove 10 44 43 1 1 1

Low mangrove 12 14 66 4 3 1

''tan'' 18 0 7 63 12 1

Savanna/rain-fed agriculture 0 0 0 1 95 3

Forest 0 0 2 0 42 56

The bold represent the no change percentage.

k and the number of bands k, the
label to be assigned to x is given by the class that maximizes
the LDC discrimination function (Kuncheva 2004 ; Hastie
et al. 2009 ). The LDC discrimination function is given,
according to Kuncheva ( 2004 ), by Eq. ( 4 ):

behaves according to the normal statistic distribution. Thus,
in this respect, it resembles the MLC. The difference is that
the LDC is based on an additional assumption which is the
homoskedasticity hypothesis, in which the classifier assumes
that each class has equal variance, and thus, all covariance
matrices are equal for every land-cover class of the nomen-
clature. Although the homoskedasticity hypothesis tends to
be unrealistic, the literature has shown that this classifier
behaves in a robust way even when there are deviations from
the hypothesis of normality and homoskedasticity (Kuncheva
2004 ). The LDC has several advantages, in that it requires less
training samples than the MLC and also it is easy to fine-tune
and robust to noisy data (Hastie et al. 2009 ). In this sense, the
LDC is a preferable classification algorithm for land-cover
mapping (Carrão et al. 2008 ), especially when the image
analyst does not have a reliable reference database to collect
representative training samples. For the classification, a set of
training sites and ground truth data were required. A sample
set of 50 training sites was established. They characterize the
six typical land-cover classes occurring in the study area. The
sample plots were digitized on screen, and then, a supervised
LDC was applied using a stack of the six (without band 6)
original bands of the Landsat image and the remote sensing
technique (NDVI) to generate a land-cover map.

From the pixel x 2 R

h i ðÞ ¼ 1

2 l i R 1 l i þ l i R 1 x

ð 4 Þ

where R is the common variance-covariance matrix, esti-
mated by the weighted average of the separately estimated
class variance-covariance matrix, i.e.,

R ¼ X

k

n i
n

R i

ð 5 Þ

i¼1

where n i is the number of training units assigned to the ith
class of X and n is the total number of training units.
The Landsat data of 1984, 1992, 1999, and 2010 are
classified into six spectral classes using the LDC. Savanna
and rain-fed agriculture are merged in the same class. This
classification algorithm is a supervised parametric classifier,
i.e., it requires a training sample in order to classify the pixel
from a given image and it assumes that each land-cover class