PE&RS February 2016 - page 149

A Region-Line Primitive Association Framework
for Object-Based Remote Sensing Image Analysis
Min Wang and Jie Wang
Abstract
In this study, we propose a novel region-line primitive association
framework (
RLPAF
) for
OBIA
. In this framework, segments (region
primitive) and straight lines (line primitive) are obtained by im-
age segmentation and straight line detection, respectively, before
their corresponding intra-primitive features are extracted. An as-
sociation model is built on inter-primitive topology and direction
relationships. Several region-line collaborative features are also
derived. Image analysis is then performed based on both region
and line primitives. The advantage of
RLPAF
is the collaborative
utilization of complementary information between regions and
lines throughout the entire
OBIA
process: from image segmenta-
tion, to feature extraction, and finally, object recognition. To
validate this framework,
RLPAF
is applied on road network extrac-
tion from high spatial resolution (
HSR
) remote sensing images.
Experiments show that the proposed framework and methods re-
fine primitive shape and spatial relationship analyses, as well as
obtain higher method accuracy, than
OBIAs
based on only regions.
Introduction
Object-based image analysis (
OBIA
) has become a routine
technique for extracting information from
HSR
images. Repre-
sentative commercial
OBIA
software systems include Trimble
Worldwide’s eCognition
and the
ENVI
Feature Extraction
module. Compared with traditional pixel-based image analysis
(
PBIA
), the minimum analyzing units in
OBIA
are generally re-
gions (or segments from the viewpoint of image segmentation)
that are composed of mutually related pixels. In this study, we
mainly use the term “region” because the analyzing units in
the proposed technical framework can also be obtained using
other methods aside from image segmentation, e.g., by utiliz-
ing geographic information system data. Compared with
PBIA
,
OBIA
utilizes more abundant features (e.g., texture, shape, and
spatial relationship) among objects. In addition, it facilitates
the fusion of knowledge rules in image processing and analy-
sis. Thus,
OBIA
may be a better choice than
PBIA
for extracting
information from
HSR
images (Benz
et al.
, 2001; Benz
et al.
,
2004; Blaschke, 2010; Blaschke
et al
., 2014).
Image segmentation plays a critical role in
OBIA
because
high-quality segments serve as the foundation for succeeding
analyses. Meinel and Neubert (2004) compared segmentation
methods used in several
OBIA
software systems and found
that the best among them is multi-resolution segmentation
(
MRS
) (Baatz and Shäpe, 2000). Moreover, many studies have
been conducted to improve the performance of
MRS
. Typical
schemes include applying automatic approaches to a best-
scale selection (
Dr
ǎ
gu
ţ
et al.
, 2010 and 2014; Tullis
et al.
, 2010;
Tong
et al.
, 2012) or using several auxiliary data sources,
including geographic information system databases, digital
maps, and LiDAR data (Smith and Morton, 2010; Anders
et
al.
, 2011). Furthermore, different kinds of segmentation meth-
ods (e.g., combining edge and region information) have also
been suggested (Kermad and Chehdi, 2002; Li
et al.
, 2010; Lin
and Chen, 2010; Chen
et al.
, 2012; Yu
et al.
, 2012). Thus, we
proposed the hard-boundary constrained image segmentation
(
HBC-SEG
) method, which exhibited many advantages over
MRS
, particularly in region boundary precision (Wang and Li,
2014). We further reduced the over-segmentation errors of
HBC-SEG
through a novel collinear and ipsilateral neighbor-
hood (
IPSL
-neighborhood) model based on region and straight
line relationship modeling (Wang
et al.
, 2015).
Abundant features can be obtained based on regions after
image segmentation. Commonly used features include region
spectra, shapes, textures, and spatial relationships at the same
scale or at different scales. Based on these features and flexible
classification rules,
OBIA
can potentially achieve more accurate
classification than traditional
PBIA
for
HSR
images.
OBIA
gener-
ally uses supervised or rule-based classification schemes to
classify images. Common tools include nearest neighbors, sup-
port vector machines (
SVM
s), and fuzzy rule-based classifiers.
Numerous studies on
OBIA
applications have been con-
ducted based on the aforementioned technical framework.
Typical cases include image classification (Gao
et al.
, 2011;
Laliberte
et al.
, 2012; Salehi
et al.
, 2012; Du
et al.
, 2013; Rasi
et al.
, 2013), thematic information extraction or object recog-
nition (Walter 2004; Huang and Zhang, 2008; Hu and Weng,
2011; Sebari and He, 2013; Benarchid and Raissouni, 2013;
d’Oleire-Oltmanns
et al.
, 2014), and change detection (Johan-
sen
et al.
, 2010; Lu
et al.
, 2011; Hebel
et al.
, 2013). Man-made
object classification or recognition is common when
HSR
images are used. This phenomenon is jointly decided by the
booming of
HSR
images, which changes application require-
ments, and the technical features of
OBIA
, which are suitable
for such kinds of application. High-quality image segmenta-
tion, robust classification rules, and minimum algorithm
parameter dependency are important issues that should be
solved to promote the practicability of these methods.
Despite their diverse technical details, common
OBIA
ap-
plications generally follow the “segment and then classify”
framework. In this framework, the minimum analyzing unit
(region) is determined after an image is segmented. Then,
subsequent analyses are implemented in regions. Several
OBIA
studies have combined image segmentation and classification
in flexible schemes, which weakens the role of initial segmen-
tation. For example, in Tiede
et al
. (2010 and 2011), segmen-
tations were also tailored at a later stage for specific classes or
regions in the image when required in the classification pro-
cess. However, the region-based analysis framework has been
completely adopted in the aforementioned studies. Although
Key Laboratory of Virtual Geographic Environment (Nanjing
Normal University), Ministry of Education, Nanjing, Jiangsu,
P.R. China, 210023; and the Jiangsu Center for Collaborative
Innovation in Geographical Information Resource Develop-
ment and Application, Nanjing, Jiangsu, P.R. China, 210023
(
;
).
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 2, February 2016, pp. 149–159.
0099-1112/16/149–159
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.82.2.149
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2016
149
71...,139,140,141,142,143,144,145,146,147,148 150,151,152,153,154,155,156,157,158,159,...171
Powered by FlippingBook