PAN-R-NIR1
combination with a closeness index of 0.15 (rela-
tive accuracy of 89 percent). There is a slight improvement in
tree crown delineation when incorporating the
DSM
with opti-
cal imagery. OverSegmentation (correctness) represents the
ratio of the area of geographic intersection of reference objects
and tree crowns to the total area of reference objects. Hence,
the
PAN-R-NIR1
method got the highest error rate (15 percent).
Completeness (UnderSegmentation) represents the ratio of the
area of geographic intersection of reference objects and tree
crowns to the total area of corresponding objects (Clinton
et
al
., 2010). Completeness of both methods is very good, show-
ing 0.02 for UnderSegmentation.
The visual appearance of results obtained from the
IWS
method was poor (Plate 3e and 3f), and the quantitative accu-
racy assessment confirmed this in Table 3. A closeness index
value of 0.94 indicated very poor completeness due to high
UnderSegmentation value. The OverSegmentation value was
close to zero indicating that the most objects delineated using
the
IWS
method coincided with reference tree crowns.
Discussion
Various approaches have been developed for detecting indi-
vidual tree crowns from remotely sensed data. However, most
of these were developed with deciduous or coniferous forests.
Trees within these forests have crowns with conical shapes.
These methods are also generally successful in plantation for-
ests where there are regular distances between trees (Kaartinen
et al
., 2012; Leckie
et al
., 2005; Tiede
et al
., 2005). According
to Vauhkonen
et al
. (2012), the success of tree crown delinea-
tion algorithms was significantly affected by the forest struc-
ture, especially tree density and clustering. To confirm that,
their study investigated six different tree crown delineation
algorithms for different forest types such as boreal forests in
Norway and Sweden, coniferous and broad-leaved forests in
Germany, and a tropical pulpwood plantation in Brazil using
laser scanning data. Due to the development of tree crown de-
lineation algorithms for specific forest types and species com-
positions, they are less likely to be applied to mangroves. For
example, Hirata
et al
. (2010) segmented fewer mangrove tree
crowns using QuickBird panchromatic satellite images and a
watershed segmentation than the reality. The presence of inter-
mingled mangrove tree crowns and the structural complexity
of mangrove forests make them unique and densely clustered.
In this study, tree top detection using local maxima, and
crown boundaries identification using a region growing
method was tested on different data layer combinations from
high-resolution optical images. This approach is more flexible
compared to the inverse watershed method in detecting tree
crowns with varying sizes as it uses the local maxima as the
starting point of the growing seed, and the local minima as
the ending point. For example; the
PAN-R-NIR1-DSM
combina-
tion used local maximum values of the
DSM
and
NIR1
band as
the starting point of the seeds, grew them until the ratio of the
NIR
1 spectral value of the seed object and neighboring objects
reached to 0.9, and the height difference between the seed
object and the neighboring objects was up to 2 m. Further,
instead of only investigating the spectral variations of images,
incorporating height variations of canopies was fruitful as
the generated
DSM
identified minor height variations within
canopies (Plate 2).
The
PAN-R-NIR
1 method showed a good visual appearance
of mangrove tree crowns when overlaid on the panchromatic
image (Plate 3a). This is supported by the detailed study of
Kamal
et al
. (2014) who confirmed the possibility of iden-
tifying a single mangrove tree crown from remotely sensed
images with pixels smaller than 2 m. However, the number
of trees detected using the
PAN-R-NIR
1 method was less than
the
PAN-R-NIR1-DSM
method, and had relatively large gaps
between demarcated crowns. This was due to the similarities
in spectral reflectance causing some trees to be aggregated and
demarcated as one. Therefore, it can be concluded that the
homogeneity within each mangrove species, a percentage of
crown overlap and limited height variations of canopies can
affect the accuracy of delineation. For example: if neighboring
trees were as tall as the seed trees, brightness variations in the
image were unlikely to be seen, and they were therefore ag-
gregated as one crown (Plate 3a). However, conically shaped
crowns and crowns having considerable distance between
each other were accurately delineated.
Successful tree crown delineation is strongly influenced by
the shape of tree crowns. Gougeon and Leckie (2006) stated
that the detection of tree locations using local maxima pro-
vides good results for medium to dense coniferous stands with
conical shape crowns in conjunction with
GEOBIA
. However,
this is not the case with mangroves. Most mangrove species
tend to be rather flat as opposed to conical in shape (Duke,
2006; Wightman, 2006), and therefore the local maxima ap-
proach is less suitable at some places. The best approach is to
divide the area into homogenous forest stands before apply-
ing crown delineation methods. This finding supports that of
Larsen
et al
. (2011), who evaluated six different individual
tree crown delineation algorithms in varying forest conditions.
When the mangrove tree crowns obtained from the
PAN-R-
NIR1-DSM
combination were overlaid with the high-resolution
panchromatic image, the visual appearance was better than
the
PAN-R-NIR1
combination (Plates 3c and 3d). The shapes of
the crown boundaries were improved due to incorporation of
both image reflectance and height variations for region grow-
ing. The number of trees identified using the
PAN-R-NIR1-DSM
method was larger than the other methods (Table 2). Since the
DSM
is sensitive to height variations (spatial resolution of the
DSM
was 15 cm), trees with small height variations compared
to neighboring trees can also be detected. Plate 4c shows some
examples of low laying mangrove tree identification. How-
ever, there might be some false tree identifications as well. If
the
DSM
is sensitive to small height variations within cano-
pies, multiple upward pointing branches of the same tree are
identified as different trees. Hence, more than one tree crown
can be counted per tree. Although there was a treetop filter-
ing technique applied in terms of average distance between
trees, it was not possible to avoid false tree identification
completely. The filtering distance was calculated according
to field observations and generalized to the Rapid Creek area.
To achieve higher accuracy, this could be done on a species-
specific level although that in itself is a challenge (Heenkenda
et al
., 2014). Further, there might be some inaccuracies with
respect to the
DSM
generation. For example; the homogeneous
forest coverage within communities and the repetitive pattern
of mangroves can significantly impact on the image matching
accuracy (Baltsavias
et al.
, 2008; Gruen, 2012). Introducing
ground control points and manual editing processes cannot
avoid such irregularities completely. One of the alternatives is
to use laser scanning data with high point density to generate
canopy height models.
Although the
IWS
method detected tree tops as success-
fully as the other two methods (Plate 4c and Table 2), once
we implemented the processing steps, the method clustered
neighboring mangrove trees with similar characteristics
together and delineated them as a single tree crown (Plate 3e
and 3f). The key input for the
IWS
algorithm is a flow direc-
tion map that shows direction of flow out of each pixel of the
inverted surface raster. If the height irregularities to the neigh-
boring pixels are negligible with respect to the introduced
control algorithms, the system cannot identify corresponding
flow direction (ArcGIS Resources - Esri, 2012). Although the
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
477
