4 DATA ACQUISITION 
5 PREPROCESSING 
In order to ensure optimum effectiveness of an air 
borne campaign great attention is paid to internal 
and external calibration of the radar and choice of 
flight dates and -geometry. 
4.1 Calibration 
Early in 1985 the Dutch digital SLAR was fitted with 
an internal calibration mechanism. In short, the 
procedure may be visualized as follows. A small part 
of every transmitted pulse is delayed and weakened, 
then positioned at the beginning of each received 
backscatter signal. From the amplitude of the 
recorded calibration pulse the amplitude of the ori 
ginal, transmitted pulse is calculated. Hence each 
received and recorded radar return signal may be 
corrected for variations in the corresponding trans 
mitted pulse. 
As an additional measure of validating the calibra 
tion external corner reflectors are placed in the 
target area. From the raw received signals the 
return of the corner reflectors may be calculated 
and compared with the theoretical reflection from a 
well-defined corner reflector. 
It is quite necessary for further processing and 
classification to start out with a calibrated system 
as the transmitted power of the X-band SLAR is ex 
pected to fluctuate with elapsed time. 
4.2 Optimization of flight-date and -geometry 
The choice of the optimum flight days and the deter 
mination of the flight geometry is essential in 
planning an airborne campaign. The main purpose is 
to identify dates and angles whereby the possibility 
of discriminating various crops is the greatest. 
When identifying the ROVE ground truth data base it 
was found in 1983 and 1984 that an airborne SLAR 
campaign was to be conducted between May and August, 
as potatoes do not appear above the soil until late 
May and may be harvested as early as late July in 
The Netherlands. Furthermore the data base and crop 
growth parameters showed that winter wheat and 
potatoes tend to exhibit similar backscatter charac 
teristics in all months except May and that potatoes 
plus winter wheat can be discriminated from most 
other crops in July (Figure 2, 3). Thus, in 1983 and 
1984, it was decided to conduct an airborne campaign 
with the following objectives: 
- distinguish potatoes from winter wheat in May and 
- distinguish potatoes from other crops in June/ 
July. 
The above objectives could only be achieved with 
specific viewing angles. Thus an aircraft altitude 
of 500 meters above ground to obtain 5°-14° and 
another run at 2700 meters above ground to enable 
25°-45° grazing angles were deemed necessary. 
In 1985 the ROVE group "Crop Identification" chose 
to adhere more to expected satellite programmes, 
where steeper angles for satellite simulation are a 
prerequisite. 
Again the ROVE data base was consulted and this 
resulted in a flight altitude of app. 4000 meters in 
order to obtain grazing angles of 40°-70° for one 
track and 30°-48° (grazing) for a laterally trans 
posed track at the same altitude (Fig. 4). The 
corresponding appropriate flight dates of early June 
and early July 1985 gave good confidence to be able 
to adequately discriminate between the various crop 
types. 
In order to enable steep grazing angles the SLAR 
antenna had to be rotated by almost' 20° in order to 
make use of the linear part of the antenna gain 
function [Ref. 6, 7]. 
To overlay and register data from multitemporal air 
borne sensors may be very difficult due to random 
variations for attitude, velocity and position of 
the aircraft during the execution of the flight 
track. These deviations usually distort the resul 
ting data set sufficiently to disable further multi 
temporal processing. Also, several effects influence 
the signal received by the radar antenna; e.g. the 
antenna gain is a function of the depression 
angle, the received signal strength is dependent 
on slant range distance. Hence, the National Aero 
space Laboratory NLR in close collaboration with 
the Physics and Electronics Laboratory TNO, the 
Delft University of Technology and the Survey 
Dept, of Rijkswaterstaat, undertook to implement a 
preprocessing procedure (PARES) on the central 
computer of NLR. PARES enables raw acquired data 
to be corrected for variations in aircraft motion 
and -attitude through simultaneously recorded INS 
parameters in the NLR laboratory aircraft (INS = 
Inertial Navigation System). For every acquired 
line its exact position in a terrestrial 
coordinate system is calculated, thus enabling 
later multitemporal registration of data via an 
earth-bound coordinate system. The PARES-procedure 
also accounts for radiometric system corrections. 
The data preprocessed through PARES may now be 
multitemporally processed without further complica 
tions due to airborne data acquisition [Ref. 2, 3]. 
6 PROCESSING AND CLASSIFICATION 
The inherent speckle of a radar image is a distur 
bing factor in further processing and classifi 
cation. Also, most advanced classification algo 
rithms and user-defined information extraction pro 
grammes do not require information on a pixel-by- 
pixel basis but rather on a field-by-field basis. 
Many microwave remote sensing groups have devised 
ways to minimize speckle through field averaging. 
Various filtering algorithms have been tried but The 
Netherlands ROVE-group has chosen for the "split- 
and-merge" algorithm based on the Pavlidis-concept 
which has been implemented on the image processing 
system RESEDA at the National Aerospace Laboratory 
NLR [Ref. 5]. 
The algorithm has been designed to process (in 
parallel) up to 7 multitemporal images with a number 
of user-identified parameters like minimum field 
size and variance to obtain field averages where 
original pixel-values have been replaced by a field- 
dependent mean value without loss of information. 
Care must be taken to exclude fields which contain 
only a few pixels as they are likely to contaminate 
the segmentation result. Other small objects like 
power-line masts, farms, etc. will also corrupt the 
resulting data set [Ref. 4, 5]. 
After this extensive (pre-)processing the final 
classification procedure can be comperatively 
simple. An interactive parallelepiped classifier may 
be applied to the data set. The classifier allows 
for interactive classification from data space into 
feature space and vice versa. It will be obvious 
that this procedure is usually very time-consuming 
when applied to standard imagery. The preceeding 
segmentation algorithm has, however, already greatly 
reduced the dynamic range of possible pixel values, 
so now near real-time classification is possible. 
Great care must be taken, to prevent the angular 
dependency of backscatter of agricultural crops to 
be introduced into the classification result. One 
way to eliminate this effect is to process suffi 
ciently narrow strips where the angular dependence 
may be neglected.