..
Extrapolating forest biomass dynamics over large
areas using time-series remote sensing
A thesis submitted in fulfilment of the requirements for the degree of Doctor of Philosophy
Huy Trung Nguyen
Bsc (Hons), Thai Nguyen University, Vietnam
Msc Environmental Science, Thai Nguyen University, Vietnam
School of Science
College of Science, Engineering and Health
RMIT University
February 2020
Declaration
I certify that except where due acknowledgement has been made, the work is that of the
author alone; the work has not been submitted previously, in whole or in part, to qualify for
any other academic award; the content of the thesis is the result of work which has been
carried out since the official commencement date of the approved research program; any
editorial work, paid or unpaid, carried out by a third party is acknowledged; and, ethics
procedures and guidelines have been followed.
I acknowledge the support I have received for my research through the provision of an
Australian Government Research Training Program Scholarship
Trung Nguyen
26 February 2020
i
Abstract
Forest biomass, accounting for over 80% of global vegetation biomass, is considered a key
factor in terrestrial ecology, atmospheric processes and the water and carbon cycles. Forest
biomass has been recently recognised as a Global Climate Observing System (GCOS)
Essential Climate Variable (ECV), which is an important input to the United Nations’
Reducing Emissions from Deforestation and forest Degradation-plus (REDD+) program and
Earth system models. Reducing carbon emissions from forest changes is one of the core
requirements to mitigate the impacts of climate change on Earth. Consequently, monitoring
forest biomass dynamics is an international concern which has attracted attention from
government (at local, regional, national and international levels), academics and the general
public. According to the Global Forest Resources Assessment 2015, deforestation and forest
degradation have been persisting in tropical developing countries where demand for
exploiting natural resources are high and significantly increasing. Thus, these countries
urgently need a robust and cost-effective national forest biomass monitoring system that can
support their policy-making processes that aim to protect ecosystem integrity in forests and
reduce greenhouse gas emissions while simultaneously maintaining their social-economic
development needs. While improving the quality of carbon reporting is needed, it is
challenging for most developing countries due to their low capacities to perform national
forest inventory on a regular basis. Forest inventory data may be available in these countries,
but they are often out-of-date. Using remote sensing data, such as Landsat satellite imagery,
is one of the most practical and cost-effective alternatives to enable developing countries to
overcome this current challenge. Landsat satellites are unique as they have been creating the
longest continuously-acquired, space-based and moderate-resolution data collection since
1972. The free access and use data policy of the Landsat archive since 2008 has
revolutionized the use of Landsat data for worldwide forest research and monitoring
activities, especially forest biomass monitoring.
This research first comprehensively reviewed the state and improvements of current
approaches using Landsat time-series (LTS) for characterising forest biomass dynamics.
This literature review indicated that the use of LTS not only enables production of spatially
and temporally explicit estimates of biomass but also can improve the quality and accuracy
of biomass models. Many innovative approaches for estimating forest biomass across space
and time from LTS have been recently demonstrated. However, most of these methods have
ii
been developed for areas that are supported by comprehensive forest inventories and/or
Lidar datasets. Therefore, it is important to demonstrate an approach that is more possible
for applications in developing countries where forest inventory data are measured for a
single-time step which is often out-of-date.
This research develops a robust and consistent Landsat-based framework that can support
developing countries improve their capacities in monitoring and reporting forest biomass
and carbon stocks and changes across large areas. The framework is developed by utilising
a 30-year annual time-series of Landsat images (1988-2017) and one-off inventory data,
which are commonly available in developing countries. The study area comprised over 7.1
million ha of public forests in Victoria, south-eastern Australia. Although Victoria is not a
country, its size / forest inventory scenario is similar to many developing countries, making
it a good case study. LTS data were processed through several steps to produce a stack of
cloud-free, annual mosaic composites. This dataset was then used as a foundation input in
further analyses for characterising forest disturbance and recovery and estimating forest
biomass dynamics across space and time.
In the first stage, LTS data were utilised for developing a robust approach for mapping forest
disturbance and recovery at a landscape scale. Forest changes were detected through pixelbased change detection process using the LandTrendr temporal segmentation algorithm. A
two-phase classification process was then developed using the Random Forest (RF)
algorithm to predictively map disturbance and recovery levels (high, medium and low) and
disturbance causal agents (including wildfire, planned burns, clear-fell logging, selective
logging) for multiple detected disturbance events (both primary and secondary events).
