DESIGNING FORECAST THEMATIC MAPS USING TIME SERIES REMOTELY
SENSED IMAGES
A.V. Zamyatin, N.G. Markov
Tomsk Polytechnic University
84, Sovetskaya street, Tomsk, 634034, Russia,
e-mail: ZamyatinA V@prim.ce.cctpu.edu.ru
KEY WORDS: Remote Sensing, Land Use Mapping, Prediction, GIS, Neural Networks
ABSTRACT:
In the paper a number of original algorithms implemented as the program system are considered. The algorithms allow to perform
the automated forecasting of land use/cover change using time series remotely sensed (RS) images. In the framework of the proposed
approach the forecasting performs in two stages. In the first stage time series RS images are classified by the original classification
algorithm, which is based upon both statistical nonparametric and artificial neural network (ANN) classification and separate
processing of spectral and spatial features of RS images. The complex classification allows significantly decreasing the “noisiness”
of the final thematic maps and to increase the classification accuracy in comparison with classification methods of traditional
processing,
RS and mapping software. To make the interpretation of RS images more flexible and effective it is proposed to perform
classification using ANN with original way of forming feature space. In the second stage, the designed raster thematic maps are
processed by the enhanced algorithm of time series analysis, which is based upon Markov chains for transition matrix calculation
and cellular automata application allowing to take into account not only probability of transition from one class to another but also
spatial interclass correlation.
The results of effectiveness investigation of the proposed algorithms, obtained with use of model RS images and real time series RS
images for Uymon steppe area (Altay Region, Russia) obtained from satellite RESURS-O1 are discussed.
1. INTRODUCTION
Every year high accuracy and operative forecasting necessity of
land cover change has been growing. At the same time the
operative and accurate forecasting is not possible without joint
pplication of the comprehensive technologies for time series
(TS) RS data interpretation, geoinformation systems (GIS)
performing complex spatial analysis of interpreted data, and
comprehensive methods of spatial forecasting.
lot of researchers are solving the task of designing
forecast maps with use of simple classification methods and
highly specialized models of land use/cover change.
Application of simple methods and algorithms of classification
lcads to inadequate thematic maps. Saying about highly
pecialized models of land use/cover models with their
idvantages and disadvantages, we may say that these models
ften require a large amount of additional information (digital
vation model, dependability of different land types change,
of migration etc.) that makes their application field
nited. The all mentioned shortcomings restricts the
| of existing program facilities, that does not allow to
At present a
orecast maps with the appropriate accuracy using TS RS
tioned above declare about the imperfections for using
approaches and algorithms for RS images
ntation in the task of forecasting land use/cover
Also it says about urgent need for new methods and
hms allowing to solve all mentioned problems and
trictions more effectively
492
Let's consider the possibilities and restrictions of methods and
algorithms applied for tasks of RS images interpretation and
designing forecast maps more detailed.
Now the task of interpretation of TS aerospace images, which
in their turn can be used for forecasting, traditionally is solving
with image processing, RS and mapping software such as ER
Mapper (Earth Resource Mapping), ERDAS Imagine
(ERDAS), Idrisi 32 (Clark University). This software is based
upon either parametric statistical methods using assumption
about normal distribution of features (traditional maximum
likelihood algorithm) or nonparametric methods that can
produce acceptable results in few cases only. Besides the actual
image processing RS, and mapping software do not use in full
measure spatial (texture) information about classes on
aerospace images. Application of such simple methods and
approaches allow to use RS information to some extent only.
Therefore for obtaining more accurate interpretation results it is
required new and significantly more complicated scheme of RS
images classification.
While forecasting the behaviour of complex systems such as
mapping nature territorial complexes to which influence a lot of
stochastic processes, basically have been modelled by
stochastic forecasting methods (Baker W. L, 1989). The
widespread among them are methods using Markov chains.
Markov chains have been used in a variety of fields and have
modelled changes on a variety of spatial scales (Baker W. L,
1989). In order to Markov model considers spatial interaction
between classes on thematic map cellular automata (CA) are
often applied (e.g. Park, S. and Wagner,
CA is the distance of the neighbourhood from the central grid-
cell. In majority cases of CA application for land use/cover
change modelling this parameter is taken equal for all types of
1997). A parameter of
Inter
————
class
of sp
et al.
[n tl
origi
syste
forec
algol
algo
imag
Proc
prop
sequ
by tl
proc
allo
incre
inter
imag
inter
case
lead
In t
proc
upot
optii
algo
tran:
corr
Let
adve
anal
2:1
Adv
app!
min
p
whe
clas
At 1
per!
eaci
con
acci
prol
Spe
Bes
or a