1
1
Biomass estimation in humid tropical forest using a combination of ALOS PALSAR and
2
SPOT 5 satellite imagery
3
4
Jonathan Y. Goh1, Jukka Miettinen1*, Aik Song Chia1, Ping Ting Chew2 and Soo Chin Liew1
5
6
1
Centre for Remote Imaging Sensing and Processing (CRISP),
7
National University of Singapore (NUS),
8
10 Lower Kent Ridge Road
9
Singapore 119076
10
11
2
Central Nature Reserve,
12
National Parks Board (NParks),
13
1 Cluny Road,
14
Singapore 259569
15
16
* Corresponding author,
17
Email address: jimietti@yahoo.com, scliew@nus.edu.sg
18
Phone: +65 65165069
19
Fax: +65 6775 7717
20
2
1
Biomass estimation in humid tropical forest using a combination of ALOS PALSAR and
2
SPOT 5 satellite imagery
3
4
Abstract
5
In this study we investigated the usability of a combination of SPOT 5 (Satellite Pour
6
l'Observation de la Terre) optical satellite imagery and Advanced Land Observing Satellite
7
(ALOS) Phased Array type L-band Synthetic Aperture Radar (PALSAR) data for above
8
ground biomass estimation in humid tropical forest. The study area covered around 2500 ha in
9
the Central Nature Reserve of Singapore. Linear regression models were developed between
10
biomass measured in 25 field sample plots and remotely sensed parameters using a stepwise
11
regression approach. The best performing regression model which included NIR band from
12
SPOT 5 and HV radar backscatter from ALOS PALSAR achieved adjusted r2 of 0.46 and
13
RMSE of 152 t/ha (36%) with essentially no bias. The results suggest that a combination of
14
optical and radar remote sensing data supported by field sampling can be used to estimate
15
biomass in large humid tropical forest areas using empirical regression models in a rather
16
homogeneous environment. However, the study also exposes the large pixel level errors of
17
such models and highlights the unsuitability of empirical models for biomass estimation
18
outside the vegetation type they have been developed for.
19
20
21
22
Keywords: Above ground biomass, Carbon accounting, Radar, Singapore
3
1
1. Introduction
2
Deforestation and forest degradation have been identified as major contributors to global
3
climate change producing nearly 20% of global anthropogenic greenhouse gas emissions
4
(IPCC 2007). Due to the high carbon content of vegetation biomass, accurate quantification
5
and monitoring of forest biomass is necessary for evaluation of schemes aiming to reduce
6
carbon emissions from changes in forest areas (e.g. the REDD+ scheme; Reducing Emissions
7
from Deforestation and Forest Degradation). Remote sensing based forest biomass estimation
8
has received much attention in recent years due to the relative ease and low cost of acquiring
9
remotely sensed data over wide and often nearly inaccessible forest areas (see e.g. Goetz et al.
10
2009). Remote sensing methods available for forest biomass estimation can be coarsely
11
divided into two categories: 1) passive (in this context optical) and 2) active (most typically
12
radar) remote sensing. While optical remote sensing relies on the detection of radiation
13
originating from the sun and reflected by the surface of the earth, active remote sensing
14
devices send and receive their own radiation impulse. Datasets acquired by both optical and
15
active remote sensing technologies have been used for forest biomass estimation.
16
17
With its complex structure combined with high level of biomass, humid tropical forest
18
presents a challenging environment for optical remote sensing based biomass estimation. Not
19
only do the techniques capture merely radiation reflected from the top of the vegetation
20
canopy but they are also sensitive to topographic and atmospheric influences. Regardless of
21
the rather positive results obtained in boreal areas (e.g. Tuominen et al. 2010), optical remote
22
sensing based biomass estimation efforts have faced severe problems and limitations in
23
tropical forests (Nelson et al. 2000, Steininger 2000, Foody et al. 2001, 2003, Lu 2005, 2006,
24
Sarker and Nichol 2011). It has been shown that spectral reflectance and vegetation indices
4
1
alone are not reliable indicators of biomass in tropical forests. Foody et al. (2003) found that
2
vegetation indices were weakly related to biomass and that the direction of their relationship
3
was inconsistent in the tropical forests of Thailand, Brazil and Malaysia. Steininger (2000), on
4
the other hand, showed that spectral reflectance was not sensitive to biomass changes in areas
5
with higher than 150 t/ha biomass in the tropical forests in Manaus, Brazil.
6
7
The poor performance of optical remote sensing methods in the tropics has been mainly
8
attributed to the multi-layered closed canopy structure of tropical forest. Nevertheless, some
9
authors have achieved positive results. Nelson et al. (2000) successfully estimated tropical
10
forest biomass with optical remote sensing data by including the age of the forest into their
11
Landsat TM based analysis. Lu (2005) and Sarker and Nichol (2011), on the other hand,
12
improved biomass estimation results in tropical forests by incorporating texture information
13
into the process. Lu (2005) concluded that for forests with more complex stand structure,
14
image texture features are more important than spectral reflectance for biomass estimation.
