gruss 
The average inertia image was quantized into five 
empirically determined levels which represent the ee i 
following crops: he d 
Texture Level 1 (Black): Wheat * Barley speci 
Texture Level 2 (Dark Grey): Grass3» Corn at ai 
Texture Level 3 (Medium Grey): Corn?» Sugar Beets was al 
Texture Level 4 (Light Grey): CornssSugar Beet» using 
Trees solut 
Texture Level 5 (White): Sugar Beetes Trees duce ; 
(Also Facory and explo; 
Urban Area) data | 
comme! 
minat 
Figure 7. Filtered Average The image resulting from this level slicing is 
Inertia Image For Data shown in Figure 8. 
Set. 09 Owl MS J : The at 
| Team v 
matior 
Because the textural pattern people 
of wheat and barley are quite Dr.Que 
alike at this time of-the year Golliv 
June 10), it is not surprising prepar 
that we are unable to separate Center 
them automatically. Similar res 
marks hold for corn and sugar 
beets, because both are in an 
early development stage. A de- | 1. Aus 
tailed analysis of the crop F 
responses, their overlaps and 2 
their differences was beyond 2. ERI 
the scope of this investi- Figure 8. Classification Map D 
gation. Such an analysis should 
then take into account factors such as crop development stages, percentage of 3. Jou 
vegetation cover and variations within the fields, soil moisture content, row 4. Ste 
directions, lodging, weeds etc. S 
Similar results were obtained for the X,, data sets from the other two flight 5. Lou 
directions (data sets 710 and 711). However, the values used for the five level C 
slice intervals were slightly different in each case and the areas of heavy M 
shadow due to the side view of the tree line were included in the lowest inter- 
val. In all three data sets, agricultural fields down to as small as 10 meters 6. Ste 
in narrowest dimension could still be clearly discriminated provided their P 
texture was significantly different from neighboring fields. The use of the M 
local tone image instead of the inertia image for discrimination didn't result 7. Pre 
in better crop separability. The reason is found in the statistics of the in- d 
dividual fields which obviously have narrower standard deviations but their 
means are also closer together precluding reliable separability. 8. Nüe 
C. 
( 
CONCLUSIONS AND RECOMMENDATIONS > H s 
. ar 
The work reported here represents a first attempt at automatic digital crop T 
classification of SAR data using only texture. In spite of the tremendous 
computational burden - generation of a single 1024 x 1024 inertia image from 
the texture image required 150 CPU minutes on a VAX 11/780 - experience gained 
during the investigation suggests that considerable information is contained in 
238