Model explanatory data included a range of trajectory-based change metrics derived from
the LandTrendr analysis, while model training and validation data were derived from a
human-interpreted reference dataset. In addition, a space-time data cube concept was
introduced to simultaneously report on both newly detected disturbance events (detected
disturbances) as well as events that have previously occurred but are ongoing (ongoing
disturbances), which has been often under-reported. RF classification models obtained high
overall accuracies (73-81%). The data cube analysis revealed that although annual
disturbance area was dominated by newly detected disturbances, ongoing disturbances
accounted for a considerable area (over 50% of newly detected disturbances). These results
iii
indicate the utility of LTS in accurately capturing and mapping forest disturbance and
recovery, facilitating further analyses on biomass estimates.
The second stage of this research tested and compared different modelling approaches for
estimating forest biomass using Landsat time-series and inventory data. This analysis used
the outputs from the first stage (i.e., spectral change metrics, predicted disturbance and
recovery levels and causal agents) in combination with data extracted from forest inventory
field plots. In particular, 18 k-nearest neighbour (kNN) imputation models were tested to
predict three aboveground biomass (AGB) variables (total AGB, AGB of live trees and AGB
of dead trees). These models were developed using different distance techniques (RF,
Gradient Nearest Neighbour (GNN), and Most Similar Neighbour (MSN)) and different
combinations of response variables (model scenarios). Direct biomass imputation models
were trained according to the biomass variables while indirect biomass imputation models
were trained according to combinations of forest structure variables (e.g., basal area, stem
density and stem volume of live and dead-standing trees). The results show that RF
consistently outperformed MSN and GNN distance techniques across different model
scenarios and biomass variables. The indirect imputation method generally achieved better
biomass predictions than the direct imputation method. In particular, the RF-based kNN
model trained with the combination of basal area and stem density variables was the most
robust for estimating forest biomass. As the kNN imputation method is increasingly being
used by land managers and researchers to map forest biomass, this analysis helps those using
these methods to ensure their modelling and mapping practices are optimized.
The last stage presented a consistent approach for estimating forest AGB dynamics across
space and time using LTS and single-date inventory data. This approach consisted of three
components: (1) a modelling method for creating annual forest AGB maps from Landsat
time-series and one-off inventory data; (2) evaluation of the robustness and transferability
of applying a single model through time to estimate AGB dynamics; (3) a spatial and
temporal analysis of AGB dynamics according to forest disturbance and recovery histories,
from which to inform jurisdictions as to how these ecological changes impact AGB
dynamics. These analyses were based on the findings of the first two stages. A RF-based
kNN imputation model, which was defined as the most accurate method in the second stage,
was developed to produce annual maps of AGB for 30 years (from 1988 to 2017 over 7.2
million ha of forests in Victoria, Australia). Annual predictions of AGB and its change were
iv
independently evaluated using multi-temporal Lidar data. These obtained relatively high
accuracies, indicating the robustness and transferability over time of the developed
modelling method. Temporal trends of AGB were analysed according to forest disturbance
and recovery levels and causal agents (derived in the first stage) in order to understand how
AGB responds to both natural and anthropogenic processes. Specifically, change metrics
(e.g., AGB loss and gain, Years to Recovery - Y2R) were calculated at the pixel level to
characterise the patterns of AGB dynamics resulting from forest changes. AGB change
metrics showed that changes in AGB values associated with forest disturbance and recovery
(decrease and increase, respectively) were captured by predicted maps. Results also
indicated that AGB loss and Y2R varied across the states’ biogeographic regions and were
highly dependent on the level of disturbance severity (i.e., a greater loss and longer recovery
time were associated with a higher severity disturbance).
The framework presented in this research has potential for application in different forest
areas to support forest managers and policy makers to measure and report on forest biomass
changes. This research focuses on providing a solution for developing countries, where only
single-date (often out-of-date) and sparse inventory data are available, to improve their
capacities in monitoring and reporting forest carbon stocks and changes. The findings from
this research also demonstrate the utility of Earth Observation satellite data in monitoring
forests across large areas (a difficult task when only reliant on field-based methods).
Furthermore, regular and consistent observations acquired through LTS can provide us with
a better understanding of the complexity and dynamic nature of forested systems and help
us meet forest related sustainable management and development goals.
v
Acknowledgement
I would like to take the opportunity to specifically thank those who have contributed to this
research and support me throughout my PhD. Without your help, it could not be completed.
My first gratitude goes to my panel of supervisors Prof Simon Jones and Dr Mariela SotoBerelov from RMIT, and Dr Andrew Haywood from the European Forest Institute. For all
of you, I would like to thank for your patience and understanding my strengths and weakness.
Your supports through the last four years are unwavering and invaluable. Also, I would like
to thank you for adding me in the LandFor project team that allowed me to conduct me PhD
research in a collaborative approach and to achieve high quality outputs. Simon, thank you
for accepting me onto this PhD from a very early date (nearly five years ago) and for your
on-going support and encouragement since then. Mariela, thank you for being not only my
supervisor but also one of my best friends in Australia. Your advice has been always
invaluable. Andrew, your industry perspective and high-level strategic advice have been of
great benefits to my PhD research. I also thank to my PhD companion, Samuel Hislop, for
his support and contribution throughout our shared PhD journeys.