15
16
In theory, radar technology can be used to categorize the structure of the forest. The radar
17
impulse penetrates into the canopy and thereby delivers information on the canopy structure.
18
There have been a number of studies on the usability of Synthetic Aperture Radar (SAR)
19
imagery to estimate various forest variables such as age, height and biomass (e.g. Luckman et
20
al. 1997, Bergen et al. 1999, Kuplich et al. 2005, Champion et al. 2008, Hajnsek et al. 2009).
21
Results have indicated that L-band SAR data has the most potential for forest biomass
22
estimation as it carries mainly information about larger components of vegetation such as
23
trunks and branches (Luckman et al. 1997, Lu 2006, Goetz et al. 2009, Wang et al. 2008
24
Wolter et al. 2011). However, L-band sensitivity to biomass has a saturation point which
5
1
some studies have put at around 40-90 t/ha (Fransson and Israelsson 1999, Luckman et al.
2
1997, Salas et al. 2010) and others at up to 100-150 t/ha (Dobson 1992, Sandberg et al. 2011).
3
The relatively low saturation level causes dramatic limitations on the applicability of radar
4
methodology for biomass estimation in humid tropical regions due to the typically high levels
5
of forest biomass in these areas. Ratios of SAR backscatter and polarization as well as various
6
SAR texture measures which aim to isolate the contribution of biomass to backscatter and
7
reduce the effect of forest structure have been proposed as a means to extend the range of
8
biomass estimation from SAR data (Salas et al. 2002a, 2002b, Kuplich et al. 2005, Wang and
9
Qi 2008). These approaches, however, are yet to be fully investigated.
10
11
Several studies have shown that the integration of optical and SAR data for biomass
12
estimation is promising due to the complementary strengths of the sensors (Araujo et al. 1999,
13
Moghaddam et al. 2002, Townsend 2002, Wang and Qi 2008, Amini and Sumantyo 2009,
14
Wolter and Townsend 2011). At least two studies have reported that the range of validity of
15
SAR signals can be extended by including optical data into canopy scattering models for
16
biomass estimation (Wang and Qi 2008, Moghaddam et al. 2002). Araujo (1999) used a
17
combination of Landsat TM and JERS-1 data to develop a multiple regression model to
18
estimate forest and savanna biomass in Brazil. He concluded that while SAR data did not
19
contribute significantly to the biomass estimation model, it was useful for characterisation of
20
different forest types. Townsend (2002) showed that integrating NDVI with SAR data
21
improved the performance of empirical models for estimating forest basal area. Very recently,
22
optical and SAR data fusion techniques have been tested for forest biomass estimation in
23
northern Iran (Amini and Sumantyo 2009) and for estimating forest species composition and
24
abundance in Northern Minnesota (Wolter and Townsend 2011).
25
6
1
In this study we investigated the usability of a combination of optical and radar satellite
2
imagery for above ground biomass estimation in humid tropical forests. We derived a set of
3
remotely sensed parameters from SPOT 5 (Satellite Pour l'Observation de la Terre) and
4
Advanced Land Observing Satellite (ALOS) Phased Array type L-band Synthetic Aperture
5
Radar (PALSAR) datasets and investigated their correlation with biomass measured on 25
6
field sample plots. The strongest remotely sensed parameters for biomass estimation were
7
sought using stepwise multiple regression modelling. The performances of the models were
8
evaluated using 10 independent validation field sample plots.
9
10
2. Materials and methods
11
2.1. Study Area
12
The study area consisted of the Central Nature Reserve (CNR) of Singapore, a forest area
13
covering some 2500 ha located in the central part of the highly urbanised island of Singapore
14
(Figure 1). CNR enjoys legal protection as a nature reserve and is under the management of
15
the National Parks Board of Singapore (NParks). Approximately 200 ha of nearly pristine
16
patches of primary dipterocarp forest remain in fragments within CNR, engulfed by secondary
17
forests in various stages of succession (Teo et al. 2003). These secondary forests were
18
formerly used for plantation agriculture but they were quickly abandoned due to unsustainable
19
agriculture practices which reduced soil fertility. Forest succession has taken place since the
20
early 20th century until today. These up to 100 year old secondary forests are now largely
21
dominated by Rhodamnia cinerea and still have less biodiversity than the areas of CNR that
22
are covered by primary forests (Turner et al. 1996, Teo et al. 2003).
