. Istanbul 2004
|| efficiency of
the application
ed. The spatial
essary point in
nd to calculate
f reflection a
| space by the
to the point of
‘oaches to the
et, Hilbert etc.
is based upon
F the proposed
t types of the
rat the density
s more robust
e is performed
effective is à
sity estimation
ithm should be
data dimension
thm is applied,
thm is applied.
ies of couple
5, allows the
imes compared
n the additional
used (Haralick
id method for
al approach to
exture features
x of a defined
en used, which
ing the rate of
1g box. Further
(Haralick R.M.
acteristics have
vness, kurtosis,
this hopeful
difficulties of
e.
e authors used
In the first step
(S. image with
culated. In the
ve features are
ng to algorithm
n criterion of
/-distance). In
—
Un
Nr
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
In case the data in the processed pairs of samples fit the normal
distribution, the following expression is applied:
JM, =2(1-e e):
where
515]; [&, «z,)2]
i 7 ! I *,
rad
B18 | (2-112 a
| Tahki Fit]
at that 4, Hj and X >. — parameters of normal
distribution (means vectors and covariance matrices) of i and J
samples.
The selection of more informative texture features is carried out
in the following way. For all obtained textures band the average
JM-distance is calculated. Then these bands are ranged by the
average JM-distance from the best to the worst separability and
the samples agreements with the normal distribution are defined
(to apply parametric density estimation if possible). After that,
in the band with the best separability, the worst separating
(target) pair of samples is defined. Next, according to the
ranged order, another band to this current band is added. First
the second band is added, then the third etc. The separability of
each two bands is calculated but only for target pairs samples. It
significantly increases the computational performance of the
best features selection algorithm. Two features with the best
separability are taken as the intermediate complete combination
and for them again the worst separating (target) pair is defined.
Then the addition of band from the ranged bands to the
complete bands combination is again performed and the
separability of target pair samples is calculated. This time the
best combination of three features is defined. This procedure is
repeated up to the moment the five best bands to be selected
and we obtained final complete combination of features.
Iterative increasing of feature set and calculation of separability
only for target pair of samples is needed for the maximum
increase of computational performance of the procedure that is
very important for nonparametric computation of JM-distance
according to (5).
ANN: The way of forming feature vector for ANN
classification so called context-spectral is differed from the way
of forming feature vector one for statistical classification by the
significant simplicity.
Each component of the feature vector contains the focal and its
neighbor elements from all bands of RS image. Such way of
forming feature space allows to consider the interband and
pixels correlation (texture information) without special
calculation of texture features and feature space optimization.
3.3 Enhanced forecast maps designing by TS analysis
algorithm
The procedure of obtaining forecast maps with use of
interpreted RS images described in (A.V. Zamyatin & N.G.
Markov, 2004), therefore only the main stages are to be
discussed:
— determination of CA optimal neighborhood for each
classes;
— constructing transition matrix with use of first-order
and high-order Markov chains (if we have more then
two TS RS images);
— making primary forecast map with the use of obtained
transition matrix;
— processing of primary forecast map with use of CA
with optimal neighborhood size.
One of the key moments is the determination of CA optimal
neighborhood size for each class. It is determined by so called
the enrichment factor (Verburg P.H. et al, 2003), which is
defined by the occurrence of a land use type in the
neighborhood of a location, relative to the occurrence of this
land use type in the study area as a whole:
UH di Injj-
PB uu =
i.k.d N, IN
Fi characterizes the enrichment of neighborhood d of
location i with land use type K. The shape of the neighbourhood
and the distance of the neighbourhood from the central grid-cell
i is identified by d (for instance d = 1 means grid-cell 3x3). 1,4;
is the number of cells of land use type Æ in the neighbourhood d
of cell i, nj; the total number of cells in the neighbourhood
while N; is the number of cells with land use type À in the
whole raster and N all cells in the raster. The algorithm of
enrichment factor calculation is repeated for different
neighbourhoods located at different distances (in this case d
1,2,...,10) from the grid cell to study the influence of distance
on the relation between land use types. The average
neighbourhood characteristic for a particular land use type l
( F i44) is calculated by taking the average enrichment factors
for all grid cells belonging to a certain land use type 1.
following:
= 1 i
FE; d = F J ?
k.d N, i.k.d
iel
where L — the set of all locations with land use type / and N, the
total number of grid-cells belonging to this set. The grid-cell of
size d for each class type is fixed in case of maximum of the
average enrichment factor. These values are to be used for
every class in CA. In most land cover and land use change
model first-order Markov chains and only two classified images
are used.
4. RESULTS AND DISCUSSION
To investigate the efficiency of proposed algorithms the set of
experiment with model and real RS images are carried out.
4.1 Aerospace image models used
In the conducting research model RS images of two types are
applied. The multispectral images of the first type are images
with implicit texture of classes and arbitrary distribution in
classes (Figure 1). The multispectral images of the second type