I would like to extend my gratitude to my RMIT fellow PhD and postdocs: Sam (Hislop),
Chithra, Ahmad, Nenad, Luke, Sam (Hillman), Bryant, Daisy, Chats, Shirley, Eloise, Jenna,
Fiona and Jing. I appreciate your friendship and support for the last four years. I was not
alone on my PhD journey as we were always together. I would like to acknowledge the
Victorian Forest Monitoring Program team (Salahuddin Ahmad and Liam Costello) at the
Department of Environment, Land, Water and Planning, who provided forest inventory data
and support for this research.
I would like to acknowledge the Australian Award Scholarship (AAS) for providing the
funding that made my PhD in Australia possible. My thank goes also to Jamie Low, AAS
coordinator at RMIT, for her assistance in various matters. I also thank FrontierSI (formally
CRCSI) for providing me a top-up scholarship to improve the quality of this research.
I greatly appreciate the constant support of my friends (in both Vietnam and Australia)
during the last four years. Finally, to my family (bố Quang, mẹ Lan, bố Mẫn, mẹ Xuân, chị
Hiền, Trang, và Hiếu Hạnh), without you I was not able to achieve this PhD. Mom and Dad,
I know you will never read and understand what I am writing here (and I will also never tell
you) but you are always my greatest motivation. Most importantly I would like to thank my
wife, Hòa, and my two daughters, Chi and little Cherry; the reasons I get out of bed in the
morning and come back home in the evening! Thank you for always with me, for your
unwavering love and patience. Chi, you had obtained your first master’s with your mom,
and now your first PhD with me. We are so proud of you!
Thank you, everyone.
vi
Contents
Declaration .........................................................................................................................i
Abstract ......................................................................................................................... ii
Acknowledgement ............................................................................................................vi
Contents ........................................................................................................................vii
List of figures.................................................................................................................... ix
List of tables ....................................................................................................................xii
List of publications ........................................................................................................ xiii
Chapter 1. Introduction ...................................................................................................... 1
1.1. Context ................................................................................................................... 2
1.2. Methods for estimating forest biomass .................................................................... 4
1.3. Satellite remote sensing time-series for forest monitoring ....................................... 9
1.4. Objectives and research questions ......................................................................... 12
1.5. Study area............................................................................................................. 13
1.6. Thesis structure .................................................................................................... 14
Chapter 2. Landsat time-series for large area estimating of forest aboveground biomass
dynamics: A review ......................................................................................... 15
2.1. Introduction .......................................................................................................... 17
2.2. Advanced preprocessing and change detection methods for LTS .......................... 18
2.3. How has LTS been utilised to improve the estimation of AGB? ............................ 24
2.4. What LTS-based approaches have been demonstrated for estimating AGB and its
dynamics across space and time? .......................................................................... 29
2.5. Conclusions and future opportunities .................................................................... 45
Chapter 3. A spatial and temporal analysis of forest dynamics over large areas using Landsat
time-series........................................................................................................ 47
3.1. Introduction .......................................................................................................... 49
3.2. Study area............................................................................................................. 52
3.3. Methods ............................................................................................................... 54
3.4. Results .................................................................................................................. 66
3.5. Discussion ............................................................................................................ 74
3.6. Conclusion ........................................................................................................... 79
vii
Chapter 4. A comparison of imputation approaches for estimating forest biomass using
Landsat time-series and inventory data ............................................................. 80
4.1. Introduction .......................................................................................................... 82
4.2. Materials and methods .......................................................................................... 85
4.3. Results .................................................................................................................. 97
4.4. Discussion .......................................................................................................... 103
4.5. Conclusions ........................................................................................................ 108
Chapter 5. Monitoring aboveground forest biomass dynamics over three decades using
Landsat time-series and single-date inventory data ......................................... 109
5.1. Introduction ........................................................................................................ 111
5.2. Study area........................................................................................................... 114
5.3. Materials and methods ........................................................................................ 115
5.4. Results ................................................................................................................ 122
5.5. Discussion .......................................................................................................... 132
5.6. Conclusion ......................................................................................................... 137
Chapter 6. Synthesis ....................................................................................................... 138
6.1. Research questions ............................................................................................. 139
6.2. Application in developing countries .................................................................... 146
6.3. Future directions and opportunities ..................................................................... 148
Bibliography .................................................................................................................. 152
Appendices .................................................................................................................... 173
viii
List of figures
Figure 1.1. Timelines of major Earth observation satellites with optical/multispectral sensors
(Modified and adapted from Kuenzer et al. (2014)) ......................................... 10
Figure 2.1. A common concept for estimating AGB dynamics using LTS data. ................ 35
Figure 3.1. Study area in Eastern Victoria, Australia, covered by four Landsat WRS-2 scenes.