23
24
<Insert Figure 1 somewhere here.>
7
1
2
2.2. Field Measurements
3
A total of 35 permanent forest inventory plots (25 m radius circular plot) located in the CNR
4
were used in this study. These permanent plots had been established within the past 30 years
5
in the CNR for the purpose of ecological research and forest health monitoring. 25 plots
6
ranging from young secondary forest to primary lowland tropical forest were used to analyse
7
the correlation between biomass and remotely sensed parameters. Further 10 additional plots
8
were used for model validation. In all of the plots, all individual trees with a diameter at breast
9
height (DBH) exceeding 9 cm were measured for DBH and identified by species in a field
10
campaign between October 2009 and May 2010. The biomass of each tree was calculated
11
using an allometric model (Equation 1) developed by Chave et al. (2005). The particular
12
model used in this study was developed for humid tropical forest including sampling sites in
13
Malaysia and Indonesia, the neighbouring countries of Singapore:
14
15
AGB = ρ × exp (-1.499 + 2.148 ln (D) + 0.207 (ln (D))2 - 0.0281 (ln (D))3)
(Equation 1)
16
17
where ρ is the wood density (oven dry mass divided by green volume) (g/cm3) of the tree
18
species and D is the DBH (cm) of the tree. Wood density values were obtained from the
19
Global Wood Density Database (Chave et al. 2009, Zanne et al. 2009) which is the largest
20
wood density database collected to date.
21
22
2.3 Remotely Sensed Data
23
A set of remote sensing data from both optical and radar remote sensing sensors were
24
compiled for this study. An image from the SPOT 5 HRG (High Resolution Geometric)
8
1
sensor acquired on the 14th of April 2010 formed the optical dataset. In addition, a SAR scene
2
in L-band HH and HV polarization from the ALOS PALSAR sensor acquired on the 1st of
3
July 2010 was used.
4
5
The SPOT 5 HRG image had a spatial resolution of 10 m and contained four wavelength
6
bands; green (Band 1; 0.50 – 0.59 µm), red (Band 2; 0.61 – 0.68 µm), near infra-red (Band 3;
7
0.79 – 0.89 µm) and shortwave infra-red (Band 4; 1.58 – 1.75 µm). The raw digital numbers
8
(DN) were first converted to top of the atmosphere (TOA) reflectance and then corrected for
9
Rayleigh scattering and molecular absorption using routines in the 6S package (Vermote et al.
10
1997), assuming a standard tropical atmosphere with considerations of the spectral response
11
of each spectral band of the sensor. In addition to the four wavelength bands a set of other
12
parameters were derived from the optical image to be used in the analysis (Table 1).
13
14
<Insert Table 1 somewhere here.>
15
16
For the ALOS PALSAR data, the Level 1.0 images (HH and HV) were first multi-looked by a
17
factor of three in the azimuth direction. They were then geo-registered (using a nearest-
18
neighbour scheme to preserve the original data values) to the same image geometry as the
19
SPOT 5 mosaic to facilitate later comparison. Using SRTM DEM data to calculate the local
20
incidence angle at each pixel, topographic normalization (Leclerc et al. 2001) was carried out,
21
before obtaining the final output data in σ0 (normalized radar backscatter cross-section, also
22
called sigma-nought) values, with units in dB (decibels) (Shimada et al. 2009). Correlation
23
between HH and HV backscatter and biomass was then analysed (Table 1).
24
9
1
2.4 Above ground biomass estimation procedure
2
Vegetation biomass is a comprehensive parameter that is related to many factors such as
3
vegetation stand structure, vegetation density and vegetation species composition. Simple
4
linear and multiple regression models were tested to empirically relate measured field
5
biomass values to the remote sensing parameters. The remotely sensed parameter values were
6
derived by calculating the mean values of image pixels within each of the 25 biomass plots.
7
First, the correlation between each individual parameter and biomass was evaluated.
8
Subsequently, using field biomass values as the dependant variable and the mean parameter
9
value of each plot as the independent variable, stepwise backward regression analysis was
10
used to find the best parameter combinations to estimate forest biomass. In addition to the
11
basic statistics of regression model fitness, 10 independent validation plots were used to
12
evaluate the performance of the models.
13
14
3. Results
15
Table 2 summarises the results of the correlation analysis between field biomass and the
16
remotely sensed parameters. Note that only the best performing occurrence and co-occurrence
17
parameters among the numerous texture measures tested (Table 1) have been presented in
18
Table 2. In general, the level of correlation between biomass and the remotely sensed pixel
19
level parameters was rather low. The highest correlation was observed in Band 3 reflectance
20
with r = -0.69. Both the vegetation indices (NDVI, SAVI, EVI) and the linear combinations of
21
bands (Albedo, 1st PCA component) performed on either around the same level or worse than
22
Band 3, with correlation ranging from r = -0.23 to r = -0.67 for NDVI and the 1st PCA
23
component respectively. The texture measures indicated somewhat higher correlation than the
24
pixel level parameters, with the mean occurrence and co-occurrence texture measure (5x5
10
1
kernel) of Band 3 having an r value of -0.71. As for SAR data, rather weak correlation was
2
witnessed between HH and HV backscatter and biomass, having r of only 0.35 and 0.48
3
respectively.
