ently 
stand 
jority 
the 
ative 
y on 
tree 
arcel- 
nplex 
ixed, 
nand 
tially 
| is 
nage 
5 to 
owns 
jf 6m 
ated 
SIDO 
to be 
tree 
lernal 
d a 
ta or 
matic 
ment 
single 
Jown. 
olive 
i-byte 
other 
ted in 
10x10 
- an 
rms. 
on is 
, with 
from 
S. In 
ıltural 
ilable 
ortho- 
r the 
2ece). 
ing to 
model 
ion of 
First, the data source most commonly used (at 
present) is scanned aerial photography: 
problems of view angle, variable Sun-target- 
sensor geometry (time of day, flight-line 
orientation, view azimuth), imaging season 
(phenological considerations) and image 
radiometric characteristics (different films, 
spatial resolutions, analogue development 
processes, scanning parameters) all play an 
important part in ensuring that few stable 
radiometric parameters can be defined for 
automatic recognition. Indeed, reflectance 
properties of crowns themselves can be 
expected to be extremely variable. 
Second, the trees themselves can have a 
range of crown shapes, sizes and densities; 
transmittance properties of the crown can be 
high, and understorey can strongly influence 
brightness values. Therefore, the varied 
agronomic management conditions of olive 
tree stands (ploughed, grassed or cultivated 
understorey, pruning practices, soil mineral 
content, drainage conditions, etc.) cause 
further confusion in the definition of reliable 
parameters. The development of a geometric- 
optical model (cf. Larsen, 1997, applied to 
spruce stands) was rejected as too complex at 
this stage. 
During this project, however, it was quickly 
identified that in relative terms the tree crown 
can be usually modelled according to a number 
of basic principles: 
a) the crown (plus shadow) is locally darker 
than its surrounding background (Figure 1); 
b) crown diameters range typically between 3 
and 12m; 
c) the shape of the crown (aspect ratio, 
roundness) is typically regular (but may be 
distorted by shadow); 
d) crown densities are usually high on the 
image data used. 
Two further characteristics were investigated 
but discarded after detailed evaluation: 
e in many areas, tree crowns present a "hole" 
of greater brightness, due to pruning 
techniques; however, this funnel-shape is 
not unique to olive tree cultivation and is 
not visible for smaller trees on 1m pixel 
imagery; 
e attempts to use regular planting spacing in 
the object-identification ^ analysis are 
thwarted by the many irregular stands in 
the Mediterranean basin. 
Briefly, the algorithm works by (Figure 2): 
  
4 
Frequency 
  
  
9 -———e 255 
Thresholding range 
Figure 1: Simplified image histogram for an 
olive tree stand 
e Identifying the range of Digital Number (DN) 
values in the target image for tree crowns 
e application of a binary threshold to the 
target image 
e region-growing of linked candidate pixels 
e testing - against a series of shape and size 
criteria - whether these regions are likely to 
be trees 
e Storing a list of trees in the (sub)image 
e cycling through the DN value range, 
e returning this information back to the GIS 
environment 
A single advantage in the operational use of the 
algorithm is that a count for each stand 
generally constrains some of the variable 
conditions. 
Digital Pancro Aerial Photo 
se € —— : 0 Image Pixel 
| Labelling 
  
Blob Map 
Classification 
7 : + 
rr ad] 
+ 
+ 
Olive Trees List 
Figure 2: Olicount program flow 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 359