Review on High Spatial Resolution
Remote Sensing Image Segmentation Evaluation
Yangyang Chen, Dongping Ming, Lu Zhao, Beiru Lv, Keqi Zhou and Yuanzhao Qing
Abstract
Image segmentation is a key technique involved in informa-
tion extraction from high spatial resolution remote sensing
images. Studying the impact of the evaluation method on
the segmentation result is equally as important as study-
ing the segmentation algorithm itself. However, research
in segmentation evaluation is behind that of segmentation
algorithms. Only a few review articles about segmentation
evaluation were published in computer vision field. There-
fore, reviewing segmentation evaluation methods used for
high spatial resolution remote sensing images is of great
significance. This paper summarizes widely used evaluation
methods in remote sensing field, analyzes their advantages
and shortcomings, and discusses their application range.
Especially this paper uses series of experiments to demon-
strate the supervised and unsupervised image segmenta-
tion evaluation process and analyzes the performance of
some commonly used supervised and unsupervised evalu-
ation indexes. Further, potential applications and possible
future direction for high spatial resolution remote sensing
image segmentation evaluation are finally summarized.
Introduction
Recently, with the rapid development of remote sensing tech-
niques, various high spatial resolution remote sensing images
can be conveniently and efficiently acquired by satellites or
aircraft. High spatial resolution remote sensing images can
now be used for land resource management, urban planning,
traffic planning, natural disaster monitoring, and military
target recognition, and more. However, compared to the prog-
ress that remote sensing imaging techniques has made, image
processing and analyzing techniques have developed more
slowly. Thus, the potential of high spatial resolution remote
sensing imaging has not yet been fully unleashed. Therefore,
how to extract and efficiently analyze information from high
spatial resolution remote sensing images is currently the focal
point of the field of remote sensing and will be in the future.
High spatial resolution remote sensing images contain a
massive amount of detailed spatial information compared
to low and medium resolution images, which make spatial
analysis and information extraction much more difficult.
Traditional pixel-based analysis methods can only extract and
use spectral statistics from a single pixel, while ignoring the
spatial information between targets (Blaschke, 2001). Thus,
using a pixel-based analysis method for high spatial resolu-
tion remote sensing images will lead to misclassification of
ground features and cause the salt-pepper effect (Ming
et al
.
2015; Yu
et al
. 2006).
GeOBIA
(Geographic Object-Based Image
Analysis) method has become increasingly popular in the re-
mote sensing field (Blaschke 2010). Using the
GeOBIA
method
for high spatial resolution remote sensing image analysis is far
superior to the traditional pixel-based analysis method, both
for the results and precision (Blaschke
et al
. 2014), because it
can take full advantage of the richer size, shape, texture, topo-
logical information and expert knowledge into consideration
for classification (Ming
et al
., 2012).
GeOBIA
works by using image segmentation as the key
technique for image information extraction, by partitioning
the image into segmented objects based on intra-homogeneity
and inter-heterogeneity criteria (Haralick and Shapiro, 1992;
Pekkarinen, 2002). It also represents the image as discrete seg-
mented objects which better conforms to human visual per-
ception and understanding (Goodchild
et al
. 2007). With spa-
tial resolution of 5.0 m and finer (Wulder
et al
., 2004), pixels
in traditional pixel-based analysis methods can be replaced
by segmented objects in subsequent analysis and processing
(Benz
et al
. 2004; Zhang
et al
. 2015c; Zhang 1996a), such as
image classification, target extraction, etc. Feature extraction
and object expression, based on image segmentation, can
provide the ability to transform the original image into a more
abstract and compact form, which makes high-level image
analysis and understanding possible (Ming
et al
., 2006).
So far, scholars have proposed a huge number of segmen-
tation algorithms specific to high spatial resolution remote
sensing images. Unlike segmentation in the field of computer
vision, in remote sensing field, not only the suitable segmen-
tation algorithm needs to be chosen, but also one or more
scale parameters. Comparing different segmentation algo-
rithms, or the segmentation result based on different combi-
nations of scale parameters, is one of the current difficulties
that directly affects the result and accuracy of the subsequent
process for the realization of high spatial resolution remote
sensing images (Dorren
et al
., 2003; Estrada and Jepson, 2005;
Palus and Kotyczka, 2001; Yang
et al
., 2015b). Hence, pro-
posing an evaluation system and evaluation method for the
segmentation results of high spatial resolution remote sensing
images is significant.
Research about segmentation evaluation is far behind that
of segmentation algorithms, only a few review articles about
segmentation evaluation were published. Zhang (1996b)
firstly published a review article about segmentation evalua-
tion methods in 1996, inventively established an evaluation
system for image segmentation. However, it was published 20
years ago and some of the summarized methods are already
dated. Besides, segmentation evaluation was mainly used for
proving the superiority of segmentation algorithms quanti-
tatively at the time, which was very different from current
application in remote sensing field. After that, Zhang
et al
.
(2008) summarized the segmentation evaluation method
based on current research, but mainly focused on unsuper-
vised evaluation methods in the domain of computer vision.
More attention about segmentation evaluation were paid
to tradition computer vision than remote sensing field,
School of Information Engineering, China University of
Geosciences (Beijing), 29 Xueyuan Road, Haidian, Beijing,
100083, China. (
).
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 10, October 2018, pp. 629–646.
0099-1112/18/629–646
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.10.629
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2018
629
