XIX-B8, 2012
HODS
Mountains National
order with Czech
covers an area of
7.79 km?. Forested
ce (Picea abies L.).
or health condition.
the adverse effects
s forests originating
in small fragments
d during July 2009
cular sample plots
n a systematic grid
area of the Stolowe
lot’s centre were
ind dead trees with
red.
008 using Altman's
f Cessna 206 plane.
vith LIDAR sensor
mote display. The
ing the flight was
510, which consists
tic dual-frequency
ler deflection and
MU).
tics
Trees Vis Software
ted using formulae
st management in
ne was assumed to
nd the lower dbh
me — divided into
e of dead trees’ —
. Results obtained
eference data.
ec software, in the
'HM) was loaded;
adius);
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
d) Segments which shape is not similar to tree were connected
to neighboring segments, according to the condition:
diameter ratio of the ellipse described on crowns extends
beyond the range 0.3-3);
e) Layer heights are determined (automatically or manually);
f) Primary segments are allocated to the height layers;
g) After allocation the segments are combined;
h) While there are more than one height layer, those from one
height layer are grouped as separate spatial objects (groups
of primary segments — stands/parts of the stands with
similar height/crown size);
i) Height layers are re-filtered using Gaussian filter, of varying
sizes - the higher layers of larger and lower smaller;
j) In this case three groups were established - the height limit
was 28 m and 16.6 m, filter size of 7, 5 and 3 pixels,
respectively, for layer heights;
k) After this segmentation average area of each tree segment
were determined. If the crown had a surface area greater
than the mean + standard deviation, it was again filtered
Gaussian filter of 2 pixels less at first and segmented one
again.
1) Segments which shape is not similar to tree were connected
to neighboring segments, according to the condition:
diameter ratio of the ellipse described on crowns extends
beyond the range 0.33-3.5);
m) For every final segment, based on maximum height from
CHM was determined and pixels below 0.7xHinax were
removed
For final segments such parameters as maximum height (Huy),
minimum height (H,;,) and crown radius (Cua) were
calculated.
Based on these characteristics, the mapped forest fragments,
which corresponded to the size and location of the ground
sample plots, the following characteristics were calculated: the
number of trees (NT), total tree height (SUMH), average tree
height (HA), the sum of crown projection area (CAR), the sum
of crown volume (VC). The crown volume was taken as a cone
with base equal to the projection of the crown and height such
as the length of the crown. All the features were related to trees
with a minimum 7 m height. Dead trees were removed based on
orthophoto interpolation.
2.6 Variants comparison
In the study the extent of the crown was associated with the
height of the tree (Fig. 1).
CoixH,. (1)
where:
C 7 crown extend CHM value
H,, = maximum CHM value inside final single tree
region
i = tested threshold height
In the presented study the test based on 34 sample plots was
carried out. Two variants of the assigning separate crowns in the
sample were used: (1) according to the centroid position
(‘centroid’), (2) according to the location of any fragment of the
crown inside the sample plot boundary (‘touch’) (Fig. 3). In
each of the variants five series of measurements with different
relative height of the range 0.65-0.8 (with a gap of 0.05) were
carried out.
For each selected height threshold and sampling rule calculated
correlation between the volume of live trees (field
measurements) and the features measured based on LIDAR
data. Set of features was determined on the basis stepwise
regression with the backward elimination model. The best
solution was chosen based on the highest value of correlation
(R).
Relationship between the stock volume of trees measured on the
ground and LIDAR parameters based on CHM was determined
in first set. Due to impact of dead trees, in second analysis, dead
trees were excluded from analysis. All calculations were
conducted using STATISTICA 8 (StatSoft, Inc.)
0 5 10 20
Meters
Figure 3. Results of two sample rules based on centroid
(polygons (trees segments) with solid line) and ‘touch’
(polygons (trees segments) with dashed line and with solid line)
based on sample plot area (gray circle with 12.62 m radius)
3. RESULTS
3.1 Segmentation accuracy
Accuracy evaluation was carried out in 9 different stands, on 5
sample plots. Photogrammetric measurements were performed
in the DEPHOS photogrammetric station, and were used as a
reference data. The spatial resolution of the images was 0.15 m.
The accuracy of the segmentation obtained for spruce was
82.3%. The species specific segmentation process correctly
detected 1156 of 1404 trees located in the stands from 27 to 110
years old. The 186 trees (13.3 %) were not detected and the 62
tree peaks (4.4 %) identified during stereo photogrammetric
observations were removed due to the occurrence of two
vertices of a tree in a single segment. In presented test
segmentation properties were not species specific, but general
(described in 2.5 section). This can cause a little worse
detection rate, and accuracy about 75 % (visual comparison
both segmentation results).
3.2 Volume analysis
Strength of the correlation between the volume of trees (defined
on the ground) and the characteristics of the LIDAR data set
was greater in the variant "touch". Optimal cut was in this case
the relative tree height of 0.7-0.75 (Tab. 2). In ‘centroid’
variant, strength of association was slightly smaller, and it was
the best cut for the relative height of 0.75.
