PE&RS January 2016 - page 42

The objective of this paper therefore, is to develop a meth-
od based on multivariate relevance vector regression ( MVRVR)
as a kernel-based Bayesian model for above-ground biomass
(AGB) estimation, and then to compare the results with those
of three methods, MLR, MLPNN, and SVR . In this study, we use
ALOS PALSAR Fine Beam Dual (FBD) imagery with HH and HV
polarizations acquired in 2007, 2008, and 2010. The study
area is that of the western Hyrcanian forests of Iran (south
west of the Caspian Sea). Field AGB data relating to 110 plots
are provided from 2007. The paper is organized as follows.
The next Section provides details of the materials and meth-
ods used in the research, including a brief description of the
study area, field measurements, ALOS PALSAR data, modeling,
and validation of the models; followed by the results and
corresponding discussion, respectively. The final Section
presents the conclusions.
Materials and Methods
Study Area
The study area is located within the Hyrcanian forests of Iran
[Plate 1), which are situated southwest of the Caspian Sea
between 37.3° to 37.65° N and 48.7° to 49.1° E (upper panel
in Plate 1), in the temperate zone. The forests grow at eleva-
tions ranging from 600 m to 950 m above mean sea level,
and the region is one of the rainiest areas in Iran. As such, it
is a suitable habitat for broadleaved trees, and is home to four
main tree species: Fagus orientalis, Alnus serrulata, Carpinus
betulus, and Ulmus glabra.
Field Measurements
In the summer of 2007, field work was undertaken in the
study area. This consisted of measuring tree height and diam-
eter at breast height (DBH) in 110 plots. The DBH was measured
using a simple tape scale, known as a diameter tape, at a
height of 1.3 m above the ground. Tree heights were calculat-
ed using a hypsometer based on the geometric method (West,
2009). The geographic coordinates of each plot were obtained
using a Global Positioning System (GPS), which enabled loca-
tion of the plots (errors up to 3 to 5 m) (lower panel in Plate
1(lower panel).
The plots each measured 50 m × 50 m, and contained four
dominant species: Fagus orientalis, Alnus serrulata, Carpinus
betulus, and Ulmus glabra. The number of species in each plot
was measured to achieve a desired precision level. Table 1
summarizes the field measurements and resulting calculations.
Table 1. Height, Diameter At Breast Height (DBH), And Above-Ground
Biomass (AGB) For Each Forest T ype W ith Minimum (Min), Maximum (Max),
M ean, And S tandard Deviation (Std.) Of AGB Values
Dominant
Species
# of
plots
Mean
height
(m)
Mean
DBH
(cm)
Mean AGB (Mg.ha-1)
Fagus orientalis 39 27 58 39.2 324.3 128.7 53.6
Alnus serrulata 28 22 46 16.9 288.5 93.8 31.8
Carpinus betulus 23 21 45 27.1 307.9 102.4 42.3
Ulmus glabra 20 19 35 17.2 210.6 74.4 22.7
Sharifi et al. (2015) reported that using the Chave model
with calibrated coefficients delivers better results than other
generalized allometric models in relation to the Hyrcanian
forests of Iran. In this respect, the tree biomass was calculated
for each plot using the calibrated Chave allometric equation
(Chave et al., 2005) as shown in Equation 1:
AGB=0.117 (ρ DBH2. H)0.928 (1)
where DBH is the diameter at breast height, H is height, and
ρ is density of the wood. The density of the wood of Fagus
orientalis, Alnus serrulata, Carpinus betulus, and Ulmus gla-
bra is 633, 535, 621, and 755 Kg.m-3, respectively (Kiaei and
Samariha, 2011).
ALOS PALSAR Data
Three ALOS PALSAR ne beam dual (FBD ; HH and HV polariza-
tions) images taken on 17 July 2007, 19 July 2008, and 25 July
2010 were acquired to enable analysis of the study area. Each
was taken with a 34.3° off-nadir angle. Single-look complex
(SLC) data were produced using the PALSAR raw data (level 1.0)
with SAR processor software developed by the remote sensing
department of Tehran University. To obtain a ground resolu-
tion of 12.5 m both in range and azimuth, a factor of 1 for
range and 5 for azimuth was used.
The number of speckles was first reduced using the pro-
posed algorithm by Sharifi et al. (2015), and the equivalent
number of looks (ENL s) (Equation 2) was then estimated over
a set of homogeneous areas belonging to the species classes
(Woodhouse, 2005):
ENL =μ2⁄σ2 (2)
where μ and σ2 are the mean and variance of the backscatter
intensity values. The estimated ENL mean values were 3.12
and 3.35, for the HH and HV polarizations, respectively.
The ALOS images were first geocoded using tie points and
orbit parameters. The study area is very mountainous, and
therefore without topographic normalization, the backscatter
values would be related to both slope and backscatter (Guo et
al., 2010). The topography normalization was undertaken us-
ing an extracted DEM from a 1:25 000 scale topographic map.
The RMSE values of geolocation for three images were 0.24,
1.59, and 0.87 pixels, respectively. Plate 1 (lower panel) shows
the processed PALSAR image acquired in 25 July 2010. Also,
topographic normalization of backscatters values can be calcu-
lated using Equation 3), as reported by Carreiras et al. (2006):
A c σ σ θ
θ f
flat iref
slope iloc
A cos
os
0 0 = .
. (3)
where σf
0 and σ0 represent the topographically normalized and
uncorrected backscatter intensity, respectively; θθiref is the in-
cidence angle at a reference location (e.g., mid-swath); θθiloc is
the local incidence angle, Aflat and Aslope are the SAR pixels size
for at terrain and the true local SAR pixel size, respectively.
Digital numbers (DNs) were converted to sigma naught
values to obtain backscatters using the following Equation 4
(Shimada et al., 2009):
σ 0(dB )=10(log 10DN2 )+CF (4)
where CF is the calibration factor, set at -80.2 for HV polariza-
tion and at -83.2 for HH polarization.
The ALOS PALSAR backscatter coef cients were then tempo-
rally averaged, because the average values of multi-temporal
backscatter of the dry season are very stable, and there is little
variation with the climatic conditions (Englhart et al., 2011;
Townsend, 2001). Figure 1 shows the average of the HH- and
HV-polarized backscatters versus AGB. It can be seen from this
figure that the backscatters increase with the AGB up to a point
of approximately 150 Mg.ha-1, and thereafter do not increase in
line with higher values of AGB.
AGB Estimation Modeling
In order to investigate the relationship between PALSAR
intensity backscatter and the AGB, four different models were
42 January 2016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING