You are on page 1of 13

Journal of Environmental Management 292 (2021) 112816

Contents lists available at ScienceDirect

Journal of Environmental Management


journal homepage: www.elsevier.com/locate/jenvman

A novel approach for estimation of aboveground biomass of a carbon-rich


mangrove site in India
S.M. Ghosh a, M.D. Behera a, ∗, B. Jagadish a, A.K. Das b, D.R. Mishra c
a
Centre for Oceans, Rivers, Atmosphere and Land Sciences; Indian Institute of Technology Kharagpur, West Bengal, 721302, India
b
Space Applications Centre, ISRO, Ahmedabad, India
c
Department of Geography, University of Georgia, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Mangroves can play a crucial part in climate change mitigation policies due to their high carbon-storing capacity.
Aboveground biomass However, the carbon sequestration potential of Indian mangroves generally remained unexplored to date. In this
Sentinel 1 and 2 study, multi-temporal Sentinel-1 and 2 data-derived variables were used to estimate the AGB of a tropical
Machine learning regression
carbon-rich mangrove forest of India. Ensemble prediction of multiple machine learning algorithms, including
Ensemble modeling
BhitarKanika wildlife sanctuary
Random Forest (RF), Gradient Boosted Model (GBM), and Extreme Gradient Boosting (XGB), were used for AGB
Uncertainty assessment prediction. The multi-temporal dataset was used in two different ways to find the most suitable method of using
them. The results of the analysis showed that the modeling field measured AGB with individual date data values
results in estimates with root mean square errors (RMSE) ranging from 149.242 t/ha for XGB to 151.149 t/ha for
the RF. Modeling AGB with the average and percentile metrics of the multi-temporal image stack improves the
prediction accuracy of AGB, with RMSE ranging from 81.882 t/ha for the XGB to 74.493 t/ha for the RF. The
AGB modeling using ensemble prediction showed further improvement in accuracy with an RMSE of 72.864 t/ha
and normalized RMSE of 11.38%. In this study, the intra-seasonal variation of Sentinel-1 and 2 data for
mangrove ecosystems was explored for the first time. The variations in remotely sensed variables could be
attributed mainly to soil moisture availability and rainfall in the mangrove ecosystem. The efficiency of Sentinel-
1 and 2 data-derived variables and ensemble prediction of machine learning models for Indian mangroves were
also explored for the first time. The methodologies established in this study can be used in the future for accurate
prediction and repeated monitoring of AGB for mangrove ecosystems.

1. Introduction Aboveground Biomass (AGB) serves as a good indicator of vegetation


carbon storage in forests, which almost maintains a thumb rule of 50
Mangrove forests are one of the most carbon-rich ecosystems in the percent of the total biomass (Houghton et al., 2009). Therefore, the
tropics as they can accumulate almost four times carbon per unit area estimation of AGB of mangrove forests would provide more insight into
compared to other forests (Donato et al., 2011). Although mangroves its understudied carbon dynamics. Most of the mangrove forests are
cover only 0.7% of the world’s tropical forests, their high carbon-storing distributed in low-income developing or underdeveloped countries.
capacity emphasizes mangroves’ importance in alleviating climate Therefore, they are understudied concerning their carbon sequestration
change (Giri et al., 2011). Mangrove deforestation is happening at a rate potential, and more so, using multi-sensor and multi-temporal satellite
of 0.39% per year globally (Hamilton and Casey, 2016), contributing to data combined with extensive field measurements. AGB estimation
around 10% of emissions from deforestation alone (Donato et al., 2011). studies increasingly integrate field measurements to remote
Interestingly, the Indian mangrove forests have increased by 181 sensing-based models that provide estimates over time and space (Dash
km2–4921 km2 during 2015–2017 as per the Forest Survey of India es­ et al., 2020). Studies have reported a strong correlation between foliage
timate (FSI, 2017), in contrast to the decrease in mangrove forests of the biomass and trunk biomass and utilized it to predict AGB using vege­
Indian Ocean region such as in Indonesia and Malaysia (Murdiyarso tation indices (VIs) (Roy, 1989; Günlü et al., 2014). However, vegetation
et al., 2015). canopy reflectances used to compute VIs get saturated at higher biomass

∗ Corresponding author.
E-mail address: mdbehera@coral.iitkgp.ac.in (M.D. Behera).

https://doi.org/10.1016/j.jenvman.2021.112816
Received 3 March 2021; Received in revised form 11 May 2021; Accepted 16 May 2021
Available online 21 May 2021
0301-4797/© 2021 Elsevier Ltd. All rights reserved.
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

levels making empirical VI-based AGB estimations erroneous (Mutanga vegetation parameter estimations (Dube et al., 2014; Elith et al., 2008;
and Skidmore, 2004). Longer wavelength L-band SAR data demon­ Pham et al., 2020; Samat et al., 2020). Different machine learning al­
strated a good correlation for tropical forest AGB (Behera et al., 2015). gorithms often produce slightly different but comparable results (Gar­
In contrast, C-band SAR data suffers from the saturation problem at a cía-Gutiérrez et al., 2015; Ghosh et al., 2020). However, an ensemble
low biomass range (Imhoff, 1993). model can combine the predictions from different algorithms and pro­
In recent years, satellite-based AGB estimation research received a duce more robust results, such as for classification (Du et al., 2012;
substantial boost with the availability of high-resolution Sentinel 1 and Tinoco et al., 2013) and forest parameter estimation (Zhang et al.,
2 data. The Sentinel-2 derived VIs demonstrated better relationships 2018).
with AGB in comparison to Sentinel-1 SAR backscatter values (Ghosh Mangroves can play a crucial role in UNFCCC’s Reducing Emissions
and Behera, 2018; Castillo et al., 2017; Chang and Shoshany, 2016). from Deforestation and forest Degradation (REDD+) program to miti­
Berninger et al. (2018) used Sentinel-1 and PALSAR data together for gate climate change (Ahmed and Glaser, 2016). Aziz et al.’s (2016)
AGB estimation of mangroves in Kalimantan (Indonesia) and observed study on Malaysia’s Matang Mangrove Forest Reserve showed the
up to 353 t/ha AGB with a correlation of 0.63. Castillo et al. (2017) implementation of REDD + could help in maintaining ecological and
measured up to 346 t/ha AGB in Palawan (Philippines) with a high socio-economic sustainability. Preparation of AGB maps is a crucial
correlation (0.84) using Sentinel-1 and 2 data synergy and machine component in that aspect as they can help identify the main factors
learning algorithms. However, they did not consider the impact of the behind carbon emission from the forest (Corona-Núñez et al., 2021).
high intra-seasonal variability of the satellite data on AGB estimate. There is an urgent need to prepare the AGB maps for the mangroves of
The VIs derived from optical sensors have considerable intra- India to make them an integral part of the REDD + program. The
seasonal variability (Gonzalez Del Castillo et al., 2018; Ardö et al., methodological framework of this work was aimed at achieving multiple
2014; Chikoore and Jury, 2010). Multiple studies conducted on Indian objectives that are crucial for accurate estimation of mangrove forest
tropical forests showed a rise in VIs values in the post-rainy season (Al AGB, such as (i) an uncertainty assessment of the field measured data
Balasmeh and Karmaker, 2020; Nischitha et al., 2014; Chandrasekar with 304 stratified random quadrats, (ii) finding a suitable method for
et al., 2006), while the VIs of mangrove forests have shown a prominent the use of the multi-temporal data to address field estimated
negative trend (Songsom et al., 2019; Pastor-Guzman et al., 2018). As AGB-remote sensing variable correlation, and (iii) use of ensemble
field sampling for AGB data collection can last from a few days prediction of machine learning models in AGB estimation. The study was
(Kachamba et al., 2017) to months (Gillerot et al., 2018) to years (Mauya undertaken in the mangrove forests of Bhitarkanika Wildlife Sanctuary
et al., 2015; Hamdan et al., 2014) depending on the accessibility and (BWS), Odisha, located on India’s eastern coast at the delta of Brahmani
extent of the study site, there could be several remote sensing datasets and Baitarani rivers (Parida and Kumari, 2020).
available for correlation over the period. As there could be high intra-
and inter-seasonal variations of satellite data due to varied weather 2. Materials and methodology
conditions over a period, the selection of an appropriate method for
using multi-temporal data in AGB estimation modeling is challenging 2.1. Study area and field data collection
and may affect the estimation accuracy.
Sentinel sensors provide several overpasses during a period of field Bhitarkanika Wildlife Sanctuary (BWS) is an estuary located in
measurements. Thus, the accuracy of AGB estimation using Sentinel eastern India adjoining the Bay of Bengal in the delta of Brahmani and
series data would depend on the temporal data use. Huang et al. (2018) Baitarani rivers (Figure 1). BWS accommodates 75 plant species, of
have shown multi-temporal Sentinel-1 data use by averaging to improve which most are mangroves, covering an area of 130 km2 (Pattanaik
the AGB estimate’s accuracy. However, the arithmetic mean of the et al., 2008). BWS experiences a typical warm and humid tropical
multi-temporal data may provide misleading information about the data climate with maximum temperature in May and minimum temperature
trends in the absence of any prior knowledge on data distribution values in January. It receives an average annual rainfall of about 1642 mm,
(Derrible and Ahmad, 2015) and is also affected by extreme values most of which are received in the monsoon season from June to October.
(Gupta, 2012). Forkuor et al. (2020) used single-year time series values BWS is surrounded by 81 villages, triggering anthropogenic pressure on
of Sentinel-1 and Sentinel-2 data for a tropical savannah forest to show the mangrove forests (Pattanaik et al., 2008). Bhitarkanika Wildlife
that dry season images are the most crucial variables in AGB estimation. Sanctuary accommodates furious crocodiles and poisonous snakes such
Wittke et al. (2018) showed that multi-temporal features did not as King Cobra and Python, making field measurements extremely risky.
improve the biophysical parameter retrieval accuracy for a boreal forest. The diurnal tidal effect leads to periodic flooding and devoid of
However, the temporal variation of Sentinel-2 data-derived indices and ground vegetation. A reconnaissance survey, done prior to the field
any appropriate method to handle the data has not been studied yet for campaign, was used to stratify the study areas to cover the heterogeneity
AGB estimation of mangrove forests. in terms of dominant compositions with varied density and girth classes.
AGB estimation using remote sensing variables is generally achieved A field survey was conducted from November to December 2018, and
using parametric regression (Ali et al., 2015). Most studies assume that measurements from 304 stratified random quadrats (each of 20 × 20 m2,
AGB and its optical response follow a standard empirical relationship, 0.04 ha) were collected on the circumference at breast height (CBH) of
either linear, logarithmic, or polynomial (Feliciano et al., 2017; Heis­ each plant. CBH was converted to the diameter at breast height (D), and
kanen, 2006; Sinha et al., 2016). Though these methods are simple to wood density was used from the global wood density database (Chave
implement and produce reasonably accurate results, the AGB and its et al., 2009) to estimate AGB as per Equation (1) (Komiyama et al.,
reflectance response captured in remote sensing images tend to be more 2005). Komiyama et al. (2005) used field data for ten mangrove species
convoluted (Ali et al., 2015). Recent advances in machine learning al­ with a total of 104 stems from five study sites to establish the relation­
gorithms have enabled researchers to accurately model the complex ships, which is commonly used for mangrove AGB estimation (Pham
relationship between AGB and remote sensing variables (Ghosh and et al., 2018; Kamruzzaman et al., 2017; Pham and Brabyn, 2017; Donato
Behera, 2021; Martins Silva et al., 2019; Wu et al., 2019). Decision et al., 2011).
tree-based models such as Random Forest have shown great promise in
AGB = 0.251∗ρ∗D2.46 (1)
forest biophysical parameters estimation (Breiman, 2001), including
AGB (Ghosh et al., 2020; Ghosh and Behera, 2018; Furtado et al., 2016).
where ρ = wood density (in gm/cm3) obtained from the specific gravity
Two other decision tree-based machine learning models, gradient
information as per Chave et al. (2009); and D = Diameter at breast
boosted models (Friedman, 2002) and extreme gradient boosting (Chen
height (in cm).
and Guestrin, 2016), have been used successfully in multiple studies for

