Research paperAboveground biomass estimates of tropical mangrove forest using Sentinel-1 SAR coherence data - The superiority of deep learning over a semi-empirical model
Introduction
One of the tropics' major carbon sinks is the mangrove forests (Bouillon et al., 2008; Cahoon et al., 2003; Nellemann et al., 2009). Mangroves showed to have high aboveground biomass primarily due to its high productivity (Alongi et al., 2004; Matsui, 1998; Putz and Chan, 1986), and high carbon sequestration rates (Alongi, 2012; Breithaupt et al., 2012; McLeod et al., 2011) in multiple field-based studies. There are growing efforts to a more accurate mapping of global carbon stocks, and fluxes have gained significant momentum in recent years (Saatchi et al., 2011; Baccini et al., 2012). However, due to their small extent and the challenging conditions, mangroves have been mainly ignored in these analyses. Aboveground biomass (AGB) may range from about 8 t/ha for dwarf mangroves to >500 t/ha in riverine and fringe mangroves of the Indo-Pacific region (Kauffman et al., 2011; Kauffman and Cole, 2010).
Forest AGB is one of the major indicators of its carbon content, which can be easily measured in the field (Temesgen et al., 2015). Field-based measurements have an essential role in establishing and validating areal or satellite data-based AGB estimation models (Molto et al., 2013). The longer wavelength SAR satellite gives AGB more direct measurements as the signals get backscattered directly from the stem by penetrating the canopy cover. Multiple studies have shown the effectiveness of longer wavelength L-band SAR data in tropical forest AGB estimation (Behera et al., 2016; Carreiras et al., 2013; Cartus et al., 2012). Nevertheless, it is difficult to measure and regularly monitor the AGB of a large area using only L-band SAR data as L-band data from Advanced Land Observing Satellite (ALOS), and Satélite Argentino de Observación COn Microondas (SAOCOM) satellites are not available freely. Though C-band SAR backscatter is not suitable for AGB estimation of tropical forests, C-band SAR interferometric coherence has shown great promise in estimating biophysical parameters like growing stock volume while employing semi-empirical models, e.g., Water Cloud Model (WCM) and the Interferometric WCM (Askne et al., 2013; Kumar et al., 2012).
SAR interferometry (InSAR) is the process of getting the elevation information about the earth's surface features by correlating two SAR images obtained through two slightly different positions. The interferometric pair's correlation is generally computed via producing an interferogram comprising the interferometric coherence and phase difference among the interferometric images. The density of the vegetations and elevation of the scatterers govern the interferometric phase's magnitude and coherence for an interferogram of a forested area (Santoro et al., 2018). The interferometric coherence, i.e., the measurement of similarity between interferometric images, changes depending on multiple factors. Temporal decorrelation, i.e., a decrease in coherence, occurs due to an increase in the interferometric pair's time interval. Another reason for decorrelation is an increase in the baseline length, which results in spatial and volume decorrelation. If the baseline distance crosses the critical value, the interferometric phase reduces to pure noise (Ferretti et al., 2007), and coherence falls to zero (Richards, 2009). An interferometric pair with a shorter time gap and a considerably smaller baseline can estimate AGB with better accuracy (Askne et al, 1997, 2013).
The semi-empirical Water Cloud Model (WCM) estimates AGB by decomposing the forest SAR backscatter into vegetation canopy and ground contributions (Attema and Ulaby, 1978). Askne et al. (1999) proposed the improved semi-empirical Interferometric Water Cloud Model (IWCM) that utilizes SAR interferometric coherence and tree height information for growing stock volume or AGB retrieval (Santoro et al., 2002). IWCM model can stretch the saturation limit of AGB estimation using SAR data to a great extent (Askne et al., 2013).