....................................................................................................................... 52
Figure 3.2. Australian forest structural definitions ............................................................ 53
Figure 3.3. Overall research methodology flowchart for characterising forest dynamics using
Landsat time-series ......................................................................................... 54
Figure 3.4. LandTrendr-derived fitted trajectory of NBR and extracted disturbance and
recovery metrics .............................................................................................. 56
Figure 3.5. Disturbance and recovery maps of public forests in Eastern Victoria. (a) and (b)
onset years (grouped in 4 year intervals) of primary and secondary disturbances,
respectively (the black box is the insert shown in Figure 3.10 and Figure 3.11).
(c) and (d) the primary disturbance and recovery levels (see Table 3.3 for
description of categories) and the associated causal agents, respectively. ........ 67
Figure 3.6. Rankings of variable importance as reported by the RF models of disturbance
and recovery levels (phase one). Importance is defined by the mean decrease
accuracy. ........................................................................................................ 69
Figure 3.7. a) Forest disturbance and recovery in 2003 (at the local scale) extracted from the
FDDC. b) Annual disturbance rates combining yearly detected and ongoing
disturbance. .................................................................................................... 71
Figure 3.8. Average annual disturbance rates by different (a) causal agents and (b)
disturbance levels ........................................................................................... 72
Figure 3.9. Annual disturbance rates by (a) wildfire and (b) clear-fell disturbances. ......... 72
Figure 3.10. Tracking 30-year history of pixels of interest using the FDDC. (a) Prediction
maps of disturbance and recovery ingested into the FDDC (at the local scale,
insert box in Figure 3.5a). (b) A Hovmoller graph displays the time-series arrays
(Mxy) of pixels along a 12 km transect (the black line in the maps). The vertical
axis is the distance along the transect, horizontal axis is time. It is important to
note that a “Full/Partial Recovery” status should be interpreted with its associated
time period. For example, a “Full Recovery” labelled for a 10-year period
following a fire means that it took 10 years for fully recovering after the fire. . 73
ix
Figure 3.11. Examples of disturbances followed by partial or no recovery. (a) Disturbance
and recovery patterns at the local scale (black box in Figure 3.5a) with two
marked examples of HD-NR and HD-PR. Images from Google Earth show the
pre-disturbance and current condition of forests on the ground: (b) a clear-fell
logged area (2013) not recovered yet; (c) an example of partial recovery
following a high intensity fire (2003), the current condition is clearly sparser than
pre-event condition. ........................................................................................ 76
Figure 4.1. Overall flowchart of steps used for developing and comparing biomass
imputation approaches. ................................................................................... 86
Figure 4.2. Study area in Victoria, Australia. (a) Bioregions and VFMP inventory plots with
a local map showing examples of random points selected around an inventory
plot; (b) public land forest extent and Landsat scenes; (c) map of Australia. .... 87
Figure 4.3. Example of a trajectory of NBR time-series and extracted change metrics. ..... 91
Figure 4.4. Importance scores of predictor variables (scaled from 0 to 100) to response
variables, reported by the LASSO model. Each box plot associated with a
predictor variable shows the dispersion of importance scores of that predictor
variable to 9 response variables (listed in Table 4.1). ......................................97
Figure 4.5. Generalized root mean squared difference of biomass imputations reported by
kNN models (BM = biomass, BA = basal area, TD = stem density, VL = tree
volume). ......................................................................................................... 98
Figure 4.6. Relative mean deviation (%) of biomass imputations reported by kNN models
(BM = biomass, BA = basal area, TD = stem density, VL = tree volume). ......99
Figure 4.7. (a) The imputation map of total AGB for 2016 across public land forests in
Victoria. (b–d) Imputation maps of total AGB, AGB of live tree and dead tree,
respectively, at a local scale. (e) A Google Earth image (un-scaled) showing the
same area as the local maps. Predictions of AGB are consistent with forest
conditions displayed on the Google Earth image, with total AGB at a medium
level. Live trees predominate in the top-left corner while dead trees are dominant
in the bottom-right corner as a consequence of a 2007 fire. ........................... 100
Figure 4.8. Imputed versus observed biomass values from leave-one-out cross validation (n
= 633), reported by the RF-based BA-TD model. .......................................... 101
Figure 4.9. Boxplots of scaled imputed AGB values by different disturbance severity levels
associated with fire and logging disturbances occurring between 2013 and 2016.
