he
he
gh
ng
Oo
he
ue
st
th
he
S.
n,
to
ee
nd
ry
WO
re
e=
to
nd
of
to
of
of
s-
me
er
ty
es
i-
th
of
4.4.1. Pixel Extractions Algorithms
The algorithms for range and azimuth compression (pixel
extraction) of SAR raw data are essentially based on the di-
rect convolution in time or frequency domains or de-chirping
followed by spectral analysis. These algorithms have been pro-
ved to be computationally efficient in a number of SAR sy-
stems. Pixel extraction for ARTS-IP needs the additional fea-
ture of flexibility in. conjunction with computational effi-
ciency. Flexibility is required to accomodate a variable fre-
quency modulation rate of the input data and a variable geo-
metrical radiometric resolution of the various operational
modes. Computational efficiency is mandatory to process in
real time the large amount of raw data. Polyphase networks,
which have been developed in fields not related to radar, seem
a valuable solution to both problems. It "has already been
pointed out that the convolution of SAR data with a reference
chirp can be carried out in an extremely efficient way if the
whole signal bandwidth is divided into several contiguous band-
pass portions so as to decompose the high bandwidth convolu-
tion in several narrow bandwidth partial convolutions. The
convolution process is now the cascade of a multiplexing net-
work, a bank of partial narrow band convolution processors
and a de-multiplexing network to recombine the partial convo-
lution results. This technique has not been used because of
the burden posed by filter banks and the decimation/interpola-
tion (multiplexing and de-multiplexing) networks. It has been
shown the polyphase networks overcome this problem. This algo-
rithm has been successfully tested by processing SEASAT raw
data.
4.4.2 Image Processing
Image processing applied to SAR data is a young technique.
ARTS-IP is revising the state of the art in view of the diffe-
rent application and to establish basic tools.
The SAR sensor operation and processing result in an image
which is particularly difficult to interpret. This "is due to
the imaging projection utilised and the phenomenon of speckle.
In order to understand the scene sensed by the SAR several
processing steps have to be made on the image. They "can be
classified in low and high level image processings. The low
level image processing segments the image in a number of homo-
geneous regions, register the image with either earlier images
or to a cartographic reference, then generates attributes for
each region which can be used as primitives in the high level
processing.
Segmentation comprises three stages, namely: smoothing (to
reduce image noise or speckle), edge detection (to locate. ob-
ject boundaries), and some form of boundary tracing {to inter-
polate missing edge fragments).
Registration of a SAR image tO a standard cartographic
frame is required to locate the image and to integrate two Or
more processed images for multitemporal and multisensor stu-
353