MRV Manual Chapter5

REDD+ MEASUREMENT,
REPORTING AND
VERIFICATION (MRV) MANUAL
VERSION 2.0
FOREST CARBON, MARKETS AND COMMUNITIES
(FCMC) PROGRAM
DECEMBER 2014
This publication was produced for review by the United States Agency for
International Development. It was prepared by Tetra Tech.
This publication was produced for review by the United States Agency for International Development by Tetra Tech ARD,
through a Task Order under the Prosperity, Livelihoods, and Conserving Ecosystems (PLACE) Indefinite Quantity
Contract Core Task Order (USAID Contract No. EPP-I-00-06-00008-00, Order Number AID-OAA-TO-11-00022).
Tetra Tech
159 Bank Street, Suite 300
Burlington, Vermont 05401 USA
Telephone: (802) 658-3890
Fax: (802) 658-4247
E-Mail: international.development@tetratech.com
www.tetratechintdev.com
Tetra Tech Contacts:

Ian Deshmukh, Senior Technical Advisor/Manager
Email: ian.deshmukh@tetratech.com
Forest Carbon, Markets and Communities (FCMC) Program

1611 North Kent Street
Suite 805
Arlington, Virginia 22209 USA
Telephone: (703) 592-6388
Fax: (866) 795-6462
Stephen Kelleher, Chief of Party

Email: stephen.kelleher@fcmcglobal.org
Olaf Zerbock, USAID Contracting Officer’s Representative

Email: ozerbock@usaid.gov
Editors:
• Jennifer Hewson, Conservation International
• Marc Steininger, FCMC & Conservation International
• Stelios Pesmajoglou, Greenhouse Gas Management Institute
Contributing authors:
• Angel Parra, Consultant; GHG inventory & LULUCF sector expert
• Gordon Smith, Greenhouse Gas Management Institute
• David Shoch, TerraCarbon, LLC
• John Musinsky, National Ecological Observatory Network
• Fred Stolle, World Resources Institute
• Kemen Austin, World Resources Institute
• Irene Angeletti, Greenhouse Gas Management Institute
The US Agency for International Development (USAID) has launched the Forest Carbon, Markets and Communities
(FCMC) Program to provide its missions, partner governments, local and international stakeholders with assistance in
developing and implementing REDD+ initiatives. FCMC services include analysis, evaluation, tools and guidance for
program design support; training materials; and meeting and workshop development and facilitation that support US
Government contributions to international REDD+ architecture.
Please cite this report as:

Hewson, J., M.K. Steininger and S. Pesmajoglou, eds. 2014. REDD+ Measurement, Reporting and Verification (MRV) Manual,
Version 2.0. USAID-supported Forest Carbon, Markets and Communities Program. Washington, DC, USA.
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 i

REDD+ MEASUREMENT,
REPORTING AND
VERIFICATION (MRV)
MANUAL, VERSION 2.0
FOREST CARBON, MARKETS AND COMMUNITIES

(FCMC) PROGRAM
DECEMBER 2014
DISCLAIMER
The author’s views expressed in this publication do not necessarily reflect the
views of the United States Agency for International Development or the United
States Government.
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 ii

TABLE OF CONTENTS
TABLE OF CONTENTS....................................................................................... III
ACRONYMS AND ABBREVIATIONS ...................................................................... V
ACKNOWLEDGEMENTS ................................................................................... VIII
1.0 INTRODUCTION....................................................................................... 1
1.1 PURPOSE, SCOPE AND STRUCTURE ...................................................................................... 1
1.2 BACKGROUND ............................................................................................................................... 4
1.3 REFERENCES...................................................................................................................................... 8
2.0 INSTITUTIONAL ARRANGEMENTS.............................................................. 9
2.1 INTRODUCTION ............................................................................................................................ 9
2.2 ELEMENTS OF A MRV SYSTEM FOR REDD+ .......................................................................13
2.3 KEY FUNCTIONS AND COMPONENTS OF NATIONAL ARRANGEMENTS ......... 17
2.4 STEPS IN ESTABLISHING INSTITUTIONAL ARRANGEMENTS.....................................19
2.5 EXAMPLES ........................................................................................................................................22
2.6 EPA NATIONAL SYSTEM TEMPLATES ...................................................................................26
2.7 REFERENCES....................................................................................................................................30
3.0 ESTIMATING GREENHOUSE GAS EMISSIONS AND REMOVALS .................... 31
3.1 INTRODUCTION ..........................................................................................................................31
3.2 IPCC GUIDANCE...........................................................................................................................34
3.3 INVENTORY AND REPORTING STEPS .................................................................................38
3.4 DEFINITIONS OF CARBON POOLS AND LAND USES ..................................................41
3.5 METHODOLOGIES FOR ESTIMATING EMISSIONS AND REMOVALS ....................... 42
3.6 REFERENCES....................................................................................................................................49
4.0 FIELD-BASED INVENTORIES .................................................................... 50
4.1 INTRODUCTION ..........................................................................................................................50
4.2 CARBON POOLS AND THEIR MEASUREMENT ................................................................54
4.3 CONCEPTS AND CONSIDERATIONS IN INVENTORY DESIGN ................................ 57
4.4 THE FOREST CARBON INVENTORY TEAM .......................................................................66
4.5 FIELD WORK AND ANALYSIS .................................................................................................67
4.6 CALCULATING CARBON STOCKS FROM FIELD DATA ...............................................69
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 iii
4.7 DATA CHECKING ........................................................................................................................75
4.8 CONSOLIDATING INVENTORY DATASETS ......................................................................76
4.9 THE GAIN-LOSS METHOD ........................................................................................................78
4.10 REFERENCES .................................................................................................................................79
4.11 SELECTED RESOURCES ............................................................................................................81
5.0 REMOTE SENSING OF LAND COVER CHANGE ........................................... 82
5.1 INTRODUCTION ..........................................................................................................................82
5.2 LAND USES AND CATEGORIES IN THE UNFCCC ..........................................................84
5.3 OVERALL STEPS AND NEEDS...................................................................................................90
5.4 REMOTE SENSING OVERVIEW ................................................................................................93
5.5 EMERGING AREAS OF RESEARCH....................................................................................... 109
5.6 REFERENCES................................................................................................................................. 114
5.7 COMMON SATELLITE DATA SOURCES FOR LAND-USE MONITORING ............ 118
5.8 SELECTED RESOURCES............................................................................................................ 121
6.0 REPORTING AND VERIFICATION: ELEMENTS AND GUIDANCE .................. 125
6.1 INTRODUCTION ....................................................................................................................... 125
6.2 REPORTING.................................................................................................................................. 127
6.3 VERIFICATION ............................................................................................................................ 141
6.4 REFERENCES................................................................................................................................. 151
7.0 THEMATIC REVIEWS ............................................................................. 152
7.1 HISTORY OF REDD+ UNDER THE UNFCCC .................................................................. 152
7.2 COMMUNITY-BASED MONITORING ................................................................................ 165
7.3 NEAR-REAL TIME MONITORING AND ALERT SYSTEMS ........................................... 181
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 iv

ACRONYMS AND
ABBREVIATIONS
ACR American Carbon Registry
AD Activity Data
AFOLU Agriculture, Forestry and Other Land Use
AGB Aboveground biomass
BCEFs Biomass conversion and expansion factors
BRDF Bi-directional reflectance distribution function
BURs Biennial Update Reports
CH4 Methane
CI Conservation International
CMP Conference of the Parties serving as the Meeting of the Parties to the Kyoto Protocol
CO2 Carbon dioxide
COP Conference of the Parties
CV Coefficient of Variation
DBH Diameter at Breast Height
DEM Digital Elevation Model
DTs Decision Trees
EFDB Emissions Factor Database
EFs Emissions Factors
EM Electromagnetic
EOS Earth Observation System
EPA Environmental Protection Agency
FAO Food and Agriculture Organization
FAS Fire Alert System
FCMC Forest Carbon, Markets and Communities Program
FCPF Forest Carbon Partnership Facility
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 v

FIRMS Fire Information and Resource Management System
FREL Forest Reference Emission Level
FRL Forest Reference Level
FSI Forest Survey of India
FUNCATE Foundation of Space Science, Applications and Technology
GEF Global Environmental Facility
GFIMS Global Fire Information Management System
GFOI MGD Global Forest Observation Initiative Methods and Guidance Documentation
GFW Global Forest Watch
GHG Greenhouse gas
GHGMI Greenhouse Gas Management Institute
GIS Geographic Information System
GLAS Geoscience Laser Altimeter System
GOFC-GOLD Global Observation of Forest and Land Cover Dynamics
GPG-LULUCF Good Practice Guidance for Land Use, Land-use Change and Forestry
GPS Global Positioning System
IDEAM Colombian Institute for Hydrology, Meteorology and Environmental Studies
ILUA Integrated Land Use Assessment
INPE Brazilian National Space Research Institute
IPCC Intergovernmental Panel on Climate Change
KCA Key Category Analysis
LDCM Landsat Data Continuity Mission
LEDS Low Emission Development Strategies
LiDAR Light Detection and Ranging
LUC Land-use Change
MADS Colombian Ministry for Sustainable Development
MCT Brazilian Ministry of Science, Technology and Innovation
MMU Minimum-mapping unit
MRV Measurement, Reporting and Verification
N20 Nitrogen oxide
NAMA Nationally Appropriate Mitigation Strategies
NASA National Aeronautics and Space Agency
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 vi

