compared to that of the pixel-based methods (Table 4). A
multi-scale image segmentation, in which multiple scale
parameters are used to create multiple layers of image objects,
may be an appropriate solution. Thus, while cottonwood trees
in dense stands were not correctly segmented using the pa-
rameters adopted in this study, the stand as a whole could be
(a)
(b)
(c)
(d)
(e)
(f)
Plate. 2 Two examples of the smoothing effect of large texture windows: (a) and (d) are a subset of the panchromatic image; (b) and (e)
are the SVM classification results on spectral-texture data with a window size of 39 pixels; (c) and (f) are the SVM classification results on
spectral-texture data with a window size of seven pixels.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2015
803
