' Remote THEMATIC CLASSIFICATION OF LANDSAT TM IMAGERY BY A NEURO-FUZZY METHOD
al Image Árpád Barsi
Department of Photogrammetry
The use Technical University of Budapest
remotely H-1111 Budapest, Müegyetem rkp.3
!R :
molt Commission VII, Working Group 4
ERN KEYWORDS: Thematic Classification, Landsat TM, Neural Networks, Fuzzy Logic
:., 1988.
| T
Lorton ABSTRAC
' Remote
The discovery of artificial neural networks in 1958 and the first paper about the fuzzy logic in 1965 count as milestones
b. of new computation research. These tools of artificial intelligence can be used for problems which were till nowadays
ification not or just partly solvable. The main task of digital image processing and computerized image interpretation can as well
rmation be renewed using these achievements. Neural networks and fuzzy logic can be combined luckily, in the paper I'll show |
incering it in a land cover mapping example in Budapest. |
er, JP; |
of forth- 1. INTRODUCTION 2. TOOLS OF ARTIFICIAL INTELLIGENCE |
e sensor
tric En-
).
The researchers have used numerous well-proved and
new methods for processing the satellite images sent to
the Earth's surface. Maybe the mostly used imaging
satellite is the Landsat. The relatively high geometric
resolution (pixel size 30 m) of TM is coupled with
adequate spectral one (7 bands). The correct processing
of information can produce good quality thematic maps.
To fulfill the quality conditions also accuracy analysis of
the used method is indispensable.
The paper shows an experiment with Landsat TM image
covering the capital of Hungary, Budapest. The image
was made in august 1989. Bands 1, 2, 3, 4, 5 and 7 were
used, the thermal infrared channel wasn't important for
the experiment. The size of the image is 301 by 460
pixels thus take nearly 125 km’. The six bands were
preprocessed (systematic error corrections etc.).
For about 1.8 % of the area is ground truth information
available. I used 2/3 of this amount for training, 1/3 for
testing purposes. In the classification the following
categories were defined:
Coniferous forest
Deciduous forest
Parkland
Brush land
Meadow
Plow land
Fallow lad
Urban area
City area
Water.
The theoretical development was done in a mathematical
modeling environment MATLAB. then it was
implemented in a raster GIS package (Intergraph MGGA)
and in an image processing software (ERDAS Imagine).
2.1. Neural networks
Feed-forward neural networks are a sort of function
approximators. These networks are built up of several
layers of - mostly — sigmoid-based neurons. Such a
neuron calculates its output by transferring the sum of its
weighted inputs and the bias through the logistic sigmoid
function. The answer of a 3-layer network is the
following with the notation of vector algebra:
o - F(W, -F(W, -F(W, -p+b,)+b,)+b,)
(1)
where F is the logistic sigmoid transfer function, W,;, W,
and W; are the weight matrices and b;, b, and b; the bias
vectors of the neuron layers. The criteria to be able to
calculate the output is to have correctly defined weight
and bias values. These parameters are determined in the
training phase. There's a need of adequate input-output
data pairs to train a network what are collected from the
training areas.
The usual Rumelhart-McClelland learning method
seemed to be slowly for my experiment so I've chosen
the Levenberg-Marquard algorithm known from the
mathematical optimization. The following equation is the
rule of learning:
AW = (JJ + ul)" Je
2)
where AW is the set of network parameters (change of
the weights and biases), ju is a scalar, I an identity matrix,
e the error vector and J the Jacobian matrix of every
errors for all weights.
Neural networks are suitable tools for thematic
classification where the pixels of an image must be
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 323