2
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Fig. 1. Location of Bhitarkanika Wildlife Sanctuary (BWS) in Eastern India with field quadrate locations shown as yellow coloured dots, overlaid on a natural color
composite of Sentinel 2A images; A network of creeks connect the mangrove forests. (For interpretation of the references to colour in this figure legend, the reader is
referred to the Web version of this article.)

2.2. Uncertainty in field estimated data


2.3. Satellite data and pre-processing
Field estimated AGB are susceptible to multiple sources of errors
The Sentinel-1 and Sentinel-2 data were procured from Copernicus
(Chave et al., 2004). It was not feasible to directly follow the methods of
scientific data hub (Table 1) for post monsoon dry winter season as it
Chave et al. (2004) in the current study as they had a plethora of
matches with the field data collection period. It is also the time when
multi-temporal field measured data and harvested tree data, which are
Sentinel-1 and Sentinel-2 variables were found to be mostly correlated
not available for this study. However, two sources of errors were
with the AGB (Forkuor et al., 2020). Six sets of dual-polarized (VV and
addressed, such as (i) error due to measurement of individual trees, (ii)
VH) ground range detected Sentinel-1 data were procured from
error due to the use of a specific allometric equation. Chave et al. (2004,
November 5, 2018 to January 4, 2019, with 12 days temporal resolu­
2014) showed that measurement errors for trees with >10 cm diameter
tion. All tiles of Sentinel-1 data were calibrated, terrain corrected, and
generally stand around 50% in the tropics. Chave et al. (2014) proposed
resampled to a pixel size of 20 × 20 m2 using SNAP (Sentinel Application
a method to calculate the quadrat level uncertainty based on the indi­
Platform) software. Multi-temporal speckle filtering was applied to the
vidual tree measurement error (equation (2)).
stacked data as it has the advantage of obtaining a better result than
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ 2 single date filtering while preserving the edge information and essential
i [AGBest (i)]
CVplot = CV × ∑ (2) details. Gamma MAP filtering was chosen due to its suitability for
i AGBest (i)
vegetated areas (Lopes et al., 1990; Huang and Genderen van, 1996).
The intensity values were converted into decibel for further utilisation.
where AGBest (i) is the AGB of the i-th tree in the quadrat, CV is the
The Level-1C orthorectified Sentinel-2 cloud-free images acquired on
average uncertainty of each tree, i.e., 50% and CVplot is the quadrat-
16th, 21st, 26th November, 11th, 26th, December 31, 2018, and
based uncertainty. The uncertainty also depends on the allometric
January 5, 2019 were subjected to atmospheric correction using the
equation. In this study, the Komiyama et al. (2005) model was chosen as
SEN2COR processor (Louis et al., 2016). As Sentinel-2 data has 13 bands
it has been successfully implemented in several mangrove studies. Two
at different spatial resolutions, they were resampled to 20 × 20 m2. This
global models (Equations (3) and (4)) by Chave et al. (2005) and another
datasets were used for further processing and eventually AGB modeling
model (Equation (5)) developed by Ray et al. (2011) for the Sunderbans
(see Figure 2).
mangrove were selected to estimate the AGB. Only the tallest canopy
height was measured in the field, and that was used in the equations as
2.4. Generation of vegetation indices (VIs)
the H. The standard deviation from these AGB measurements was ob­
tained to get the uncertainty in the allometry. The total uncertainty
Eight VIs including normalized difference vegetation index (NDVI),
combined from two sources was estimated as per the method followed
soil adjusted vegetation index (SAVI), green NDVI (GNDVI), normalized
by Saatchi et al. (2011), assuming that all the errors are independent and
random (Equation (6)).
Table 1
AGB = 0.0509∗ρ∗D2 ∗H (3) Details of the Sentinel 1 and 2 dataset used is AGB estimates.
( ) Data Satellite Processing Acquisition date
AGB = ρ∗exp − 1.349 + 1.980 ∗ ln D + 0.207 ∗ (ln D)2 − 0.0281 ∗ (ln D)3 type level
(4) SAR Sentinel- GRD 5th, 17th, 29th November; 11th,
1A December 23, 2018; January 4, 2019
AGB = 1.0471∗ρ− 1.37
∗D0.864 ∗H 0.635 (5) Optical Sentinel- L1C 16th, 26th November; December 26,
2A 2018; January 5, 2019
( )1/2 Sentinel- L1C 21st November; 11th, December 31,
Uncertaintytotal = Umeasure 2 + Uallometry 2 (6) 2B 2018

3
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Fig. 2. Methodological flow diagram for AGB estimation of BWS showing the steps of data pre-processing, data processing and AGB modeling using multiple machine
learning models.

