WE EEE EEE 
THEMATIC INFORMATION EXTRACTION IN A NEURAL NETWORK CLASSIFICATION OF MULTI-SENSOR DATA 
INCLUDING MICROWAVE PHASE INFORMATION. 
Gerrit Huurneman, Rüdiger Gens, Lucas Broekema 
International Institute for Aerospace Survey and Earth Sciences (ITC) 
P.O.Box 6, 7500 AA Enschede, The Netherlands 
Tel. +31-53-4874358, Fax. +31-53-4874335 
E-mail: HUURNEMAN@ITC.NL, GENS@ITC.NL 
Commission ll, Working Group 4 
Keywords : 
Abstract 
ERS-tandem mode, multi source data, coherence map, neural network, land use classification 
Microwave data (ERS-1 and ERS-2) and optical data (SPOT-XS) were used for the classification of an area with different 
land use classes. Classifications were executed for the optical data alone and for a combination of the three data sets. Two 
classifiers, one based on the maximum likelihood algorithm and the other on a neural network approach, were applied. From 
the ERS tandem mode SAR data a coherence map was created and included in the classifications in the form of an 
additional dimension in the feature space. The accuracy and reliability of the four classifications are presented and the 
results discussed. 
1. INTRODUCTION 
Optical and microwave data can provide complementary 
information about objects that cover the Earth surface. 
Optical data contains the information about the reflection of 
the solar energy in selected parts (bands) of the spectrum. 
The image elements (pixels) of optical sensors can be seen 
as vectors of which the components represent the 
reflection in the different bands. Image elements of 
microwave data consist of two components, the magnitude 
and the phase which are stored as complex numbers in two 
"layers". If optical and microwave data sets are correctly 
combined, the resulting product will convey more 
information and could prove to be more useful then either 
image alone. 
The information contained in the multi-sensor data can be 
extracted visually or by computer supported methods. In 
computer analysis, a suitable classifier is needed to handle 
the optical and microwave sensor data. Since the nature of 
the two data sources is different, which results in different 
frequency distributions of the data, it makes sense to use 
the neural network approach to analyze this data. Neural 
network classifiers are able to handle the multi-nature data 
from different sources. 
Phase information of microwave image data is mainly used 
in the field of interferometry. Interferometric processing of 
SAR data from space combines images from two passes of 
a sensor system or combines the data from two sensor 
systems in tandem mode. This process derives precise 
measurements of the differences in path length to the two 
sensor positions. The main output of interferometry of SAR 
data is topographic information related to terrain heights or 
the monitoring of positional changes of the Earth surface. A 
strong relation exists between the quality of these products 
and the correlation of the complex data sets. In those areas 
where high correlations exist an accurate Digital Terrain 
Model can be realized. On the other hand, height 
information cannot be extracted in areas with low 
correlation. So the quality of the products is characterized 
by the "interferometric correlation”, which is a measure of 
the variance of the interferometric phase estimate. The 
amount of correlation is a function of the system noise, the 
volume scattering, baseline configuration and temporal 
change. Consequently, the interferometric (de)correlation 
itself contains significant thematic information that can be 
useful for several other applications. 
In SAR images, the magnitude and the phase of each 
element are the coherent summations of the back scattering 
and phase of the individual scatterers inside a resolution 
cell. The variation in the overall phase and the overall 
magnitude of cells with equal cover type will appear as 
speckle. However two images taken from the same position 
at the same moment will be identical (neglecting system 
noise). If two images are taken from different positions 
and/or instants of time, variations in pixels representing the 
same surface cell will appear. 
If the structure or chemical composition of the ground-cover 
changes, then the amount of temporal decorrelation will 
vary. This variation can contain information about the type 
of ground-cover and/or the situation in which it exists. 
Influence on temporal changes, in case of space borne 
sensor-systems, can be minimized by a high temporal 
resolution of the system or by a combination of two 
"identical" systems in a tandem mode. The ERS-1 and 
ERS-2 systems satisfy this last requirement. 
Correlation between the data from a cell is expressed in 
terms of the summated phase and intensity of the resulting 
back scatter. A cell is considered to contain a set of 
individual back scatterers distributed over the cell. The 
amount of energy back scattered by the individual 
scatterers can be equal or can vary and the positional 
170 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996 
disti 
If th 
san 
Sm: 
grot 
ima 
In € 
havi 
sign 
Apa 
con: 
betv 
redu 
orbi 
of tc 
are 
heic 
line 
to re 
extr: 
to tl 
relat 
is Cc 
of tk 
In p 
basi 
diffe 
tem| 
addi 
imac 
1.1 | 
Neu 
usin 
con: 
orde 
one 
elen 
with 
outf 
cont 
the r 
men 
able 
nee 
con\ 
1,2 
For 
data 
com 
phas 
fluct 
Winc 
Acc 
factc 
UI +