study area
difficult.
r research
f a priori
likelihood
) reported,
lities of
che maximum
ves the
ndcovers in
a ratios of
lowing the
The area
irea can be
1 es of
a, applying
n of mixed
landcover
a maximum
the area
cea.
The first
gorithm of
applied to
as a priori
The second
ilities of
1l by this
that the
ilities of
ssification
landcover
rea
landcovers
je equal to
the study
area. Equation 1 shows that the mean vector
of the study area can be expressed as the
summation of the weighted mean vectors of the
training data. The weights of the mean
vectors of the training data are the a priori
probabilities of the training categories.
Io[k] - X X[i.j.k]/*,
= V Y a[h,i,j] *It[h, k] /
ij h=1 15)
XD (Xa[h,i,j)/ 2 * It[h,K]
h=1 ij ij
Y, A[h] + I[h,K] oon. (eq.1)
=1
where,
H
X[ijk].- Y a[h,ij] * It[h,k]---(eq.2)
h=1
IIh.kl- HET... rt n (eq.3)
H
alh,i,j1=0, X albhi,jl=1
h=1
Alh]= X alb,ij1 /X
ij
A[h] 2 0, Y A[h]=1
(Mean vectors of landcovers in the study area
are very close to those of the training data.)
Note:
H: Number of landcovers
K: Number of bands of image data
h: Landcover ID No. (1 to H)
i,j: X,Y-coordinates of the image in the
study area
k: Band ID No. (1 to K)
Io[k]: Mean vectors of the study area
I[h,k]: Mean vectors of Landcover [h] in
the study area
It[h,k]: Mean vectors of landcover
training data [h]
A[h]: Area ratio of landcover[h] in
the study area(unit: IFOV)
a[h,i,j]: Area ratio of landcover[h] in
the IFOV of pixel[i,j] (unit: IFOV)
X[i,j,k]: Image data vector pixel[i,j]
The area ratios of landcovers in the study
area can be estimated from the image by
applying the algorithm of the decomposition
of mixed data. We can improve the landcover
classification accuracy by introducing the
area ratios of landcovers into the Bayes'
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
classifier as a priori probabilities of
landcovers in the study area.
Methods to decompose mixed data,
especially mixels, have been presented
[Inamura 1987, Ito 1987,Matsumoto 1991]. The
algorithm which was presented by Matsumoto
(1991) was applied to decompose mixed data
into landcover categories in this reseach.
This algorithm performs better than other
algorithms, because it takes account variances
of data distribution.
The method to decompose mixed data which
has been presented by Matsumoto [1991] is
shown in the Appendix A.
3.Results of the estimation of the area ratios
of landcovers
Innba district, an area near Tokyo, was
selected as the study area. The ground truth
was conducted on March 16, 1994 around this
area. The seven bands of TM image data(path
107 row 35,radiometrically calibrated,
resampled by cubic convolution) of Landsat-5
acquired on April 22, 1994 were analyzed.
Seven training categories(Paddy field,
Forest,Upland field, Waste land, Water body,
Residential area, Sports ground) were
selected. These training areas were selected
from the whole study area, and then their
signatures were calculated in this area.
Moreover, separability of those training data
was repeatedly examined. Seven small study
sites, where the area ratios of landcovers
were also measured using the ground truth
data, were selected from the study area. Each
small study site had two to five landcovers
(Table 1). After the mean vectors of each
small study site were computed, they were
decomposed into landcover categories by the
method of the maximum likelihood estimation.
Landcovers on the ground truth map were
digitized and those areas were computed on the
high resolution digital image data. On the
other hands, classified pixels to landcovers
in the study sites were counted, considering
pixels on the boundary of study sites. The
pixel on the boundary was assumed that that
pixel contained 50$ of the landcover category
to which that pixel was classified. Then, the
area of each landcover was computed by
multiplying the area of IFOV on the ground.
The estimation errors of the landcovers in
each small study site are shown in Table 2.
The area ratios of landcovers in each small
study site was estimated well.