difference index produced using band 4 and 5 of Sentinel-2 data used together to establish the models (VSall). Next, the mean, median,
(NDI45), inverted red-edge chlorophyll index (IRCEI), transformed and percentile metrics of the multitemporal image stack were used
normalized difference vegetation index (TNDVI), enhanced vegetation together as predictor variables (VSmeanp). Apart from the remote sensing
index (EVI), and two bands enhanced vegetation index (EVI2) were variables, two other variables were included in the analysis. First, the
generated using the surface reflectance data for individual date images distance from the boundary, which can account for the external influ­
(Table 2). The backscatter value ratio, average, and the square root of ence on AGB content. Second, the impact of salinity by including the
the product were generated using Sentinel-1 VV and VH polarised data. distance from the nearest creek. The root mean square error (RMSE) and
Temporal coverage of both Sentinel-1 and 2 data, including VIs and the coefficient of determination between field-measured and model-
SAR backscatter images, were matched with the period of field mea­ estimated AGB values were obtained to evaluate the model accuracy.
surements. VIs and SAR backscatter values for all the quadrat locations The normalized RMSEs were also derived for comparison with the re­
were incurred by averaging all pixels extracted with a 30 m radius center sults from other studies. F-test was conducted by comparing the sum of
of quadrat locations to accommodate any positional error of GPS. The squared errors of the model predicted values based on both VSmeanp and
variable values from single date images were extracted at first. The mean VSall variables to check if the results from models with VSmeanp and VSall
and the median values represented by the 50th percentile were extracted variables have a statistically significant difference. The RF, GBM, and
further using the multi-temporal image stacks for the VIs and the XGB were implemented individually using the CARET package in R
backscatters. The 10th and 90th percentile metrics of the multi-temporal statistical software (Kuhn, 2008). Then, based on the accuracy, the
image stack were also extracted to accommodate extreme values of prediction of each model was given a weight. The final predictions were
either end. made using the individual model prediction and associated weight. An
uncertainty of prediction was also prepared using the coefficient of
2.5. AGB estimation using ensemble modeling variation between predictions. AGB maps were generated using indi­
vidual models. The final map was prepared by using the weights
The modeling was done twice with two different sets of input vari­ calculated as:
ables. At first, all single date values for VIs and SAR backscatter were ∑ Ri Pi
B= ∑N
i=1 Ri
Table 2
List of vegetation indices selected for AGB estimation. The number followed ’B’ where B is the final prediction; Pi is the prediction of model i; Ri is the
is the actual location of the band in Sentinel-2 data. The numbers in subscripts coefficient of determination of the model i; and N is the total number of
represent the central wavelength value of the spectral band. models in the ensemble analysis, i.e., 3 in this case.
Index Formulation (R = reflectance) Reference

NDVI [(B8)0.842 – (B4)0.665)/((B8)0.842 + (B4)0.665] (Rouse et al., 1973) 3. Results


GNDVI [(B8)0.842 – (B3)0.560)/((B8)0.842 + (B3)0.560)] (Gitelson and
Merzlyak, 1996) 3.1. Field data
IRCEI [(B7)0.783 – (B4)0.665)/((B5)0.705/(B6)0.740] (Frampton et al.,
2013)
NDI45 [(B5)0.705 – (B4)0.665)/((B5)0.705 + (B4)0.665] (Delegido et al., 2011)
Five species dominate the BWS mangrove forest, such as Heritiera
TNDVI [((B8)0.842 – (B4)0.665)/((B8)0.842 + (B4)0.665) (Tucker, 1979) fomes, Excoecaria agallocha, Avicennia officinalis, Ceriops decandra
+ 0.5]1/2 Cynometra irripa, of which the first three forms the top canopy. In
SAVI (1+L)*[(B8)0.842 – (B4)0.665)/((B8)0.842 + (Huete, 1988) contrast, Ceriops decandra and Cynometra irripa generally remain
(B4)0.665 + L]; L = 0.5
understorey. Heritiera fomes occur as the most dominant species, fol­
EVI 2.5*[(B8)0.842 – (B4)0.665)/((B8)0.842 + 6* (Huete et al., 1999)
(B4)0.665–7.5*(B2)0.490 + 1] lowed by Excoecaria agallocha, which comprises nearly 90 percent of
EVI2 2.5*[(B8)0.842 – (B4)0.665)/((B8)0.842 + 2.4* (Jiang et al., 2008) the species distribution (Figure 3a). Avicennia officinalis have massive
(B4)0.665 + 1] growth with high AGB and sparse distribution in the top canopy.

4
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Fig. 3. Charts showing basic statistics of field data - (a) Dominant species and their frequency (b) DBH statistics, maximum (left), mean (middle), and minimum
(right), of different species (c) AGB range and the related number of plots.

Sonneratia apetala are found abundantly along river banks. Avicennia 3.2. Uncertainty in the field measured data
officinalis has the highest average diameter (8–87 cm), followed by
Heritiera fomes (3–50 cm), Excoecaria agallocha (3–50 cm), Cynometra The uncertainty in the field measured AGB data were estimated using
irripa (3–32 cm), and Ceriops decandra (Figure 3b). Tree density per two primary sources of possible errors – (i) individual tree measurement
quadrat varied from 26 to 140, with seventy-seven quadrats accom­ error (ii) allometric equation error. The uncertainty at the quadrat level
modated 60–100 trees per 0.04ha quadrat (Figure 3c). The AGB density varied with an average of 15% (5%–37%), which could be due to the
of quadrats was generally low (Figure 3d). The majority of the quadrats erroneous CBH measurement of trees in the field. The highest average
have an AGB up to 400 t/ha, with fewer quadrats accounting for very AGB of 253.08 t/ha was found using Equation (4) (Chave et al., 2005),
low (<100 t/ha) or high (>500 t/ha) AGB values. Interestingly, it was the lowest average AGB of 117.5 t/ha was estimated using Equation (5)
noticed that few individual trees of Avicennia officinalis species in a (Ray et al., 2011), resulting in an allometric equation uncertainty of
quadrat contribute to very high AGB. In contrast, several individuals of 30%. Thus, the overall uncertainty of field estimated AGB was up to 33%
Excoecaria agallocha contribute to low biomass in a quadrat. As ex­ as per Equation (6) (Saatchi et al., 2010). The field estimated AGB un­
pected, the tree density and AGB values per quadrat followed an inverse certainty was combined with the ensemble model predicted AGB as per
relation owing to tree age and maturity. the Vorster et al. (2020) methodology to get the final propagated

Fig. 4. Importance of top 20 variables while using individual date image values (VSall) as predictor in (a) GBM (b) RF and (c) XGB model. TNDVI, GNDVI and NDVI
shows more importance than other VIs and backscatter values.

5
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

uncertainty. error bars (Figure 6). The prediction error varied widely between plots
from 1.261 t/ha to 107.411 t/ha.

3.3. Ensemble modeling of AGB


3.4. Generation of AGB maps for BWS
The variable importance plots for VSall variables showed VIs gener­
ally worked as crucial variables for all three models. The influence of The AGB maps were prepared only for models with VSmeanp variables
NDVI, GNDVI, and TNDVI was more prominent among the VI variables as they showed better performance in AGB prediction. Most of the
(Figure 4). In the models with VSmeanp variables, the average values of Bhitarkanika Wildlife Sanctuary (BWS) demonstrated high AGB density
NDVI and TNDVI multi-temporal stack come across as the most useful by all models, and they showed a similar trend (Figure 7). The majority
variables. Their influence was greater than the rest of the variables. The of the BWS showed AGB density between 200 and 600 t/ha, while low
percentile metrics of multi-temporal image stacks also showed some AGB areas could be found across edges between forest and non-forest,
importance. However, SAR variables were primarily absent from the top and forest and water. The areas with very high AGB were at the core
20 essential variables list for all models (Figure 5). The non-remotely of the forest with the highest distance from the boundary. The variation
sensed variables, i.e., distance from boundary and distance from the in model prediction was the measure of prediction uncertainty. It was
creek, also did not appear in the list of essential variables, signifying measured as the coefficient of variation in percentage. The final un­
their lack of importance. certainty was calculated as per the Vorster et al. (2020) to include the
The modeling with the VSall variables showed poor results for AGB error propagated from field estimation of AGB (Figure 8). The total
prediction (Table 3). The model training RMSE remained considerably uncertainty in AGB predictions was found to be within 30–50%. Most
high, ranging from 158.04 t/ha for the RF model to 160.57 t/ha for the parts of the study area showed uncertainty in the lower range. Only a
XGB model. The model validation RMSE was also similar, ranging from small portion of the area exhibited a higher uncertainty, mainly near the
149.242 t/ha for the XGB model to 151.149 t/ha for the RF model. The forest boundary or the creeks.
coefficient of determination between field estimated and model pre­
dicted AGB also remained low for model training and validation. GBM 4. Discussion
model showed the best coefficient of determination during model
training with a value of 0.18. The coefficient of determination value for 4.1. High AGB in BWS mangrove forest and uncertainty estimation
validation remained at 0.14 for all the models.
The model training results were also similar for all models estab­ BWS accommodates high average AGB as observed from both field
lished using VSmeanp variables (Table 4). RF showed the highest coeffi­ data analysis and remote sensing-based estimation, wherein AGB den­
cient of determination and RMSE with values of 0.75 and 75.584 t/ha, sity varied from 53 t/ha to 741 t/ha for field-collected data. Pandey et al.
respectively. The model training results of RF and XGB were similar. In (2019) reported that the Sonneratia apetala and Cynometra iripa species
model validation results, RF showed the least RMSE of 74.493 t/ha and have the highest AGB of 643.12 t/ha and 652.14 t/ha, respectively,
the highest coefficient of determination of 0.74. Between the other two while Bal and Banerjee (2020) reported total biomass of 866.67 ±
algorithms, GBM showed results similar to RF, but XGB showed 166.10 t/ha in BWS. These numbers contrast to a very low AGB of
considerably poor results. The coefficient of determination values for the 3–50.9 t/ha (Joshi and Ghose, 2014; Manna et al., 2014), up to 113.67
model validation was taken into consideration for the calculation of the t/ha (Banerjee et al., 2013) reported from proximal Sundarbans
weights for the final prediction. The RMSE and coefficient of determi­ mangrove forest in India. The high AGB density of BWS can be attributed
nation improved to 72.864 t/ha and 0.76, respectively, when weighted to its species composition, as indicated by Pandey et al. (2019). The low
average predicted values were plotted against the field estimated AGB AGB reported by Manna et al. (2014) for Sundarbans attributed to a
values. The result of the F-test shows that the difference in predicted young Avicennia marina plantation, while Joshi and Ghose (2014)
values for VSmeanp and VSall variables based models are statistically conducted the study on smaller trees. BWS is dominated by mature
significant with a p-value less than 0.01 (Table 5). Heritiera fomes (DBH up to 87 cm) and large Avicennia officinalis (DBH
As multiple predictions were considered to get the final estimated up to 50 cm) trees, accounting for a higher AGB in contrast to the
AGB, the prediction uncertainty was calculated as the standard devia­ Sunderbans mangroves. However, AGB density close to or higher than
tion of all the measurements. The uncertainties in the field estimated 500 t/ha is quite common in the mangroves of the Indo-Pacific region
and model predicted AGB were shown in the correlation plot through (Kauffman and Donato, 2012). The AGB density of mangroves reported