..................................................................................................................... 102
x
Figure 4.10. AGB predictions (scaled values) according to disturbance severity and time
since disturbance (TSD). ............................................................................... 102
Figure 5.1. Study area in Victoria, Australia. (a) The public forest extent, Landsat scenes,
and extent of Lidar capture; (b) IBRA bioregions and forest inventory plots from
the Victorian Forest Monitoring Program (VFMP). ....................................... 115
Figure 5.2. Change metrics extracted from a fitted AGB trajectory................................. 121
Figure 5.3. Relationship between Lidar-based and Landsat-based AGB values across the
8210 validation plots, with the 1:1 line in red. Point density is indicated by a
colour gradient from light yellow for high-density to purple for low-density. 123
Figure 5.4. Validation of AGB change according to the history of forest disturbance and
recovery. The 1:1 line is shown in red and the intercept of x and y axes in black.
..................................................................................................................... 125
Figure 5.5. Predicted AGB maps of 2003 (a) and 2007 (b) in eastern Victoria, Australia;
AGB change over 15-year periods: 2003-1988 (c) and 2017-2003 (b). .......... 126
Figure 5.6. AGB change metrics across Victoria’s public forests during 1988-2017. (a) and
(b) show AGB loss and gain (∆AGB loss and ∆AGBgain) as a result of the greatest
disturbance and subsequent recovery, respectively. (c) and (d) show the RI and
Y2R at a local scale (black box in (a))........................................................... 127
Figure 5.7. Temporal patterns of AGB loss caused by (a) fire and (b) logging ................ 129
Figure 5.8. Distribution of rAGBloss and Y2R by disturbance levels with results from variance
tests, (**** is noted for the significance level of p < 0.0001). ......................... 129
Figure 5.9. Means of AGB loss and Y2R associated with fire disturbance across bioregions,
with 95% confidence intervals. Notes: H = High, M = Medium, and L = Low
disturbance level).......................................................................................... 131
Figure 5.10. Means of AGB loss and number of Y2R associated with logging disturbance
across bioregions. Notes: H = High, M = Medium, and L = Low disturbance
level) ............................................................................................................ 132
Figure A.1. AGB dynamics in un-disturbed forests across bioregions from 1988 to 2017. p
and z statistics are reported by Mann-Kendall trend tests. ............................. 174
xi
List of tables
Table 2.1. Common publicly available and automate change detection algorithms using
LTS. See Table A.1 for a description of spectral indices. ................................ 22
Table 2.2. Benefits of using LTS to improve the estimation of forest AGB (both single-date
and over time). ................................................................................................ 24
Table 2.3. A summary of studies that used LTS-based approaches for estimating forest
AGB dynamics. See Table A.1 for a description of spectral indices. ............... 30
Table 3.1. Predictor variables used for modelling phases (1+2). For each phase, these
variables were used to derive both primary and secondary disturbance events. 57
Table 3.2. Definition of disturbance and recovery levels used in this study. ...................... 61
Table 3.3. Categories describing disturbance and recovery patterns. ................................. 62
Table 3.4. The OOB accuracy assessments of modelling disturbance and recovery patterns
(phase one). The number in each cell corresponds to the number of reference
pixels in that category (HD = High disturbance, MD = Medium disturbance, LD
= Low disturbance, FR = Full recovery, PR = Partial recovery, NR = No
recovery; PA = producer’s accuracy, UA = user’s accuracy). .......................... 68
Table 3.5. Out-of-bag accuracy assessments for modelling disturbance agents (phase two).
The number in each cell corresponds to the number of patches in that category.
....................................................................................................................... 69
Table 3.6. The most important variables for predicting disturbance causal agents (phase two).
....................................................................................................................... 70
Table 4.1. Forest biomass and structure variables extracted from inventory data. .............89
Table 4.2. Predictor variables derived from LTS and topographic and climatic data. ........ 90
Table 4.3. Model scenarios developed for each distance technique (BM = biomass, BA =
basal area, TD = stem density, VL = tree volume, X = denotes response variable
group in each model). ..................................................................................... 94
Table 5.1. Internal assessment of the RF-based kNN model via bootstrapping (n = 633) 123
Table 5.2. Time-series validation of AGB predictions using un-changed Lidar pixels. The
colour ramp dark to light grey corresponds with higher to lower accuracies,
respectively. ................................................................................................. 124
Table 5.3. Spatial summary of AGB dynamics by bioregion and at the state level during
1988-2017 .................................................................................................... 128
Table A.1. Landsat spectral indices commonly used for forest AGB estimates. .............. 173
xii
List of publications
Published peer-reviewed journal articles
Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., & Hislop, S. (2020).
