' to perform 
he proposed 
lassification 
nd separate 
“noisiness” 
f traditional 
] to perform 
ic maps are 
Calculation 
her but also 
ne series RS 
nethods and 
retation and 
ages, which 
ly is solving 
such as ER 
S Imagine 
are is based 
assumption 
| maximum 
Is that can 
es the actual 
t use in full 
classes on 
nethods and 
extent only. 
1 results it 1s 
:heme of RS 
ems such as 
ence a lot of 
odelled by 
1989). The 
rkov chains. 
ds and have 
Baker W. L, 
| interaction 
ita (CA) are 
parameter of 
central grid- 
id use/cover 
all types of 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
classes. But such approach does not take into account features 
of spatial interaction of classes on thematic map (Verburg P.H. 
et al., 2003). 
In the paper the approach, with combined application of 
original algorithm of RS images interpretation, geoinformation 
system (GIS) and also algorithm of enhanced designing of 
forecast maps is developing. The features of the proposed 
algorithms are considered and investigation results of these 
algorithms using modeled aerospace images and real TS RS 
images are discussed. 
2. PROCESS DESCRIPTION 
Processing and interpretation of TS RS images using the 
proposed methods and approaches are based upon two 
sequential stages. In the first stage TS RS images are classified 
by the original advanced interpretation algorithm with separate 
processing of spectral and spatial features of RS images. It 
allows to obtain the less noisiness final thematic map and to 
increase its classification accuracy compared to traditional 
interpretation, RS and mapping software. The first stage of RS 
images interpretation is a key stage, because on the basis of 
interpretation results the forecast maps will be designed. In that 
case more accurate and adequate of interpretation results will 
lead to more accurate and adequate forecast results. 
In the second stage the obtained raster thematic maps are 
processed by the original TS analysis algorithm, which is based 
upon Markov chains and CA processing algorithm, includes the 
optimal neighborhood size determination. This TS analysis 
algorithm allows to consider not only the probability of 
transition one class to the another, but the spatial interclass 
correlation. 
Let's consider the algorithm applied in the stages of the 
advanced interpretation and the algorithm of the enhanced TS 
analysis in details. 
2.1 First stage — advanced interpretation 
Advanced interpretation in the framework of the proposed 
approaches is based upon Bayes decision rule of empirical risk 
minimization: 
p(o,)p(X |œ,) ‘ (1) 
pio xy M : 
S p(œ,)p(X | œ, ) 
where p(o,) — prior probability of class i, M — the number of 
classes, pO 1o.) - conditional probability density of class i. 
At the same time according to (1) interpretation in two steps is 
performed. In the first step the posterior probability maps for 
each classes are designed, at that the feature space is 
considering spatial characteristics. In the second stage 
according to (1) designed maps are to be used as prior 
probabilities of classes and the feature space consist only of 
Spectral features. 
Besides, it should be noted that in the first step cither statistical 
or artificial neural network (ANN) classification could be used. 
493 
Statistical classification is based upon combined application of 
parametric and nonparametric algorithm of density estimation 
depends on the agreement of the data with the normal 
distribution according to the chi-square criterion, and also 
statistical classification includes standard parallelepiped 
classification algorithm. The simple and fast parallelepiped 
classification algorithm is applied in case the sample data range 
is not intersected by data range of any other samples. 
ANN classification is implemented together with the approach 
to the storage and the search of the ANNs in a database. The 
general purpose of the approach is to make the process of 
ANNs topology and parameters definition easier and also to 
make the learning process of ANN faster. The search might be 
done with test of sign-rank correlation between the investigated 
data sample and the ANN train data sample stored in the 
database. The possibility of ANN search makes the ANN 
learning more predictable and robust. That is why in case of 
successful search of the appropriate stored ANN for 
investigated data sample the designing of prior probability 
maps is performed by ANN classification. 
2.2 Second stage — forecast maps designing 
TS thematic maps, designed in the first stage, and other 
additional data, obtained by including GIS are carried out by 
enhanced TS analysis algorithm. The algorithm includes 
iterative performing of three operations. 
The first operation of the considering process is the analysis of 
the neighborhood characteristics of the raster interpreted 
thematic maps. At that the optimal scale of classes 
representation is being determined. The optimal scale needs 
further for the effective CA application. 
The second operation is the constructing of transition matrix 
(TM). The operation includes the analysis of two and more 
thematic maps using first-order or high-order Markov chains 
respectively. The use of suitability maps, which show the 
probability of change one class to another, allows to range all 
image elements from high disposed to change till low disposed 
to change. 
The final operation is processing of primary forecast map by 
CA with the optimal neighborhood size, defined in the first 
tep. 
N 
The final result of the enhanced TS analysis algorithm is a 
forecast map for the further time step. To perform longer 
forecast it is needed to pass the obtained forecast map to the 
input of the algorithm as the TS input thematic map. 
3. FEATURES OF APPLIED ALGORITHMS 
One of the original algorithms applied for interpretation of RS 
images is the density estimation. nonparametric algorithm, 
which is based upon Rosenblatt-Parzen (RP) algorithm and k- 
nearest neighbour (k-NN) algorithm. The original algorithm 
provides a computational cost over dozens times compared to 
existing nonparametric algorithms. Moreover the ways of 
forming feature space for both statistical and neural network 
approaches are original. Also the approach to processing of 
thematic maps by CA with optimal neighbourhood size defined 
by the enrichment factor is original. It is proposed to consider 
these features more detailed.