October 2019 Layout Flipping Full - page 765

A CNN-Based Subpixel Level DSM Generation
Approach via Single Image Super-Resolution
Yongjun Zhang, Zhi Zheng, Yimin Luo, Yanfeng Zhang, Jun Wu, and Zhiyong Peng
Abstract
Previous work for subpixel level Digital Surface Model (
DSM
)
generation mainly focused on data fusion techniques, which
are extremely limited by the difficulty of multisource data ac-
quisition. Although several
DSM
super resolution (
SR
) methods
have been developed to ease the problem, a new issue that
plenty of
DSM
samples are needed to train the model is raised.
Therefore, considering the original images have vital influence
on its
DSM
’s accuracy, we address the problem by directly im-
proving images resolution. Several
SR
models are refined and
brought into the traditional
DSM
generation process as an im-
age quality improvement stage to construct an easy but effec-
tive workflow for subpixel level
DSM
generation. Experiments
verified the validity and significance of bringing
SR
technology
into this kind of application. Statistical analysis also con-
firmed that a subpixel level
DSM
with higher fidelity can be
obtained more easily compared to directly
DSM
interpolation.
Introduction
With the development of earth observation technologies, we
have entered an era of a remote sensing data explosion (Chi
et
al.
2016; Li
et al.
2018; Ma
et al.
2015), where the requirements
to geographic information products are increasingly higher.
Digital Surface Model (
DSM
), as one of the fundamental digital
representations to the Earth’s surface, is extremely important
for both research and practical applications. And it is widely
used in a variety of applications, such as building detection
(Rottensteiner
et al.
2005), city planning (Chai 2017; Yan
et al.
2015), and environment surveillance (Sadeghi
et al.
2016).
As high-resolution
DSMs
or Digital Elevation Models (
DEMs
)
are always confidential, it is impractic
these Geodata (Xu
et al.
2019). So how
high accuracy and high resolution
DSM
/
DEM
common topic in remote sensing fields (Aguilar
et al.
2014).
In recent years, instead of using high-resolution stereo/
multiview image acquisition equipment (with the existing
low-resolution data to obtain high quality
DSM
/
DEM
results)
seems much more economical and practical. Therefore, recent
work for high-quality
DSM
/
DEM
generation mainly focused on
data fusion techniques (Taud
et al.
1999; Zhou
et al.
2013).
For example, under the framework of sparse representation,
Shen
et al.
(2016) investigates
DEM
generation from contours.
In their proposed method, a lower spatial resolution
DEM
and
sparsely distributed contours of the same geographical area
are integrated to generate a higher resolution
DEM
. However,
in actuality, it is not easy to obtain multisource data, and it
is quite complicated to register the data among various data
sources, since different data sensors focus on different operat-
ing ranges and environmental conditions.
To overcome the problem of acquiring multisource data,
DEM
super resolution, a new concept, has been discussed in
recent years. Motivated by the success of convolution neural
network (
CNN
)-based image super resolution (
SR
), Xu
et al.
(2015) introduced a nonlocal algorithm to improve the resolu-
tion of an original
DEM
based on its partial new measurements
obtained with high resolution. Later, they brought transfer
learning into the task of
DEM
super resolution and designed a
deep gradient prior network (Xu
et al.
2019). Using this kind
of method can avoid the problem of multisource data acquisi-
tion, while a new issue that model training needs abundant
DEM
samples is raised.
When it is impractical to obtain multisource data or enough
DSM
samples, how can we obtain a high quality and high-reso-
lution
DSM
with low-resolution image data? The main purpose
of this paper is to explore a simple but effective solution for
this problem. Specifically, it takes a
CNN
-based
SR
method
into traditional
DSM
generation process to improve image
resolution, and then followed by a pixel level
DSM
generation
process to achieve the goal of subpixel level
DSM
generation.
SR
has been widely used in vision-related tasks. Dai
et al.
(2016) have verified the effectiveness and usefulness of
SR
in
some vision applications, such as edge detection, semantic
segmentation, digit, and scene recognition. Besides, dense
image matching, a kind of method that directly operates on
ages, is an advisable and mainstream
DSM
g
eneration. And it is obvious that images
r to acquire compared to the aforemen-
tioned multisource data or
DEM
samples. Therefore, the pro-
posed method is much simpler and more practical compared
to the aforementioned two kinds of methods in the task of
generating high quality and subpixel
DSM
. The main contribu-
tions of this paper are as follows:
1. A simple but effective method for generating high quality
subpixel
DSM
is proposed, which can work well without
the demand of multisource data or abundant
DSM
samples.
2. As there is not a similar work that has been done before,
some existing
SR
methods are refined and used in this
paper to verify the feasibility of the proposed method. By
comparing the experimental results, we identify the devel-
opment trends of this kind of applications.
Specifically, to validate the effectiveness and usefulness of the
proposed method, two aerospace image datasets from World
Yongjun Zhang and Zhi Zheng are with the School of Remote
Sensing and Information Engineering, Wuhan University,
Wuhan, 430079, China (
).
Yimin Luo is with the Division of Imaging Sciences and
Biomedical Engineering Research, King’s College London,
London SE1 7EH, U.K.
Yanfeng Zhang is with the Riemann Laboratory, Huawei
Technologies Co., Ltd. Wuhan 430079, China.
Jun Wu and Zhiyong Peng are with the School of Electronic
Engineering and Automation, Guilin University of Electronic
Technology, Guilin, 541004, China.
Photogrammetric Engineering & Remote Sensing
Vol. 85, No. 10, October 2019, pp. 765–775.
0099-1112/19/765–775
© 2019 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.85.10.765
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2019
765
699...,755,756,757,758,759,760,761,762,763,764 766,767,768,769,770,771,772,773,774,775,...778
Powered by FlippingBook