Estimation of Forest Biomass Using Multivariate
Relevance Vector Regression
Alireza Sharifi, Jalal Amini, and Ryutaro Tateishi
Abstract
The objective of this study is to develop a method based on
multivariate relevance vector regression (
MVRVR
) as a kernel-
based Bayesian model for the estimation of above-ground
biomass (
AGB
) in the Hyrcanian forests of Iran. Field
AGB
data
from the Hyrcanian forests and multi-temporal
PALSAR
back-
scatter values are used for training and testing the methods.
The results of the
MVRVR
method are then compared with other
methods: multivariate linear regression (
MLR
), multilayer per-
ceptron neural network (
MLPNN
), and support vector regression
(
SVR
). The
MLR
model showed lower values of R2 than the three
other approaches. Although the
SVR
model was found to be
more accurate than
MLPNN
, it had the lowest saturation point of
224.75 Mg/ha. The use of
MVRVR
model significantly improves
the estimation of
AGB
(R
2
= 0.90;
RMSE
= 32.05 Mg/ha), and the
model showed a superior performance in estimating
AGB
with
the highest saturation point (297.81 Mg/ha).
Introduction
Biomass is the total mass of living matter within a given unit
of an environmental area. It is thus a measure of the carbon
stock of an ecosystem, and the importance of estimating the
amount of biomass has been reported in a number of stud-
ies (Heimann and Reichstein, 2008; Le Quere
et al.
, 2009;
Peregon and Yamagata, 2013; Simard
et al.
, 2006; Tanasea
et
al.
, 2014). In this respect, biomass mapping is an important
practical tool for use in forest management, and in particular
for use in forest monitoring and making assessments of defor-
estation processes. The most accurate way of
AGB
retrieval are
forest inventories which use eld-based measurements (e.g.,
tree height, diameter at breast height) to calculate the biomass
on the basis of allometric equations (Chave
et al.
, 2005; Lu,
2006). Satellite observations can also be used, and although
they have the advantage of enabling monitoring of large and
remote areas, their measurements are less accurate (Lu, 2006).
Polarimetric Synthetic Aperture Radar (Pol
SAR
) provides
information on changes in the polarization state of electro-
magnetic waves reflected from the earth surface that can be
exploited to extract information for identification and clas-
sification of different natural features, as each polarization is
sensitive to different surface characteristics and properties.
In forestry, one of the most important applications of Pol
SAR
data processing is its use in biomass estimation (Carreiras
et
al.
, 2012; Englhart
et al.
, 2012; Huang
et al.
, 2009; Lucas
et al.
,
2010; Minh
et al.
, 2014; Ormsby
et al.
, 1985; Sandberg
et al.
,
2011; Soja
et al.
, 2013; Ticehurst
et al.
, 2004). Previous studies
have indicated that radar backscatter of Pol
SAR
data at lower
frequencies has a positive correlation with the
AGB
of forests,
especially in the cross-polarized Horizontal-Vertical (
HV
)
backscatter (Englhart
et al.
, 2012; Dobson
et al.
, 1992; Imhoff
and Gesch, 1990; Imhoff, 1995; Fransson and Israelsson, 1999;
Santos
et al.
, 2002; Saatchi
et al.
, 2007; Sader, 1987; Sartori
et
al.
, 2011; Lucas
et al.
, 2006). These studies have shown that
when used in conjunction with a suitable type of polarization
and a robust estimation model, multi-temporal
SAR
data re-
corded during the dry season can provide an improvement in
the accuracy of
AGB
estimations (Hame
et al.
, 2013; Townsend,
2001). Thus, in order to find a mapping function between
AGB
of forests and
SAR
backscatters, linear regression methods
(Kasischke and Bourgeau-Chavez, 1997; Mitchard
et al.
, 2009;
Watanabe
et al.
, 2006), arti cial neural networks (
ANN
) (Amini
and Sumantyo, 2009; Del Frate and Solimini, 2004), and sup-
port vector regression (
SVR
) (Camps-Valls
et al.
, 2006; Englhart
et al.
, 2012; Monnet
et al.
, 2011) have all been proposed.
Although multivariate linear regression (
MLR
) is a compara-
tively comprehensible model, some studies (Mitchard
et al.
,
2009; Watanabe
et al.
, 2006) have indicated that the results of
MLR
model may have been underestimated because of its lin-
ear nature and improper target results. In addition, the multi-
layer perceptron neural networks (
MLPNN
) model proposed
for
AGB
estimation by Amini and Sumantyo (2009) tended to
overfit data when few observations were used in the training
data set. Wang
et al.
(2008) reported that deep neural nets
with a large number of observations are very powerful ma-
chine learning systems; overfitting can be a serious problem
in such networks when only a few observations are included.
In addition, large networks are slow to use, making it difficult
to deal with overfitting by combining the predictions of many
different large neural nets at test time. Englhart
et al.
(2012)
reported that the
SVR
model is the best machine learning
method for biomass estimation, but the results have not yet
been acceptable at high biomass levels (Amazon rainforest,
Pacific temperate rainforest, Congo rainforest, and Sinharaja
forest are well-known forests which have high biomass level)
where radar backscatters eventually reach saturation level.
The introduction of a new technique in relation to a model
that used fewer observations would be very useful in mak-
ing forest
AGB
estimations, as it would deliver the advantage
of being less time-consuming and would simplify the need
for ground data collection. Therefore, in consideration of the
limitations of the regression models mentioned above, we
propose a Bayesian kernel-based regression model known
as multivariate relevance vector regression (
MVRVR
) (Thay-
ananthan
et al.
, 2006), which is a multivariate version of the
relevance vector machine proposed by Tipping (2001). This
approach utilizes a number of relevance vectors (scarcity con-
cept) to synthesize the output values used for image analysis,
and in addition, uses a limited amount of ground data (Pal
and Foody, 2012; Sharifi and Amini, 2015).
Alireza Sharifi and Jalal Amini are with the Department of
Remote Sensing, Faculty of Survey Engineering and Spatial
Information, University of Tehran (
).
Ryutaro Tateishi is with the Center of Environmental Remote
Sensing, Chiba University, Chiba, Japan.
Photogrammetric Engineering & Remote Sensing
Vol. 82, No. 1, January 2016, pp. 41–49.
0099-1112/16/41–49
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.1.41
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2016
41