4
5
<Insert Table 2 somewhere here.>
6
7
The best combinations of parameters to estimate biomass was sought by means of stepwise
8
linear regression analysis. The results indicated that Band 3 mean co-occurrence texture
9
measure alone was the best predictor of biomass with the regression model explaining 49% of
10
the variance in field biomass data (adjusted r2 = 0.49) and producing a root mean square error
11
(RMSE) of 149 t/ha or 35% of the average biomass in the field sample plots. As only one
12
parameter was identified, this indicated that inclusion of other parameters would not
13
contribute positively to the success of the model. However assessment of the model using the
14
validation plots revealed that in some cases the difference between the biomass predicted by
15
the model and the biomass measured in the field was quite significant. The plotwise errors
16
reached as high as 113% of the measured field value (Table 3: Model 4). It is important to
17
note, however, that the estimation bias (i.e. the percentage difference between the average
18
biomass of the plots and the estimated biomass of the plots) was only 2.5% (Table 3, Figure
19
2). Thus, the results indicate that the performance of the model was low when applied on a
20
plot level basis and thus inappropriate for biomass estimation at the pixel level. But when
21
applied on large contiguous forest areas such as the entire CNR, the model could be expected
22
to give reasonable biomass estimates.
23
24
<Insert Table 3 and Figure 2 somewhere here.>
11
1
2
Due to the discrepancy between the plotwise vs. large scale reliability of empirical forest
3
biomass estimation highlighted above, bias is an important indicator for the practical usability
4
of any given model. The usability of the model to derive estimates of a total biomass in large
5
forest areas depends primarily on the bias and not so much on the pixel level accuracy.
6
Therefore, we further compared the performance of four alternative regression models (Table
7
4) which produced the best results in the stepwise regression process and all of which had
8
performed nearly on a similar level based on r2 values. We evaluated the performance of these
9
four models by calculating their estimation bias using 10 independent validation plots (Tables
10
3 and 4). Two of the tested models included only optical remote sensing parameters, while
11
two of the models incorporated HV backscatter information into the estimation.
12
13
<Insert Table 4 somewhere here.>
14
15
All of the four regression models tested achieved a low level of estimation bias ranging from
16
0.8% (Model 3) to 2.5% (Model 4). Note that both of the models incorporating radar data
17
(Models 1 and 3) showed a slightly smaller bias than the models using the same optical
18
parameters without radar data (Models 2 and 4). It is important to remember at this point,
19
however, that the radar data alone did not allow meaningful estimations of above ground
20
biomass in high volume tropical forests, as shown by the weak correlation of HV backscatter
21
and biomass (r = 0.48: Table 2). Nevertheless, comparison of the estimation bias suggests that
22
radar data complements optical remote sensing parameters resulting in slightly improved
23
biomass estimation in the case of our study.
24
12
1
Among all the evaluated models, Model 1 was regarded as the most suitable option for
2
practical biomass estimation efforts. Although Model 3 had slightly lower bias, Model 1 was
3
considered superior for practical implementation due to the very simple remote sensing
4
parameters used. Texture measures used in Model 3 are not only computationally expensive
5
but they tend to be scene dependent. To evaluate the practical usability of the biomass
6
estimation process, Model 1 was applied to the study area (Figure 3). Although in general the
7
model produced a meaningful biomass map over the study area, there were some points where
8
the regression model returned negative biomass values when applied to the remote sensing
9
data. Closer inspection revealed that these areas were in fact patches of shrub and grassland
10
found within the CNR. These vegetation types were not represented in any of the sampling
11
plots. This very concretely highlights that the use of these types of empirical biomass
12
estimation models cannot be extended beyond those vegetation types that have been sampled
13
during the creation of the model.
14
15
<Insert Figure 3 somewhere here.>
16
17
4. Discussion
18
In this study we investigated the usability of a combination of optical and radar satellite
19
imagery for biomass estimation in humid tropical forest. Among the remotely sensed
20
parameters optical near infra-red reflectance and radar HV backscatter provided the most
21
recommended combination for biomass estimation. The regression based data fusion approach
22
enabled biomass estimation for the study area with essentially no bias. However, the results
23
revealed high pixel level errors. Similar discrepancy in pixel level vs. large scale accuracy of
24
biomass estimation with remote sensing data has been noticed also outside tropical areas (see
13
1
e.g. Tuominen et al. 2010). Nevertheless, even the pixel level errors reported in this study
2
(~35% of sample plot average) were well comparable with the results published by other
3
researchers who have worked with tropical forest biomass estimation (Foody et al. 2001,
4
2003, Lu 2005).
5
6
The HV backscatter radar data did not allow reliable biomass estimation on their own nor did
7
they consistently improve the pixel level biomass estimation results compared to models
8
using merely optical remote sensing data. This lack of correlation between radar backscatter
9
and biomass is expected to be due to high biomass content of humid tropical forests (see e.g.
10
Salas et al. 2010, Sandberg et al. 2011). However, the use of radar data was noticed to cause a
11
slight reduction of the bias over the entire study area when used in combination with optical
12
remote sensing data. This reduction of estimation bias may have been related to the canopy
13
penetrating capabilities of active remote sensing. Penetration of forest canopy allows radar
14
based technologies to obtain information on the vegetation structure. This may improve the
15
separation of varying types of forest cover as indicated by Araujo (1999). Subsequently, the
16
variation of forest types is likely to be correlated to variation in biomass. We believe that this
17
may at least partly explain why a synergistic use of optical and radar data was noticed to
18
result in a slightly lower estimation bias than the use of only optical data although the radar
19
backscatter information alone had very low correlation with biomass.
