PE&RS November 2018 Full - page 682

homogeneity (Espindola et al., 2006; Johnson and Xie, 2011;
Ming et al., 2012; Ikokou and Smit, 2013; Yang et al., 2015b)
or empirical goodness using multi-measures based on both
region and edge (Ming et al., 2009). In general, the most used
measures in unsupervised image segmentation evaluation are
intrasegment homogeneity and intersegment heterogeneity. In
current research, intrasegment homogeneity is mainly mea-
sured by variance within segmented parcels and intersegment
heterogeneity is often indirectly measured by correlation
between segmentation parcels, which is originally proposed
by Espindola et al. (2006).
It is recognized that low variance (high intrasegment
homogeneity) and low spatial autocorrelation (high interseg-
ment heterogeneity) are desirable (Johnson and Xie, 2011).
High intrasegment homogeneity of image objects can main-
tain the internal consistencies of classes. That means average
intrasegment homogeneity should be maximized to get global
mutual best fitting (Baatz and Schape, 2000). Meanwhile,
high intersegment heterogeneity is expected to maintain
high separability between image objects. That means aver-
age intersegment heterogeneity should be maximized to get
global separability (Ming et al., 2012). Therefore, most studies
use a Global Score (GS) (Johnson and Xie, 2011) or something
similar (Espindola et al., 2006; Ming et al., 2012; Ikokou and
Smit, 2013) to measure the segmentation accuracy. According
to Johnson and Xie (2011), GS is the sum of the normalized
average intrasegment homogeneity and normalized average
intersegment heterogeneity, which is equivalently based on
the assumption that intrasegment homogeneity and interseg-
ment heterogeneity are both equally important.
However, when the geo-objects within the image area are
uniform, large amount of experiments (Ming et al., 2012;
Ming et al., 2015) have proved that the image objects are
generally merged with the increase of scale parameter, which
is accompanied by the decrease of intrasegment homogeneity
and the increase of intersegment heterogeneity. That means
it is difficult to synchronously get both high intrasegment
homogeneity and high intersegment heterogeneity, from
which some issues have stemmed. First, of the two measures,
intrasegment homogeneity and intersegment heterogeneity,
which one is more important to the terminal optimal clas-
sification? Second, what is the relationship between the two
measures? Third, are there any coupling relationships among
scale parameter, image-object size or amount, intrasegment
homogeneity, intersegment heterogeneity, segmentation ac-
curacy and classification accuracy? The goal of this paper is
to quantitatively find the answers or ideas to the above ques-
tions. Taking the mostly used multi-resolution segmentation
(MRS), mean-shift segmentation (MS) and edge-constrained
watershed segmentation (ECW ) as examples, this paper adopts
series of image object classifications to quantitatively analyze
the coupling relationship between multi-scale image segmen-
tation and classification accuracy in GeOBIA. The research
findings are in a general sense and practically meaningful for
scale processing in GeOBIA .
Methodology
Scale Parameter in Multi-Scale Segmentation
Scale is a broadly used term in geoscience and has a variety of
meanings in different contexts; however a specific definition
of an object-based scale of remote sensing images has not been
given. Generally, the object-based scale is considered the size
of the meaningful unit (image primitive, which is also named
image object in eCognition®). From the viewpoint of using an
algorithm to extract the image objects, the object-based scale
corresponds to the scale parameters in the multi-scale image
segmentation (Ming et al., 2015). Spatial attribute and spectral
attribute are general attributes for spatial data, and they are
used in most segmentation algorithms, so scale parameter
can be basically divided into two groups, the spatial param-
eter and the spectral parameter. Although the principles of
different segmentation algorithms vary from each other, the
essence of scale effect is always the heterogeneity controlled
by spatial parameter and spectral parameter, which directly
presents the change of image object number or size with
change of spatial or spectral parameter. So, based on three
commonly used multi-scale image segmentation algorithms,
multi-resolution segmentation (MRS), mean-shift segmentation
(MS) and edge-constrained watershed segmentation (ECW), this
paper testify the coupling relationship between multi-scale
image segmentation and classification accuracy in GeOBIA.
Multi-Resolution Segmentation
MRS is a bottom-up region-merging technique. It starts with
one-pixel objects and merges similar neighboring objects to-
gether in subsequent steps until a heterogeneity threshold, set
by a scale parameter (SP), is reached (Benz et al., 2004). SP, an
important parameter of the multi resolution segmentation al-
gorithm, means the upper limit for a permitted change of het-
erogeneity throughout the segmentation process and directly
determines the average image object size. The object homo-
geneity to which the scale parameter refers is defined in the
composition of homogeneity criterion field. The homogeneity
criterion can be customized by weighting shape and com-
pactness criteria. Actually, SP is defined by the composition
of homogeneity criterion comprising two aspects, spectral
homogeneity and shape homogeneity or more concretely, the
weights of spectral heterogeneity (1-w1) and shape heteroge-
neity (w1); furthermore, shape heterogeneity is defined by the
weights of smoothness and compactness (w2). For additional
information about the segmentation algorithm, please refer to
(Baatz and Schäpe, 2000; Benz et al., 2004; Definiens Image
GmbH, 2007 and 2009).
Mean-Shift Segmentation
MS is a robust and adaptive clustering algorithm with non-
parametric density estimation and has been successfully
used in image segmentation (Comaniciu et al., 2001; Ming et
al., 2012; Ming et al., 2015). In mean-shift-based multi-scale
segmentation, there are three scale parameters (hs , hr, M). Spa-
tial bandwidth hs is the spatial distance between classes in
the spatial domain, and it indicates the spatial window size
in the segmentation. Attribute bandwidth (or spectral band-
width) hr represents the spectral difference between classes in
the spectral domain. An optional step of eliminating spatial
regions containing less than M pixels is another approach to
performing multi-scale segmentation. As proposed by Ming et
al. (2016), fixing hs and hr as statistically optimal values, the
parameter M greatly impacts the segmented image object size,
so M is usually used as the main scale parameter in multi-
scale segmentation, as well as in this paper.
Edge-Constrained Watershed Segmentation
In ECW based segmentation (Wang and Li, 2014; Wang et al.,
2015), initial small segments, also called sub-object primitives
(sub-OPs), are obtained using edge-constrained watershed
segmentation and edge allocation. Then, these segments are
gradually merged into a larger segment until the edge-con-
trolled limits are reached, thereby creating the initial OPs. At
last, non-constrained merging is conducted on the OPs, which
results in final segmentation. This segmentation method
has the advantages of higher segmentation accuracy and OP
boundary precision. The main scale parameter in this seg-
mentation method has similar meaning (maximal heterogene-
ity in merging process) as that in MRS (Benz et al., 2004). For
682 November 2018 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING