In case of time domain matched filtering one would combine the RCMC and 
the azimuth compression; hence both processes together require 19.8 mil- 
lion complex operations. 
Comparing the computation effort between time domain processing and fast 
convolution techniques shows that the ratios are: 
for range compression: 43 to 1, 
for RCMC and azimuth compression: 76 to 1. 
2.3 How can fast SAR processing be achieved? 
As an ultimate goal for fast SAR processing one would consider through- 
put in real time. For ERS-1 this implies with a pulse repetition fre- 
quency of about 1680 Hz: 
using fast convolution techniques: 1680 * 364264 = 612 * 106 complex 
operations per second, and 
in time domain computation: 1680 * 23.5 mill. = 39.5 * 109 complex oper- 
ations per second. 
Supposing that the preparatory computations such as derivation of match- 
ed filters and range cell migration from geometry dependent parameters 
can be done sufficiently fast, the use of fast convolution techniques 
presents already a demanding throughput requirement. The time domain 
approach seems to be unrealistic, since it exceeds capabilities of a 
modern array processor still by a factor of 1000. 
There are two principle ways to make processing like the one for SAR 
fast: 1. Taking a supercomputer (this is convenient but rather expen- 
sive), 2. Making advantageous use of the pecularities of the required 
computations for a dedicated processor design. 
The second approach implies for SAR processing to make maximum use of 
parallelism in the computation. 
Fast convolution techniques have the disadvantage that they force the 
processing to be performed in blocks dictated by the FFT size. Moreover, 
parallelism implies that due to the range cell migration data have to be 
split into overlapping subsets. This leads to considerable data man- 
agement overhead which complicates a fast SAR processing architecture 
(ref. 5, 6 and 14). 
On the other hand, time domain convolution is a straightforward process 
which can be performed continuously. One can produce an output sample 
per input sample (after lead time of the filter length) if all the mul- 
tiplications are done in parallel and the summation is implemented as a 
cascade operation. 
Based on this concept a time domain convolution machine has been devel- 
oped by W. Dillen and G. Kluge (ref. 1). It makes use of semi-custom 
CMOS gate array technology and achieves a processing throughput of up to 
5 complex Megasamples per second (for comparison: the ERS-1 raw data in- 
put is about 9 complex Megasamples per second). A contribution to this 
conference is devoted to the subject. 
438