NCs National Communications
NFMS National Forest Monitoring System
NGGIP National Greenhouse Gas Inventories Program
NGO Non-governmental organization
NNs Neural Networks
NRT Near-real Time
PCA Principal components analysis
PRODES Projeto De Estimativa De Desflorestamento da Amazoni (Brazilian Amazon deforestation
monitoring program)
QA/QC Quality Assurance and Quality Control
QUICC Quarterly Indicator of Cover Change
RADAR Radio Detection and Ranging
REDD+ Reducing emissions from deforestation and forest degradation, plus the role of conservation,
sustainable forest management and enhancement of forest carbon stocks.
SBSTA Subsidiary Body on Scientific and Technical Advice
SES Social and Environmental Soundness
SINA Colombian National Environmental System
SLR Side Looking RADAR
SRTM Shuttle Radar Topography Mission
TACCC IPCC principles of transparency, accuracy, completeness, comparability, and consistency
TOA Top-of-atmosphere
UMD University of Maryland
UNDP United Nations Development Programme
UNEP United Nations Environment Programme
UNFCCC United Nations Framework Convention on Climate Change
USAID United States Agency for International Development
USGS United States Geological Survey
VCS Verified Carbon Standard
WGs Working Groups
WMO World Meteorological Organization
WRI World Resources Institute
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 vii
ACKNOWLEDGEMENTS
The authors thank the various colleagues in our organizations who have commented on the text, including
Colin Silver, Maggie Roth, Rishi Das, and Leif Kindberg of FCMC; Karyn Tabor, Mario Chacon and Johnson
Cerda of Conservation International; Deborah Lawrence of the University of Virginia; Andrew Lister, Ronald
McRoberts, and Charles Scott of the US Forest Service; Chris Potter of the US National Aeronautics and
Space Administration’s Ames Research Center; Christine Dragisic of the US State Department; Megan
McGroddy, SilvaCarbon Consultant; Michael Gillenwater of the Greenhouse Gas Management Institute;
Asim Banskota of the University of Minnesota; Axel Penndorf of BlackBridge; Frank Martin Seifert of the
European Space Agency's Centre for Earth Observation ESRIN; Carly Green and Jim Penman of the Global
Forest Observations Initiative; Brice Mora of the Global Observation of Forest Cover and Land Dynamics;
and members of the USAID Climate Change office. The authors also thank those who have made graphics
available for use in this Manual, including Ned Horning of the American Museum of Natural History, John
Rogan of Boston University, and Jamie Eaton of TerraCarbon LLC. Additional figures are from the websites
of the United Nations Framework Convention on Climate Change, the International Panel on Climate
Change and the Landsat program.
REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 viii
5.0 REMOTE SENSING
OF LAND COVER
CHANGE
Authors: Marc Steininger and Jennifer H ewson
5.1 INTRODUCTION
This chapter focuses on the application of remote sensing-based approaches to forest cover and change
monitoring. Section 3.3, Inventory and Reporting Steps, of this Manual outlines the sequence of steps required for
generating a national greenhouse gas (GHG) inventory. This chapter is relevant to the activities highlighted
on the following page.
REDD+ MRV MANUAL: CHAPTER 5.0 – REMOTE SENSING OF LAND COVER CHANGE 82
STEP 0: Establish Institutional Arrangements.
STEP 1: Estimate the land areas in each land-use category, through stratification and other methods,
for the time period required, for representing areas in the GPG-LULUCF.
STEP 2: Conduct key category analysis (KCA) for the relevant categories. Within the categories
designated as key, assess significant non-CO2 gases and carbon pools and prioritize such pools in
terms of methodological choice.
STEP 3: Design a forest carbon inventory to generate Emissions Factors (EFs), if using the gain-loss
method; ensuring that the requirements in terms of emission and removal factors are met. EFs
represent coefficients that quantify the emissions/removals per unit area.
STEP 4: Generate Activity Data (AD); appropriate to the tier level identified. AD represents the
extent over which a human activity occurs.
STEP 5: Quantify emissions and removals; estimating the uncertainty in each estimate. Emissions
and removals estimates represent the product of the AD by the associated EFs.
STEP 6: Report emissions and removals estimates; using the reporting tables, and worksheets
where appropriate. Document and archive information used to produce the national emissions
and removals estimates following specific instructions under each land-use category, carbon pool
and non-CO2 source.
STEP 7: Perform verification and implement quality control checks; including expert peer review of
the emission estimates following specific guidance under each land-use category, pool or non-CO2
gas.
In the context of the activities highlighted above, this chapter discusses:
• The context of land uses within the United Nations Framework Convention on Climate Change
(UNFCCC);
• A brief review of remote sensing for forest monitoring;
• Overall steps and needs for consideration in developing a satellite-based forest monitoring system;
and
• An overview of emerging areas of remote sensing-based research for forest monitoring.
Remote sensing, especially satellite-based approaches, provides the most practical option for monitoring land
cover change over large areas. This chapter emphasizes optical satellite remote sensing of deforestation.
Optical satellite remote sensing is the most heavily used type of remote sensing for this application, and
deforestation represents the largest source of GHG) emissions from the land-use sector in most tropical-
forest countries. Another important use of remote sensing in a National Forest Monitoring System (NFMS) is
to produce a forest benchmark map. This is needed to define the national forest area at the beginning of a
reporting period and within which carbon stocks and forest changes will be monitored. Finally, remotely-
sensed data represent a key input to a stratification of forest types and can be used to characterize the
seasonality of leaf cover, inundation, and spectral variations due to very different canopy structures.
Stratification should seek to identify forest types with potentially significantly different levels of biomass to
assist field sampling strategies (see Chapter 4). Additional forest strata could be of interest for national
management and planning purposes and such stratification activities can be facilitated through the use of
remotely-sensed data.
The Intergovernmental Panel on Climate Change (IPCC) Good Practice Guidance for Land Use, Land-Use
Change and Forestry (GPG-LULUCF) is a key resource for countries. However, it provides limited
information on specific approaches to remote sensing of land use. The information in this chapter
summarizes remote sensing issues for a NFMS. As mentioned in Chapter 1, other valuable resources include
the Global Observation of Forest and Land Cover Dynamics (GOFC-GOLD) Sourcebook (GOFC-GOLD,
2013), the Global Forest Observations Initiative (GFOI) Methods and Guidance Document (MGD) (GFOI,
2013), and the Forestry and Forest Products Research Institute’s REDD-plus CookBook (Hirata et al., 2012).
Links to additional resources for training on remote sensing are provided in Section 5.8.
5.2 LAND USES AND CATEGORIES IN THE UNFCCC
LULUCF, within the context of the United Nations Framework Convention on Climate Change (UNFCCC),
refers to land-use change or persistence among the six broad uses defined by the IPCC: Forest Land,
Cropland, Grassland, Wetlands, Settlements, and Other Land (IPCC 2006, Vol. 4; see Chapter 2). Possible
types of land-use change among, or persistence within, these six broad uses are called Categories. Sub-
categories can be defined within a category to more precisely define changes and emission sources.
For land use, the IPCC recognizes two methods to estimate carbon emissions: the stock change method 23 and
the gain-loss method (IPCC, 2003), as noted in Chapter 3. This chapter discusses remote sensing of land-use
change in the context of the gain-loss method, where estimates of changes between uses is a specific input to
emissions estimates. However, the remote sensing techniques and issues discussed in this chapter are also
relevant to application in the Stock-Change method, particularly in the stratification of forests for field
sampling.
23 The GPG-LULUCF (IPCC, 2003) uses the term Stock-Change, while the Guidance from 2006 uses Stock-Difference
(IPCC, 2006). There has been no decision for non-Annex regarding use of the 2006 Guidelines, and thus in this
Manual we use the former term throughout, even though the 2006 Guidelines are more up-to-date and use the
latter.
Activity data (AD), or the extent over which a human activity occurs, are data on the area of a Category that
potentially results in GHG emissions or removals, over a given period of time. As illustrated in Chapter 3
(Figure 3.2), AD are combined with Emissions Factors (EFs). EFs are coefficients that quantify the
emissions/removals per unit area, thus providing the data on differences in the carbon stocks before and
after the cover change has occurred. Multiplying the AD, the extent over which a human activity occurs, by
the EFs, the coefficients quantifying the emissions/removals per unit area, provides the estimated associated
GHG emissions for each category.
The IPCC (2006) describes three overall approaches, not to be confused with tiers, for the representation of
land use (see Chapter 3). These approaches, listed in Box 5.1, are used to estimate AD for each Category:
Box 5.1: Approaches

Approach 1: Net area of each land use reported at different time period, yet no tracking of
specific conversions among them
Approach 2: Tracking of land-use conversions, yet on a non-mapped manner
Approach 3: Tracking of land-use conversion on a mapped manner
Approach 3 is most informative and applicable to a mechanism for reducing emissions from deforestation
and forest degradation, plus the role of conservation, sustainable forest management and enhancement of
forest carbon stocks in developing countries (REDD+). However, it is acceptable to use a mix of the three
approaches among regions or categories in a country. Case studies of countries that have used different
approaches are provided in Annex 2A.1 of the GPG-LULUCF. For example, existing data available for the
Argentine Pampas were sufficient for either Approach 1 or 2. Agricultural census data, documenting the area
of each land use over time and with full coverage, existed for the entire region, thus enabling Approach 1.
Data existed on land-cover change, documenting transformations between natural grasslands to pasture and
cropland, thus enabling Approach 2. In Australia, the creation of a multi-temporal map of change in forest
cover as well as some sub-categories enabled Approach 3 for those categories.
It is important to consider the characteristics of land-use parameters that will be monitored and the cost
implications of a full-coverage mapping versus a sampling-based method. While satellite-based remote
sensing is a valuable tool for monitoring several parameters of land use, some types of land-use categories
(e.g., forest degradation), or regions (e.g., mountainous areas), may be more effectively monitored through
airborne or field-based data collection approaches. The costs associated with these approaches could be
significant and thus necessitate a sampling-based approach.
5.2.1 Definition of national forest and other classes

The 2006 IPCC Guidelines for National Greenhouse Gas Inventories consolidated both LULUCF and
Agriculture into the Agriculture, Forestry, and Other Land Use (AFOLU) in Volume 4 (IPCC 2006; See
Chapter 3). Throughout this chapter, definitions have been adapted from the 2006 IPCC Guidelines and the
GPG-LULUCF 2003, unless noted otherwise. While countries must report on land use, satellite monitoring is
more suited to detecting land cover, as it is based on relationships between observed spectra in the images
and the structural characteristics of the soil and vegetation covering land. Land use, however, can usually be
inferred based on local context and a general knowledge of the area.
Forest definition
A fundamental step in the development of a MRV system is the national definition of forest. Countries have
some flexibility in developing their forest definition, yet they are constrained by certain criteria. The definition
must be developed based on both the physical structure of the present and potential vegetation as well as how
the land is used. The physical criteria for forest, and the range countries can select for their definition, are
included below in Box 5.2.
Box 5.2: Definition of forest criteria

Potential to reach a minimum canopy height at maturity of 2m to 5m;
Minimum tree-crown cover of 10 percent to 30 percent; and
Minimum patch size of 0.05 ha to 1 ha.
Tree-crown cover is not the same as leaf cover, as tree crown cover is defined by the periphery of the crown,
while leaf cover is the proportion of leaf versus canopy-gaps. A site is defined as forest if it meets the above
criteria and if its main use is assumed to be forest-related. For example, while an urban park or agricultural
fallow may meet the physical criteria of forest, these areas have urban and agricultural uses (i.e., non-forest
uses), thus, they belong to a non-forest category. Agricultural fallow is a particularly important example for
many tropical countries, as much of their agricultural land is in some stage of fallow. While in terms of
structure these are young, regrowing “forests,” they are part of an agricultural cycle with a defined temporal
period, and are expected to be re-cleared after that period. Therefore, they are part of a non-forest use.
Considering agricultural fallows as non-forest greatly facilitates reporting on deforestation and associated
GHG emissions, since a country would not be required to estimate rates of the appearance of new fallows
and their re-clearance when reporting changes in forest area 24.
According to the IPCC Report Definitions and Methodological Options to Inventory Emissions from Direct Human-
induced Degradation of Forests and Devegetation of Other Vegetation Types, forest degradation could be defined as “a
direct human-induced long-term loss (persisting for X years or more) of at least Y percent of forest carbon
stocks [and forest values] since time T and not qualifying as deforestation or an elected activity under Article
3.4 of the Kyoto Protocol” (IPCC 2003). For example, selective logging may occur in a site defined as forest.
If the tree cover was not reduced enough to pass below the threshold of the forest definition, then the site
remains forest, but has undergone degradation. Conversely, another site that has been logged and did cross
below this threshold could be classified as deforestation. However, in addition to the change in physical
structure, the use of the land must also change. If the site is still under a forest use (i.e., forest concession
subjected to some selective-logging cycle), it would still be defined as forest despite the structural change. In
this case, there are carbon stock losses in the ‘forest remaining forest’ class. This will likely necessitate a
subclass for ‘intact to degraded forest,’ and this subclass should be sampled to estimate carbon stock change.
The GFOI MGD provides a further discussion of forest degradation in Chapter 2.2.2.
Forest benchmark, stratification and change estimation
The development of forest mapping for REDD+ activities can be considered as a trio of components:
creation of a forest benchmark map, stratification of forests within the map extent, and estimation of changes
within the map extent. It is also preferable to conduct these three activities one at a time; attempting to do all
in a single process can lead to slow and overly-complicated processes.
The forest benchmark map should be created first, as this defines the geographical extent for conducting the
other two activities. The development of the forest benchmark map should reflect the national forest
definition, and use as much field information, aerial photos, very high resolution imagery and expert
knowledge as possible to facilitate remote sensing imagery interpretation. For example, if agricultural fallows
24 Examples of national forest definitions are available at http://cdm.unfccc.int/DNA/cdf/index.html
are excluded from the national definition, aerial photos or very high resolution imagery could be obtained
over sample areas to allow analysts to train themselves in the interpretation of fallow versus mature forest in
different parts of a country. This is a critical step for analysts to correctly train and assess classifications of
forest for the benchmark period. Other characteristics that analysts should pay special attention to include
gradations to lower-stature vegetation that fall below a threshold in the national forest definition, such as
montane shrubs, or open-canopy vegetation, such as open woodland.
The benchmark map should be based on satellite data from dates closest to the start date of a monitoring
period. This could be the date of initiation of REDD+ activities, and thus it defines the extent within which
REDD+ crediting may be possible and where monitoring is required. It could also be the start date of a
historical analysis for use in estimating a Forest Reference Emissions Level or Forest Reference Level
(FREL/FRL; see Chapter 2). In the latter case, it will eventually need to be updated to define the extent of
forest at the beginning of the implementation of REDD+ activities. Satellite images for an exact date can be
difficult to acquire due to cloud cover or gaps in data collection. Therefore, image dates used for the
benchmark map may vary. A conservative approach is to avoid using images from before the start date, as
this may result in the inclusion of forests that were cleared just before the start date.
A forest benchmark map should have as few data gaps as possible; gaps can occur as a result of cloud cover,
satellite sensor problems, etc. This may necessitate the acquisition of multiple images for the same area and,
thus, analysts will need to identify an appropriate methodology for analyzing multiple images. See, for
example, Section 5.4.2. of this Manual, or Section 3.5.2 of the GFOI MGD (GFOI 2013); Step 5.2 of the
GOFC-GOLD SourceBook (GOFC-GOLD 2013). The creation of a forest benchmark map will require a
careful and iterative process. Generally, it will necessitate several iterations of image analysis followed by
reviews from experts and stakeholders, such as sub-national governments and REDD+ project implementers.
It is important that a final benchmark map is communicated to stakeholders with any concerns addressed, as
this sets an important precedent for where REDD+ activities may or may not be implemented. Once the
forest benchmark map is finalized, any areas defined as unmanaged forest (see Chapter 3) can be omitted in
order to provide a final definition of where REDD+ can be implemented and where monitoring must be
conducted.
Stratification of forests within the benchmark area should be conducted in coordination with the field-based
inventory team (See Chapter 4) and with the guidance of an expert statistician. Stratification is not required,
although it is recommended as a method to reduce both field-survey costs and carbon-stock uncertainties, as
discussed in Chapter 4. Even the creation of a few broad classes is very useful for these purposes.
Stratification based on satellite imagery can benefit from the inclusion of seasonal information, e.g., where
more deciduous forests may have significantly lower biomass levels. Seasonal information could be obtained
from coarser resolution data, such as monthly image composites from MODIS, or from very high resolution
imagery, if acquired for particular leaf-on and leaf-off months. Other characteristics often discernable from
satellite data are major differences in canopy structures, such as those of inundated forests, liana forests, and
palm forests. Stratification can also make use of ancillary GIS data, such as data on climate, soils, and
elevation. Different potential strata can be assessed by merging with any existing data on carbon stocks, and
those strata with little difference in stocks may be combined, resulting in fewer total strata. If such field data
do not exist, strata can be created based on expert opinions of forest types that should have different biomass
levels, and these can be re-combined later if field data indicate otherwise.
The creation of a forest benchmark map facilitates the estimation of change within forests. Many approaches
exist for estimating change within forests. Section 5.4.2. of this Manual, Section 3.5.2 of the GFOI MGD
(GFOI 2013), and Step 5.2 of the GOFC-GOLD SourceBook (GOFC-GOLD 2013) provide useful
overview information on the range of approaches and methodologies available. A key consideration is the
type of change to estimate. At a minimum, forest clear-cutting (removal of all trees) should be estimated, as
this is the main emissions source for many countries. A decision should also be made on the minimum size of
clearance to be estimated. While the inclusion of the smallest clearances, i.e., smaller than a few hectares, can
potentially yield a more correct and higher rate, the time and resources involved may be costly, could lead to
greater classification error, and may not reflect a significant source of emissions.
Degradation from selective logging is more difficult to detect and often only detectable for more intensive
forms and with data obtained soon after the logging event. Other forms of degradation, such as wildfire, may
be considered. Again, not all of these must be estimated in the same process. For example, a country could
first produce an estimate of forest clear-cutting using one approach, and then use a different approach, such
as sampling of higher-resolution data or even targeted field surveys, for degradation.
Other classes
It may be important to further stratify the six broad use classes into sub-classes where carbon stocks vary
significantly. This should be assessed as part of both the national Key Category Analysis (KCA) discussed in
Chapter 3 and the forest stratification process discussed in Chapter 4. Including sub-classes may provide data
that are very useful for REDD+ national strategies and management policies. However, countries should
have strong justification for including sub-classes since this will require methodologies that consume more
resources. The expected increase in the accuracy and overall usefulness of emissions estimates should be
clearly assessed, and countries should consider the trade-off that usually exists between thematic precision of
a land-use change study and the accuracy of the change estimates (e.g., Mather, 1999; Foody, 2000). Tables
5.1 and 5.2 illustrate this below. Table 5.1 presents an example of a land-cover change matrix with three
broad categories (forest, degraded forest, nonforest). Conversely, Table 5.2 presents an example with greater
thematic precision where the broad categories have been divided into sub-classes.
Issues related to sub-classes are somewhat different for forest versus non-forest. For forests, a country will
conduct a stratification for a national forest inventory. This can be performed one time, following the
creation of the forest benchmark map, as it can be assumed there are no transitions among these naturally-
occurring vegetation types over the required reporting periods. In contrast, transitions among different post-
deforestation land uses do occur over short time periods. Spectral distinction of these uses is often difficult,
especially when one cannot be very selective about the season of the imagery used for analysis because of
frequent cloud cover. For example, managed grassland, cropland, plantations and fallows may be difficult to
distinguish, depending on the season and stage of crop development at the acquisition times of available
images. Again, strong justification in terms of improvement of emissions estimates is needed to justify
attempts to include transitions among these classes.
A step-wise process may be worth exploring that uses different methods and levels of detail for different
monitoring aspects. For example, an initial step could be to produce a forest benchmark map, with forest
sub-classes that have significantly different carbon stocks. A second step could be to produce a map of a
single, broad deforestation class that occurs anywhere within this benchmark. By combining the two,
deforestation can be attributed to different forest sub-classes. A third step could be to use samples of
airborne or other very-high resolution data to estimate the proportions of sub-classes of non-forest following
deforestation, as well as any important transitions among those.
Combining approaches like these can provide all of the necessary estimates to complete a full land-cover
change matrix, while not requiring a very difficult process of spectral classification of all transitions among
sub-classes. This is an area where there are many options, and many different opinions within the research
community.
T2
a) Forest Degraded Non-forest Sum T1 b) Forest Degraded Non- %

Forest Forest forest T1
Forest 9,940 40 20 10,000 Forest 99.4 0.4 0.2 100
Degraded 5 1,970 25 2,000 Degraded 0.3 98.5 1.3 100

T1 T1
Forest Forest
Non- 4,000 4,000 Non- 100.0 100

forest forest
Sum T2 9,945 2,010 4,045 % T2 99.7 98.9 101.5

Table 5.1: Example of a land-use change matrix with few land-use classes and change categories. “Forest” in this table is non-degraded
forest only. “Non-forest” includes all non-forest, both naturally-occurring and anthropogenic. Values in (a) are in absolute units, such as
hectares, and in (b) are percentages. T1 and T2 are the first and second time periods, referred to in the IPCC as “Initial land-use class”
and “Land use during reporting year.” Values in Sum T1 and Sum T2 are total area and percent change for each class. Values inside the
matrix are areas and percent change for each category of persistence or change. In this example, gross deforestation plus forest
degradation is 0.6 percent (adding values 0.4 and 0.2 in the first row of (b)).
a) T2
Degraded Degraded
Lowland Montane Natural
Lowland Montane Fallow Cropland Pasture Sum T1
Forest Forest Grassland
Forest Forest
Lowland Forest
7945 35 3 5 7 7995
Montane Forest
1995 5 2 3 2005
Degraded Lowland
Forest 5 1500 2 6 12 1525
Degraded Montane
Forest 470 1 4 475
T1
Natural Grassland
993 3 4 1000
Fallow
350 50 150 550
Cropland
200 700 100 1000
Pasture
50 1400 1450
Sum T2
7950 1995 1535 475 993 608 771 1673
b) T2
Degraded Degraded
Lowland Montane Natural
Lowland Montane Fallow Cropland Pasture % T1
Forest Forest Grassland
Forest Forest
Lowland Forest
99.4 0.4 0.1 0.1 100
Montane Forest
99.5 0.2 0.1 0.1 100
Degraded Lowland
Forest 0.3 98.4 0.1 0.4 0.8 100
Degraded Montane
Forest 98.9 0.2 0.8 100
T1
Natural Grassland
99.3 0.0 0.3 0.4 100
Fallow
63.6 9.1 27.3 100
Cropland
20.0 70.0 10.0 100
Pasture
3.4 96.6 100
% T2
99.7 99.5 98.8 99.2 99.3 87.5 80.8 135.1
Table 5.2: Example of a land-use change matrix with more precise land-use classes and change categories. “Forest” here means intact,
non-degraded forest, according to the national forest definition. Natural Grassland, Fallow, Cropland, and Pasture represent non-forest
classes. Values in (a) are in absolute units, such as hectares and in (b) are in percent. T1 and T2 are the first and second time periods,
referred to in the IPCC as “Initial land-use class” and “Land use during reporting year.” Values in Sum T2 are total area and percent
change for each class. Values inside the matrix are areas and percent change for each type of category. In this example, the majority of
forest occurs in the lowlands, the majority of deforestation (to fallow, croplands, and pasture) and forest degradation also occurs in the
lowlands. A high degree of rotational land use is also indicated by, for example, the large areas of change from cropland to fallow (200) or
pasture (100). The 12.5 percent reduction in fallow indicates intensification of land use, either via a shortening of fallow cycles or an
increase in permanent pasture. The 35.1 percent increase in pasture indicates an increasing importance of this use.
5.3 OVERALL STEPS AND NEEDS

Figure 5.1 illustrates the key decisions a country should consider when developing an effective and efficient
satellite-based forest monitoring system after identifying which categories and sub-categories to monitor and
the scale of monitoring. Criteria include the type and resolution of satellite data and the degree to which a full
coverage, or sampling-based, approach should be applied to monitoring different land cover classes. The
appropriateness of different monitoring methodologies will need to be assessed, including the types and
availability of different satellite data, pre-processing, classification algorithms, level of automation and analyst
expertise. Where automation is not possible, it is important to consider how consistency will be achieved and
what methods will be used to effectively combine data from different time periods. Each of these decisions is
discussed below.
Figure 5.1: Key considerations in the development of a NFMS
1) What categories are most important to monitor?

A KCA involves identifying the major land-use-based sources of GHG emissions. A KCA should be
performed as part of the development of a REDD+ strategy within the national development-planning
context. For MRV, the process should extend to defining the types of land-cover changes that are major
GHG contributors and to aligning these definitions with the land-cover change categories defined by the
GPG-LULUCF. Finally, a country must determine the geographical extent of managed land, and thus where
monitoring should be conducted (see Chapter 2).
2) What are the appropriate scales and/or sampling approaches for monitoring?
Once the categories and classes to monitor have been assessed, it is necessary to consider the appropriate
scale and approach. For example, do change events occur in small patches of several hectares, or are they
much larger? Different types of changes may also be most appropriately monitored with different sources of
data. For example, some land-use dynamics may be very appropriate for satellite-based monitoring, whereas
other dynamics, particularly some forms of degradation and post-deforestation land-use changes, may require
airborne or field-based monitoring. These latter land-use dynamics may require more costly data -collection
processes and thus necessitate a sampling approach. Further, some vegetation types, such as deciduous
woodland, may require data from particular or multiple seasons within each year, again necessitating increased
data and analysis demands.
3) What methodological aspects should be considered?
A country should consider a range of methodological options. This will avoid either: (i) the use of methods
with little justification based on in-country testing, or (ii) spending too much time investigating issues that can
be well-informed by existing literature or are not significant to the potential accuracy of the final emissions
estimates. Many differing views exist regarding optimal methods for monitoring land-use change and,
therefore, a country should seek its own cadre of experts with strong fundamental backgrounds in remote
sensing to fully access and understand the relevant literature and options. A country should obtain opinions
from a range of international experts, and conduct assessments with national data, focusing on the categories
identified in the KCA.
Some of the main questions to consider within the methodological options are:
1) What types of satellite data are most appropriate for monitoring the classes identified?
2) What type of classification approach should be used?
3) What types of pre-processing are needed for the particular method of image analysis being considered,
and what level of analyst expertise is required?
4) How much of the process can be automated, and for those parts that are dependent on analyst
interaction, how can consistency and reliability be assured?
5) How should data from different time periods be combined to produce change estimates?
6) What post-classification processing steps should be applied?
7) What validation approach should be used, including data sources and sampling?
Some of the most important considerations are whether to use optical data versus Radio Detection and
Ranging (RADAR) data, what spatial resolution is needed and whether the data source has an appropriate
archive and acquisition strategy.
For question 4, a country should seek to produce the most accurate estimates possible for key categories while
using an approach that is “replicable.” This is a fundamental requirement of the GPG-LULUCF, although it is
only vaguely defined in the context of satellite monitoring. Question 5 includes both the approach to processing
the satellite imagery from multiple dates and the approach to estimating change rates from completed land-use
change maps or sample estimates.
By distinguishing mature forest versus fallow areas, either during creation of the forest benchmark map or
during the stratification of forests in this map, a country can minimize confusion between mature forest
clearing and fallow cycles in later monitoring. While older secondary forest can be difficult to distinguish
from mature forest, most fallows younger than 10 years are distinguishable, and the great majority of fallow
periods are shorter than 10 years. Question 6 includes merging of temporary sub-classes, possibly combining
information from multiple dates into a single multi-date product, and often some type of filtering to a defined
minimum-mapping unit (MMU). The MMU should be smaller than the minimum patch size included in the
national forest definition, or a case should be made that using a larger MMU does not significantly affect
resulting area estimates. However, while using a larger MMU may not significantly affect area estimates for
static areas of classes, estimates of change can be very sensitive to the MMU. Question 7 should consider
various sampling schemes and the availability of very high-resolution satellite or aerial observations, as well as
an independent team of analysts to interpret validation data and conduct error calculations.
5.4 REMOTE SENSING OVERVIEW

This section provides a summary of remote sensing fundamentals. Numerous text books are also available on
remote sensing of land-cover. Links to selected internet resources are provided in Appendix 5B.
5.4.1 Types and characteristics of remote sensing data

Remote sensing is the process of sensing energy emitted or reflected at some wavelength along the
electromagnetic (EM) spectrum by an object, rather than being in direct contact with it. The human eye, for
example, senses a relatively small portion of the total spectrum of energy emitted by the sun; this is the visible
portion of the EM spectrum. The amount and type of energy sensed is usually recorded in digital form; the
amount representing the strength of the signal, and the type representing the recording of the signal across a
spectrum. Fundamental assumptions, though not always valid, are that different land-cover types can be
distinguished based on this recorded information and that land use can be inferred from land cover.
Satellite-based remote sensing is most common because of the full, repeated coverage offered by one or more
satellite data sources; thus enabling national monitoring for terrestrial-based applications. Airborne remote
sensing capacities are also of interest, as such systems could be applied over large regions, or entire countries,
depending on the type of equipment, sampling approach, and resources available. At the highest level, two
broad types of remote sensing for monitoring land-cover exist: passive and active.
Passive remote sensing

The majority of remotely-sensed data used for monitoring land use is passive. Passive remotely-sensed data
are acquired by a sensor that passively receives energy originating from another source; the instrument does
not emit its own signal. The sun is the source for visible and shortwave-infrared spectral regions of the earth,
the feature itself is the source for thermal-infrared regions (Figure 5.2). The portion of the sun’s energy across
these spectral regions reflected by the land surface is often indicative of the structural and chemical
characteristics of the surface features (Figure 5.3). Different spectral regions are represented by relatively
narrow “spectral bands” (Figure 5.4) and, by combining images of energy measured in different spectral
bands and assigning a separate color for display, “multi-spectral” images are produced, as illustrated in Figure
5.5.
Figure 5.2: Optical satellite remote sensing. Shortwave energy is emitted by the sun, passes through the
atmosphere, reflects off a surface, passes again through the atmosphere and reaches a sensor on board a satellite.
The signal detected is dependent not only on the reflectance properties of the surface but also on the sun angle,
topography, view angle and atmospheric properties.
Figure 5.3: Generalized spectral curves of fundamental features in remote sensing of land-cover. Most types
of land-cover are a mixture of these features, plus non-green vegetation and shadows caused by the geometry
of terrain and vegetation.
A) B)
Figure 5.4: Example of spectral resolution. Both (a) and (b) represent the entire visible range of the electromagnetic
spectrum. Spectral bands are defined by a range of wavelengths, and in the example here they are divided by white lines. A
single channel of a multi-spectral sensor is sensitive to energy only within a certain band. In (a), the bands cover a wide
range of energy, and a sensor with such bands would be considered a broad-band sensor. In (b), the bands are narrow, and
a sensor with channels along these bands would have a high spectral resolution. A sensor, such as illustrated in (b), with so
many channels and bands would be considered hyper-spectral.
Figure 5.5: Image data combined from three sensor channels to produce a multi-spectral image. Such color-composite
images aid visualization and interpretation of the land-cover. Brightness levels, shown as grey tones, represent values in the
individual channels.
In addition to the visible and near and shortwave-infrared regions, passive remote sensing systems also
acquire data in the thermal region. Thermal energy is emitted by the land surface itself and, while rarely used
for distinguishing types of land-cover, it facilitates the detection of clouds, active fires, and urban heat islands,
as well as modeling various ecosystem processes and vegetation-climate interactions.
Active remote sensing

In active remote sensing, an instrument sends out a signal at certain wavelengths and measures the return
time and strength of the back-scattered signal. RADAR and Light Detection and Ranging (LiDAR) are the
most commonly used active remote sensing techniques for terrestrial applications.
In forest environments, RADAR information is primarily related to structural features at the wavelength scale
of the energy being sensed, versus optical sensors which measure reflected energy that is largely a function of
canopy architecture, leaf pigments, and soil background. RADAR data provide information related to the
density of leaves in the canopy, or branches and tree trunks, depending on the wavelength used. RADAR data
are also sensitive to canopy and soil moisture, and are extremely influenced by topography. One major
advantage of RADAR systems is their ability to penetrate clouds due to the longer wavelengths, the
microwave portion of the EM spectrum. Because of their sensitivity to the geometric properties of forests,
RADAR data have potential for relating to forest biomass. RADAR data, and data from other satellite
sources, for example, were used to produce two recent maps of global forest biomass (Saatchi et al., 2011;
Baccini et al., 2012).
Until recently, all RADAR sensors on board satellites collected measurements in only one wavelength band
and one polarization. The resulting images did not have the dimensionality that multi-spectral images have,
and thus yielded limited potential for classification of land-cover types. Several recent satellites carry RADAR
sensors that collect data in multiple bands and in different polarizations, thus extending their utility for
classification of land-cover types. RADAR is further discussed in Section 5.5.
As with RADAR, LiDAR instruments emit a pulse of energy, some of which is scattered back to the sensor
by the target. The distance between the sensor and the target is then calculated from the elapsed time for the
LiDAR signal to make a complete round trip. However, in contrast to RADAR, LiDAR operates in the
visible and near infra-red portions of the EM spectrum and, thus, does not penetrate clouds. Applications of
LiDAR in forestry have mainly focused on measurement of canopy height, sub-canopy topography, and the
horizontal and vertical distribution of vegetation; these parameters can be used to model estimates of
aboveground biomass (see, for example, Clark et al., 2004; Lefsky et al., 2002; Lim et al., 2003; Nelson et al.,
2009).
LiDAR systems are also generally classified into full-waveform LiDAR and discrete LiDAR systems (see
Figure 5.6). Full-waveform systems record the entire waveform of a returning pulse, while discrete systems
sample a discrete number of points, usually between one and five, per transmitted pulse. Both forms of
LiDAR have been shown to be useful for estimating forest biomass via comparison with field data and
modeling. While some LiDAR instruments collect data only along sampling lines, others have scanning
abilities to collect data both along and across sampling lines, enabling the creation of images.
Full Discrete
Waveform Return
LiDAR LiDAR
Figure 5.6: Example of full waveform vs. discrete waveform LiDAR (Lim et al. , 2003)
The majority of LiDAR remote sensing to-date has been airborne-based. However, one LiDAR instrument,
the Geoscience Laser Altimeter System (GLAS) on board the Ice, Cloud, and land Elevation Satellite
(ICESAT) was satellite-based. Though the ICESAT satellite is no longer operational, GLAS provided full-
form LiDAR information for linear tracks along the satellite path, with a ground resolution of 70m. The
linear samples from GLAS were inputs to the two global biomass studies noted above.
Resolution and other considerations

In addition to the type and spectral characteristics of different images, consideration must be given to: spatial
and temporal resolution, data collection strategy, cost of procuring the imagery and image archive length.
Spatial resolution is important as the resolution must be fine enough to detect the changes of interest (i.e., at
least half the size of the scale of changes). Publically available data have spatial resolutions ranging from 0.7m
to 1km. Data used for land-cover monitoring have resolutions ranging from 5m to 30m. For example, A 30m
resolution observation, or pixel, represents a ground area of 900 m2 (30m by 30m); a ground area of one
hectare would be represented by 11 pixels at this resolution.
Data with coarser spatial resolutions are used for global studies and are not generally suitable for land-cover
monitoring, as such data will not detect smaller-scale changes. The use of such data for land-cover
monitoring yields an inherent bias in such derived estimates. Conversely, high resolution data, less than 10m,
has traditionally only been used over small areas because of cost and frequency of availability. However, as
high resolution data continue to become available and more affordable, the use of these data over large areas,
especially via sampling, is becoming more practical.
Temporal resolution refers to the frequency with which data are collected. Many satellites, such as Landsat,
have defined orbits that dictate how frequently the satellite will return to view the same location on the earth
and acquire a new image. The re-visit time for Landsat is 16 days, so the most that an area could be
monitored is every 16 days. However, persistent cloud cover often reduces the frequency of acquiring useable
images. Other satellites, including many of the high resolution sensors such as RapidEye, Quickbird,
IKONOS, WorldView-2, SPOT HRV series, CBERS HRC, GeoEye-1 & -2, the DMC constellation,
KOMPSTAT-2 or RESOURCESAT-1 are pointable, meaning they can be tilted to view a location that is at
an angle to their defined orbit. While this can result in the same area being repeatedly imaged at much higher
frequency, such acquisitions occur for short periods and some require tasking. While this has historically
limited the practicality of using such satellites in a monitoring capacity, Box 5.3 highlights two examples
where very high-resolution imagery is being used extensively in the development of NFMSs.
Box 5.3. Examples of the use of high resolution imagery in NFMSs

While many countries are generating forest benchmark maps and deforestation monitoring products
using Landsat imagery, some countries are investigating the utility of other sources of, for example,
higher resolution imagery for monitoring purposes, especially in areas where degradation pervades.
Guyana, for example, has implemented a national-level, annual monitoring activity using wall-to-wall
RapidEye imagery. For Guyana, RapidEye, with five MSI bands of 6.5m, has proven particularly useful
for assessing and addressing the impacts of forest degradation.
Mexico utilizes Landsat in its MRV system and has also incorporated RapidEye. It uses RapidEye
imagery from the dry and wet season to better differentiate between seasonal biomass coverage.
Data archive length is another important consideration for developing historical analyses, even for periods as
brief as the past decade. To facilitate consistent monitoring and ease of logistics, it is preferable to work with
a single source of data throughout a study period when possible. The Landsat series is the most common data
source for monitoring land-cover change, as data extend back to 1972 for the multi-spectral scanner (MSS)
and 1982 for Thematic Mapper. Figure 5.7 illustrates the archive history of the Landsat satellite series.
Further, as the Landsat satellites have a defined orbit, roughly the same area is acquired each time the satellite
returns to view the same location on the earth, meaning image pairs from multiple dates mostly overlap.
Figure 5.7: Landsat archive timeline. Landsat 1 – 3 carried only the MSS instrument; Landsat 4 – 5 carried both the MSS &
TM instruments; Landsat 7 carries the ETM+ instrument but since 2003 has experienced data gaps due to mechanical
failure; Landsat 8, carrys the OLI and TIRS instruments. From http://landsat.usgs.gov/about_ldcm.php
Finally, a satellite program’s future sensor launch and data acquisition strategy, as well as cost policies, are
important considerations when planning a monitoring program. The Landsat Data Continuity Mission
(LDCM), for example, ensures that future National Aeronautic and Space Agency (NASA) satellites will
continue to provide a long-term data record, including the successful launch of Landsat 8 in February 2013.
Landsat’s no-cost data policy allows flexibility in data use. Costs of other data sources are also trending
downwards. The upcoming launch of the Sentinel-2 series, with 2A scheduled for launch in 2015, will provide
an additional source of imagery for consideration in the development of NFMS (see Box 5.4). RapidEye, with
five satellites in its current constellation, has a follow-on mission planned for 2019-2020; thus providing
another source of data for an extended period. A table of current and future satellite data options, including
characteristics, is listed in Section 5.6.
Box 5.4: Additional future data sources for consideration in the development of NFMSs;
the example of Sentinel-2
The European Space Agency (ESA) is responsible for the space component of the Copernicus
programme, under which a series of dedicated satellites, the Sentinels, are being constructed. The
Sentinel-2 series of multi-spectral imaging satellites, with 13 bands located in the VIS, NIR and SWIR
wavelengths and spatial resolutions of 10m, 20m and 60m, will provide additional data options for
consideration in the development of NFMS. The first satellite, Sentinel-2A, will be launched between
May-July 2015 and will offer a revisit time of 10 days. When the second unit (Sentinel-2B) is launched
in 2016, a 5-day revisit time will be possible. The interoperability between these sensors and the
Landsat missions will further enhance the revisit time and improve overall data availability. Based on
the free, full and open data policy adopted for the Copernicus programme, the Sentinel data products
available to all users.
In summary, key data characteristics to consider are:

• What geographical, phenological, and atmospheric (especially persistent cloud cover) conditions
exist?
• What are the spectral regions, and bands within them, where data are collected, and how do these
relate to the potential for distinguishing the land-cover types of interest, and changes among them?
• What is the spatial resolution of the data and how appropriate is it, relative to the scale of the land-
cover changes to monitor?
• What is the temporal resolution in terms of potential frequency of acquisition of non-cloudy
observations compared to the desired frequency of monitoring?
• What is the longevity of the image archive length – does this meet the historical mapping needs?
• What are the cost implications of these data in terms of purchase and analysis?
• What are the future satellite development and launch commitments?
5.4.2 Image pre-processing, analysis, and post-processing
Image pre-processing refers to any step that is applied to an image in preparation for the image analysis step,
and image analysis is the process of generating a land-cover class for all parts of an image. Post-processing
occurs after the image analysis step, and enables the estimation of rates and patterns of land-cover change to
be generated.
Pre-processing usually includes geometric registration and co-registration, atmospheric correction, and
occasional data transformation. Atmospheric correction may be necessary depending on the image analysis
approach that will be used. Data transformation, though useful, is optional depending on the image analysis
approach. As previously outlined, post-processing activities may include a number of steps. Finally, the
calculation of change rates and error estimates are required. The summary of pre-processing, analysis, and
post-processing steps below is based on optical data and approaches to classification, using examples of
Landsat data analysis.
Image pre-processing
Geometric registration and co-registration
Geometric registration is the process of mapping data in a geographical coordinate system. This is to
understand the geographical area represented and is applied when importing the image into a GIS or image-
analysis format for processing.
However, geometric registration may have errors up to 100s of meters. Therefore, although images have been
geometrically registered, it does not mean that images of the same area acquired from different dates will
overlay well enough to avoid errors in change estimates resulting from poor co-registration. Therefore, co-
registration may still be necessary. Co-registration is a standard, simple process that takes a modest amount of
time and involves the identification of one image to use as the base image to which the remaining images will
be co-registered. Automation is increasingly available for processing numerous images, but traditional analyst-
driven methods are also sufficient. The United States Geological Survey (USGS) has been reprocessing much
of the Landsat archive, resulting in the creation of a L1T precision and terrain corrected product 25. These data have
already been geometrically corrected using precision ground control points and Shuttle Radar Topography
Mission (SRTM) Digital Elevation Model (DEM) information, yielding a dataset with accuracies within 30m
and, thus, eliminating the need for further geometric correction. Co-registration among images should be
reviewed and may require adjustments.
Atmospheric correction
The atmosphere has several effects on visible and infrared energy as it passes through the atmosphere from
the sun to the land and back to a satellite or airborne sensor (Figure 5.8). Atmospheric correction is
frequently performed in combination with a bi-directional reflectance distribution function (BRDF)
correction. BRDF defines how light is reflected from a surface, and is dependent on both the incident and
reflected directions.
25 http://landsat.usgs.gov/descriptions_for_the_levels_of_processing.php
Atmospherically-corrected images contain data representing surface reflectance, in unitless values from zero
to one, as opposed to the digital numbers of the raw image data. Most atmospheric correction algorithms are
applied to satellite images prior to mapping, and use a single correction algorithm for the entire image. These
usually assume constant atmospheric conditions across an image, although there is active research on
accounting for variability within an image.
Performing atmospheric correction depends on the image analysis approach used (Song et al., 2001).
Approaches to classification that involve the creation of sub-classes for each type of land use and change can
yield accurate maps without atmospheric correction because sub-classes can account for different
atmospheric conditions. Conversely, methods that apply constant class signatures over images with variable
atmospheric conditions should include atmospheric correction. Some semi-automated methods apply
constant signatures over multiple images or image dates, and these methods are highly dependent on careful
atmospheric correction.
Figure 5.8: Atmospheric effects on optical data. The electromagnetic radiation source is the sun, and this radiation can be
blocked or scattered by clouds in addition to being affected by a “clear” atmosphere. A “clear” atmosphere still causes
scattering and absorption of the radiation as it is transmitted from the sun to the earth and back to the satellite. Sun and sensor
view angles also impact the effects of the atmosphere. BRDF characterizes how an object illuminated by a source, such as the
sun, appears brighter or darker depending on the angle of the source and the angle at which it is viewed by a satellite sensor.
Several programs exist to perform atmospheric corrections over entire images. LOTRAN and 6S are the most
common, and several tools have been created to facilitate their application. One example, developed by
NASA, is the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) tool (Masek et al.,
2008). LEDAPS uses information on water vapor, atmospheric pressure, ozone, a topography-dependent
Rayleigh scattering correction and an aerosol optical thickness component based on Kaufman et al. (1997) to
generate a surface reflectance value for each pixel. LEDAPS also generates water, cloud, cloud shadow and
snow masks.
These corrections can be applied to several partially-cloudy images of the same area. The images can then be
combined to produce a single, “gap-filled” composite mosaic. While the corrections are applied to all the
images and the resulting composite should therefore appear seamless, atmospheric artifacts may remain,
appearing as darker or brighter patches (Figure 5.9), and requiring each atmospherically-corrected image to be
classified separately and then combined. Alternatively, additional algorithms based on, for example, local
histogram matching could be applied to further reduce artifacts. In addition, beginning in summer 2013, all
Landsat data, including those from the LDCM, will be available with top-of-atmosphere (TOA) reflectance
corrections applied 26.
26 http://landsat.usgs.gov/about_LU_Vol_7_Issue_2.php#3b
A) B)
C) D)
E) F)
Figure 5.9: Example of atmospheric correction of Landsat data from San Martin, Peru. A) an unstretched “true-color
composite” where the red, green and blue bands are assigned to the red, green and blue colors in the display. The black
lines are SLC-off data gaps. B) the same image, but with a Gaussian stretch applied to the data histogram. C) the near-
infrared, middle-infrared and red bands assigned to the red, green and blue colors, a common assignment for a “false-color
composite”, allowing visual exploration of the infrared data. D) the same, but post-atmospheric correction and with a
cloud/cloud-shadow mask applied using LEDAPS. Note that while the linear gaps have been filled, some of the cloud
gaps remain since they were also cloud in the second image, thus requiring additional images. E) a mosaic of two
atmospherically corrected images, but with no histogram matching between them applied; note the orange-tone artifacts
that appear to the left of the remaining cloud gaps in the upper-left of the image. F) mosaic of the same two images ,but
with histogram matching applied; the artifacts in E) are no longer visible.
Data transformation
Some analysis methods include data transformation techniques, such as various forms of ordination, prior to
classification. Principal components analysis (PCA) is one example of a transformation technique involving
ordination. These techniques alter the information to facilitate interpretation. The “Tassled Cap”
transformation has been used extensively in the classification of vegetation types. Spectral mixture analysis
(SMA) is another data transformation approach. SMA utilizes estimated spectral reflectances of a set of
“pure” features that aim to represent the observed surfaces. In vegetated lands, for example, these are sunlit
leaf, soil, and woody vegetation or litter. Soil theoretically could be split into multiple soil types with different
reflectance properties, and water could be included. Shadow is also generally included, as this is an important
feature of most spectral images because of the geometry of the vegetated canopies.
SMA involves defining the spectral reflectance of each main feature representing the landscape under study
and, based on these, estimating the proportions of each component for each pixel. Definition of the pure
features may be via laboratory analysis, field analysis, or literature. When applied in SMA, they are referred to
as “spectral end members,” since they are located at the outer ends of the multi-dimensional distribution of
the spectral data. End members can also be defined by simply selecting the extreme pixels in the multi-
dimensional data; these are termed “image end members.” However, if image end members are used, the
resulting SMA analyses are relevant only to that image. An output could estimate, for example, that a
particular pixel represents an observation of a piece of land that is 30 percent sunlit leaf, 20 percent soil and
50 percent shadow. These can be visualized as “fractional images” and used as inputs to classifications. SMA
can be a useful approach in understanding the spectral data contained in the image data, as it explains the data
in terms of physical features. Like PCA, and other types of data ordination, SMA does not add to the
information content. Depending on the classification approach used, these techniques may help to produce
more accurate or efficient land-use classification.
Classification
Land-cover classification produces a thematic representation of land by categorizing pixels based on their
spectral signatures. Two broad types of classification exist: supervised and unsupervised.
In supervised classification, the analyst identifies “training sites” and creates training data by delineating areas
known to be of each class. Statistics of the pixel data within these areas are calculated and, at a minimum,
these include the means, variances and co-variance matrix of the spectral data, defining the “spectral
signature” of each class. The level of statistical separability among the classes can be evaluated, and this may
suggest a need to merge or add additional sub-classes. Based on these statistics, various algorithms can be
used to estimate the most likely class of the remaining, unidentified pixels, yielding a classified image.
Often, output classifications are evaluated and, based on conspicuous errors, training data are modified and a
new iteration of the classification is run. Some of the common algorithms in remote sensing software
packages are, in increasing complexity: Parallelepiped, Minimum Distance, Maximum Likelihood, and
Mahalanobis Distance. Figure 5.10 illustrates the supervised classification approach.
A) B)
C) D)
Figure 5.10: Example of a supervised classification of two dates of images in a single process, from Liberia. Spectral
images (A) from two dates can be combined and viewed to more easily observe spectral changes (B) indicative of land-use
change. Training sites can be drawn (C) based on field and aerial data as well as an analyst’s knowledge of an area and
expertise in interpretation. These are the basis for class statistics used to classify the rest of the image. The final product
(D) is often filtered to eliminate small and spurious errors.
In unsupervised classification, no training process is applied. Instead, algorithms identify spectrally similar
pixels and then assign them to a user-specified number of groups. The output of an unsupervised
classification is then reviewed by an analyst, and each group is labeled to a class based on the analyst’s visual
interpretation of the spectral data, the location of the pixels, and available ancillary field or other data.
Additional iterations are typically run to further split groups that overlap different land-cover types. The
ISODATA algorithm is common in most software packages.
An assumed advantage of supervised over unsupervised classification is that the analyst directs the process
based on a priori knowledge of the area being classified. Conversely, an assumed advantage of the
unsupervised approach is the algorithm evaluates the distribution of the data itself. Supervised algorithms that
explore the distribution of the data while still allowing the analyst to direct the process via training are
increasingly used. Two such algorithms are Decision Trees (DTs) and Neural Networks (NNs). DTs operate
by iteratively seeking a binary split in the data in each of the bands, based on the data in the training sites
identified by the analyst. The split is one that optimizes accuracy at that stage in the development of the DT.
The final tree is often composed of hundreds of splits and terminal nodes representing the land-cover classes
contained in the training data. “Boosting” and “pruning” processes can be applied to DTs in order to
improve the efficiency and reduce the number of final splits; the resulting DT is a set of rules that is applied
to the rest of the pixels to produce a classified image. Numerous studies have used DTs to generate robust
classification results in many regions (Friedl et al., 1999; Hansen et al., 2000; Pal and Mather, 2003; Rogan et
al., 2003; Hansen et al., 2008b).
NNs attempt to mimic the human learning process to associate a class with the image data. Many variants of
NNs exist, though NNs can run more slowly than DTs. Both DTs and NNs have become favored over
maximum likelihood and other classification algorithms. Most recently, Random Forest ensemble
classification methods have been successfully applied to land-cover and land-cover change classification (Pal,
2005; Gislason et al., 2006; Rodriguez-Galiano et al., 2012). This approach, unlike DTs, randomly selects
some, but not all, of the variables to build the resulting tree and identifies resulting splits based only on this
subset of variables. Such methods do not suffer from overfitting, which can be a problem with DTs, and
generally perform efficiently.
Other approaches under exploration include learning classification techniques, such as Random Forests 27, in
which multiple decision trees are constructed and an output class is selected based on the majority of votes
from all the trees. Random Forests, for example, could be particularly useful in areas where the spectral
separation of vegetation types is limited.
All of the above approaches are examples of “per-pixel” classification, meaning the pixel is classified solely
based on its spectral characteristics. Any of these methods can be expanded to become part of a contextual
classification. In contextual classification approaches, a pixel is classified based on its own spectral
characteristics as well as those of surrounding pixels. One type of contextual classification is textural
classification. In this approach, the variance of the pixels within a certain window around the center pixel,
e.g., a five-by-five pixel window, is used to inform the classification. Another type could use the average, or
some other metric, of the pixels within the window. Weighting can also be used to apply different weights to
pixels that are closer to, or further from, the central pixel being classified.
Image segmentation, another contextual approach, is a statistical method that groups contiguous pixels into
areas (segments) that are relatively homogeneous. Segmentation generally represents an intermediate step
prior to classification, and segmentation algorithms allow an analyst to specify the relative size and shape of
the segments. The resulting segmented image can then be classified at the segment level, rather than the pixel
level, providing additional information that can be utilized by the classification algorithm, or the analyst while
developing the training data sites.
Each of the above approaches can be applied to a single image at a time, or to mosaics of images of the same
area and time period. They can also be applied to multi-temporal image data, i.e., images from the beginning
and end of a study period. This enables a direct estimation of change and persistence from the multi-temporal
imagery. Some form of “direct change estimation” process is usually recommended for change estimation.
This process also includes a single classification step that yields a two-date classification, rather than the
classification of two individual images and two single-date classification outputs, both of which may contain
errors. These errors would be compounded when the two maps are combined during post-processing.
Some recent semi-automated approaches use much more of the data archive than a single image from a start
date and another one from an end date. These approaches may be based on the seasonal signal of different
types of vegetation and estimate changes based on where anomalies in these seasonal signals are detected (see,
for example, Friedl et al., 2010 or Jiang et al., 2012). Other approaches mine all available data, such as the
entire Landsat archive, and generate many multi-temporal metrics, such as “linear trend in red reflectance” or
“maximum middle-infrared reflectance recorded since the initial date” (e.g., Hansen et al., 2008a; 2008b).
These are powerful because short-lived signals of land-use change are more likely to be captured, and all
available data are employed, which may be critical in cloudy areas.
Replicability and analyst interaction versus automation

In the case of estimating deforestation, many studies with analyst interaction have produced accurate
estimates of national forest cover. Accuracies have often been reported over 90 percent (e.g., Harper et al.,
2007; Lindquist et al., 2008; Evans et al., 2010; Longépé et al., 2011); accuracies for land-use classes such as
agriculture and grassland tend to be lower, generally 70 to 80 percent, and these estimates are generally
derived from local rather than national studies.
27 http://www.stat.berkeley.edu/~breiman/RandomForests/cc_home.htm
In recent years there has been valuable research on automated methods for processing satellite data. This has
mostly been in the pre-processing steps, although in some cases it has also included the classification step.
For example, there are well-published approaches that use automation for a series of pre-processing steps,
then the actual change estimation is conducted using a set of rules or digital classification assisted by analyst
interpretation (Souza et al., 2005; Masek et al., 2008). Conversely, the recently published Deforestation Atlas of
the Democratic Republic of the Congo was produced by an entirely automated approach (Hansen et al.,
2008a).
There are a wide range of options to apply automation in the classification step itself. For example,
classification algorithms could be rule-based, in the form of thresholds applied to reflectance data, or derived
data in some other units. In this case, the validity of the results would be very dependent on precise
correction and normalization of the images in the pre-processing step. Further, if relatively few rules are used,
the assumption that accurate results can be achieved over large areas, using few rules, must be valid. This is
often not the case and should therefore be tested. Such rules, for example, may only yield accurate results
over certain parts of the study area where the cover types are most easily distinguished with the spectral data,
or to classes that are spectrally distinct, such as dark or clear water, snow, and bright non-forest areas such as
urban areas and exposed sand and soil. While rapid estimation of cover and change based on these rules
could be achieved for much of the country, the remaining areas or classes would need to be estimated via
other approaches.
Another example of automating the classification step could be to automate the process of collecting data on
training sites. Hansen et al. (2008b) sampled an existing vegetation map to generate training points. While the
results are encouraging, further testing should be conducted in other regions, especially mountainous regions
or areas with more deciduous vegetation. A related approach is to use traditional interpretation methods to
identify training sites for classes, as is typically done in a supervised classification approach. A large set of
training sites could be built for the entire country, or for various strata within it. Once it is confirmed that this
set can be used to produce an accurate map, the same training sites could be applied to new data in later years
to calculate new class spectral signatures to be used with the new imagery. The approach could be automated
once a national training data set is defined, as the spectral variations in the new data are accounted for each
time these new data are combined with the training site locations.
Countries could potentially automate all but the final steps of a methodology to estimate change. Another
approach could be to automate the estimation of the most conspicuous changes (e.g., clear-cutting of forest)
while applying a less-automated method to estimate the less conspicuous ones. Alternatively, a country could
choose to monitor certain parts of the country that are more appropriate for automated monitoring, like areas
with modest topography and cloud cover. Difficult areas may require more direct analyst interaction to obtain
accurate results. As the scientific community itself uses a broad range of approaches, this indicates there is no
single best answer and countries should evaluate options themselves. In doing so, countries should seek an
optimal balance between the accuracy of the final estimates, the replicability of the methodology, and the
cost.
Post-processing
Post-processing refers to any step conducted after the classification step, and the post-processing steps
required will vary depending on the classification approach and attributes desired in the final map product
used to calculate areas for categories.
If the classification methodology included the creation of sub-classes merged into the final desired classes,
this should be done first. Each class in the digital output file of the classification has an assigned number, and
merging can be accomplished by recoding the values of all the sub-classes to a value that represents the final
class. If a two-date classification was not conducted, the following step should be applied to the two
classifications to create a two-date change map. The values of the classifications from both dates should be
recoded to form the basis for a final class map that records categories of change and persistence. However,
note that the previous section recommends directly estimating change from multi-temporal images. With
direct change detection, yielding a two-date classification, the output classifications will already have values
representing classes of change and persistence and, therefore, the previous recoding step would not be
required. Further, errors present in each of the single-date classifications would be compounded in the
merged classification output.
After recoding to the final class values, some filtering of the product is usually desirable. Filtering is generally
performed for two reasons. First, to eliminate small errors associated with micro-topography and other very
local effects that produce a speckled pattern of mis-classified raster cells and, second, to eliminate patches
smaller than a desired MMU or the minimum patch size in a national class definition. Note that we use the
term cell instead of pixel when referring to classification outputs rather than spectral images. Some
classification methods are more prone to producing these errors than others, but they are common artifacts
that should be removed.
Two broad types of filters are commonly used, and can be used in sequence. “Local filters,” are based on the
class values around (within a three-by-three window) a center cell. A common type is a local majority filter,
where the center cell is re-assigned to the most common cell value within the window. This not only removes
the speckled pattern but also smoothes jagged edges, which may or may not be desired. For small windows,
such as a three-by-three window, this is subtle. The second type of filter is a “sieve filter.” In this, patches of
cells with the same value are identified, and patches smaller than a user-defined size eliminated. As mentioned
above, this is useful because the final product can have a defined MMU that meets a country’s national
definition of forest.
Calculating change rates

Several factors must be carefully addressed when calculating change rates. First, the source images may not be
from the exact beginning and end of the time period reported, especially where cloud cover limits the
coverage of optical images. For example, many studies that report changes over five or ten years use images
that are within one or two years of each target date. In this case, the study areas should be divided into areas
where the image pairs representing the start and end time have different lengths of time separating them. For
example, for a 2000 – 2010 study, one part of the analysis may be based on images from 2001 and 2009, while
another is from 1999 and 2011. These areas should be defined and recorded as having an eight and twelve-
year difference in dates, and rates of change calculated for each area with a given difference, and for each
forest stratum. In this case, each forest stratum experiencing change between the eight year difference would
have an entry, and each forest stratum experiencing change between the twelve year difference would have an
entry. This allows for a temporal extrapolation for each area in each stratum, in this case to a ten year period.
Data gaps, from cloud cover or other reasons, must also be addressed. If reporting in units of percent, one
should consider if the sampled area, which may be the great majority of the study area, is representative of the
entire study area. If reporting in absolute units, extrapolation is needed. This may warrant another
stratification so that percent rates are not extrapolated into very different areas and thus not well-represented
by those where data exist. After extrapolating the percent rates, rates can be converted to absolute values by
combining with the forest area at the beginning of the reporting period. In this step, data on change can be
combined with those on the forest strata, in order to report change for each forest stratum.
Third, if the analysis was not for a single-year period, the total rate should be reported for the entire time
period, not in per-year units. In the above example of 2000 – 2010, if the aim is to report in units of percent
per year, then a correction must be applied. This is calculated based on areas at the start and end date:
Annual change rate = [ (area t2 / area t1) ˄ (1 / (date t2 – date t1)) ] – 1
where t1 and t2 are the beginning and end of the time period of the study, in years (Puyravaud, 2003).
Accuracy assessment
Uncertainty is the error in a particular estimate, in this case the estimate of an area of change over time.
Accuracy equals one minus the error value, and thus accuracy assessment also refers to assessment of error or
uncertainty. Performing an accuracy assessment of a thematic map represents a very important component of
the process. An accuracy assessment allows errors in the map to be estimated and uncertainty quantified, thus
providing additional explanation of, and validity to, the results. . The main elements of an accuracy
assessment are the error matrix, or confusion matrix, and associated statistics (Congalton, 1991). The error
matrix is generated by comparing the classification results with reference data. The statistics include overall
accuracy, and the producer’s and user’s accuracy for each class in the product. The Kappa coefficient can also
be calculated, but many articles highlight the limitations of this statistic (Olofsson et al., 2012; Pontius and
Millones, 2011; Foody, 2002).
In the example error matrix in Table 5.3, the columns contain verified land uses and the rows contain
estimated uses from the classification. The values along the diagonal are the number of correctly classified
pixels, and those off the diagonals are errors of omission and commission. Overall accuracy is the portion of
the total number of correctly mapped pixels.
The producer’s accuracy indicates how often a pixel is correctly assigned to a specific class. This statistic is
based on errors of omission, i.e., how often a pixel was incorrectly omitted from the class. The user’s
accuracy indicates how often a pixel was incorrectly assigned to a given class. This is based on errors of
commission, i.e., how often a pixel was incorrectly included in a class. In the example table, the producer’s
accuracy for degraded forest is: 100 x 1,890 / 2,040 = 92.6 percent. The user’s accuracy for the same class is:
100 x 1.890 / 2,000 = 94.5 percent.
Reference
Forest Degraded Forest Non-forest Map total

Forest 9,880 90 30 10,000
Land-use map
Degraded Forest 70 1,890 40 2,000
Non-forest 10 60 3,930 4,000
Reference total 9,960 2,040 4,000 16,000
Producer's accuracy (%) User's accuracy (%)
Forest 99.2 Forest 98.8

