Pixel-based Classification Using Spectral-Texture Images
Figure 5 shows the overall classification accuracies using
spectral-texture images, with texture-window sizes ranging
from seven pixels (corresponding to a 4.2 m ground distance)
to 41 pixels (corresponding to a 24.6 m ground distance).
SVM
outperformed
MLC
for all texture-window sizes. As sug-
gested by the higher user’s accuracy of cottonwood in the
SVM
results, fewer pixels were misclassified as cottonwood when
SVM
was used as the classification algorithm (confusion matri-
ces not shown). However, compared to results derived using
only spectral bands, neither classifier improved classification
accuracy much by adding texture information to the spectral
data. For both
MLC
and
SVM
, overall classification accuracy
decreased as the texture-window size increased, but then ac-
curacy increased after the window size reached 17 pixels.
Semi Object-based Classification
Based on a manual comparison among various segmentation
results, a scale parameter of 25 was chosen for image segmen-
tation, as shown in Figure 6. Shadow objects, which were
used to mark candidates as cottonwood objects (as previously
detailed in the Methods Section), were created based on the
multispectral
SVM
classification result with the highest pro-
ducer’s accuracy for the shadow class.
As expected, many non-cottonwood objects, such as soil,
saltcedar, and Sophora that neighbored cottonwood, were
included as cottonwood tree crown candidates (Figure 6). To
exclude these non-cottonwood pixels, we overlaid onto the
candidate objects the spectral-texture
SVM
classification result
with a spectral window size of 35 pixels, which had the high-
est user’s accuracy for the cottonwood.
Figure 5. Overall accuracy of MLC and SVM classification results using various texture-window sizes.
Figure 6. Image segmentation result (scale parameter 25) and cottonwood object candidates after proximity analysis overlay on the
QuickBird panchromatic image.
800
October 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