Fig. 5. Importance of top 20 variables while using average and percentile values (VSmeanp) as predictor in the (a) GBM (b) RF and (c) XGB model.

6
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Table 3
Performance of the models with the training data and the validation data while using individual date image values as predictor.
Method Model R2 Model RMSE (t/ha) Model nRMSE (%) Validation R2 Validation RMSE (t/ha) Validation nRMSE (%)

XGB 0.13 160.57 22.26 0.14 149.242 22.07


GBM 0.18 158.42 21.96 0.14 150.427 22.25
RF 0.16 158.04 21.91 0.14 151.149 22.35

Table 4
Performance of the models with the training data and the validation data while using mean, median and percentile values as predictor.
Method Model R2 Model RMSE (t/ha) Model nRMSE (%) Validation R2 Validation RMSE (t/ha) Validation nRMSE (%)

XGB 0.73 79.853 11.07 0.69 81.882 12.11


GBM 0.74 76.495 10.6 0.74 75.102 11.11
RF 0.75 75.584 10.48 0.74 74.493 11.02

AGB. It may happen due to the high temporal variation of remote


Table 5 sensing variables which was observed in the study area. Even for the
F-score of the models for the predicted AGB values with validation data to test
cloud-free, atmospherically corrected data from the same sensor, there
the statistical significance of the AGB prediction by different models.
are considerable variations in VIs and backscatter values within 50 days
Method F-score p-value (from F-table) (Figures 9 and 10). The vegetation reflectance varies depending on
XGB 2.612 <0.01 multiple factors such as moisture content, plant type, and orientation of
GBM 3.388 <0.01 vegetation canopy (Roy, 1989). As a result, the remote sensing param­
RF 3.506 <0.01
eters captured by sensors would also differ from time to time. In this
study, Sentinel-1 backscatter and Sentinel-2 derived VIs showed
from other regions in Southeast Asia includes Vietnam from 30 to 596 considerable variation, even though the observation period was only
t/ha (Pham and Brabyn, 2017), Philipines from 1.1 to 346 t/ha (Castillo around two months. In the winter season, VIs generally had a negative
et al., 2017), Indonesia and Malayasia around 400 t/ha (Hamdan et al., trend due to mangrove phenology (Pastor-Guzman et al., 2018; Songsom
2014; Wicaksono et al., 2016). Thus, the AGB density of BWS mangroves et al., 2019). However, the change was considerably less. Roy et al.
is on par with other studies conducted in other parts of the Indo-Pacific (2017) showed that for Sentinel-2 MSI, changes in sensor viewing and
region. sun angle affect the surface reflectance. However, the sensor viewing
The accuracy of remote sensing-based AGB estimates was evaluated angle and sun angle for the study area remains almost the same
by the field-based AGB estimates considering they are the true values. throughout the study period (Table 6). As the time interval is too small
However, the field-based methods are prone to several errors, which for any significant shift in species composition and its orientation, much
introduces uncertainty in the field estimated AGB (Chave et al., 2004). of the change in indices values can be traced to the rainfall events in
Therefore, these errors propagate to the remote sensing-based AGB es­ BWS. Pastor-Guzman et al. (2018) analyzed the variation of EVI, NDVI,
timate when using a model (Chen et al., 2015; Vorster et al., 2020). GNDVI for the Yucatan Peninsula, Mexico, over three years and found
Thus, estimation of uncertainty of field measurements offers an insight that mangrove greenness reacts to rainfall with a lag. The VI curves
into the accuracy assessment of the AGB prediction using remote sensing showed alternate high and low values from November to early January.
data. The high uncertainty (up to 33%) with the field-based AGB esti­ Guan et al. (2015) and Suepa et al. (2016) have shown that the variation
mate was attributed to two sources of errors that are either avoidable or
should be taken care of in future studies. Measurements of accurate AGB
density of the forest with low uncertainty depend on the quadrat size.
Larger the quadrat size lesser the error (Chave et al., 2004). Hetzer et al.
(2020) showed that smaller quadrats could be efficient if their numbers
are large enough. Chave et al. (2004) used the standard deviation of AGB
for 1 ha quadrats and their smaller subquadrats to examine the effect of
quadrat size on AGB estimation uncertainty. In the absence of
bigger-size quadrats or subquadrats, this uncertainty cannot be quanti­
fied. Smaller quadrats (0.04 ha) ought to be replaced with either larger
ones (≥1ha) or more representative field quadrats to minimize errors in
the AGB estimate (Chave et al., 2004; Hetzer et al., 2020). However,
risky ground conditions with crocodiles and poisonous snakes often
hinder field measurement in these protected areas where many other
conservation impositions prevail. As the primary source of allometric
uncertainty was the sample of harvested trees used to build the equation,
Chave et al. (2004) concluded that these uncertainties could not be
investigated further without detailed information on harvested trees. It
is difficult to build new allometric equations for a protected site like
BWS, as the harvest of trees is strictly prohibited. Therefore, the un­
certainty due to the use of allometric equations will remain.

4.2. Temporal variation of remote sensing variables


Fig. 6. Correlation plot of ensemble model predicted AGB and field estimated
The individual single date images do not work as a good predictor of AGB with uncertainties showed as error bars.

7
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Fig. 7. Maps showing the AGB values (in t/ha) estimated using (a) GBM model (b) RF model (c) XGB model and (d) Ensemble model. Most of the area show a AGB
between 200 and 600 t/ha.

of VIs is closely related to the seasonal rainfall in tropical regions, of­


fering a distinct intra-seasonal variation of VIs.
Similarly, it was observed here that the amount of precipitation and
soil moisture affected the Sentinel-2 based VI values. The Global Pre­
cipitation Measurement (GPM) mission (Skofronick-Jackson et al.,
2017) data revealed that the study area received very little rainfall in
November 2018 (Figure 11a). As a result, VIs values were also reduced
to the lowest level in the early days of December 2018. Further, a sig­
nificant amount of rainfall in mid-December 2018 resulted in an in­
crease in greenness for the subsequent days. Yet, due to no rain in late
December 2018 and early January 2019, indices values decreased
further. The soil moisture values obtained from Soil Moisture Active
Passive (SMAP) mission (Entekhabi et al., 2010) data showed a declining
trend for November 2018 (Figure 11b). However, in December 2018, the
pattern stabilized for a period coinciding with precipitation, reflecting
an increase in the VIs values. Again in early January 2019, soil moisture
value decreased along with the VIs values. The fluctuations in soil
moisture values could vary as per the diurnal fluctuation in creek areas
in BWS. The effect of spatial heterogeneity on the soil moisture retrieval
from SMAP data was pointed out by Abbaszadeh et al. (2019). As the
BWS has a complex surrounding, which includes mangrove forest,
agricultural land, human settlements, and a considerable portion of
water bodies, it affects the soil moisture retrieval and results in high
variability.

4.3. Efficiency of sentinel data in AGB estimation


Fig. 8. AGB prediction uncertainty map combining the field measurement and
model predicted uncertainty. Most of the area show an uncertainty between 30
and 40%. In this study, VI-based model estimates showed a lower RMSE value
and a better coefficient of determination between observed and pre­
dicted AGB than SAR data. Previous works have demonstrated that VIs
generally saturate at higher AGB (Reddersen et al., 2014; Tilly et al.,
2015). However, Sentinel-2 data combines better spatial, temporal, and

8
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Fig. 9. Variation pattern of (a) EVI, (b) EVI2, (c) SAVI, (d) NDI45, (e) GNDVI, (f) NDVI, (g) IRECI, and (h) TNDVI over field measured AGB location plots during Nov
17, 2018 to January 05, 2019. The solid line represents the average value of all the location. Error bars shows the range as twice standard deviation.