20
21
The low pixel level accuracy noticed in this study as well as other similar studies (Foody et al.
22
2001, 2003, Houghton et al. 2001, Lu 2005) is expected to be partly caused by a lack of direct
23
correlation between remotely sensed parameters and vegetation biomass and partly by
24
variation in remotely sensed parameter values due to other effects than variation in vegetation
14
1
(e.g. remnants of atmospheric effects in surface reflectance values). The lack of direct
2
correlation between the remotely sensed parameters and biomass leads to derivation of
3
biomass estimates based on secondary features (e.g. canopy closure and the intensity of self
4
shadowing). We believe that the negative correlation between near infra-red reflectance and
5
biomass documented in this study was mainly caused by the increasing unevenness in the
6
canopy structure of aging humid tropical forests and the subsequently increasing self
7
shadowing. This emphasizes the highly empirical nature of biomass estimation in high
8
volume forest areas based on correlation between remote sensing parameters and field
9
sampling. The general level of biomass is entirely derived from the field measurements, while
10
the small variations and their spatial distribution within the area are estimated by means of
11
remote sensing. This was well highlighted by the negative biomass values produced for the
12
grassland and shrub areas inside our study area. It is therefore clear that the usability of this
13
type of empirical models is limited to the type of forest they have been developed for.
14
15
The results of this study also highlighted problems related to the use of texture measures in
16
biomass estimation. Studies so far have not found a consistent texture measure for extracting
17
forest biomass information (Sarker and Nichol 2011, Lu 2006). Likewise, in this study the
18
best texture measure identified for biomass estimation (co-occurrence texture measure, mean
19
of near infra-red band) was different from earlier studies. This disagreement between studies
20
regarding the most suitable texture measures may also be related to the secondary nature of
21
the vegetation characteristics on which the biomass estimation is based. As these parameters
22
do not have a direct connection with the amount of biomass, the most suitable parameters may
23
vary dramatically depending on the structural characteristics of the vegetation in the area of
24
interest, if not even depending on the quality of the remotely sensed image as well. In
15
1
different areas entirely different textural features may correlate with the amount of biomass. It
2
is therefore quite obvious that such results cannot be directly applied outside the ecosystems
3
they have been developed for.
4
5
Regardless of the limitations discussed above, models such as the ones developed in this
6
study seem to be suitable for monitoring biomass in specific study areas. Utilisation of a set of
7
permanent sample plots allows creation of a relationship between remotely sensed parameters
8
and biomass in a specific case of the area of interest. This relationship can then be propagated
9
throughout the area of interest using a wall-to-wall remote sensing dataset. Repetition of this
10
same procedure in the future would allow reliable estimation of the changes in biomass.
11
Thereby, regardless of the low pixel level accuracies and other limitations witnessed in this
12
study, these types of empirically derived biomass estimation models show potential for
13
practical biomass monitoring. This is very important e.g. for the ever more numerous REDD+
14
projects and similar schemes aiming to promote sustainable forest management and preserve
15
biomass/carbon in the ecosystem. Based on the findings in this study, remote sensing may be
16
used in these initiatives to evaluate the success of the carbon conservation measures taken.
17
18
5. Conclusion
19
In conclusion, the results of this study suggested that a combination of optical and radar
20
remote sensing data supported by field sampling can be used to monitor biomass in limited
21
interest areas covered by rather homogeneous humid tropical forests using empirical
22
regression models. These types of empirical biomass estimation models have a range of
23
suitability where they can be reliably applied to but the suitability of such models for areas
24
outside this range should be examined in case-by-case basis using local field data. We believe
16
1
that empirically derived biomass estimation models combined with extensive network of
2
permanent sample plots would be of use for monitoring purposes in REDD+ and other similar
3
projects aiming to promote preservation of biomass/carbon in humid tropical forest
4
ecosystems.
5
6
Acknowledgements
7
This project was primarily funded by the Ministry of National Development, Singapore. The
8
authors from the Centre of Remote Imaging, Sensing and Processing (CRISP) acknowledge
9
also research funding from the Agency for Science, Technology and Research (A*STAR) of
10
11
12
Singapore.
17
1
References
2
Anys, H., Bannari, A., He, D.C. and Morin, D. (1994). Texture analysis for the mapping of
3
urban areas using airborne MEIS-II images. In First International Airborne Remote Sensing
4
Conference and Exhibition, 11-15 September 1994, Strasbourg, France (pp. 231-245). Ann
5
Arbor, MI: Environmental Research Institute of Michigan.
6
7
Araujo, L.S., Santos, J.R., Freitas, C.C. and Xaud, H.A.M. (1999). The use of microwave and
8
optical data for estimating aerial biomass of the savanna and forest formations at Roraima
9
State, Brazil. In International Geoscience and Remote Sensing Symposium IGARSS 1999, 28
10
June - 2 July, Hamburg, Germany (pp. 2762-2764). New Jersey: IEEE international.
