Research paper
Aboveground biomass estimates of tropical mangrove forest using Sentinel-1 SAR coherence data - The superiority of deep learning over a semi-empirical model

https://doi.org/10.1016/j.cageo.2021.104737Get rights and content

Highlights

  • High temporal decorrelation affects Sentinel-1 InSAR coherence.

  • IWCM is not effective in estimating above ground biomass using Sentinel-1 data.

  • Deep neural network models are more accurate in estimating above ground biomass.

Abstract

The availability of advanced Machine Learning algorithms has made the estimation process of biophysical parameters more efficient. However, the efficiency of those methods seldom compared with the efficiency of already established semi-empirical procedures. Aboveground biomass (AGB) of mangrove forests is a crucial biophysical parameter as it is positively correlated to the carbon stocks and fluxes. The free availability of Sentinel-1 C-band SAR data and machine learning algorithms hold promises in estimating AGB of tropical mangrove forests. We reported high AGB (70 t/ha to 666 t/ha) using 185 field quadrats of 0.04ha each from Bhitarkanika Wildlife Sanctuary, located on the eastern Indian coast that could be attributed to species composition. The AGB maps generated using Interferometric Water Cloud Model (IWCM) and Deep Learning models were different from each other as they rely on different variables. IWCM was more dependent, especially on ground and vegetation components of coherence, while canopy height acted as the most crucial variable in the Deep Learning model. However, the negligible variations in Deep Learning-based AGB maps can be attributed to interpreting the importance of coherence and VH backscatter. Due to low canopy penetration power of C-band SAR, high temporal decorrelation resulting from longer time gap between interferometric image pairs, and high spatial heterogeneity of mangrove forests, IWCM found as an unsuitable method for AGB estimation. Interestingly, a Deep Learning algorithm could translate the exact relationship between predictor variables and mangrove AGB in Bhitarkanika Wildlife Sanctuary. The AGB estimation studies in mangrove forests using Sentinel data should focus more on using machine learning algorithms like Deep Learning rather than semi-empirical models.

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)

  • M. Mura et al.

    Exploiting the capabilities of the Sentinel-2 multi spectral instrument for predicting growing stock volume in forest ecosystems

    Int. J. Appl. Earth Obs. Geoinf.

    (2018)
  • O. Mutanga et al.

    High density biomass estimation for wetland vegetation using worldview-2 imagery and random forest regression algorithm

    Int. J. Appl. Earth Obs. Geoinf.

    (2012)
  • G. Nagababu et al.

    Estimation of technical and economic potential of offshore wind along the coast of India

    Energy

    (2017)
  • L.T.H. Pham et al.

    Monitoring mangrove biomass change in Vietnam using SPOT images and an object-based approach combined with machine learning algorithms

    ISPRS J. Photogramm. Remote Sens

    (2017)
  • S. Puliti et al.

    Combining UAV and Sentinel-2 auxiliary data for forest growing stock volume estimation through hierarchical model-based inference

    Remote Sens. Environ

    (2018)
  • F.E. Putz et al.
    (1986)
  • Y. Sadeghi et al.

    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.

    (2018)
  • M. Santoro et al.

    Stem volume retrieval in boreal forests from ERS-1/2 interferometry

    Remote Sens. Environ

    (2002)
  • M. Santoro et al.

    Forest stem volume estimation using C-band interferometric SAR coherence data of the ERS-1 mission 3-days repeat-interval phase

    Remote Sens. Environ

    (2018)
  • W. Xing et al.

    A novel cellular automata model integrated with deep learning for dynamic spatio-temporal land use change simulation

    Comput. Geosci.

    (2020)
  • M. Abadi et al.

    TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems

    (2015)
  • D.M. Alongi

    Carbon sequestration in mangrove forests

    Carbon Manag.

    (2012)
  • D.M. Alongi et al.

    Influence of roots and climate on mineral and trace element storage and flux in tropical mangrove soils

    Biogeochemistry

    (2004)
  • J.I.H. Askne et al.

    Model-based biomass estimation of a hemi-boreal forest from multitemporal tanDEM-X acquisitions

    Remote Sens

    (2013)
  • J.J.I.H. Askne et al.

    C-band repeat-pass interferometric SAR observations of the forest

    IEEE Trans. Geosci. Rem. Sens.

    (1997)
  • E.P.W. Attema et al.

    Vegetation modeled as a water cloud

    Radio Sci.

    (1978)
  • M.D. Behera et al.

    Above-ground biomass and carbon estimates of Shorea robusta and Tectona grandis forests using QuadPOL ALOS PALSAR data

    Adv. Space Res.

    (2015)
  • T. Beysolow

    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

    (2017)
  • S. Bouillon et al.

    Mangrove production and carbon sinks: a revision of global budget estimates

    Global Biogeochem. Cycles

    (2008)
  • L. Breiman

    Random forests

    Mach. Learn.

    (2001)
  • J.L. Breithaupt et al.

    Organic carbon burial rates in mangrove sediments: strengthening the global budget

    Global Biogeochem. Cycles

    (2012)
  • D.R. Cahoon et al.

    Mass tree mortality leads to mangrove peat collapse at Bay Islands, Honduras after Hurricane Mitch

    J. Ecol.

    (2003)
  • J.M.B. Carreiras et al.

    Estimating the above-ground biomass in miombo savanna woodlands (Mozambique, East Africa) using L-band synthetic aperture radar data

    Remote Sens

    (2013)
  • J. Chave et al.

    Towards a worldwide wood economics spectrum

    Ecol. Lett.

    (2009)
  • F. Chollet et al.

    R Interface to Keras

    (2017)
  • C. Cortes et al.

    Support-vector networks

    Mach. Learn

    (1995)
  • P.-A. Esseen et al.

    Boreal forests

    Ecol. Bull.

    (1997)
  • T.R. Feldpausch et al.

    Tree height integrated into pantropical forest biomass estimates

    Biogeosciences

    (2012)
  • A. Ferretti et al.

    Part C. InSAR processing: a mathematical approach

    Eur. Sp. Agency

    (2007)
  • A. Fisher et al.

    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.

    (2019)
  • A. Ghatak

    Deep Learning with R, Deep Learning with R

    (2019)
  • Cited by (36)

    View all citing articles on Scopus
    View full text