PE&RS July 2016 Public - page 575

method. This approach determines the “presence” or “ab-
sence” of classified objects (i.e., delineated dead trees) utilizing
a 50 percent rule. If a mapped dead tree object and a corre-
sponding validation dead tree object overlap by more than 50
percent then the object is considered in agreement or correct; if
the overlap is less than 50 percent then the object is considered
to be erroneous. The resultant accuracy is the total number of
classified objects meeting this 50 percent threshold condition,
divided by the total number of validation objects (50).
Results
Final parameter selections for the segmentation step of the
object-based approach are found in Table 3, which lists the
scale, shape/color, and compactness factors selected for
classification trials based on test images. A total of 74 clas-
sification productes were created and evaluated for each
approach, which consist of each added transformation test for
each study area, and for each imagery type (Table 4). When
comparing the results of the object-based classification with
varying products derived with spectral transformation as
inputs, four products had higher accuracy levels compared
to the spectral bands alone. Up to a 5 percent increase in
accuracy and up to a 13 percent reduction in commission
error was achieved with the products from the Combination
approach. Specifically, these products are the combination
of all inputs, PCA + spectral bands, NDGR + spectral bands,
and NDVI + spectral bands. A paired t-test was performed to
Table 3. Object-based (eCognition ) Segmentation Parameters.
Imagery Date Scale Shape/Color Compactness
2002 20 0.3/0.7 0.5
2005 50 0.3/0.7 0.5
2007 150 0.3/0.7 0.5
Table 4. Dead Tree C lassification Results for the Object-Based Classification Accuracy Results (eCognition) Classifier Comparing Combinations of Addition of
Transforms to Spectral Bands and S pectral Transforms Alone, for Each Date of imagery, and for Each Study Site (A = Accuracy; C = Commission Error; O
= Omission Error). Note that the 2002 Imagery is CIR (G, R, NIR bands), the 2005 Imagery is T rue Color (B, G, R Bands ), and the 2007 I magery is Four
Bands (B, G, R, NIR). The Top Two Accuracy Values are in Bold for Emphasis, for Each I magery Date.
Table 4A. Palomar
Palomar
2002 2005 2007
Inputs to
Classifier A C O A C O A C O
spectral bands 46 46 54 60 32 40 76 38 24
Combination 46 52 54 64 32 36 80 50 20
PCA + bands 54 46 46 68 26 32 80 42 20
PCA transform 48 40 52 60 58 40 72 68 28
NDGR + bands 42 54 58 64 32 36 82 32 18
NDGR transform 38 56 62 48 45 52 56 42 44
VARI + bands ** ** ** 58 32 42 58 50 42
VARI transform ** ** ** 46 38 54 30 22 70
NDVI + bands 44 48 56 * * * 74 46 26
NDVI transform 46 44 54 * * * 68 40 35
SAVI + bands 56 50 44 * * * 74 46 26
SAVI transform 38 46 62 * * * 34 40 66
SR + bands 42 54 58 * * * 76 46 24
SR transform 34 42 66 * * * 42 46 58
*= no NIR band available to construct these particular transforms
**= no blue band available to construct this particular transforms
Table 4B. Volcan and L aguna
Volcan Laguna
2002 2005 2002 2005
Inputs to
Classifier A C O A C O A C O A C O
spectral bands 88 54 12 66 70 34 44 44 56 42 62 58
Combination 96 24 4 70 64 30 50 44 50 54 64 46
PCA + bands 92 24 8 70 20 30 42 56 58 46 38 54
PCA transform 80 30 20 60 54 40 40 52 60 44 38 56
NDGR + bands 92 46 8 66 24 34 44 48 56 54 64 46
NDGR transform 72 38 28 50 36 50 40 44 60 46 40 54
VARI + bands ** ** ** 68 50 32 ** ** ** 48 64 52
VARI transform ** ** ** 48 56 52 ** ** ** 34 48 66
NDVI + bands 80 34 20 * * * 44 60 56 * * *
NDVI transform 66 58 34 * * * 44 56 56 * * *
SAVI + bands 76 34 24 * * * 46 56 54 * * *
SAVI transform 46 50 54 * * * 40 58 60 * * *
SR + bands 78 50 22 * * * 42 56 58 * * *
SR transform 48 40 52 * * * 36 48 64 * * *
*= no NIR band available to construct these particular transforms
**= no blue band available to construct this particular transforms
Table 5. Spatial Contextual C lassification Accuracy Results Comparing Combinations of Spectral B ands and Spectral T ransforms for Each Date of Imagery
and Each Study Site Comparing Addition of T ransforms to Spectral B ands and Transforms Alone, for Each Date of Imagery, for Each Study Site (The T op Two
Accuracy Values are in B old for Emphasis, for Each Imagery Date)
Table 5A. Palomar
Palomar
2002 2005 2007
Inputs to
Classifier A C O A C O A C O
spectral bands 60 34 40 60 70 40 88 32 22
Combination 72 34 28 78 54 22 76 42 24
PCA + bands 78 40 22 70 52 30 54 42 46
PCA transform 70 46 30 62 54 38 80 80 20
NDGR + bands 84 36 16 74 54 26 86 32 14
NDGR transform 62 38 38 44 56 56 42 46 58
VARI + bands ** ** ** 62 58 38 88 28 12
VARI transform ** ** ** 24 20 76 80 50 20
NDVI + bands 78 56 22 * * * 50 42 50
NDVI transform 66 72 34 * * * 40 28 60
SAVI + bands 66 42 34 * * * 46 30 54
SAVI transform 26 32 74 * * * 34 48 66
SR + bands 60 50 40 * * * 66 54 34
SR transform 66 66 34 * * * 60 42 40
*= no NIR band available to construct these particular transforms
**= no blue band available to construct this particular transforms
Table 5B. Volcan and L aguna
Volcan Laguna
2002 2005 2002 2005
Inputs to
Classifier A C O A C O A C O A C O
spectral bands 64 34 36 76 70 24 46 42 54 66 46 34
Combination 68 44 32 70 46 30 62 32 38 60 38 40
PCA + bands 76 34 24 66 44 34 76 50 24 56 48 44
PCA transform 70 56 30 66 52 34 66 48 34 60 50 40
NDGR + bands 86 38 14 70 40 30 82 34 18 72 36 28
NDGR transform 60 58 40 58 54 42 46 48 54 32 44 68
VARI + bands ** ** ** 78 40 22 ** ** ** 56 54 44
VARI transform ** ** ** 74 54 26 ** ** ** 52 46 48
NDVI + bands 78 40 22 * * * 66 52 34 * * *
NDVI transform 66 74 34 * * * 44 36 56 * * *
SAVI + bands 66 46 34 * * * 66 50 34 * * *
SAVI transform 46 58 54 * * * 44 50 56 * * *
SR + bands 68 48 32 * * * 68 40 32 * * *
SR transform 68 64 32 * * * 46 40 54 * * *
*= no NIR band available to construct these particular transforms
**= no blue band available to construct this particular transforms
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING July 2016 575