REMOTE SENSING IMAGE SEGMENTATION BASED ON WILCOXON RANK SUM TEST
AND MEAN ABSOLUTE DEVIATION
Libao Zhang*, Shuang Wang, Qiaoyue Sun, Aoxue Li
College of Information and Technology, Beijing Normal University
ABSTRACT
In this paper, a novel threshold segmentation method for
remote sensing images is proposed. The proposed method is
based on Wilcoxon rank sum test and mean absolute
deviation (MAD) model with color feature and can segment
roads and residential areas from vegetation more accurately.
Three steps are used to realize the new method. First, we use
blue and green color components as paired sample on
Wilcoxon rank sum test to partition the vegetation. Second,
a road-residential area map is constructed by mean absolute
deviation on an improved two dimensional histogram to get
the optimal threshold for segmentation. Finally, we fuse
vegetation and residential map to get the final segmentation
result. Compared with several existing algorithms, the
proposed method presents a more accurate segmentation.
Index Terms—Remote sensing, image segmentation,
Wilcoxon rank sum test, mean absolute deviation
1. INTRODUCTION
Image segmentation, as a critical step for image analysis and
understanding, has been introduced into the remote sensing
image analysis field and has become an important technical
approach for improving analysis accuracy in image
processing [1]. In order to get accurate segmentation results,
many different types of algorithms have been proposed.
Among the algorithms, thresholding is the most simple and
effective one [2]. Different types of thresholding have been
proposed in the literatures [3-6]. In [3, 4], the maximum
between class variance criteria was proposed. Hou et al [6]
proposed minimum class variance thresholding (MCVT)
using the relative distance and average distance. In addition,
a median minimum error thresholding (M-MET) was
proposed by Xue with expanding maximum log-likelihood
estimation in [4]. These algorithms were all based on the
one-dimensional (1D) intensity histograms of images.
However, in practice, influenced by noise and other
interference factors, there are no significant peaks and
troughs in the 1D intensity histograms, so it is necessary to
consider the spatial information to construct a two-
dimensional (2D) histogram for threshold selection.
However, 2D algorithms give rise to the exponential
increment of computation time. Therefore, by using iterative
process, Chen et al. [7] derived an improved 2D Otsu fast
algorithm. Besides, a fast algorithm that employed MCVT
into a two-dimensional intensity histogram (2D MCVT) was
proposed by Nie et al [8]. Although these results of the
above algorithms improved, there are still some drawbacks.
The segmentation results are inaccurate because the color
features of different objects were ignored and only gray
information was taken into account. For example, in fig.
1(b), the vegetation should be black; however, the
vegetation is mistakenly regarded as residential areas
because the color feature is not considered.
Fig.1 Example of mistaken segmentation (a) Origin remote sensing
image, (b) Enlarged piece of (a), (d) Segmentation results of
remote sensing image by 2D MCVT, (c) Enlarged piece of (d).
In addition, the algorithms above were proposed
mainly based on Gaussian distribution histograms, when
the histograms deviates Gaussian distribution, the
performance of the segmentation will be not satisfying [5].
In this paper, we proposed a novel remote sensing
image threshold segmentation method based on Wilcoxon
rank sum test and mean absolute deviation (MAD) model
with color features. Taking color features into consideration,
for vegetation map, we use Wilcoxon rank sum test on green
and blue components to partition vegetation. For road &
residential area map, by using MAD model on the 1D
histogram which is transformed from 2D I-G histogram via
diagonal projection, we get optimal threshold based on class
mean absolute deviations. Experiment results show that this
algorithm has more accurate segmentation results for images
characterized by Gaussian distribution histograms.
2. PROPOSED METHOD
2.1. Vegetation map
Wilcoxon rank sum test is a non parametric statistical test
method to compare the difference between location
parameters of two populations. Wilcoxon rank sum test is
insensitive with the distribution of samples but sensitive
6340978-1-5090-3332-4/16/$31.00 ©2016 IEEE IGARSS 2016