11
12
Bergen, K.M. and Dobson, M.C. (1999). Integration of remotely sensed radar imagery in
13
modeling and mapping of forest biomass and net primary production. Ecological Modeling,
14
122, 257-274.
15
16
Champion, I., Dubois-Fernandez, P., Guyon, D. and Cottrel, M. (2008). Radar image texture
17
as a function of forest stand age. International Journal of Remote Sensing, 22, 229-242.
18
19
Chave, J., Andalo, C., Brown, S., Cairns, M.A., Chambers, J.Q., Eamus, D., Fölster, H.,
20
Fromard, F., Higuchi, N., Kira, T., Lescure, J.-P., Nelson, B.W., Ogawa, H., Puig, H., Riera,
21
B. and Yamakura, T. (2005). Tree allometry and improved estimation of carbon stocks and
22
balance in tropical forests. Oecologia, 145, 87–99.
23
18
1
Chave, J., Coomes, D., Jansen, S., Lewis, S.L., Swenson, N.G. and Zanne, A.E. (2009).
2
Towards a worldwide wood economics spectrum. Ecology Letters, 12, 351-366.
3
4
Dobson, M.C., Ulaby, F.T., Letoan, T., Beaudoin, A., Kasischke, E.S. and Christensen, N.
5
(1992). Dependence of radar backscatter on coniferous forest biomass. IEEE Transactions on
6
Geoscience and Remote Sensing, 30, 412-415.
7
8
Foody, G.M., Cutler, M.E., Mcmorrow, J., Pelz, D., Tangki, H., Boyd, D.S. and Douglas, I.
9
(2001). Mapping the biomass of Bornean tropical rain forest from remotely sensed data.
10
Global Ecology and Biogeography, 10, 379-387.
11
12
Foody, G.M., Boyd, D.S. and Cutler, M.E. (2003). Predictive relations of tropical forest
13
biomass from Landsat TM data and their transferability between regions. Remote Sensing of
14
Environment, 85, 463-474.
15
16
Fransson, J.E.S. and Israelsson, H. (1999). Estimation of stem volume in boreal forests using
17
ERS-1 C- and JERS-1 L-band SAR data. International Journal of Remote Sensing, 20, 123-
18
137.
19
20
Goetz, S.J., Baccini, A., Laporte, N.T., Johns, T., Walker, W., Kellndorfer, J., Houghton, R.A.
21
and Sun, M. (2009). Mapping and monitoring carbon stocks with satellite observations: a
22
comparison of methods. Carbon Balance and Management, 4, 2.
23
19
1
Hajnsek, I., Kugler, F., Lee, S.K., Papathanassiou, K.P. (2009). Tropical-Forest-Parameter
2
estimation by means of Pol-InSAR: The INDREX-II Campaign. IEEE Transactions on
3
Geoscience and Remote Sensing, 47, 481-493.
4
5
Haralick, R.M., Shanmugan, K. and Dinstein, I. (1973). Textural Features for Image
6
Classification. IEEE Transactions on Systems, Man, and Cybernetics, 3, 610-621.
7
8
Houghton, R.A., Lawrence, K.T., Hackler, J.L. and Brown, S. (2001). The spatial distribution
9
of forest biomass in the Brazilian Amazon: a comparison of estimates. Global Change
10
Biology, 7, 731-746.
11
12
IPCC (2007). Climate Change 2007: Contribution of Working Group III to the Fourth
13
Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge and New
14
York: Cambridge University Press.
15
16
Kuplich, T.M., Curran, P.J. and Atkinson, P.M. (2005). Relating SAR image texture to the
17
biomass of regenerating tropical forests. International Journal of Remote Sensing, 26,
18
4829-4854.
19
20
Leclerc, G., Beaulieu, N. and Bonn, F. (2001). A simple method to account for topography in
21
the radiometric correction of radar imagery. International Journal of Remote Sensing, 22,
22
3553-3570.
23
20
1
Lu, D. (2005). Aboveground biomass estimation using Landsat TM data in the Brazilian
2
Amazon. International Journal of Remote Sensing, 26, 2509-2525.
3
4
Lu, D. (2006). The potential and challenge of remote sensing-based biomass estimation.
5
International Journal of Remote Sensing, 27, 1297-1328.
6
7
Lu, D., Mausel, P., Brondizio, E. and Moran, E. (2004). Relationships between forest stand
8
parameters and Landsat TM spectral responses in the Brazilian Amazon basin. Forest Ecology
9
and management, 198, 149-167.
10
11
Luckman, A., Baker, J.R., Kuplich, T.M., Yanasse, C.C.F. and Frery, A.C. (1997). A study of
12
the relationship between radar backscatter and regernaring forest biomass for space borne
13
SAR instrument. Remote Sensing of Environment, 60, 1-13.