The semi-empirical models are mainly designed for and used in the boreal forests (Table 1), as the homogeneous boreal forest canopy make them more suitable for AGB estimation (Soja et al., 2017), with good accuracy using SAR data ranging from X-band to L-band (Santoro et al., 2002; Cartus et al., 2012; Askne et al., 2013). The biomass density of those forests mainly lies around 250 t/ha. However, these models were used in some cases for high AGB forests of tropics also. Behera et al. (2016) estimated AGB for a semi-evergreen tropical forest using ALOS PALSAR data and WCM with acceptable accuracy. Kumar et al. (2012) used IWCM for a semi-evergreen forest and found the model to be quite useful for AGB measurement. Mangroves forests are devoid of any understory vegetation, unlike other tropical forests (Janzen, 1985). Therefore the use of shorter wavelength C-band data may not miss any significant portion of the vegetation.
The Sentinel-1 and Sentinel-2 data have opened up new methods for forest structural parameters and AGB estimation. Sentinel-1 C-band SAR backscatter alone does not act as a good indicator of AGB, while Sentinel-2 is useful for estimating both AGB and growing stock volume (Hawryło and Wezyk, 2018; Mura et al., 2018; Puliti et al., 2018). The capacity of Sentinel-1 C-band SAR backscatter and Sentinel-1 interferometric SAR coherence in estimating mangrove AGB has not been confirmed with an adequate number of studies.
In recent years many studies have explored the possibility of estimating the AGB using machine learning regression models, especially for tropical forests (Carreiras et al., 2012; Ghosh and Behera, 2018; Jachowski et al., 2013; Mutanga et al., 2012; Navarro et al., 2019). These models includes Random Forest (Breiman, 2001), Gradient Boosted Model (Friedman, 2002) and Support Vector Machine (Cortes and Vapnik, 1995). Deep learning algorithms have been developed as a subfield of machine learning, which can recognize intricate nonlinear patterns in data by mimicking the neural architecture of the human brain (Beysolow II, 2017). A deep learning model can interpret the different complexity levels in the data using multiple successive layers of representations (Beysolow II, 2017; Ghatak, 2019). The learning process of a deep learning model consists of finding the values for each layer's parameters. Deep learning models are generally developed by using DNN models with multiple layers (Moolayil, 2019). Deep learning models have been used in multiple geospatial analysis studies in recent times (Xing et al., 2020; Yu et al., 2020). Zhang et al. (2019) estimated the AGB of a forest using the Stacked Sparse Autoencoder network (SSAE) algorithm and found it performed better than other machine learning algorithms such as Random Forest and Support Vector Regression. However, the efficiency of deep learning algorithms for mangrove AGB estimation has seldom been used for prediction using remote sensing datasets.
A deep learning (DL) algorithm is generally applied through an existing framework that uses reusable codes to develop different constitute parts of a DL model (Moolayil, 2019). The existing frameworks are informally grouped into two parts. First is the low-level framework, where one has to write quite long codes to prepare the final DL model. Tensorflow by Google is the most popular and widely used low-level framework (Abadi et al., 2015). Keras is a popular high-level Deep Learning framework that uses Tensorflow as a back end (Chollet and Allaire, 2017). Keras is written in Python though it can be used in R interface as well. A fully functional DL model can be developed in Keras with a few lines of code.
Studies on different Indian mangroves show that their AGB varies from as low as 20.9 t/ha in the Sundarbans region (Manna et al., 2014) to 196.48 t/ha and 236 t/ha for mangroves in Kerala coast and Mahanadi delta, respectively (Vinod et al., 2018; Sahu et al., 2016). The results of machine learning models for AGB estimation of mangroves have not been directly compared with the semi-empirical models as they have never been used for AGB estimation of mangroves. A review of previous studies also shows that in the Indian tropical forests context, the application of semi-empirical and machine learning models in AGB estimation is rare.
In this work, the AGB of an Indian mangrove forest at Bhitarkanika Wildlife Sanctuary (BWS) was estimated using semi-empirical and deep learning models and their results were compared. As the first part of the work AGB was estimated in field for 185 quadrats. Then IWCM was established using Sentinel-1 SAR backscatter and coherence along with field estimated AGB and canopy height. Thereafter DL modelling was done using the same set of variables. Finally AGB maps were prepared for both the methods for the whole study area.