Monitoring aboveground forest biomass dynamics over three decades using Landsat timeseries and single-date inventory data. International Journal of Applied Earth Observation
and Geoinformation, 84. p.101952. https://doi.org/10.1016/j.jag.2019.101952.
Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., & Hislop, S. (2020). Landsat
Time-Series for Estimating Forest Aboveground Biomass and Its Dynamics across Space
and Time: A Review. Remote Sensing, 12(1), p.98. https://doi.org/10.3390/rs12010098.
Nguyen, T.H., Jones, S., Soto-Berelov, M., Haywood, A., & Hislop, S. (2018). A
comparison of imputation approaches for estimating forest biomass using Landsat timeseries
and
inventory
data.
Remote
Sensing,
10(11),
p.1825.
https://doi.org/10.3390/rs10111825.
Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., & Hislop, S. (2018). A
spatial and temporal analysis of forest dynamics using Landsat time-series. Remote
Sensing of Environment, 217, pp.461-475. https://doi.org/10.1016/j.rse.2018.08.028.
Peer-reviewed conference proceedings
Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., & Hislop, S., (2019).
“Estimate forest biomass dynamics using multi-temporal lidar and single-date inventory
data”. In IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing
Symposium
(pp.
7338-7341).
IEEE.
Yokohama,
Japan.
https://doi.org/10.1109/IGARSS.2019.8897905.
Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., & Hislop, S., (2018).
“Extrapolating single-date forest inventory attributes through space and time using
Landsat time series”. ForestSAT 2018 conference. Maryland, USA.
Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., & Hislop, S., (2017).
"Mapping forest disturbance and recovery for forest dynamics over large areas using
Landsat time-series remote sensing", In: SPIE 10421, Remote Sensing for Agriculture,
Ecosystems, and Hydrology XIX (pp. 104210W-110421-104211). Warsaw, Poland.
http://dx.doi.org/10.1117/12.2276913.
xiii
Chapter 1. Introduction
This study aims to develop a robust and consistent remote sensing-based framework that can
support developing countries to improve their capacities in monitoring and reporting forest
biomass and carbon stocks and changes across large areas. This chapter introduces the
research context, provides an overview of biomass estimation methods as well as the concept
of satellite remote sensing time-series for forest monitoring. Finally, an outline of the
research questions and the structure of the thesis is given.
1
Chapter 1
Introduction
1.1. Context
Forest biomass, accounting for over 80% of global vegetation biomass, is considered a key
feature in terrestrial ecology, atmospheric processes and the water and carbon cycles
(Houghton, R. A., Hall & Goetz 2009). Changing forest biomass is a key factor in climate
change issues. Given around 50% of forests is carbon, forest biomass provides estimates of
carbon pools in forests. As a result, biomass represents the potential amount of carbon
emissions when the forest is disturbed. Forest biomass has been recently recognised as a
Global Climate Observing System (GCOS) Essential Climate Variable (ECV) (Duncanson
et al. 2019; GCOS 2010; Le Toan, T. et al. 2011), which is an important input to the United
Nations’ Reducing Emissions from Deforestation and forest Degradation-plus (REDD+)
program and Earth system models (Herold et al. 2019; Sessa & Dolman 2008). Globally,
deforestation and forest degradation are main drivers of forest biomass dynamics and thus
carbon emissions and associated forest disturbance. It is estimated that forest disturbance is
the second-largest source of global carbon emissions, accounting for approximately 17%,
more than from the global transportation sector and second only to the energy sector.
Reducing emissions from the forest sector, therefore, is one of the core requirements to avoid
and constrain the impacts of climate change on Earth (i.e., rising temperatures and sea
levels). Consequently, monitoring forest biomass dynamics is an international concern and
has attracted attention from governments (at local, regional, national and international
levels), academics and the general public.
Most parts of the world are now facing the dual pressures of economic growth and
environmental protection. Adjusting and balancing between these pressures is, therefore, a
matter of concern in many countries, especially in less wealthy regions where demand for
exploiting natural resources are high and significantly increasing. According to the Food and
Agriculture Organization (FAO) Forest Resources Assessment 2015 (FRA2015),
deforestation and forest degradation have been persisting in poor tropical countries, despite
a positive trend observed for the global forest extent during 1990-2015 (Sloan & Sayer
2015). The FRA2015 indicates that demand for wood products (including industrial and fuelwood) increased by 35% since 1990, mainly in tropical developing countries. This resulted
in a significant loss of forest area (over 65 Mha) and extensive canopy cover reduction across
large forest areas within these countries during 1990-2015 (Sloan & Sayer 2015).
