notably the
PCA
transform by itself also outperformed the
spectral bands alone. Including transform inputs increased the
efficiency of the software program in terms of refining the
ANN
routine, without impacting processing times significantly.
For the spatial contextual classifier, the approach of mask-
ing non-forest areas prior to classifying dead and live trees
yielded substantially higher classification accuracy. This is
most likely because the dead conifer class can be routinely
confused with other land covers and materials such as se-
nescent grass and bracken ferns, bare ground, and roads, so
creating a non-forest layer to mask out these features, resulted
in less confusion and a more accurate classification. Tiede
and Hoffmann (2006) also found that masking out the back-
ground bare ground dramatically improved their object-based
classifications for single tree detection.
Based on dead tree class accuracy measures alone, the
classification products from the spatial contextual approach
yielded significantly higher accuracies than the object-based
products. In particular, the addition of transforms to the spec-
tral bands helped boost the performance of the spatial contex-
tual approach. In addition, the object-based products pro-
duced a significantly greater number of objects than the spatial
contextual approach, although the commission errors were
not significantly different between products. This is useful to
know, if an analyst needs to manually edit a product derived
by an automated image classification routine, as it would be
more efficient to edit a product that has fewer commission
objects, as is the case for the spatial contextual product.
Several qualitative characteristics are evident, when compar-
ing the object-based and spatial contextual approaches. The spa-
tial contextual approach is a straightforward supervised training
oriented program that is easy to learn. The object-based program
has a multitude of options and choices for the analyst to make
when classifying an image, and can take a while to learn. One of
the strengths of object-based routine is the rule-based approach
to classification where the analyst can customize a variety of
specific rules to help refine the classification. This approach was
found not to be effective, however, because of spectral-bright-
ness variation and illumination differences across the aerial
orthoimagery scenes. The “rules” could be satisfactorily devel-
oped for one portion of the image, but did not apply to other
portions. The greatest advantages of the eCognition object-based
program are the great amount functionality with object classifi-
cation tools and versatility of its output formats. It can produce
a raster classification with multiple classes, or if a single class
polygon map is desired, it can be exported as a vector file.
The segmentation routine is also a strength of the object-
based program, but it required considerable computing capacity,
and became problematic with larger image data sets. For exam-
ple, the Laguna data sets for 1997, 2000, and 2002, which were
102
MB
, 102
MB
, and 233
MB
in size, respectively, could be pro-
cessed as complete scenes. However, the 2005 Laguna Mountain
imagery data set, which is 967
MB
in size, had to be segmented
in sections, classified, and then the classified sections had to
be stitched back together, which generated a whole other set
of issues to contend with. The spatial contextual program had
no problems handling any of the data sets, although processing
times were slightly slower with the larger image extents.
The processing times for both software packages were
comparable. The most time-consuming element for the object-
based approach was the segmentation process. The most
time-consuming element for the spatial contextual approach
is the “clean-up clutter” process, where the analyst identifies
correct polygons and incorrect polygons for many sections of
the image and then re-runs the classification.
Despite the fairly high accuracies for the dead tree class
that were achieved from both of these approaches, the util-
ity of the classification products are limited without further
editing of the products. Lang
et al
. (2009) found that object-
based classifications could be improved effectively with the
addition of manual approaches, due to the human brain’s
enhanced cognition abilities. An accuracy of around 85
percent for a mapped class is an accepted standard in order
to consider the product reliable for use in further analysis,
(Anderson
et al
., 1976) while accuracies for many of the map
products were in the 60 to 70+ percent correct range for the
dead tree class, some were in the 30 to 50 percent range.
However, despite lower accuracies for some of the classifica-
tion products, using these semi-automated approaches are ad-
vantageous over trying to create a dead tree classification from
visual interpretation techniques alone. An attempt was made
to manually digitize dead tree polygons grid by grid for an en-
tire image, however this proved to be ineffective and mostly
intractable. For the large areas being mapped, it is much more
efficient to semi-automatically map dead trees then edit the
errors than to manually generate the map from the beginning.
For subsequent analyses the dead tree maps were manually
edited to remove false positives and then add omitted dead
tree objects. The resultant products were reassessed using the
object-based accuracy assessment. In this way, maps were cre-
ated which were sufficiently accurate for subsequent analyses
of tree mortality patterns (not reported in this paper).
Other studies have noted that object-based approaches
have both advantages and limitations. Mallinis
et al
. (2008)
found that their best maps had an accuracy rate of 80 percent,
with added transforms, when using QuickBird imagery and
an object-based classification to map forest vegetation, and
further use of these maps would need to be considered care-
fully without additional editing. Chubey
et al
. (2006) found
that object-based classification approaches (eCognition and
decision tree classifiers) were very useful for deriving certain
forest inventory parameters, such as individual forest species,
non-forest land-cover, and percent crown cover, but were not
as effective in identifying stand height and age classes. The
authors suggest that there is a growing utility of methods such
as these to meet operational forest inventory needs.
Study Sites, Imagery Types and Dates
Both the object-based and spatial contextual approaches
yielded consistent results when applied to multiple types
of imagery with different spatial resolutions. For both ap-
proaches, when considering all of the classifications products,
there were no statistically significant differences between the
accuracies for the different dates (e.g., imagery types related
to different sensors). However, the accuracies of the resultant
products did vary for both approaches by study area, with the
classifications for the Volcan study area having consistently
the highest accuracies, followed by Palomar and then Laguna.
When comparing accuracy values for each study area for the
true object-based method, there were significant differences
between all of the study sites. With the spatial contextual ap-
proach there was a significant difference between the accura-
cies for the Volcan and Laguna sites. The few classification
products with the highest (90 percent) accuracies were pro-
duced for the Volcan study area with the
CIR
(2002) imagery.
The reliability of delineated dead trees for the different
study sites varied, likely due to factors such as background
vegetation/cover and differential view angle effects for the
large format aerial imagery. Furthermore, study design and
data availability make it difficult to separate the effects of
these factors. Volcan and Palomar are much smaller in area
than Laguna, and it is possible to achieve higher accura-
cies when classifying smaller areas of imagery due to greater
homogeneity of vegetation types and conditions and less
complexity in background characteristics.
Also, the target to background contrast is less defined
in the Laguna imagery, due to the more open nature of the
578
July 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