Section snippets
Study area and field data
The mangroves of Bhitarkanika Wildlife Sanctuary consists of an area of 130 km2 on India's eastern coast (Fig. 1) and shows a high species diversity (Pattanaik et al., 2008; Reddy et al., 2006). The dominant species are Heritiera fomes, Excoecaria agallocha, Avicennia officinalis, Ceriops decandra, and Cynometra irripa. Sonneratia apetala can only be located along the banks of creeks and rivers. Species assemblages vary from double to many, with most areas as homogeneous forests. The forest has
Field data
Five species such as Heritiera fomes, Excoecaria agallocha, Avicennia officinalis, Ceriops decandra Cynometra irripa dominate the BWS mangrove forest, of which the first three forms the top canopy. The field-measured canopy height showed an average of 9 m in most of the quadrats, with the tallest canopy heights in the range of 14–16 m to the lowest of 2–3 m (Ghosh et al., 2020). BWS accommodates high average AGB as observed from the field data analysis, wherein AGB density varied from 49 t/ha
Field data
The Bhitarkanika WLS demonstrated high AGB that vary from 70 t/ha to 666 t/ha. Banerjee et al. (2013) reported AGB ranges from 27.46 t/ha to 113.67 t/ha for Sunderbans mangroves, which is considerably lower than Bhitarkanika WLS. Pandey et al. (2019) reported that the Sonneratia apetala and Cynometra iripa species have the highest AGB of 643.12 t/ha and 652.14 t/ha, respectively, while Bal and Banerjee (2019) reported total biomass of 866.67 ± 166.10 t/ha in BWS. The high AGB density of BWS can
Conclusions
The free availability of high spatial and temporal resolution Sentinel series data along with multiple machine learning algorithms has revolutionized the process of forest AGB estimation in recent years. This work was focused on the relatively unexplored area of checking the efficiency of the IWCM and Deep learning methods in estimating AGB of mangrove forests. Results showed that due to low canopy penetrating power of C-band SAR, high temporal decorrelation, and high spatial heterogeneity,
Computer code availability
All the codes written for this work are available at github (https://github.com/sujitmg/AGB_keras).
Author contribution
SM Ghosh - Conceptualization, methodology, data analysis, writing-original draft preparation; MD Behera - Conceptualization, writing-review, and editing.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
SMG acknowledges the Fellowship received from MHRD for Ph.D. research. MDB acknowledges the financial support received from SAC, ISRO, through the NISAR grant that helped in field-sampling, and thanks to the Odisha forest wildlife division for granting permission to conduct fieldwork in BWS, India. We acknowledge the facilities provided by Head, CORAL and authorities of IIT Kharagpur, to undertake this study.
References (82)
- et al.
Salinity based allometric equations for biomass estimation of Sundarban mangroves
Biomass Bioenergy
(2013) - et al.
Understanding the relationship between aboveground biomass and ALOS PALSAR data in the forests of Guinea-Bissau (West Africa)
Remote Sens. Environ
(2012) - et al.
Mapping forest aboveground biomass in the Northeastern United States with ALOS PALSAR dual-polarization L-band
Remote Sens. Environ
(2012) - et al.
Estimation and mapping of above-ground biomass of mangrove forests and their replacement land uses in the Philippines using Sentinel imagery
ISPRS J. Photogramm. Remote Sens
(2017) Stochastic gradient boosting
Comput. Stat. Data Anal.
(2002)- et al.
Aboveground biomass estimation using multi-sensor data synergy and machine learning algorithms in a dense tropical forest
Appl. Geogr.
(2018) - et al.
L-band ALOS PALSAR for biomass estimation of Matang mangroves, Malaysia
Remote Sens. Environ
(2014) - et al.