14
15
Moghaddam, M., Dungan, J.L. and Acker, S. (2002). Forest variable estimation from fusion
16
of SAR and multispectral optical data. IEEE Transactions on Geoscience and Remote
17
Sensing, 40, 2176-2187.
18
19
Nelson, R.F., Kimes, D.S., Salas, W.A. and Routhier, M. (2000). Secondary forest age and
20
tropical forest biomass estimation using Thematic Mapper imagery. Bioscience, 50, 419-431.
21
22
Salas, W.A., Ducey, M.J., Rignot, E., Skole, D. (2002a). Assessment of JERS-1 SAR for
23
monitoring secondary vegetation in Amazonia: I. Spatial and temporal variability in
21
1
backscatter across a chrono-sequence of secondary forest in Rondonia. International Journal
2
of Remote Sensing, 23, 2357-1379.
3
4
Salas, W.A., Ducey, M.J., Rignot, E., Skole, D. (2002b). Assessment of JERS-1 SAR for
5
monitoring secondary vegetation in Amazonia: II. Spatial, temporal, and radiometric
6
considerations for operational monitoring. International Journal of Remote Sensing, 23, 1381-
7
1399.
8
9
Sandberg, G., Ulander, L.M.H., Fransson, J.E.S, Holmgren, J., Letoan, T. (2011). L- and P-
10
band backscatter intensity for biomass retrieval in hemiboreal forest. Remote Sensing of
11
Environment, 115, 2874-2886.
12
13
Sarker, L.R. and Nichol, J.E. (2011) Improved forest biomass estimates using ALOS AVNIR-
14
2 texture indices. Remote Sensing of Environment, 115, 968-977.
15
16
Shimada, M., Isoguchi, O., Tadano, T. and Isono, K. (2009). PALSAR radiometric and
17
geometric calibration. IEEE Transactions on Geoscience and Remote Sensing, 47, 3915-3932.
18
19
Steininger, M.K. (2000). Satellite estimation of tropical secondary forest above-ground
20
biomass: Data from Brazil and Bolivia. International Journal of Remote Sensing, 21, 1139-
21
1157.
22
22
1
Teo, D.H.L., Tan, H.T.W., Corlett, R.T., Wong, C.M. and Lum, S.K.Y. (2003). Continental
2
rain forest fragments in Singapore resist invasion by exotic plants. Journal of Biogeography,
3
30, 305–310.
4
5
Townsend, P. (2002). Estimating forest structure in wetlands using mulit-temporal SAR.
6
Remote Sensing of Environemt, 79, 288-304.
7
8
Tuominen, S., Eerikäinen, K., Schibalski, A., Haakana, M. and Lehtonen, A. (2010). Mapping
9
Biomass Variables with a Multi-Source Forest Inventory Technique. Silva Fennica, 44, 109-
10
119.
11
12
Turner, I.M., Tan, H.T.W. and Chua K.S. (1996). Relationships between herb layer and
13
canopy composition in a tropical rain forest successional mosaic in Singapore. Journal of
14
Tropical Ecology, 12, 843-851.
15
16
Vermote, E., Tanre, D., Deuze, J.L., Herman, M. and Morcette, J.J. (1997). Second simulation
17
of the satellite signal in the solar spectrum: An overview. IEEE Transactions on Geoscience
18
and Remote Sensing, 35, 675-686.
19
20
Wang, C. and Qi, J. (2008). Biophysical estimation in tropical forest using JERS-1 SAR and
21
VNIR imagery II: Aboveground woody biomass. International Journal of Remote Sensing,
22
29, 6827-6849.
23
23
1
Wolter, P.T. and Townsend, P.A. (2011). Multi-sensor data fusion for estimating forest
2
species composition and abundance in northern Minnesota. Remote Sensing of Environment,
3
115, 671-691.
4
5
Zanne, A.E., Lopez-Gonzalez, G., Coomes, D.A., Ilic, J., Jansen, S., Lewis, S.L., Miller, R.B.,
6
Swenson, N.G., Wiemann, M.C. and Chave, J. (2009). Global wood density database, Dryad
7
Digital Repository. Available online at: http://datadryad.org/handle/10255/dryad.235
8
(accessed 9 May 2012).
9
10
11
24
1
Table 1. Remotely sensed parameters tested for biomass estimation.
2
3
Table 2. Correlation between biomass and remotely sensed parameters. Occ_Band3_mean
4
refers to the mean occurrence texture measure for Band 3 while Co_occ_Band3_mean refers
5
to the mean co-occurrences texture measure for Band 3.
6
7
Table 3. Plot level errors and overall bias of the biomass estimation models when applied to
8
the independent validation plots.
9
10
Table 4. The four best performing models compared in the final stage of the study.
11
Co_occ_Band3_mean refers to the mean co-occurrence texture measure for Band 3.
12
13
25
1
Table 1. Remotely sensed parameters tested for biomass estimation.
2
Image Parameter
SPOT 5 HRG
Bands 1 to 4
NDVI
SAVI
2-band EVI
Albedo
PCA
Occurrence texture
measures (Anys et
al. 1994)
Co-occurrence
texture measures
(Haralick et al.
