2.2 Thresholding
Panchromatic imagery is single band: and
employing the primary principle a) above, a
binary threshold can be applied to segment the
target image into “olive tree plus shadow” and
null classes (Schowengerdt, 1983).
Due to the variable brightness characteristics -
even within olive tree stands - exhibited in the
data, such an approach can only have limited
success. Therefore, following the region
growing and subsequent morphometric testing
of candidate tree "blobs" (82.3), the threshold
value is thus reduced by one digital number
value and the threshold re-applied.
In Figure 3, a schematic brightness profile
(dashed line) is shown across a tree crown;
the profile dips as the darker pixels
representing the crown are present. At
threshold value A, all pixel DN values in the
target image fall below the threshold (the profile
is not intersected), and are considered as
candidate pixels; the region produced will be
large, but nevertheless will fail morphometric
testing. At threshold value B, pixels
representing the crown fall below the threshold,
but background pixels are above; a region can
be grown and then tested for morphometric
characteristics. At threshold C, all pixels are
above the threshold, and thus no candidate
pixels will be identified.
As can be seen from Figure 3, the approach is
essentially stable and program cycles above or
below typical DN values found in the target
image are essentially non-productive. The use
of maximum and minimum thresholds is,
therefore, only intended to increase program
efficiency.
2.3 Blob generation and testing
Starting with a maximum threshold value, the
target image is scanned for what we have
DN value
bat A HA
p = m * e =» =
> ,
: Tree crown . à
Messe € ^ — e —MÀÁ— I I «——
* 4
* e
® a=
profile
ORIGINAL IMAGE LABELLED IMAGE
Eee
Figure 3: Brightness profile across an olive
tree crown
pm - rit
sd 8/10 55: | ; 21426
; 1011 9 ; 2152272
8 9 10 7 8 6 | 33 3 221258
7 8 8 2 554 | 34353 2449 7:72. 2
9 511: s "Ing 33 : t
Tan : > 53.53. 3
9 ‘PT «
x ;
4 A ‘
Blobs Background
(or Objects)
Figure 4: Blob labelling
termed "blob seeds", i.e. pixels with a value
lower than the current threshold. When a blob-
seed is found, a region growing function
generates a complete "blob", and the
corresponding pixel in a mask image (blob
mask) is updated. The blob mask is effectively
a map of processed seeds (Figure 4).
The morphological parameters are then
checked (size, aspect, density), and if the blob
satisfies all these criteria, the member set
pixels are tagged as “tree” and labelled in
sequence. Once the target image has been
scanned, a list of trees is generated. The
threshold value is then lowered by one, the
blob-mask reset to zero, and the process
begins again. A second mask image (tree
mask) is updated with all blob pixels, which
prevents searching in these regions during
subsequent cycles. The process continues until
the minimum threshold value is reached. The
final output is an ASCII file with an identifier
and barycentre coordinate for each tree.
2.4 GIS data management
In our prototype, this exchange ASCII file is the
input data source for an ArcView™ application,
which enables further manipulation (inside
polygon, buffering, on-screen editing) of the
results by the operator.
A major advantage of the GIS approach means
that a validated count for a particular tree
object can be maintained and managed
spatially, even over time. Furthermore, the
development of the algorithm in a GIS
environment minimised programming of
standard GIS functions, such as point in
polygon querying, database management, etc.
The integration of the algorithm within a
Geographic Information System (GIS)
environment is essential for any expert-system
which is intended for a real-world application
(Peedell et al. 1998).
360 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
—— a pele A amas