Fig. 10. Variation pattern of SAR backscatter values (sigma nought) for (a) VH and (b) VV polarization over field measured AGB location plots during Nov 05, 2018
to January 04, 2019 period of field visit. The solid line represents the average value of all the location. Error bars shows the range as twice standard deviation.

Table 6
Average viewing angle and sun angle obtained from the Sentinel-2 data for the study area.
November 16, 2018 November 21, 2018 November 26, 2018 December 11, 2018 December 26, 2018 December 31, 2018 January 05, 2019

View Zenith 5.912264 5.818648 5.964154 5.853759 5.9123 5.876508 5.906553


View Azimuth 97.75385 98.76692 97.92948 98.80016 97.76328 98.80896 97.74898
Sun Zenith 42.50941 43.69067 44.77271 47.20753 48.24582 48.24353 48.08
Sun Azimuth 157.1123 157.4463 157.5494 156.9239 155.0666 154.2573 153.3256

spectral resolution than other available optical datasets. It also has the influence of individual bands of Sentinel-2, as they were not used indi­
advantage of three well-placed red-edge bands that make it more suit­ vidually in this work.
able for monitoring vegetation health information AGB estimation (Han Sentinel-1 C-band SAR data have only two polarization, such as VH
et al., 2017). Sentinel-based VIs have already demonstrated their po­ and VV. In forests, the incoming vertical polarized radiation would
tential in estimating AGB in recent studies (Chen et al, 2018, 2019; mostly get depolarized and backscattered in horizontal polarization,
Quirós Vargas et al., 2019). However, it was not possible to analyze the allowing VH polarization to have a better correlation with AGB, as

9
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Fig. 11. (a) Daily Rainfall and (b) Soil Moisture variation plotted against average values of vegetation indices during November 01, 2018–January 05, 2019.

observed here as the best predictor. Since both Sentinel C-band SAR and wherein each tree output is the mean of the values in each terminal node
optical data mainly deal with leaf properties due to low penetrating of the tree. Thus, the average for a set of values must be within the value
power, their combined use demonstrated little scope for improvement in range. Accordingly, in the RF predicted map, the maximum predicted
AGB estimation accuracy. However, as VIs are a better predictor of fo­ AGB value was 644 t/ha, and areas with AGB more than 600 t/ha were
liage properties and a strong relationship between foliage biomass and very few (Fig. 7). In contrast, field-collected data showed places where
trunk biomass, the VIs emerged as better predictors of AGB. the AGB value goes beyond 700 t/ha (Fig. 3d). Therefore, in the absence
of a significantly higher and lower AGB values per the field estimate, the
training sample could have underestimated the higher AGB values and
4.4. AGB estimation accuracy compared to previous works
overestimated the lower range values in the RF model.
The correlation between observed and predicted AGB obtained in
5. Conclusions
this study is higher than most other mangroves AGB estimation studies
which did not use Sentinel-1 and 2 data. The normalized RMSE of
This study showed that Sentinel-2 data suffers from high temporal
13.82% reported by Pham and Brabyn (2017) was higher than this
variation within a short period, mainly due to a change in precipitation.
study. Hamdan et al. (2014) obtained a normalized RMSE of 10.23%,
The temporal variations in remote sensing variables have not been
but they used PALSAR L-band data, which is more suitable for AGB
addressed in earlier AGB estimation studies. The fluctuation of satellite-
estimation but not freely available. Castillo et al. (2017) study used
derived parameters indicates that estimating AGB without addressing
Sentinel 1 and 2 data for AGB mapping of a mangrove area with
this issue can result in erroneous estimation. In this study, it was shown
field-estimated AGB up to 346 t/ha. The normalized RMSE (13.09%)
that the best way to counter this problem and use them for AGB esti­
observed by Castillo et al. (2017) was marginally higher than this study.
mation is to use the average values of the multi-temporal dataset. The
Navarro et al. (2019) got the lowest RMSE and best coefficient of
ensemble prediction of machine learning algorithms measures the AGB
determination between observed and predicted. However, the field AGB
with greater accuracy than previous studies.
range for Navarro et al. (2019) was only up to 36.93 t/ha. This study
Accurate remote estimation of mangrove AGB has become feasible
showed that even for very high AGB density mangrove forests, AGB
with the free availability of high-resolution Sentinel data and advance­
could be estimated with high accuracy using only Sentinel-1 and 2 data.
ment of machine learning algorithms. Though biomass studies generally
prefer expensive longer wavelength SAR data, the combined use of open-
4.5. Suitability of the AGB estimation models sourced Sentinel data can be a suitable alternative. In the near future,
ESA’s P-band BIOMASS mission (Ho Tong Minh et al., 2015), the NASA
All the models used in this study had a different set of essential and ISRO joint L and S-band SAR mission, NISAR (Rosen et al., 2015),
variables, and their magnitudes were also different. As a result, the Argentina’s L-band SAOCOM mission (D’Aria et al., 2008), and the
predicted maps show minor differences in AGB values. However, all German Aerospace Center (DLR)’s L-band SAR instrument (Moreira
decision tree-based models suffer from the extrapolation problem et al., 2011), are expected to be operational. However, due to a lack of
(Hengl et al., 2018; Malistov and Trushin, 2019). These regression open-source data availability, dependence on Sentinel data will remain
models cannot predict values outside the range of the training data as for AGB estimation, especially from underdeveloped and developing
they are based on averaging the values of multiple outputs. Further, countries.
averaging the results of many trees, final predictions were made,