Developing countries need a robust and cost-effective national forest monitoring system
2
Chapter 1
Introduction
(NFMS) that can provide them with a comprehensive understanding of forest dynamics over
space and time. Such a system would support their policy-making processes that aim to
protect ecosystem integrity in forests and reduce greenhouse gas emissions while
simultaneously maintain their social-economic development needs, facilitating the
achievement of international agreements and conventions such as the Paris Agreement and
REDD+ (Mora et al. 2012; Tomppo et al. 2009).
Establishing and implementing a NFMS is challenging for many developing countries due
to their organisational and technical capacities. One significant challenge is the lack of a
systematic National Forest Inventory (NFI), which is considered one of the three “pillars”
of forest monitoring in the context of REDD+, for calculating country-specific emission
factors in measuring forest carbon stocks and changes (FAO 2018a). Though forest
inventory capacities of developing countries have improved over time as a result of FAO’s
supporting programs over the last few decades, many countries lack the capacity to perform
regular NFIs. In a recent analysis on the FRA2015, Romijn et al. (2015) indicates that, by
2015, only 40 of 99 tropical Non-Annex I countries (~ 40%) have capacity to develop and
maintain their NFIs on a regular basis. The other 59 countries (~ 60%), mainly in Latin
America and Africa, have low to intermediate capacities which means they are not able to
perform regular NFIs. Forest inventory data may be available in these countries but only for
a single-date and/or be incomplete: i.e. not contain full coverage for all forest areas. In
addition, using out-of-date information is a common problem that hinders reporting
capacities in many countries. Some countries such as Cameroon and Congo established a
NFI network of field plots with support from FAO. However, they are unable to monitor the
inventory plots on a regular basis for a five-year reporting cycle due to their limitations in
technical, financial and institutional capacities (FAO 2018b; Romijn et al. 2015; Tomppo et
al. 2009). As a result, it is a reasonably common phenomenon in developing nations that
they are only now generating their first iteration of comprehensive biomass carbon estimates.
Limitations in NFIs results in negative impacts on carbon reporting capacities of developing
countries. Many countries (17 of the 99 countries) are unable to report on any of the five
carbon pools, according to Romijn et al. (2015). Furthermore, the majority of developing
countries (70%) still use Tier 1 methods for biomass conversion, the least accurate method
as defined by the Intergovernmental Panel on Climate Change (IPCC) (IPCC 2006),
although they are able to report on multiple carbon pools (Romijn et al. 2015). Therefore,
3
Chapter 1
Introduction
improving the quality of carbon reporting is important and should be a current priority of
developing countries. More country-specific data of forest carbon stocks are required to
enable reporting at a higher Tier (Tier 2 or 3) that has lower uncertainties (IPCC 2006). In
the interim, this is a significant challenge for developing countries as they need to firstly
improve their inventory capacities. In other words, they need to establish a systematic NFI
and regularly update inventory data which often require long-term forest monitoring
programs (at least 5-10 years) and extensive costs.
Remote sensing data can be used as an alternative to enable developing countries to
overcome their current challenge. There has been an increasing interest in applying remote
sensing technology to estimate biomass and timely biomass changes efficiently in a cost
effective manner: using the repetitive and comprehensive observations collected by satellite
earth observation from local to large areas (e.g., Du et al. 2014; Huang et al. 2010;
Pflugmacher et al. 2014; Powell et al. 2010; Sarker & Nichol 2011; Tsui et al. 2012). A wide
range of remote sensing data sources are being used for estimating biomass in forests, ranging
from active (Radio detection and ranging (Radar), Light detection and ranging (Lidar)) to
passive (satellite imagery, aerial photos) data. Among satellite data sources, Landsat sensors
are unique since they provide the longest collection of satellite imagery (ongoing since 1972),
at a spatial resolution capable of capturing disturbances caused by both human and natural
activities (Cohen, WB & Goward 2004). As such, the Landsat series of sensors are well suited
for long-term forest change investigations. The free data policy of the Landsat archive since
2008 has revolutionized the way of using Landsat data for worldwide forest research and
monitoring activities, especially forest biomass monitoring (Wulder, MA et al. 2012).
Throughout the last decade, Landsat time-series (LTS) data have been increasingly used for
estimating forest biomass and its dynamics across space and time (Deo, RK et al. 2017;
Gómez et al. 2014; Kennedy, RE et al. 2018; Main-Knorn et al. 2013; Matasci, Giona,
Hermosilla, Wulder, White, Coops, Hobart, Bolton, et al. 2018; Pflugmacher et al. 2014;
Powell et al. 2010; Powell et al. 2013; Zald et al. 2016).
