Barsi, Arpad
THE IMPACT OF DATA COMPRESSION AND NEIGHBORHOOD INFORMATION
ON THE CLASSIFICATION ACCURACY OF ARTIFICIAL NEURAL NETWORKS
Árpád Barsi
Department of Photogrammetry and Geoinformatics
Budapest University of Technology and Economics
H-1111 Budapest, MDegyetem rkp.3. K.1.24.
barsi@eik.bme.hu
Commission VII, Working Group 5
KEY WORDS: Neural Networks, Thematic Classification, Principal Component Analysis
ABSTRACT
The artificial neural networks are nice tools in the thematic mapping. The classification procedure requires carefully prepared
training set. The research was aimed to show the effect of the generally applied principal component analysis and the coupling
Karhunen-Loeve transformation. These methods are data compression techniques developed for multispectral imagery. The next
* moment of the project was the handling of neighborhood information. It was expected that the mapping accuracy would be increased
considering this information. Two types of neighborhood were checked and they were also compared. The administration of
neighborhood leads to difficulties in the memory management, training methods and simulation algorithms. The combination of PCA
and neighborhood was found very helpful. The amount of the original data extended with neighborhood could be reduced by this
way, while few information rate is lost. Previous problems aren’t arisen. The resulting thematic map is very smooth, esthetic and has
high interpretation quality.
1. INTRODUCTION
Satellite imagery is a nice information source in thematic
mapping. For the second millennium new methods are
developed beside the “good old” traditional ones. Thematic
mapping is executed usually by human operators, who can be
supported intensively by efficient computer software. Today’s
best traditional algorithm — maximum likelihood — is basing on
the Bayes theory. The maximum likelihood method has several
implementations, faster and faster solutions are found. The
method supposes preliminary distribution information of the
participating pixels.
The artificial neural network doesn’t require such assumption.
In cases where the pixels’ normal distribution isn’t fulfilled
maximum likelihood method will produce more error. Neural
network can bring better result in this case. Of course neural
networks have disadvantage: they're black boxes, which
features must be tested intensively. In the paper I’ll present my
investigations with artificial neural networks. I'll concentrate on
the behavior of networks with normal inputs, followed by a
study when a kind of data compression and pixel neighborhood
are also taken into consideration. The outputs of the networks
are qualified by standard accuracy measures.
2. TOOLS AND METHODS
2.1. General description
The experiments of my paper need high mathematical and
computational resources. Therefore MathWorks Matlab was
chosen, which is an excellent mathematical software with
programming facilities. The connecting Neural Network
Toolbox was also applied, so I hadn't spend time with
implementing the standard training algorithms. For the image
data management the Image Processing Toolbox was very
useful.
The applied image was a subscene of a LANDSAT TM scene
covering the capital of Hungary, Budapest. The image was
captured in August 1989. The data set contains all the available
bands preceding radiometric correction. The subscene was
selected where different land cover categories are existing on
different elevation types. The image covers both urban (built-
up) and natural areas (Figure 1). The data amount to be
processed (see dimensions later) is expected high, therefore the
size of the subset was chosen for moderate (286 x 381 pixels).
Figure 1. The experimental area
The thematic classification of the current experiment was the
supervised classification. The resulting map contained the
following categories:
F1: vital, dense forests
F2: loose, partly unvital forests
M1: reach, healthy meadows
M2: thin meadows
U: urban, built-up areas
W: water (rivers and lakes).
+ + + + 9 9
140 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
ju mbit Rab i a suy 3
FF" AS
rm (79 (79 nM eM beo pee P heed 09M M TA M 68 AN 0M M o màn
Bee. MO mh m mM (mM
CO th ~~ 0 OO