10
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Credit author Chave, J., Coomes, D., Jansen, S., Lewis, S.L., Swenson, N.G., Zanne, A.E., 2009. Towards
a worldwide wood economics spectrum. Ecol. Lett. 12, 351–366. https://doi.org/
10.1111/j.1461-0248.2009.01285.x.
SM Ghosh: Conceptualization, Methodology, Formal analysis, Chave, J., Réjou-Méchain, M., Búrquez, A., Chidumayo, E., Colgan, M.S., Delitti, W.B.C.,
Writing – original draft, MD Behera: Methodology, Writing – review & Duque, A., Eid, T., Fearnside, P.M., Goodman, R.C., Henry, M., Martínez-Yrízar, A.,
editing, Funding acquisition, Supervision, B Jagadish: Investigation, Mugasha, W.A., Muller-Landau, H.C., Mencuccini, M., Nelson, B.W., Ngomanda, A.,
Nogueira, E.M., Ortiz-Malavassi, E., Pélissier, R., Ploton, P., Ryan, C.M.,
Data curation, AK Das: Writing – review & editing, DR Mishra: Meth­ Saldarriaga, J.G., Vieilledent, G., 2014. Improved allometric models to estimate the
odology, Writing – review & editing. aboveground biomass of tropical trees. Global Change Biol. 20, 3177–3190. https://
doi.org/10.1111/gcb.12629.
Chen, L., Wang, Y., Ren, C., Zhang, B., Wang, Z., 2019. Optimal combination of
predictors and algorithms for forest above-ground biomass mapping from Sentinel
Declaration of competing interest and SRTM data. Rem. Sens. 11 https://doi.org/10.3390/rs11040414.
Chen, Q., Vaglio Laurin, G., Valentini, R., 2015. Uncertainty of remotely sensed
aboveground biomass over an African tropical forest: propagating errors from trees
The authors declare that they have no known competing financial to plots to pixels. Remote Sens. Environ. 160, 134–143. https://doi.org/10.1016/j.
interests or personal relationships that could have appeared to influence rse.2015.01.009.
the work reported in this paper. Chen, T., Guestrin, C., 2016. XGBoost. Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining. ACM, New
York, NY, USA, pp. 785–794. https://doi.org/10.1145/2939672.2939785.
Acknowledgments Chikoore, H., Jury, M.R., 2010. Intraseasonal variability of satellite-derived rainfall and
vegetation over Southern Africa. Earth Interact. 14 https://doi.org/10.1175/
2010EI267.1.
SMG acknowledges the Fellowship received from MOE for Ph.D. Corona-Núñez, R.O., Mendoza-Ponce, A.V., Campo, J., 2021. Assessment of above-
research. MDB acknowledges the financial support received from SAC, ground biomass and carbon loss from a tropical dry forest in Mexico. J. Environ.
Manag. 282, 111973. https://doi.org/10.1016/j.jenvman.2021.111973.
ISRO, through the NISAR grant that helped in field-sampling. The sup­
D’Aria, D., Giudici, D., Monti Guarnieri, A., Rizzoli, P., Medina, J., 2008. A Wide Swath,
port provided by Mr. Somnath and Mr. Jayaprakash during field mea­ Full Polarimetric, L Band Spaceborne SAR. 2008 IEEE Radar Conf, pp. 5–8. https://
surements is thankfully acknowledged. MDB and SMG thank Prof. J doi.org/10.1109/RADAR.2008.4720789. RADAR 2008.
Dash of the University of Southampton, UK, for his valuable suggestions Dash, J., Behera, M.D., Jeganathan, C., Jha, C.S., Sharma, S., Lucas, R., Khuroo, A.A.,
Harris, A., Atkinson, P.M., Boyd, D.S., Singh, C.P., Kale, M.P., Kumar, P., Behera, S.
to improve the study. Thanks to the Odisha state Forest Wildlife division K., Chitale, V.S., Jayakumar, S., Sharma, L.K., Pandey, A.C., Avishek, K., Pandey, P.
for granting permission to conduct fieldwork in Bhitarkanika Wildlife C., Mohapatra, S.N., Varshney, S.K., 2020. India’s contribution to mitigating the
Sanctuary, India. impacts of climate change through vegetation management. Trop. Ecol. 61,
168–171. https://doi.org/10.1007/s42965-020-00075-9.
Derrible, S., Ahmad, N., 2015. Network-based and binless frequency analyses. PloS One
References 10, 1–29. https://doi.org/10.1371/journal.pone.0142108.
Donato, D.C., Kauffman, J.B., Murdiyarso, D., Kurnianto, S., Stidham, M., Kanninen, M.,
2011. Mangroves among the most carbon-rich forests in the tropics. Nat. Geosci. 4,
Abbaszadeh, P., Moradkhani, H., Zhan, X., 2019. Downscaling SMAP radiometer soil
293–297. https://doi.org/10.1038/ngeo1123.
moisture over the CONUS using an ensemble learning method. Water Resour. Res.
Du, P., Xia, J., Zhang, W., Tan, K., Liu, Y., Liu, S., 2012. Multiple classifier system for
55, 324–344. https://doi.org/10.1029/2018WR023354.
remote sensing image classification: a review. Sensors 12, 4764–4792. https://doi.
Ahmed, N., Glaser, M., 2016. Coastal aquaculture, mangrove deforestation and blue
org/10.3390/s120404764.
carbon emissions: is REDD+ a solution? Mar. Pol. 66, 58–66. https://doi.org/
Dube, T., Mutanga, O., Elhadi, A., Ismail, R., 2014. Intra-and-inter species biomass
10.1016/j.marpol.2016.01.011.
prediction in a plantation forest: testing the utility of high spatial resolution
Al Balasmeh, O.I., Karmaker, T., 2020. Effect of temperature and precipitation on the
spaceborne multispectral RapidEye sensor and advanced machine learning
vegetation dynamics of high and moderate altitude natural forests in India. J. Indian
algorithms. Sensors 14, 15348–15370. https://doi.org/10.3390/s140815348.
Soc. Remote Sens. 48, 121–144. https://doi.org/10.1007/s12524-019-01065-8.
Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to boosted regression trees.
Ali, I., Greifeneder, F., Stamenkovic, J., Neumann, M., Notarnicola, C., 2015. Review of
J. Anim. Ecol. 77, 802–813. https://doi.org/10.1111/j.1365-2656.2008.01390.x.
machine learning approaches for biomass and soil moisture retrievals from remote
Entekhabi, D., Njoku, E.G., O’Neill, P.E., Kellogg, K.H., Crow, W.T., Edelstein, W.N.,
sensing data. Rem. Sens. 7, 16398–16421. https://doi.org/10.3390/rs71215841.
Entin, J.K., Goodman, S.D., Jackson, T.J., Johnson, J., Kimball, J., Piepmeier, J.R.,
Ardö, J., Eklundh, L., Seaquist, J., Jamali, S., Jönsson, P., 2014. Detecting changes in
Koster, R.D., Martin, N., McDonald, K.C., Moghaddam, M., Moran, S., Reichle, R.,
vegetation trends using time series segmentation. Remote Sens. Environ. 156,
Shi, J.C., Spencer, M.W., Thurman, S.W., Tsang, L., Van Zyl, J., 2010. The soil
182–195. https://doi.org/10.1016/j.rse.2014.09.010.
moisture active passive (SMAP) mission. Proc. IEEE 98, 704–716. https://doi.org/
Bal, G., Banerjee, K., 2020. Carbon storage potential of tropical wetland forests of South
10.1109/JPROC.2010.2043918.
Asia - a case study. Environ. Monit. Assess. 191 https://doi.org/10.1007/s10661-
Feliciano, E.A., Wdowinski, S., Potts, M.D., Lee, S.K., Fatoyinbo, T.E., 2017. Estimating
019-7690-y.
mangrove canopy height and above-ground biomass in the Everglades National Park
Banerjee, K., Sengupta, K., Raha, A., Mitra, A., 2013. Salinity based allometric equations
with airborne LiDAR and TanDEM-X data. Rem. Sens. 9 https://doi.org/10.3390/
for biomass estimation of Sundarban mangroves. Biomass Bioenergy 56, 382–391.
rs9070702.
https://doi.org/10.1016/j.biombioe.2013.05.010.
Forkuor, G., Benewinde Zoungrana, J.B., Dimobe, K., Ouattara, B., Vadrevu, K.P.,
Behera, M.D., Tripathi, P., Mishra, B., Kumar, S., Chitale, V.S., Behera, S.K., 2015. Above-
Tondoh, J.E., 2020. Above-ground biomass mapping in West African dryland forest
ground biomass and carbon estimates of Shorea robusta and Tectona grandis forests
using Sentinel-1 and 2 datasets - a case study. Remote Sens. Environ. 236 https://
using QuadPOL ALOS PALSAR data. Adv. Space Res. https://doi.org/10.1016/j.
doi.org/10.1016/j.rse.2019.111496.
asr.2015.11.010.
Friedman, J.H., 2002. Stochastic gradient boosting. Comput. Stat. Data Anal. 38,
Berninger, A., Lohberger, S., Stängel, M., Siegert, F., 2018. SAR-based estimation of
367–378. https://doi.org/10.1016/S0167-9473(01)00065-2.
above-ground biomass and its changes in tropical forests of Kalimantan using L- and
FSI, 2017. State of forest report.
C-band. Rem. Sens. 10 https://doi.org/10.3390/rs10060831.
Furtado, L.F. de A., Silva, T.S.F., Novo, E.M.L., de, M., 2016. Dual-season and full-
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/A:
polarimetric C band SAR assessment for vegetation mapping in the Amazon várzea
1010933404324.
wetlands. Remote Sens. Environ. 174, 212–222. https://doi.org/10.1016/j.
Castillo, J.A.A., Apan, A.A., Maraseni, T.N., Salmo, S.G., 2017. Estimation and mapping
rse.2015.12.013.
of above-ground biomass of mangrove forests and their replacement land uses in the
García-Gutiérrez, J., Martínez-Álvarez, F., Troncoso, A., Riquelme, J.C., 2015.
Philippines using Sentinel imagery. ISPRS J. Photogrammetry Remote Sens. 134,
A comparison of machine learning regression techniques for LiDAR-derived
70–85. https://doi.org/10.1016/j.isprsjprs.2017.10.016.
estimation of forest variables. Neurocomputing 167, 24–31. https://doi.org/
Chandrasekar, K., Sai, M.V.R.S., Jeyaseelan, A.T., Dwivedi, R.S., Roy, P.S., 2006.
10.1016/j.neucom.2014.09.091.
Vegetation response to rainfall as monitored by NOAA – AVHRR. Curr. Sci. 91,
Ghosh, S.M., Behera, M.D., 2021. Aboveground biomass estimates of tropical mangrove
1626–1633.
forest using Sentinel-1 SAR coherence data - the superiority of deep learning over a
Chang, J., Shoshany, M., 2016. Mediterranean shrublands biomass estimation using
semi-empirical model. Comput. Geosci. 150, 104737. https://doi.org/10.1016/j.
Sentinel-1 and Sentinel-2. Int. Geosci. Remote Sens. Symp. 2016-Novem 5300–5303.
cageo.2021.104737.
https://doi.org/10.1109/IGARSS.2016.7730380.
Ghosh, S.M., Behera, M.D., 2018. Aboveground biomass estimation using multi-sensor
Chave, J., Andalo, C., Brown, S., Cairns, M.A., Chambers, J.Q., Eamus, D., Fölster, H.,
data synergy and machine learning algorithms in a dense tropical forest. Appl.
Fromard, F., Higuchi, N., Kira, T., Lescure, J.P., Nelson, B.W., Ogawa, H., Puig, H.,
Geogr. 96, 29–40. https://doi.org/10.1016/j.apgeog.2018.05.011.
Riéra, B., Yamakura, T., 2005. Tree allometry and improved estimation of carbon
Ghosh, S.M., Behera, M.D., Paramanik, S., 2020. Canopy height estimation using Sentinel
stocks and balance in tropical forests. Oecologia 145, 87–99. https://doi.org/
series images through machine learning models in a mangrove forest. Rem. Sens. 12,
10.1007/s00442-005-0100-x.
1519. https://doi.org/10.3390/rs12091519.
Chave, J., Condit, R., Aguilar, S., Hernandez, A., Lao, S., Perez, R., 2004. Error
Gillerot, L., Vlaminck, E., De Ryck, D.J.R., Mwasaru, D.M., Beeckman, H., Koedam, N.,
propagation and sealing for tropical forest biomass estimates. Philos. Trans. R. Soc. B
2018. Inter- and intraspecific variation in mangrove carbon fraction and wood
Biol. Sci. 359, 409–420. https://doi.org/10.1098/rstb.2003.1425.

