PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
461
Multi-scale Segmentation of High-Spatial
Resolution Remote Sensing Images Using
Adaptively Increased Scale Parameter
Xueliang Zhang, Xuezhi Feng, and Pengfeng Xiao
Abstract
The adaptively increased scale parameter (
AISP
) strategy is pro-
posed to control multi-scale segmentation based on region grow-
ing methods.
AISP
strategy contains a set of gradually increased
scale parameters to produce nested multi-scale segments.
Instead of independently assigning the set of scale parame-
ters ahead of segmentation, the contribution of this study is to
dynamically determine scale parameters during segmentation
procedure, making scale parameters adaptive to specific images
and cover meaningful segmentation scales. Furthermore, the
effectiveness of gradually increased scale parameters on seg-
mentation accuracy is analyzed, which gives a thorough under-
standing to local-oriented region growing methods. The exper-
imental results on a set of high-resolution images proved the
effectiveness of
AISP
on controlling multi-scale segmentation.
AISP
holds the application potential for object-based analysis of
high-resolution images.
Introduction
High-spatial resolution remote sensing images present geo-
graphic objects in detail, which allows an accurate geomet-
rical analysis (Benediktsson
et al
., 2012). When dealing with
high-resolution images, the geographic object-based image
analysis (
GEOBIA
) method has become the principle method as
it is less sensitive to spectral variance within objects and can
make use of both object features and spatial relations between
objects (Hay and Castilla, 2006; Blaschke, 2010; Blaschke
et
al
., 2014). Image segmentation provides objects for
GEOBIA
by
dividing an image into a set of spatially contiguous regions
with spectral homogeneity or semantic coherence, on which
the
GEOBIA
performance highly depends.
In high-resolution images, objects emerge representing var-
ious sizes, spectral heterogeneities, and shapes. It is difficult
to segment various objects as single regions in one segmen-
tation result (Borenstein and Ullman, 2008). Objects can be
represented by multi-scale segmentations with large objects
at coarse scales and small objects at fine scales (Burnett and
Blaschke, 2003; Benz
et al
., 2004; Bruzzone and Carlin, 2006;
Sharon
et al
., 2006). Furthermore, users may describe the
same object at different semantic levels, i.e., ‘tree’, ‘tree group’,
‘forest’, which also requires the multi-scale segmentation of
objects. After multi-scale segmentation, the important work is
to select meaningful segmentation scale(s) for specific appli-
cations (e.g., Carleer
et al
., 2005; Neubert
et al.
, 2008; Clinton
et al
., 2010; Drǎguţ
et al
., 2010; Zhang
et al
., 2012; Yang
et al
.,
2014). Hence, multi-scale segmentation can be viewed as pro-
viding candidates of meaningful segmentation scales.
The region growing method is a good choice to gener-
ate multi-scale segmentations by setting different stopping
rules (Dey
et al
., 2010). It can produce spatially contiguous
regions with closed boundaries intrinsically. These regions
are viewed as image objects in
GEOBIA
directly. The stopping
rule can be viewed as the scale parameter. On the control of
a scale parameter, if more growing iterations are allowed, the
coarser-scale segmentation result is produced. On the con-
trary, if fewer growing iterations are performed, the segmen-
tation result is at finer scale. Region growing is convenient to
form the horizontal-topological links within a segmentation
scale. However, to form the hierarchical-nested links between
different scales, the multi-scale segments should be nested.
Among region growing methods, the hierarchical step-
wise optimization (
HSWO
) (Beaulieu and Goldberg, 1989) and
hierarchical segmentation (HS
eg
) (Tilton
et al
., 2012) algo-
rithms can produce a segmentation hierarchy, in which the
segmentations at coarser scales are merged from regions at
finer scales. Hence, the multi-scale segmentations are nested
intrinsically. On the other hand, local-oriented region grow-
ing methods (Câmara
et al
., 1996; Baatz and Sch pe, 2000;
Castilla
et al
., 2008; Wang
et al
., 2010) search the merging pair
within local vicinity. Local-oriented methods run fast and can
take the local contexts into account. However, local-oriented
region growing cannot guarantee to produce nested multi-
scale segments just by setting different scale parameters.
To produce nested multi-scale parameters by local-oriented
region growing, the segmentation procedure can be controlled
by a set of gradually increased scale parameters (e.g., Zhang
et
al
., 2013), where each scale parameter determines a segmen-
tation scale. The merging iterations at a segmentation scale
(
k
) are performed on the segments produced at its former
segmentation scale (
k
− 1). Then, the multi-scale segments
are nested. Usually, the scale parameters are defined as the
threshold of similarity measure between adjacent regions.
Then, a set of gradually increased scale parameters is inde-
pendently assigned ahead of segmentation by setting a con-
stant gap between neighboring scale parameters. A question
raised is how to appropriately assign the set of scale parame-
ters to capture the variety among different images and cover
meaningful segmentation scales.
Xueliang Zhang, Xuezhi Feng, and Pengfeng Xiao are with
the Jiangsu Provincial Key Laboratory of Geographic Informa-
tion Science and Technology, the Key Laboratory for Satellite
Mapping Technology and Applications of State Administra-
tion of Surveying, Mapping and Geoinformation of China, and
the Department of Geographic Information Science, Nanjing
University, 163 Xianlin Avenue, Nanjing 210023, Jiangsu,
China (
).
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 6, June 2015, pp. 461–470.
0099-1112/15/461–470
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.6.461
