131 
Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986 
Identifying agricultural crops in radar images 
P.Hoogeboom 
Physics and Electronics Laboratory TNO, The Hague, Netherlands 
ABSTRACT: In 1980 a large SLAR flight program was carried out over an agricultural area in The Netherlands. A 
classification study on this multitemporal dataset (Ref. 1) showed that high accuracies are obtained from a 
simultaneous classification of 3 flights. In this paper the results of a follow-on study will be discussed. The 
goal is to obtain the best possible classification result in the earliest possible stage of the growing season. 
Therefore the SLAR flights from April, May, June and July were analyzed and the hierarchic classifier is intro 
duced. Very satisfying results were obtained from a combination of 3 flights: 1 in May, 2 in July at different 
incidence angles. 
1 INTRODUCTION 
In this paper the results will be discussed from a 
follow-on study on previous classification experiments 
(Refs. 1, 3). This study forms a part of a broader 
national remote sensing research program for agricul 
ture and forestry, carried out by the ROVE-team (Radar 
Observation on VEgetation), a collaboration of several 
institutes (Ref. 1). 
The testsite on which the study is performed is 
situated in the Flevopolder, a reclaimed land area. 
Figure 1 shows a part of this polder, including the 
testarea. The latter contains 195 agricultural fields, 
of which 164 were suitable for this experiment (crop 
type known, reasonable dimensions). Frequently used 
crop types in this area are winterwheat, potatoes and 
sugarbeets (80% of total area). Onions and peas are 
also important crop types, but grown on smaller fields 
and therefore make up only 8.5% of the area. These 5 
most occuring crop types were used in designing the 
classifier. 
Figure 1. Map of the Flevopolder with the testarea 
indicated (near 'Biddinghuizen'). The map shows an 
area of 35 x 35 km. The testarea measures 3.7 x 6.2 
km. 
The area was imaged with an X-band SLAR system, 
using digital recording, on 5 different dates 
throughout the growing season. At each flight date 
recordings were made from 3 different altitudes, 
resulting in 3 incidence angle ranges, and from 2 
opposite sides of the testarea. This flight campaign 
resulted in a multitemporal and multiangular database 
of the area. A selection of these flights is shown in 
fig. 2. The development of the radar backscatter 
through time can be viewed from this selection. The 
sampling interval is appr. 1 month. For July two 
images are shown: one is flown at 660 m altitude, 
like the other images shown, which results in a 
grazing angle range from 7.5° to 16° (right to left 
in the images). The second July image is flown at 
1600 m, resulting in grazing angles between 18° and 
35°. 
For comparison a croptype map is shown in figure 3. 
This figure results from the radar images, after 
registration and field segmentation. The segmentation 
is done manual by drawing the field boundaries in the 
image on an image processing system. Only the 3 main 
croptypes could be indicated here, because of the 
limited separability of gray tones in a black and 
white image after reproduction. However 80% of the 
area is covered by these 3 croptypes. 
The advantage of field segmentation of the radar 
data is two fold: 
1. The influence of speckle on the classification 
result is reduced to practically zero. This also 
holds for small inhomogeneities within the fields. 
2. The amount of data is tremendously reduced, 
since we end up with one value per image for every 
field, thus 164 vàlues for one image. 
For the classification experiment the radar data 
of the 6 mentioned images were combined with the 
groundtruth into one datafile. For every field there 
are 7 features, i.e. the true field label (croptype) 
and 6 average backscatter coefficients. 
2 CLASSIFICATION EXPERIMENT 
The purpose of the experiment was to design a 
classifier on basis of the radar data of the 5 most 
important croptypes. The classifier should be able 
to distinguish between these croptypes as early as 
possible in the growing season. This is different from 
the previous experiment (Ref. 1), where we used the 
flights of June, July and August for classification. 
For operational applications an early result would be 
much more useful. Certainly, an improvement over the 
older experiment should be possible, considering the 
high contrast in the early season flights (fig. 2,