11
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

specific gravity in Gazi Bay, Kenya. Ecosphere 9. https://doi.org/10.1002/ Mauya, E.W., Hansen, E.H., Gobakken, T., Bollandsås, O.M., Malimbwi, R.E., Næsset, E.,
ecs2.2306. 2015. Effects of field plot size on prediction accuracy of aboveground biomass in
Giri, C., Ochieng, E., Tieszen, L.L., Zhu, Z., Singh, A., Loveland, T., Masek, J., Duke, N., airborne laser scanning-assisted inventories in tropical rain forests of Tanzania.
2011. Status and distribution of mangrove forests of the world using earth Carbon Bal. Manag. 10, 10. https://doi.org/10.1186/s13021-015-0021-x.
observation satellite data. Global Ecol. Biogeogr. 20, 154–159. https://doi.org/ Moreira, A., Krieger, G., Younis, M., Hajnsek, I., Papathanassiou, K., Eineder, M., De
10.1111/j.1466-8238.2010.00584.x. Zan, F., 2011. Tandem-L: a mission proposal for monitoring dynamic earth
Gonzalez Del Castillo, E., Sanchez-Azofeifa, A., Paw U, K.T., Gamon, J.A., Quesada, M., processes. Int. Geosci. Remote Sens. Symp. 1385–1388. https://doi.org/10.1109/
2018. Integrating proximal broad-band vegetation indices and carbon fluxes to IGARSS.2011.6049324.
model gross primary productivity in a tropical dry forest. Environ. Res. Lett. 13 Murdiyarso, D., Purbopuspito, J., Kauffman, J.B., Warren, M.W., Sasmito, S.D.,
https://doi.org/10.1088/1748-9326/aac3f0. Donato, D.C., Manuri, S., Krisnawati, H., Taberima, S., Kurnianto, S., 2015. The
Guan, K., Pan, M., Li, H., Wolf, A., Wu, J., Medvigy, D., Caylor, K.K., Sheffield, J., potential of Indonesian mangrove forests for global climate change mitigation. Nat.
Wood, E.F., Malhi, Y., Liang, M., Kimball, J.S., Saleska, S.R., Berry, J., Joiner, J., Clim. Change 5, 1089–1092. https://doi.org/10.1038/nclimate2734.
Lyapustin, A.I., 2015. Photosynthetic seasonality of global tropical forests Mutanga, O., Skidmore, A.K., 2004. Narrow band vegetation indices overcome the
constrained by hydroclimate. Nat. Geosci. 8, 284. saturation problem in biomass estimation. Int. J. Rem. Sens. 25, 3999–4014. https://
Günlü, A., Ercanli Başkent, E.Z., Çakır, G., 2014. Estimating aboveground biomass using doi.org/10.1080/01431160310001654923.
landsat TM imagery: a case study of Anatolian Crimean pine forests in Turkey. Ann. Navarro, J., Algeet, N., Fernández-Landa, A., Esteban, J., Rodríguez-Noriega, P., Guillén-
For. Res. 57, 289–298. https://doi.org/10.15287/afr.2014.278. Climent, M., Navarro, J.A., Algeet, N., Fernández-Landa, A., Esteban, J., Rodríguez-
Gupta, S.V., 2012. Measurement Uncertainties - Physical Parameters and Calibration of Noriega, P., Guillén-Climent, M.L., 2019. Integration of UAV, sentinel-1, and
Instruments. sentinel-2 data for mangrove plantation aboveground biomass monitoring in
Hamdan, O., Khali Aziz, H., Mohd Hasmadi, I., 2014. L-band ALOS PALSAR for biomass Senegal. Rem. Sens. 11, 77. https://doi.org/10.3390/RS11010077.
estimation of Matang Mangroves, Malaysia. Remote Sens. Environ. 155, 69–78. Nischitha, V., Ahmed, S.A., Varikoden, H., Revadekar, J.V., 2014. The impact of seasonal
https://doi.org/10.1016/j.rse.2014.04.029. rainfall variability on NDVI in the Tunga and Bhadra river basins , Karnataka , India.
Hamilton, S.E., Casey, D., 2016. Creation of a high spatio-temporal resolution global Int. J. Rem. Sens. 35, 8025–8043. https://doi.org/10.1080/01431161.2014.979301.
database of continuous mangrove forest cover for the 21st century (CGMFC-21). Pandey, P.C., Anand, A., Srivastava, P.K., 2019. Spatial distribution of mangrove forest
Global Ecol. Biogeogr. 25, 729–738. https://doi.org/10.1111/geb.12449. species and biomass assessment using field inventory and earth observation
Han, J., Wei, C., Chen, Y., Liu, W., Song, P., Zhang, D., Wang, A., Song, X., Wang, X., hyperspectral data. Biodivers. Conserv. https://doi.org/10.1007/s10531-019-
Huang, J., 2017. Mapping above-ground biomass ofwinter oilseed rape using high 01698-8.
spatial resolution satellite data at parcel scale under waterlogging conditions. Rem. Pastor-Guzman, J., Dash, J., Atkinson, P.M., 2018. Remote sensing of mangrove forest
Sens. 9 https://doi.org/10.3390/rs9030238. phenology and its environmental drivers. Remote Sens. Environ. 205, 71–84.
Heiskanen, J., 2006. Estimating aboveground tree biomass and leaf area index in a https://doi.org/10.1016/j.rse.2017.11.009.
mountain birch forest using ASTER satellite data. Int. J. Rem. Sens. 27, 1135–1158. Pattanaik, C., Reddy, C.S., Dhal, N.K., Das, R., 2008. Utilisation of mangrove forests in
https://doi.org/10.1080/01431160500353858. Bhitarkanika wildlife sanctuary, Orissa. Indian J. Tradit. Knowl. 7, 598–603.
Hengl, T., Nussbaum, M., Wright, M.N., Heuvelink, G.B.M., Gräler, B., 2018. Random Pham, L.T.H., Brabyn, L., 2017. Monitoring mangrove biomass change in Vietnam using
forest as a generic framework for predictive modeling of spatial and spatio-temporal SPOT images and an object-based approach combined with machine learning
variables. PeerJ. https://doi.org/10.7717/peerj.5518, 2018. algorithms. ISPRS J. Photogrammetry Remote Sens. 128, 86–97. https://doi.org/
Hetzer, J., Huth, A., Wiegand, T., Dobner, H.J., Fischer, R., 2020. An analysis of forest 10.1016/j.isprsjprs.2017.03.013.
biomass sampling strategies across scales. Biogeosciences 17, 1673–1683. https:// Pham, T.D., Yoshino, K., Le, N.N., Bui, D.T., 2018. Estimating aboveground biomass of a
doi.org/10.5194/bg-17-1673-2020. mangrove plantation on the Northern coast of Vietnam using machine learning
Ho Tong Minh, D., Tebaldini, S., Rocca, F., Le Toan, T., Villard, L., Dubois-Fernandez, P. techniques with an integration of ALOS-2 PALSAR-2 and Sentinel-2A data. Int. J.
C., 2015. Capabilities of BIOMASS tomography for investigating tropical forests. Rem. Sens. 1–28. https://doi.org/10.1080/01431161.2018.1471544.
IEEE Trans. Geosci. Rem. Sens. 53, 965–975. https://doi.org/10.1109/ Pham, Tien Dat, Yokoya, N., Xia, J., Ha, N.T., Le, N.N., Nguyen, T.T.T., Dao, T.H., Vu, T.
TGRS.2014.2331142. T.P., Pham, Tien Duc, Takeuchi, W., 2020. Comparison of machine learning methods
Houghton, R.A., Hall, F., Goetz, S.J., 2009. Importance of biomass in the global carbon for estimating mangrove above-ground biomass using multiple source remote
cycle. J. Geophys. Res. Biogeosciences 114, 1–13. https://doi.org/10.1029/ sensing data in the red river delta biosphere reserve, Vietnam. Rem. Sens. 12, 1–24.
2009JG000935. https://doi.org/10.3390/RS12081334.
Huang, X., Ziniti, B., Torbick, N., Ducey, M., 2018. Assessment of forest above ground Quirós Vargas, J.J., Zhang, C., Smitchger, J.A., McGee, R.J., Sankaran, S., 2019.
biomass estimation using multi-temporal C-band sentinel-1 and polarimetric L-band Phenotyping of plant biomass and performance traits using remote sensing
PALSAR-2 data. Rem. Sens. 10, 1424. https://doi.org/10.3390/rs10091424. techniques in pea (pisum sativum, L.). Sensors 19. https://doi.org/10.3390/
Huang, Y., Genderen, J.L. van, 1996. Evaluation of several speckle filtering techniques s19092031.
for ERS-1 & 2 imagery. International Archives of Photogrammetry and Remote Ray, R., Ganguly, D., Chowdhury, C., Dey, M., Das, S., Dutta, M.K., Mandal, S.K.,
Sensing, pp. 164–169. Majumder, N., De, T.K., Mukhopadhyay, S.K., Jana, T.K., 2011. Carbon sequestration
Imhoff, M.L., 1993. Radar backscatter biomass saturation - observations and implications and annual increase of carbon stock in a mangrove forest. Atmos. Environ. 45,
for global biomass assessment. Igarss’93 better underst. Earth Environ. I-Iv 43–45. 5016–5024. https://doi.org/10.1016/j.atmosenv.2011.04.074.
Joshi, H.G., Ghose, M., 2014. Community structure, species diversity, and aboveground Reddersen, B., Fricke, T., Wachendorf, M., 2014. A multi-sensor approach for predicting
biomass of the Sundarbans mangrove swamps. Trop. Ecol. 55, 283–303. biomass of extensively managed grassland. Comput. Electron. Agric. 109, 247–260.
Kachamba, D.J., Ørka, H.O., Næsset, E., Eid, T., Gobakken, T., 2017. Influence of plot https://doi.org/10.1016/j.compag.2014.10.011.
size on efficiency of biomass estimates in inventories of dry tropical forests assisted Rosen, P.A., Hensley, S., Shaffer, S., Veilleux, L., Chakraborty, M., Misra, T., Bhan, R.,
by photogrammetric data from an unmanned aircraft system. Rem. Sens. 9, 1–15. Raju Sagi, V., Satish, R., 2015. The NASA-ISRO SAR mission - an international space
https://doi.org/10.3390/rs9060610. partnership for science and societal benefit. IEEE Natl. Radar Conf. - Proc.
Kamruzzaman, M., Ahmed, S., Osawa, A., 2017. Biomass and net primary productivity of 1610–1613. https://doi.org/10.1109/RADAR.2015.7131255, 2015-June.
mangrove communities along the Oligohaline zone of Sundarbans, Bangladesh. For. Roy, D.P., Li, J., Zhang, H.K., Yan, L., Huang, H., Li, Z., 2017. Examination of Sentinel-2A
Ecosyst 4. https://doi.org/10.1186/s40663-017-0104-0. multi-spectral instrument (MSI) reflectance anisotropy and the suitability of a
Kauffman, J.B., Donato, D.C., 2012. Protocols for the Measurement, Monitoring and general method to normalize MSI reflectance to nadir BRDF adjusted reflectance.
Reporting of Structure, Biomass and Carbon Stocks in Mangrove Forests. https://doi. Remote Sens. Environ. 199, 25–38. https://doi.org/10.1016/j.rse.2017.06.019.
org/10.17528/cifor/003749. Roy, P.S., 1989. Spectral reflectance characteristics of vegetation and their use in
Komiyama, A., Poungparn, S., Kato, S., 2005. Common allometric equations for estimating productive potential. Proc. Indian Acad. Sci. Plant Sci. 99, 59–81.
estimating the tree weight of mangroves. J. Trop. Ecol. 21, 471–477. https://doi. Saatchi, S.S., Harris, N.L., Brown, S., Lefsky, M., Mitchard, E.T.A., Salas, W., Zutta, B.R.,
org/10.1017/S0266467405002476. Buermann, W., Lewis, S.L., Hagen, S., Petrova, S., White, L., Silman, M., Morel, A.,
Kuhn, M., 2008. Building predictive models in R using the caret package. J. Stat. 2011. Benchmark map of forest carbon stocks in tropical regions across three
Software 28, 1–26. https://doi.org/10.1053/j.sodo.2009.03.002. continents. Proc. Natl. Acad. Sci. Unit. States Am. 108, 9899–9904. https://doi.org/
Lopes, A., Nezry, E., Touzi, R., Laur, H., 1990. Maximum A posteriori speckle filtering 10.1073/pnas.1019576108.
and first order texture models in sar images. 10th Annu. Int. Symp. Geosci. Remote Samat, A., Li, E., Wang, W., Liu, S., Lin, C., Abuduwaili, J., 2020. Meta-XGBoost for
Sens. 2409–2412. https://doi.org/10.1109/IGARSS.1990.689026. hyperspectral image classification using extended MSER-guided morphological
Malistov, A., Trushin, A., 2019. Gradient boosted trees with extrapolation. Proc. - 18th profiles. Rem. Sens. 12 https://doi.org/10.3390/rs12121973.
IEEE Int. Conf. Mach. Learn. Appl. ICMLA 2019 2019-Janua 783–789. https://doi. Sinha, S., Jeganathan, C., Sharma, L.K., Nathawat, M.S., Das, A.K., Mohan, S., 2016.
org/10.1109/ICMLA.2019.00138. Developing synergy regression models with space-borne ALOS PALSAR and landsat
Manna, S., Nandy, S., Chanda, A., Akhand, A., Hazra, S., Dadhwal, V.K., 2014. Estimating TM sensors for retrieving tropical forest biomass. J. Earth Syst. Sci. 125, 725–735.
aboveground biomass in Avicennia marina plantation in Indian Sundarbans using https://doi.org/10.1007/s12040-016-0692-z.
high-resolution satellite data. J. Appl. Remote Sens. 8, 083638 https://doi.org/ Skofronick-Jackson, G., Petersen, W.A., Berg, W., Kidd, C., Stocker, E.F., Kirschbaum, D.
10.1117/1.JRS.8.083638. B., Kakar, R., Braun, S.A., Huffman, G.J., Iguchi, T., Kirstetter, P.E., Kummerow, C.,
Martins Silva, J.P., Marques da Silva, M.L., Ferreira da Silva, E., Fernandes da Silva, G., Meneghini, R., Oki, R., Olson, W.S., Takayabu, Y.N., Furukawa, K., Wilheit, T., 2017.
Ribeiro de Mendonça, A., Cabacinha, C.D., Araújo, E.F., Santos, J.S., Vieira, G.C., The global precipitation measurement (GPM) mission for science and Society. Bull.
Felix de Almeida, M.N., Fernandes, M.R. de M., 2019. Computational techniques Am. Meteorol. Soc. 98, 1679–1695. https://doi.org/10.1175/BAMS-D-15-00306.1.
applied to volume and biomass estimation of trees in Brazilian savanna. J. Environ. Songsom, V., Koedsin, W., Ritchie, R.J., Huete, A., 2019. Mangrove phenology and
Manag. 249, 109368. https://doi.org/10.1016/j.jenvman.2019.109368. environmental drivers derived from remote sensing in Southern Thailand. Rem.
Sens. 11 https://doi.org/10.3390/rs11080928.