Mangrove biomass estimation in Southwest Thailand using machine learning
Appl. Geogr.
(2013) - et al.
Effects of soil moisture and water depth on ERS SAR backscatter measurements from an Alaskan wetland complex
Remote Sens. Environ.
(2009) - et al.
Biology of mangroves and mangrove Ecosystems
Exploiting the capabilities of the Sentinel-2 multi spectral instrument for predicting growing stock volume in forest ecosystems
Int. J. Appl. Earth Obs. Geoinf.
High density biomass estimation for wetland vegetation using worldview-2 imagery and random forest regression algorithm
Int. J. Appl. Earth Obs. Geoinf.
Estimation of technical and economic potential of offshore wind along the coast of India
Energy
Monitoring mangrove biomass change in Vietnam using SPOT images and an object-based approach combined with machine learning algorithms
ISPRS J. Photogramm. Remote Sens
Combining UAV and Sentinel-2 auxiliary data for forest growing stock volume estimation through hierarchical model-based inference
Remote Sens. Environ
Mapping boreal forest biomass from a SRTM and TanDEM-X based on canopy height model and Landsat spectral indices
Int. J. Appl. Earth Obs. Geoinf.
Stem volume retrieval in boreal forests from ERS-1/2 interferometry
Remote Sens. Environ
Forest stem volume estimation using C-band interferometric SAR coherence data of the ERS-1 mission 3-days repeat-interval phase
Remote Sens. Environ
A novel cellular automata model integrated with deep learning for dynamic spatio-temporal land use change simulation
Comput. Geosci.
TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems
Carbon sequestration in mangrove forests
Carbon Manag.
Influence of roots and climate on mineral and trace element storage and flux in tropical mangrove soils
Biogeochemistry
Model-based biomass estimation of a hemi-boreal forest from multitemporal tanDEM-X acquisitions
Remote Sens
C-band repeat-pass interferometric SAR observations of the forest
IEEE Trans. Geosci. Rem. Sens.
Vegetation modeled as a water cloud
Radio Sci.
Above-ground biomass and carbon estimates of Shorea robusta and Tectona grandis forests using QuadPOL ALOS PALSAR data
Adv. Space Res.
Introduction to Deep Learning using R - a Step-by-Step Guide to Learning and Implementing Deep Learning Models Using R, Introduction to Deep Learning Using R
Mangrove production and carbon sinks: a revision of global budget estimates
Global Biogeochem. Cycles
Random forests
Mach. Learn.
Organic carbon burial rates in mangrove sediments: strengthening the global budget
Global Biogeochem. Cycles
Mass tree mortality leads to mangrove peat collapse at Bay Islands, Honduras after Hurricane Mitch
J. Ecol.
Estimating the above-ground biomass in miombo savanna woodlands (Mozambique, East Africa) using L-band synthetic aperture radar data
Remote Sens
Towards a worldwide wood economics spectrum
Ecol. Lett.
R Interface to Keras
Support-vector networks
Mach. Learn
Boreal forests
Ecol. Bull.
Tree height integrated into pantropical forest biomass estimates
Biogeosciences
Part C. InSAR processing: a mathematical approach
Eur. Sp. Agency
All models are wrong, but many are useful: learning a variable's importance by studying an entire class of prediction models simultaneously
J. Mach. Learn. Res.
Deep Learning with R, Deep Learning with R
Cited by (36)
Characterising the distribution of mangroves along the southern coast of Vietnam using multi-spectral indices and a deep learning model
2024, Science of the Total EnvironmentDevelopment of forest aboveground biomass estimation, its problems and future solutions: A review
2024, Ecological IndicatorsCombining GEDI and sentinel data to estimate forest canopy mean height and aboveground biomass
2023, Ecological InformaticsHigh-resolution mapping of forest structure from integrated SAR and optical images using an enhanced U-net method
2023, Science of Remote SensingChallenges and prospects of climate change impact assessment on mangrove environments through mathematical models
2023, Environmental Modelling and Software