source 
hy with 
(81.3). 
were 
;ts, but 
chnical 
ers of 
ndomly 
mixed 
trained 
nethod. 
entire 
er-tree 
quality 
projects 
| as a 
: these 
retation 
t (April 
  
  
|e 
  
pt / Reject 
| 
  
LXJ 
4 / 86 
  
  
  
1998) consists of: 
e some 72,000 sample plots; i.e. randomly 
selected georeferenced areas of nominally 
100m diameter, located throughout the 
most important olive producing areas of 
Europe; 
e field validation results for more than 7,000 
of the sample plots; 
e a database containing the georeference of 
each tree and correction information derived 
from field data. 
In terms of olive tree information, and probably 
even in forestry terms, the dataset is virtually 
without precedent, and presents an ideal 
opportunity for the validation of the Olicount 
algorithm. An added value of the data set is its 
breadth of tree conditions, geographical 
contexts, photography dates and cultivation 
practice. 
4.2 Use of Olicount by JRC 
Although the Olicount algorithm was a source 
of inspiration for some of the Olistat tree 
counting, the code created by the JRC was not 
directly applied by any of the contractors 
executing the projects. 
Nevertheless, within the JRC, the authors have 
adapted the Olicount application to act as a 
photointerpretation quality assurance tool. An 
example of the interface developed for Olistat 
is given in Figure 11. 
While the primary objective of the checking 
procedure is to determine the quality of the 
photointerpretation, the use of Olicount in this 
context has been adapted to keep track of: 
e automatically identified trees 
e operator-added trees 
e operator-deleted trees 
These results have then been compared and 
tested statistically with those delivered by the 
Olistat projects contractors. 
4.3 External trials with Olicount 
In addition to the use of Olicount by the JRC, 
two MS administrations have made preliminary 
use of the software. These trials provide an 
important - albeit more subjective - testing of 
the algorithm by end-users. 
In Italy, the core algorithm was integrated into a 
customised GIS environment, and used in a 
trial with some 50 local offices to assist in the 
application process by farmers for olive oil 
subsidies. Early results have shown relatively 
good performance of the algorithm. 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
In France, a series of detailed tests carried out 
by SIDO (SIDO, 1998) using the prototype as 
developed by the JRC, have shown that the 
tool can improve the efficiency over manual 
tree counting on photographic imagery. 
4.4 Preliminary results 
Early results with Olicount show that it performs 
well as a checking tool, helping to reduce 
operator bias and allowing the technician to 
concentrate on the task of interpretation, and 
not counting. Different interpretation scenarios 
can be tried out, until an optimal solution is 
determined. 
These trials highlight the benefit of using an 
semi-automatic counting algorithm, particularly 
for large stands of trees. Nevertheless, the 
shortcomings of the imagery used, as well as 
the possibility of species confusion and other 
photointerpretation problems, must not be 
ignored. For example, coalesced crowns, 
heavy shadowing or the non-resolution of 
young plants will always introduce errors into a 
remote sensing-derived count (Howard, 1991). 
Full details of the Olistat dataset check will be 
given in the oral presentation of this paper; in 
general, the Olicount tool has been judged to 
be a benefit to the photointerpreter by the 
different organisations using it. 
4.5 Modifications 
An number of modifications are planned for the 
Olicount application: 
e User demand has arisen for a more 
sophisticated expert-system identification 
approach. In this context, candidate objects 
would be categorised according to their 
likelihood of correct identification. The use 
of ancillary parameters - regional context, 
local presence of typical commission 
species, etc. - as well as returning 
information on the sureness of blob 
identification, would be integrated into the 
GIS environment and the information 
displayed to the operator. In this manner, 
ground work can focus on stands most likely 
to present problems. 
e While the algorithm performance on 
medium sized (10MB) images at present is 
acceptable, image size can be a limiting 
factor to execution speed. With scanned 
aerial photos typically ranging from 80MB to 
120MB, techniques need to be developed to 
deal with larger image files. 
e As yet, no exploration has been made to 
determine whether multi-spectral (colour, 
false CIR) images provide a significant level 
363