12
S.M. Ghosh et al. Journal of Environmental Management 292 (2021) 112816

Suepa, T., Qi, J., Lawawirojwong, S., Messina, J.P., 2016. Understanding spatio-temporal Wicaksono, P., Danoedoro, P., Hartono Nehren, U., 2016. Mangrove biomass carbon
variation of vegetation phenology and rainfall seasonality in the monsoon Southeast stock mapping of the Karimunjawa Islands using multispectral remote sensing. Int. J.
Asia. Environ. Res. 147, 621–629. https://doi.org/10.1016/j.envres.2016.02.005. Rem. Sens. 37, 26–52. https://doi.org/10.1080/01431161.2015.1117679.
Tilly, N., Aasen, H., Bareth, G., 2015. Fusion of plant height and vegetation indices for Wittke, S., Yu, X., Karjalainen, M., Hyyppä, J., Puttonen, E., 2018. Comparison of two-
the estimation of barley biomass. Rem. Sens. 7, 11449–11480. https://doi.org/ dimensional multitemporal Sentinel-2 data with three-dimensional remote sensing
10.3390/rs70911449. data sources for forest inventory parameter estimation over a boreal forest. Int. J.
Tinoco, S.L.J.L., Santos, H.G., Menotti, D., Preto, O., Santos, a B., Horizonte, B., Santos, J. Appl. Earth Obs. Geoinf. 76, 167–178. https://doi.org/10.1016/j.jag.2018.11.009.
a, 2013. Approach Based on Linear Programming and Weighted Linear Combination Wu, C., Chen, Y., Peng, C., Li, Z., Hong, X., 2019. Modeling and estimating aboveground
4082–4085. biomass of Dacrydium pierrei in China using machine learning with climate change.
Vorster, A.G., Evangelista, P.H., Stovall, A.E.L., Ex, S., 2020. Variability and uncertainty J. Environ. Manag. 234, 167–179. https://doi.org/10.1016/j.jenvman.2018.12.090.
in forest biomass estimates from the tree to landscape scale: the role of allometric Zhang, C., Denka, S., Cooper, H., Mishra, D.R., 2018. Quantification of sawgrass marsh
equations. Carbon Bal. Manag. 15, 1–20. https://doi.org/10.1186/s13021-020- aboveground biomass in the coastal Everglades using object-based ensemble analysis
00143-6. and Landsat data. Remote Sens. Environ. 204, 366–379. https://doi.org/10.1016/j.
rse.2017.10.018.

13

You might also like