1.2. Methods for estimating forest biomass
Forest biomass is generally divided into above-ground and below-ground biomass (AGB
and BGB, respectively), and live and dead biomass; though there are various definitions
(Brown 1997; IPCC 2006; Tomppo et al. 2009). According to IPCC (2006), AGB often
4
Chapter 1
Introduction
includes all living biomass above the soil (stem, stump, branches, bark, seeds and foliage),
but deadwood and litter are also accounted as AGB in other definitions. BGB includes all
living biomass of live roots, but fine roots with a diameter less than 2mm are often excluded
(IPCC 2006). Generally, BGB can be estimated using field methods, however, from the
remote sensing perspective, it is difficult to estimate BGB using remote sensing data (Herold
et al. 2019; Lu, D 2006). Remote sensing data often provide insufficient information for
estimating BGB, though they can be used to capture different properties of forest biomass
(Herold et al. 2019). Remote sensing applications, therefore, mainly focus on estimating
forest AGB.
1.2.1. Field methods for AGB estimation
Destructive sampling and allometric equations are commonly-used field methods for
measuring forest biomass. The destructive sampling method calculates forest biomass as
the dry weight of all plant materials within a sample plot (Fournier et al. 2003). This method
is an extremely labour-intensive, time-consuming, and destructive technique as it involves
harvesting, drying, and weighing a large number of trees (Catchpole & Wheeler 1992). In
general, using destructive sampling in high biomass density areas is not practical and
repeating these measurements is not feasible (Houghton, R. A., Hall & Goetz 2009).
Indirect estimation methods such as allometric equations have been developed to eliminate
these problems. Allometric equations estimate forest biomass based on other tree
measurements such as diameter at breast height (DBH) and height (Nelson et al. 1999;
Overman, Witte & Saldarriaga 1994). Particularly, DBH and/or height data are collected on
tree species basis; the biomass of each tree (and then forest plot) is calculated using speciesspecific equations.
Field methods are the most accurate approaches for estimating forest biomass. However,
they are unsuitable for estimating biomass across large areas and over time due to the lack
of spatial and temporal coverage, especially for large jurisdictions or remote areas (Soenen
et al. 2010; Wulder, MA et al. 2004). Methods for extrapolating biomass for a large forested
area based on representative plots only, such as using mean biomass density (Woodwell &
Whittaker 1968), biomass expansion factor (Birdsey 1992) or spatial statistics, often produce
significant statistical errors (Dixon et al. 1994; Shi & Liu 2017). Moreover, to estimate
biomass change, field plots need to be remeasured regularly, which is often time, cost, and
5
Chapter 1
Introduction
labour intensive. This is challenging, especially for developing nations with large and remote
forest areas (Herold et al. 2019; Romijn et al. 2015). In the context of REDD+, reporting
carbon emission using only NFI data may result in inconsistencies between countries due to
different field inventory methods.
1.2.2. Remote sensing for forest AGB estimation
Remote sensing data have been commonly used as an alternative for estimating forest
biomass across large areas due to their wide coverage and high spatial resolutions. Remotely
sensed observations, including satellite imagery, are systematically acquired on a regular
basis, enabling the monitoring of biomass dynamics over time. The utility of remote sensing
data in up-scaling forest biomass estimations across space and/or time has been
demonstrated in various contexts. Remotely sensed observations do not directly measure
forest biomass, but the radiometry/signals acquired by sensors are sensitive to vegetation
structure and texture that are highly correlated with biomass (Shi & Liu 2017). Forest
biomass can be accurately estimated using active (Radar and Lidar) or passive optical
(satellite images) remote sensing data.
- Passive optical remote sensing for AGB estimation
Optical remote sensing observes the amount of radiation in the electromagnetic spectrum
reflected or emitted by objects on the ground including vegetation canopy layers. The
reflectance of electro-magnetic radiation (EM-R) is sensitive to forest foliage and is useful
for the estimation of forest AGB over large areas. Green leaves strongly reflect green and
near-infrared (NIR) wavelengths (0.500 - 0.578 µm and 0.750-1.400 µm, respectively) while
chlorophyll in leaves strongly absorb radiation in red and blue wavelengths (Hoffer 1978;
Ripple 1986). Therefore, measuring the response of spectral ranges (or bands) to canopy
layers allow us to determine forest properties. A combination of multiple spectral bands
results in a vegetation index, which can be derived in various ways including rationing,
differencing, and normalised differences. Vegetation indices can enhance vegetation signals
and minimise the negative impacts of environmental conditions on reflectance (e.g.,
atmospheric conditions, sun view angles), improving the correlation between spectral data
with biomass (Crist 1985; Huang et al. 2002; Huete 2012).
Several passive optical satellite systems across a range of spatial resolutions are often utilised
for estimating forest biomass. At the coarse spatial resolution (> 100m), NOAA’s Advanced
6
- Xem thêm -