1973)
HH backscatter
HV backscatter
3
4
Description
Atmospherically corrected SPOT 5 band values converted to surface
reflectance measures.
B3  B 2
, Band numbers refer to SPOT 5 bands.
B3  B 2
B3  B 2
(1  L) , Band numbers refer to SPOT 5 bands, L = 0.5.
B3  B 2  L
2.5 * ( B3  B 2)
, Band numbers refer to SPOT 5 bands.
(B3  2.4 * B2  1)
B1+B2+B3+B4, Band numbers refer to SPOT 5 bands.
1st Principal Component of Principal Component Analysis.
Data range, mean, variance, entropy, and skewness measures for all
four bands in the optical (SPOT 5) data within 5x5 pixel kernel.
Mean, variance, homogeneity, contrast, dissimilarity, entropy, second
moment, and correlation measures for all four bands in the optical
(SPOT 5) data within 5x5 pixel kernel.
HH backscatter of ALOS PALSAR sensor presented in sigma-nought
values.
HV backscatter of ALOS PALSAR sensor presented in sigma-nought
values.
26
1
Table 2. Correlation between biomass and remotely sensed parameters. Occ_Band3_mean
2
refers to the mean occurrence texture measure for Band 3 while Co_occ_Band3_mean refers
3
to the mean co-occurrences texture measure for Band 3.
4
Image Parameter
HRG Band 1
HRG Band 2
HRG Band 3
HRG Band 4
NDVI
SAVI
PCA
Albedo
2-band EVI
Occ_Band3_mean
Co_occ_Band3_mean
HH backscatter
HV backscatter
5
6
r
-0.25
-0.13
-0.69
-0.56
-0.23
-0.54
-0.67
-0.60
-0.67
-0.71
-0.71
0.35
0.48
27
1
Table 3. Plot level errors and overall bias of the biomass estimation models when applied to
2
the independent validation plots.
3
4
Validation Measured
plot
biomass
(t/ha)
1
359
2
788
3
706
4
475
5
365
6
368
7
522
8
224
9
536
10
299
Estimation bias (%)
5
6
7
8
9
10
11
Model 1
(t/ha) %
598 -66
588
25
589
16
497
-5
569 -56
352
4
301
42
305 -36
330
38
467 -56
-1.0
Estimated biomass
Model 2
Model 3
Model 4
(t/ha) % (t/ha)
% (t/ha) %
626 -74
732 -104
765 -113
563
29
612
22
603
24
634
10
624
12
656
7
490
-3
522
-10
522
-10
534 -46
552
-51
532
-46
372
-1
351
5
362
2
323
38
292
44
302
42
320 -43
263
-18
266
-19
355
34
274
49
281
48
493 -65
455
-52
469
-57
1.5
0.8
2.5
28
1
Table 4. The four best performing models compared in the final stage of the study.
2
Co_occ_Band3_mean refers to the mean co-occurrence texture measure for Band 3.
3
Regression Models
Model 1
2848.029 - 6196.912 * Band3
+ 25.071 * HV
Model 2
2826.491 - 7180.492 * Band3
Model 3
3236.006 - 84.115 *
(Co_occ_Band3_mean) +
15.145 * HV
Model 4
3299.722 - 93.026 *
(Co_occ_Band3_mean)
4
5
Adjusted
r2
0.46
Variables
Band3
HV
0.45
Band 3
0.47
Co_occ_Band3_mean
HV
0.49
RMSE
(t/ha)
Bias
(%)
152
-1.0
154
1.5
150
0.8
149
2.5
Co_occ_Band3_mean
29
1
Figure 1. Location of the study area in the central part of Singapore. Background is a SPOT 5
2
image (RGB:432) (SPOT image © 2010 CNES).
3
4
Figure 2. Correlation between measured and estimated (Model 4) biomass in 10 independent
5
validation field sample plots.
6
7
Figure 3. Biomass map of CNR derived from the regression model (Model 1) developed in
8
this study. Note the negative biomass values in vegetation types (shrub and grassland) which
9
had not been sampled during field sampling.
10
11
30
1
2
3
4
Figure 1. Location of the study area in the central part of Singapore. Background is a SPOT 5
5
image (RGB:432) (SPOT image © 2010 CNES).
6
7
8
31
1
2
800
Measured biomass (t/ha)
700
600
500
400
300
200
100
0
0
100
200
300
400
500
600
700
800
Estimated biomass (t/ha)
3
4
5
Figure 2. Correlation between measured and estimated (Model 4) biomass in 10 independent
6
validation field sample plots.
7
8
32
1
2
3
4
20
100
150
200
300
400
600
1400
Biomass t/ha
Negative Biomass
4
5
6
Figure 3. Biomass map of CNR derived from the regression model (Model 1) developed in
7
this study. Note the negative biomass values in vegetation types (shrub and grassland) which
8
had not been sampled during field sampling.
9
10