PE&RS November 2018 Full - page 681

Coupling Relationship Among Scale Parameter,
Segmentation Accuracy, and Classification
Accuracy In GeOBIA
Dongping Ming, Wen Zhou, Lu Xu, Min Wang, Yanni Ma
Abstract
The quality of multi-scale segmentation mainly consists of
intrasegment homogeneity and intersegment heterogeneity;
however, it is difficult to synchronously get both high. It is
crucial to make it clear which one of these two measures is
more important and what is the coupling relationship among
segmentation scale parameter, image segmentation and clas-
sification accuracy. This paper employs series of segmenta-
tion and classification to show that (1) intrasegment homo-
geneity is more important than intersegment heterogeneity
in GeOBIA; there is always highly positive correlation between
intrasegment homogeneity and classification accuracy; (2)
with the increase of spectral heterogeneity parameter, both
image object amount and the intrasegment homogeneity
decrease; however the intersegment heterogeneity increases
or increases first then decrease after the appropriate scale;
and (3) the appropriate scale means there is a compromise
between intrasegment homogeneity and intersegment hetero-
geneity. The research findings are helpful to raise awareness
among practitioners who suffer from scale issues in GeOBIA.
Introduction
With the increase of spatial resolution in remote sensing
images, misclassification, and broken patches become more
of an issue for the pixel-based multispectral image classifica-
tions. Geo-Object-Based Image Analysis ( GeOBIA ) can effec-
tively incorporate spatial information and expert knowledge
into the classification, and the classified image objects are a
useful links to integrate remote sensing and GIS, so GeOBIA has
gained considerable impetus over the last decade (Câmara et
al., 1996; Benz et al., 2004; Castilla and Hay, 2008). It seg-
ments the remote sensing image into homogeneous parcels
(image objects), and then extracts various object features for
classification. In this workflow, fine segmentation, appropri-
ate class features, high quality samples, and a well performing
classifier are all essential requirements of GeOBIA classifica-
tion with high accuracy (Castilla and Hay, 2007; Radoux and
Bogaert, 2014; Wang et al., 2015; Wang and Wang, 2016; Li et
al., 2016). Of all the requirements, the quality of segmentation
directly influences the classification accuracy of GeOBIA (Guo
and Du, 2016).
For a certain multi-scale image segmentation algorithm,
segmentation quality is limited by the dependency of param-
eter selection (Du et al., 2016). Change of scale parameter
usually corresponds to variation in polygon size. Different
polygon size affects reference data collection, sampling, and
analysis, and it affects the accuracy assessment of GeOBIA
(Stehman and Wickham, 2011; Castilla et al., 2014). In the
context of image analysis, scale is defined as the level of ag-
gregation and abstraction at which an object can be clearly
described (Benz et al., 2004). Some attempts have been made
to select the optimal scale parameters for multi-scale segmen-
tation based on multi-scale segmentation evaluation, such as
objective function method (Espindola et al., 2006), ESP tool
(Drǎgut et al., 2010), global score (GS) (Johnson and Xie, 2011),
discrepancy measures or supervised segmentation evaluation
method (Liu et al., 2012; Belgiu and Drǎgut, 2014; Yang et al.,
2015a; Zhang et al., 2015b), spatial statistics method (Ming et
al., 2012 and 2015), selfhood scale learning method (Zhang
and Du, 2016). Of these methods, image segmentation evalu-
ation was paid much attention and it has been a main part of
image classification uncertainty analysis. Earlier study on im-
age segmentation accuracy is mostly for natural digital image
segmentation and very few studies focus on remote sensing
image segmentation (Radoux and Defourny, 2007). With the
development of GeOBIA, remote sensing image segmentation
evaluation has been paid more and more attention, and it has
become the main approach to select the optimal segmentation
scale. According to whether reference polygons or a reference
image is involved, remote sensing image segmentation evalu-
ation can be divided into two classes, supervised image seg-
mentation evaluation (Zhang et al., 2008; Zhang et al., 2015a)
and unsupervised image segmentation evaluation (Espindola
et al., 2006; Belgiu and Drǎgut, 2014).
Supervised image segmentation evaluation usually uses
reference polygons or training objects to compute empirical
goodness or segmentation error to parameterize segmentation
scale (Tian and Chen, 2007; Zhang et al., 2008; Clinton et al.,
2010; Anders et al., 2011; Liu et al., 2012; Belgiu and Drǎgut,
2014; Yang et al., 2015a; Zhang et al., 2015b). The supervised
segmentation evaluation is good at identifying the optimal
scale for the target objects; however, their dependence on
reference data makes them less easy to be used in operational
settings (Belgiu and Drǎgut, 2014). In contrast, since unsu-
pervised image segmentation evaluation need no reference
data and is less subjective and more time-efficient (Belgiu and
Drǎgut, 2014), it is more suitable for practical application and
has become another important research focus in GeOBIA.
Unsupervised image segmentation evaluation considers
sole intrasegment homogeneity (Baatz and Schäpe, 2000; Hay
et al., 2005; Kim et al., 2008; Drǎgut et al., 2010; Drǎgut et al.,
2014) or both intersegment heterogeneity and intrasegment
Dongping Ming, Wen Zhou, Lu Xu, and Yanni Ma are with
the School of Information Engineering, China Universit y of
Geosciences (Beijing), Beijing 100083, China
(mingdp@cugb.edu.cn)
Min Wang is with the Key Laboratory of Virtual Geographic
Environment (Nanjing Normal University), Ministry of
Education, Nanjing, Jiangsu 210023, China.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 11, November 2018, pp. 681–693.
0099-1112/18/681–693
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.11.681
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING November 2018 681