1173
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
via Fourier transform inversion. Because image R(x, y) is an
analytic signal, R r (x, y) and R t (x, y) form a quadrature pair. The
amplitude envelopes of R(x, jy) may be recovered via
A(x,y) = ^R r 2 (x,y) + R 2 (x,y) (10)
The amplitude response A(x, y) is regarded as the texture
features in the research.
The log Gabor filter bank is used to convolute with each band
to derive texture response in each scale and orientation. Com
bining the texture information in all orientations, multi-scale
texture features of each band are obtained.
2.3 Watershed Transform
The watershed transform is a well-known powerful tool for
automated image segmentation. Because its resulting bound
aries form closed and connected regions, it becomes one of the
best choices of the segmentation of remotely sensed image (e.g.
Li et al., 1999; Hall et al., 2004; Chen et al., 2006; Li and Xiao,
2007), which need recognise all of the objects in the image. The
presented watershed algorithm is based on an immersion
process analogy (Vincent and Soille, 1991), in which the
flooding of the water in the image is efficiently simulated using
a queue of pixels.
Watershed transform often produce over-segmentation in
situations of high gradient noise, quantity error and detailed
texture. There are many solutions to the problem. A marker
based solution is chosen in this research that basins are flooded
from selected sources rather than minima. To integrate the edge
and texture information of the landscape objects in image,
watershed transform is implemented based on edge features,
and the marker image is calculated from texture features. The
texture features are segmented at first using a moving threshold
algorithm developed from Hill et al. (2003). This algorithm
calculates the mean and standard deviation of the texture
features. Then several binary images are produced at reasonably
spaced thresholds using the mean and standard deviation. For
each binary image, the number of closed and connected regions
greater than the given minimum size is calculated. The thres
hold with the maximum number of connected regions is used as
the output marker image. With this algorithm, no a priori
knowledge is required about the number of regions. One only
needs to give the size of the minimum region, which is always
constant to most images.
The edge features can be reconstructed with the marker image
and then be segmented based on the watershed transform. The
image has produced several scale texture features after the log
Gabor bank filtering. Use different frequency of the texture
features can produce segmentation results in different scale.
Generally, low-frequency texture features produce large-scale
segmentation results, whereas high-frequency texture features
produce small-scale segmentation results.
2.4 Accuracy Assessment
The decision of the best segmentation results usually relies on
the accuracy assessment. However, similar to the segmentation
theory there is no established procedure for the assessment of
its results. A general classification of assessment methods has
been proposed by Zhang (1996), but only very few studies
employ the assessment on remotely sensed image. Carleer et al.
(2005) proposed a supervised method using discrepancy
measures between a reference and the segmentation, which is
employed in the research.
The reference map is obtained interpreting from the original
image by remote sensing experts. The reference polygons then
are converted to raster. To compare the discrepancy, the seg
mentation results are overlaid on the reference map. Observing
from each segment region, the largest part in the reference map
is regarded as right segmentation of this region, the others are
regarded as mix-segmented pixels. Two discrepancy
evaluations are calculated: percentage of Right-Segmented
pixels in the whole image (RS), and ratio of Region Count in
segmented image to reference map (RC). For an excellent
segmentation, RS will be close to 100%, and RC will equal to 1.
In the assessment method based on reference map, if only RS is
used, RS will be increased with the fragmentation of the
segmentation, even it will be equal to 100% when the region
size equal to one pixel, which becomes a meaningless
segmentation. So RC is also important indicator to restrict the
segment regions close to the reference regions.
3. DATA
The data used throughout the research is a 512x512 pixel sub
image of a QuickBird-2 scene acquired in November 21, 2004
(Figure 2). Geographically, this area represents a portion of the
highly fragmented agricultural landscape typical of the
Jiangning region of Nanjing, China. QuickBird-2 provides 16bit
multispectral data in the red, green, blue and near-infrared
bands at 2.4m spatial resolution and an libit panchromatic
band at 0.6m resolution. For producing high-resolution
multispectral images, the multispectral bands are sharpened to
0.6m resolution using panchromatic band based on Pansharp
method proposed by Zhang (2002) at first. The false colour
image for shown in Figure 2 is composed with infrared, red, and
green band. In the false colour image, the red colour represents
vegetation, the black colour indicates water, and the white
colour represents roads.
Figure 2. Original image overlaid reference polygons