Tea Garden Detection from High-Resolution
Imagery Using a Scene-Based Framework
Xin Huang, Zerun Zhu, Yansheng Li, and Bo Wu, and Michael Yang
Abstract
Tea cultivation has a long history in China, and it is one of
the pillar industries of the Chinese agricultural economy.
It is therefore necessary to map tea gardens for their ongo-
ing management. However, the previous studies have relied
on fieldwork to achieve this task, which is time-consuming.
In this paper, we propose a framework to map tea gardens
using high-resolution remotely sensed imagery, including
three scene-based methods: the bag-of-visual-words (
BOVW
)
model, supervised latent Dirichlet allocation (
sLDA
), and the
unsupervised convolutional neural network (
UCNN
). These
methods can develop direct and holistic semantic representa-
tions for tea garden scenes composed of multiple sub-objects,
thus they are more suitable than the traditional pixel-based
or object-based methods, which focus on the local charac-
teristics of pixels or objects. In the experiments undertaken
in this study, the three different methods were tested on four
datasets from Longyan (Oolong tea), Hangzhou (Longjing
tea), and Puer (Puer tea). All the methods achieved a good
performance, both quantitatively and visually, and the
UCNN
outperformed the other methods. Moreover, it was found that
the addition of textural features improved the accuracy of
the
BOVW
and
sLDA
models, but had no effect on the
UCNN
.
Introduction
Tea is one of the most famous beverages in the world, and
the consumption of tea is growing faster than that of coffee
and cocoa (Cabrera
et al.
, 2006). China was not only the first
country to cultivate tea, but it is also one of the main produc-
ing countries (Dutta
et al.
, 2010). The cultivation and produc-
tion of tea plays an important part in Chinese agriculture,
and has a significant impact on the economic development
of rural areas. Tea is produced in most provinces of southern
China and is the major cash crop for many villages and towns.
Thus, it is necessary to monitor and assess the tea gardens.
However, this task is usually achieved by fieldwork, which is
labor- and time-intensive. Meanwhile, remote sensing tech-
nology has been widely employed to detect and map crops,
such as sugarcane (Vieira
et al.
, 2012), paddy rice (Qin
et al.
,
2015), and coca (Pesaresi, 2008). Although it is an effective
and convenient tool, few studies have been so far carried out
for tea garden mapping using remotely sensed imagery.
The tea plant, an evergreen bush, can be easily confused
with other woody vegetation in spectral characteristics.
However, tea gardens exhibit unique textural features, which
can be used to distinguish them from other vegetation, due to
the unique form of tea cultivation. In high-resolution optical
imagery, tea gardens are composed of multiple object types
with a certain spatial and structural pattern: (1) tea bushes
are generally planted in rows, showing obvious gaps between
the rows; and (2) tea gardens contain not only the tea plants,
but also bare soil, individual trees, and shadows (see Figure
1). In this regard, a tea garden can be considered as a se-
mantic scene consisting of multiple interrelated sub-objects,
rather than a single land-cover type. Because the pixel-based
(Damodaran
et al.
, 2017; Huang
et al.
, 2014) or object-based
(Ming
et al.
, 2016; Zhu
et al.
, 2017) image analysis methods,
modeling the scene in a bottom-up manner, have difficulty in
obtaining the holistic semantic representation of a scene, they
are not suitable for detecting tea garden scenes with complex
sub-categories and spatial patterns. On the other hand, scene-
based analysis techniques have been proved to be a more
productive approach for the interpretation of high-resolution
remotely sensed imagery (Cheriyadat, 2014; Huang
et al.
,
2015). Therefore, in this study, we propose a series of scene-
based interpretation models with spectral-textural features to
detect tea gardens from high-resolution imagery.
In recent years, various scene classification approaches
have been proposed. One particularly efficient method, the
bag-of-visual-words (
BOVW
) model (Sivic and Zisserman,
2003), has been widely used for remote sensing semantic clas-
sification (Cheriyadat, 2014; Sheng
et al.
, 2012). In the classic
BOVW
model, an image is represented by a set of visual words,
which are generated by clustering the local patches. Subse-
quently, topic models, such as probabilistic latent semantic
analysis (
pLSA
) (Hofmann, 2001) and latent Dirichlet alloca-
tion (
LDA
) (Blei
et al.
, 2003), have been adopted to extract
the latent semantic topic features from the
BOVW
representa-
tion and classify the scenes with the semantic topic features.
The
pLSA
model uses a probabilistic approach to model an
image, representing the frequency of the visual words as a
finite mixture of an intermediate set of topics.
LDA
improves
on
pLSA
by introducing a Dirichlet distribution into the topic
mixture, thus overcoming the overfitting problem of
pLSA
.
More recently, the
sLDA
model (Jon and David, 2008), which
extends
LDA
by adding a response variable to indicate the
class of the scenes in a generative process, has been proposed
and successfully utilized in image annotation and satellite
Xin Huang is with the School of Remote Sensing 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 (
).
Zerun Zhu is with the State Key Laboratory of Information
Engineering in Surveying, Mapping and Remote Sensing,
Wuhan University, Wuhan 430079, China.
Yansheng Li is with the School of Remote Sensing and
Information Engineering, Wuhan University, Wuhan 430079,
China.
Bo Wu is with the School of Geography and Environment,
Jiangxi Normal University, Nanchang 330022, China.
Michael Yang is with the Faculty of Geo-Information Science
and Earth Observation (ITC), University of Twente, Enschede
7500 AE, The Netherlands.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 11, November 2018, pp. 723–731.
0099-1112/18/723–731
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.11.723
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
November 2018
723
