Merging stops when the heterogeneity of the resulting object,
which is a weighted combination of the color and shape
of the objects, exceeds a predefined threshold (Benz
et al
.,
2004). Due to the spectral similarity between cottonwood and
saltcedar, we minimized the influence of the spectral informa-
tion by setting the weight of the object shape to the maxi-
mum value, 0.9. The shape of the objects is also a weighted
combination of “smoothness” which describes the similarity
between borders of an image object and a perfect square, and
“compactness” which describes closeness of clustered pixels
in an object when compared to a circle. The weights of both
“smoothness” and “compactness” were calculated (refer to
Benz
et al
., 2004 for the formulas) based on manually digi-
tized training samples of cottonwood trees. We tested various
heterogeneity thresholds, also known as scale parameters,
and the best scale parameter was chosen based on a manual
inspection of the segmentation results.
In order to use the relationship between cottonwood trees
and their shadows in the classification, shadow pixels were
first identified and converted to polygon objects. The dis-
tance from an image object to its closest shadow object (near
distance) was then generated using the spatial proximity tool
in ArcGIS
®
(Esri, Inc., California) for all image objects created
in the segmentation step. Finally, image objects with a near
Figure 3. Different texture characteristics of cottonwood, saltcedar, and Sophora shown in the QuickBird panchromatic image.
Figure 4. Cottonwood trees and their associated shadows in the QuickBird panchromatic image.
798
October 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
