PE&RS August 2017 Public - page 553

Retrieving Grassland Canopy Water Content
by Considering the Information from
Neighboring Pixels
Binbin He, Xingwen Quan, Dasong Xu, Changming Yin, Zhanmang Liao, Shi Qiu, Jinsong Ge, and Zhijun Zhang
Abstract
Accurate and robust retrieval of grassland canopy water con-
tent (CWC ) using a radiative transfer model (RTM) is generally
affected by the ill-posed inversion problem due to the lack of
enough available a priori information. To alleviate this prob-
lem when inversing the RTM, a two-step inversion method was
proposed. The key point of this method was to simultaneously
consider the spectral information from neighboring pixels and
the spatial dependency among these pixels, with the pur-
pose to win more information from these neighboring pixels.
The proposed methodology was then applied to retrieve CWC
using the PROSAIL RTM from Landsat-8 OLI data for a plateau
grassland in China. The results showed that the estimated
CWC using the proposed method (RMSE = 67.31 g m-2 and R2 =
0.81) was better than that from the traditional method (RMSE =
80.11 g m-2 and R 2 = 0.78) which only considered the informa-
tion of single pixel.
Introduction
Grassland ecosystem is one of the largest proportions of
terrestrial ecosystem in China, which possesses an area of ap-
proximately 400 million hectares, accounting for 41.7 percent
of the national territory area (Jin et al., 2014). Canopy water
content (CWC), defined as the mass of water per unit ground
area, is an important factor in monitoring grass growth state
and drought assessment (Ben-Gal et al., 2009; Casas et al.,
2014; Ceccato et al., 2002; Ceccato et al. , 2002; Chávez et al. ,
2013; Quan et al., 2015a; Trombetti et al., 2008; Yilmaz et al. ,
2008a; Yilmaz et al., 2008b;). Normally, this variable can also
be expressed as (Cheng et al., 2013; Cheng et al., 2014; Trom-
betti et al., 2008; Yebra et al., 2013),
CWC = LAI × Cw (1)
where LAI is leaf area index and Cw represents the leaf equiva-
lent water thickness. Water stress changes the spectral reflec-
tance and transmittance of leaves, reduces leaf area and alters
architecture, and therefore significantly influences the canopy
spectral response (Cohen, 1991). Hence, remote sensing of
canopy spectral properties can provide a direct way to esti-
mate and comprehensively monitor the grassland CWC varia-
tion in near-real time and large scale. Furthermore, to date
the remote sensing technique is the unique way to that end
due to its high temporal and spatial resolution image of large
landscape observation (Huang et al., 2015; Quan et al. , 2015a;
Riaño et al., 2005; Trombetti et al., 2008; Yilmaz et al., 2008a).
Remote sensing techniques used for estimating vegetation
canopy variables (including CWC) have either been based on
statistical approaches or on the inversion of a radiative trans-
fer model. The formers are simple and computationally effi-
cient (Barraza et al., 2014; Chávez et al., 2013; Colombo et al. ,
2003; Darvishzadeh et al., 2008b; Ingram et al., 2005; Verrelst
et al., 2016). However, such statistical relationships require
extensive field data and are applied to the specific vegetation
type, development stage, and study site, which are only valid
for a given sensor spectral configuration and acquisition geom-
etry (He et al., 2013; Huang et al., 2016; Houborg et al., 2007;
Houborg and Boegh, 2008; Houborg et al. , 2009; Laurent et al.,
2013; Ustin et al. , 2009). As a result, a fundamental problem
with the statistical approaches is their lack of generality. On
the other hand, the model-based inversion approaches, such
as the radiative transfer models (RTMs)-based methods, rely
on radiative transfer mechanism among canopy properties and
spectra, assuming that the RTMs describe the spectral varia-
tion in the canopy reflectance, as a function of canopy, leaf,
and soil background characteristics based on physical laws
(Darvishzadeh et al. , 2008a; Huemmrich, 2001; Meroni et al.,
2004; Quan et al., 2015a; Quan et al., 2016; Verhoef, 1984; Ver-
hoef, 1985). Thus, this technique has the advantage of repro-
ducibility and generality over empirical statistical approaches.
However, the drawback for the physical model-based models
is the ill-posed nature of model inversion (Atzberger, 2004;
Combal et al., 2002), which is mainly caused by the under-
determination nature of the modeling schemes (Jacquemoud
et al., 1995). First, the numbers of unknown variables are
generally more than the ones of independent radiometric vari-
ables remotely sampled by the sensors (Baret and Buis, 2008).
Second, different parameter combinations may yield almost
identical spectral signatures, which lead the various model
parameters may counterbalance with each other (Atzberger,
2004; Darvishzadeh et al., 2008b). The problem is further am-
plified because neither the RTM nor the reflectance measure-
ments are error-free (Baret and Buis, 2008), which may result
in unstable and inaccurate inversion performances when no
regularization is applied (Durbha et al., 2007).
Binbin He is with the School of Resources and Environment,
University of Electronic Science and Technology of China,
Chengdu, Sichuan 611731, China; and the Center for
Information Geoscience, University of Electronic Science and
Technology of China, Chengdu, Sichuan, China
(binbinhe@uestc.edu.cn).
Xingwen Quan, Dasong Xu, Changming Yin, Zhanmang
Liao, and Shi Qiu are with the School of Resources
and Environment, University of Electronic Science and
Technology of China, Chengdu, Sichuan 611731, China.
Jinsong Ge and Zhijun Zhang are with the Qinghai Remote
Sensing Monitoring Center for Ecology and Environment,
Xining, Qinghai 810007, China.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 8, August 2017, pp. 553–565.
0099-1112/17/553–565
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.8.553
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING August 2017 553