Unsupervised Deep Feature Learning for Urban
Village Detection from High-Resolution Remote
Sensing Images
Yansheng Li, Xin Huang, and Hui Liu
Abstract
Urban villages (
UVs
) are a typical informal settlement in China
resulting from the rapid urbanization in recent decades. Their
formation and demolition are attracting increasing interest.
In the remote sensing community,
UVs
have been detected
based on hand-crafted features. However, the hand-crafted
features just consider one or several characteristics of
UVs
,
and ignore many effective cues hiding in the image. Recently,
deep learning has been used to automatically learn suitable
feature representations from a huge amount of data, without
much expertise or effort in designing features. Motivated by
its great success, this paper aims to use deep learning for
detecting
UVs
. Because of the scarce labeled samples, this
paper presents a novel unsupervised deep learning method
to learn a data-driven feature. Experiments show the data-
driven feature obtained with the proposed method outperform
the existing unsupervised deep neural networks, and achieve
results comparable to that obtained using the best hand-craft-
ed features.
Introduction
As one of the by-products of urbanization, informal settle-
ments (e.g., slums and shanty towns) are common in the de-
veloping cities of many countries (Kuffer
et al
., 2016). While
they have physical similarities to these other informal settle-
ments, urban villages (
UVs
) are a unique product of the urban-
ization of China, and are common in the mega-cities of China.
Different from other regions in urban areas, most spaces in
UVs
are occupied by small buildings, leaving little room for
vegetation, streets, and bare ground. In addition,
UVs
are also
known as “
chengzhongcun
” or “villages in the city” (Chuang,
2010). As a special type of urban settlement,
UVs
result from
the complicated socio-economic development of China.
In the past few decades, large amounts of villages in the
urban fringes have been progressively enveloped by expand-
ing cities due to China’s rapid urbanization. The original
residential areas of these villages are left intact, but the farm-
land is used for urban development (Hao
et al
., 2013). The
original villagers legitimately own the residential areas, but
they are not allowed to expand the land. During this period,
the migration of large numbers of rural workers to cities has
created a great demand for affordable housing, along with the
rapid economic growth (Shen, 1995; Yang, 2000). Driven by
the economic interest, many villagers have built additional
dwellings in their residential areas and then rented them to
migrant workers and the poor. The original villagers have
become landlords and the enveloped villages have become
the so-called
UVs
. However, the development of
UVs
is neither
authorized nor planned. As a consequence,
UVs
suffer from
poor sanitary conditions, absent infrastructure, and various
social problems, including crime and environmental pollu-
tion. Accordingly, the
UVs
are preventing the mega-cities of
China becoming recognized as international modern cities.
Because of the aforementioned problems induced by
UVs
,
many cities of China have decided to dismantle and rede-
velop the
UVs
(Chuang, 2009; Chuang and Zhou, 2011). In
order to guarantee that the
UV
redevelopment policy is fully
implemented, an up-to-date
UV
map is necessary for planners
and policymakers. However, such a
UV
map is often incom-
plete or unavailable. In reality, the identification of
UVs
largely
relies on fieldwork and social investigation, and has rarely
been addressed in the remote sensing community. However,
as the mega-cities of China cover very large areas, timely and
complete detection of
UVs
just using fieldwork is impossible.
In the literature, high-resolution remote sensing images
have been successfully utilized in various applications, such
as urban land-cover classification (Stavrakoudis
et al
., 2011;
Persello and Bruzzone, 2014), object matching and detec-
tion (Ma
et al
., 2015; Cheng and Han, 2016), built-up areas
detection (Li
et al
., 2015a), central business district detection
(Taubenbock
et al
., 2013), private garden detection (Mathieu
et al
., 2007), and so forth. Compared with the detection of
these urban zones, the mapping of
UVs
using high-resolution
remote sensing images suffers from additional difficulties,
such as the large variance of the spectral reflectance and spa-
tiotemporal patterns because of the unplanned development
(Huang
et al
., 2015). In addition, these difficulties also exist in
the detection of informal settlements. In the existing studies,
both object-based and segmentation-based methods (Hofmann
et al
., 2008; Hofmann, 2001; Rhinane
et al
., 2011; Kuffer
et
al
., 2016) have been proposed to identify informal settle-
ments from urban areas using high-resolution remote sensing
images. These approaches are certainly one option that might
help in the mapping of
UVs
, but more intelligent approaches
are also needed. Particularly, the advent of deep learning and
scene-based analysis techniques opens up the possibility of
establishing new
UV
detection approaches.
Yansheng Li is with the School of Remote Sensing and
Information Engineering, Wuhan University, Wuhan 430079,
China (
).
Xin Huang is with the School of Remote Sensing and
Information Engineering, Wuhan University, Wuhan 430079,
China; and the State Key Laboratory of Information Engineering
in Surveying, Mapping and Remote Sensing, Wuhan
University, Wuhan 430079, China.
Hui Liu is with the State Key Laboratory of Information
Engineering in Surveying, Mapping and Remote Sensing,
Wuhan University, Wuhan 430079, China.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 8, August 2017, pp. 567–579.
0099-1112/17/567–579
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.8.567
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2017
567