Degraded
Degraded Forest 92.6 Forest 94.5
Non-
Non-forest 98.3 forest 98.3
Table 5.3: Example of an error matrix and resulting overall accuracy. In this hypothetical case, the land use totals are the
same as in the beginning time in Table 5.1
Both the land-use change and error matrices are common formats for reporting land-use change and errors.
While they differ from the reporting-table format of the IPCC, shown at the end of Chapter 3, the data can
easily be transferred. It is useful to calculate these statistics for the different strata in a study area, as this
allows one to combine errors with errors in carbon stock for each stratum. It also allows one to assess where
improvements are most needed, i.e., which parts of the GHG inventory are contributing the greatest errors in
GHG estimates and should be reviewed as part of the KCA.
The examples in Table 5.3 demonstrate error estimation in land-use cover for a single date. However,
countries must estimate errors in land-use change estimates over time. Multi-date accuracy assessments use
the information available from two dates. An appropriate approach is the use of careful, cross-checked, visual
interpretation of a combination of very high resolution imagery, along with imagery used in the classification
itself. Multiple interpreters can be used, and the consistency of their interpretation can indicate confidence of
the validation data set itself. Field surveys can be valuable for the classes that are the most difficult to
interpret even with very high resolution imagery, such as degraded forest. Once the error for a particular area
of change, or AD, is estimated, it can be combined with the error of the change in stocks per unit area
estimated via field inventories. This can be done using Equation 5.2.2 from the UNFCC “Good Practice
Guidance for LULUCF” (2003) for the propogation of errors, similarly to the application described in
Chapter 4.
Error-adjusted area estimates, such as those described in Olofsson et al. (2013), use the information available
in the matrix, together with the total area of each class identified in the map, to generate area-adjusted errors
based on the proportional area of each class and errors identified in the matrix. Error matrices and accuracy
assessments can also be extended to provide confidence interval (CI) information. This is especially important
as it quantifies the confidence of a particular class, thus providing very pertinent additional information.
Olofsson et al. (2013) describe a process for creating CI bounds based on area-adjusted error matrices
In addition, it is important to account for a rare class when developing a validation strategy (Stehman et al.,
2010). This type of proportional sampling design ensures that adequate sampling occurs in sparser yet critical
classes, such as deforestation. The sampling design could focus on areas of deforestation identified in the
map, and a stratification could be used to categorize areas of high change probability and low change
probability. Proportional sampling could then be focused in these strata to ensure each class is adequately
represented in the validation analysis.
Section 3.7 of the GFOI MGD (GFOI, 2013), provides guidance on considerations for generating reference
data and performing an accuracy assessment. These include: i) ensuring the reference data are of a higher
quality than the map data (for example, Manual interpretation of an image by an analyst is generally
considered higher quality than an automatic classification algorithm); and ii) Combining reference data
sources (i.e., field and aerial surveys which can be particularly cost effective if resources are limited). The
GFOI MGD also provides two example approaches for performing an accuracy assessment and area
estimation; one stratified and one model-based approach.
Olofsson et al. (2014) also provide a full review of good practice recommendations for producing transparent
and “scientifically rigorous” accuracy estimates and estimates of area based on change between time one and
time two. They detail three separate steps that should be undertaken to complete an accuracy assessment
including: i) the sampling design, which answers the question, ‘what is a suitable subset area to sample?’,
understanding that evaluation of the entire map is not possible; ii) the response design, which answers the
question ‘are the maps and reference data in agreement?’; and iii) analysis, which answers the question of
‘how to calculate accuracy and quantify uncertainty?’.
Accuracy assessment is very important, and MRV programs should work with expert statisticians in
developing strategies for validation sampling and combining information on uncertainties in AD with those in
EFs. This will lead to stronger estimation of the overall uncertainties in estimated GHGs.
5.5 EMERGING AREAS OF RESEARCH
Several areas of particularly active research in support of REDD+ activities include: the mapping and
monitoring of degradation; the use of other sources of remotely-sensed data, such as RADAR, in monitoring;
and the increased use of field sampling to facilitate remote sensing product validation.
Mapping and Monitoring of Degradation
Forest degradation is a substantial contributor to GHG emissions from land-use change (Nepstad et al, 1999;
Souza and Roberts, 2005; Stickler et al., 2009; GOFC-GOLD, 2012; Hirata et al., 2012), with estimates ranging
from 20 to 50 percent of total land-use GHG emissions over large regions (see, for example, Houghton and
Hacker, 1999; Lambin et al., 2003; Asner et al., 2005). Mapping and monitoring of forest degradation remains
challenging. Multiple definitions of forest degradation exist, adding to the complexity of mapping and
monitoring forest degradation. The IPCC’s definition of forest degradation is provided in section 5.2.1.
Conversely, GOFC-GOLD (2012) presents a range of human activities that result in forest degradation
including selective logging, forest fires (canopy and sub-canopy) and fuelwood collection. GOFC-GOLD
(2012) lists a range of human activities that result in forest degradation including selective logging, forest fires
(canopy and sub-canopy) and fuelwood collection. These different activities may require different monitoring
approaches, and a country should seek to understand the implications and applicability of different
approaches. Readers should also refer to the relevant sections of both the GOFC-GOLD SourceBook and
the GFOI MGD.
The use of other sources of remotely-sensed data in mapping and monitoring forest
A second area of research in support of REDD+ activities is the use of other sources of remotely-sensed
data, such as RADAR, to map and monitor forest extent and characteristics, deforestation, and degradation.
Several characteristics make RADAR an attractive source of information for such applications. First, because
RADAR sensors operate in longer wavelengths (generally 1cm to 1m) of the EM spectrum than, for example,
optical sensors, they are able to penetrate clouds and are thus useful for monitoring in areas with persistent
cloud cover. In addition, because the signals received by the sensor are less affected by atmospheric
conditions, and the properties of the emitted radiation from active sensors are controlled and well known,
RADAR images are directly comparable over time. RADAR signals are also sensitive to the geometric
properties of a forest, providing information on the distribution of aboveground biomass. Figure 5.11
illustrates a detail of a Landsat image compared to a PALSAR satellite image for an area in San Martín, Peru.
A) B)
Figure 5.11: Examples of Landsat and RADAR images, from Peru. Ground observations in the Landsat data in A) are partly
obscured by clouds, while the PALSAR image in B) is cloud free. While the brightness variations in both are affected by
terrain, this is more noticeable in the latter
The following provides a brief introduction to a selection of active remote sensing concepts. Additional Such
concepts are key to understanding the basic characteristics of active remote sensing data. RADAR measures
the distance between an object on the ground and the sensor based on the strength of radio waves that are
transmitted as pulses of microwave beams, directed by an antenna, that illuminate a strip of the earth’s surface
(swath). The intensity of the signal that is scattered back to the receiver from this transmitted energy is
recorded as the returned signal, and the distance is calculated based on the time elapsed for the RADAR
signal to make a complete round trip. The next transmitted pulse illuminates the next strip of terrain along the
swath, and a two-dimensional image is created (each pulse defines one line).
As with optical sensors, RADAR sensors exploit different wavelength bands. A shorter wavelength band,
such as an X-band (λ= 3 cm) may only penetrate the upper layer of a forest canopy, whereas a P-band (λ= 23
cm) may penetrate leaves and small branches thus providing information about both the big branches and
stems of the trees. Thus, resulting P-band images are important for measuring vegetation biomass and
aboveground carbon stocks.
The surface roughness, geometric shape and dielectric properties of an object also affect the information
received by the RADAR sensor. Surface roughness is a relative term that is dependent on the RADAR
wavelength. For example, small objects such as leaves and twigs are considered rough for small wavelength
RADAR, but smooth for longer wavelength RADAR such as P-band RADAR. Water bodies tend to be
relatively smooth, with most of the energy being reflected away from the RADAR, while trees and other
vegetation are rough, causing backscatter, and thus have a bright appearance in a RADAR image.
The difference in intensity of RADAR returns from two surfaces of equal roughness is an indication of the
difference in their dielectric properties, and these are strongly affected by their moisture. For example, the
brightness of areas covered by bare soil may vary depending on the roughness and moisture content of the
soil. For soil types of similar roughness, the surface with a higher moisture content will appear brighter.
A key characteristic of RADAR data not found in most optical data is polarization; this is described below.
Apart from polarization, several additional characteristics distinguish RADAR instruments and data from
their optical counterparts and are useful to understand. These include phase, side-looking observation,
Synthetic Aperture RADAR (SAR), interferometry, and polarimetry; these are outlined in Box 5.5
Polarization refers to the orientation of the electric field with respect to the direction of propagation. In
active remote sensing, the electric field of the resulting radiation has a preferred orientation. Linear is the
most common polarization used in RADAR remote sensing where a radiated electric field is oriented either
horizontally (horizontal polarization) or vertically (vertical polarization) with respect to the direction of
propagation, as shown in Figure 5.12. A sensor that can transmit either horizontally (H) or vertically (V)
polarized waves and receive both will result in the following four polarized images:
• HH: horizontal transmission and horizontal reception;
• VV: vertical transmission and vertical reception;
• HV: horizontal transmission and vertical reception; and
• VH: vertical transmission and horizontal reception.
Some space-borne satellites including RADARSAT-1 and ERS-1/2 have only single polarization
(RADARSAT-1 with HH and ERS-1/2 with VV), while other satellites, including RADARSAT-2, ENVISAT
and ALOS/PALSAR acquire data with all four polarizations (“quad-pol”) or two polarizations (“dual-pol”).
The below discussion includes a selection of SAR applications that are relevant to REDD+ activities.
5.12: Horizontal and vertical polarizations
Box 5.5. Additional characteristics of RADAR

Phase describes the relationship of the lead, or lag, of an electromagnetic wave with respect to a
reference wave of the same wavelength, and is expressed in degrees. 360 degrees represents one
complete cycle and, therefore, a wave that is lagging one quarter of a wavelength behind the reference
has a phase of 90 degrees.
Most RADAR sensors are also side-looking instruments, unlike many optical sensors which acquire
imagery at nadir (i.e., observing a location directly below the sensor). This across-track capability,
termed Side Looking RADAR (SLR), introduces a range of geometric distortions including
foreshortening, layover, and shadows that require full, or partial, correction.
Some RADAR remote sensing systems, such as Synthetic Aperture RADAR (SAR) systems, are
able to achieve a relatively high resolution without the use of a large antenna. Data from such systems
are generally processed by data distributors to Single Look Complex level data, containing both
amplitude and phase information, and a range of derived products which are usually geocoded, ortho-
rectified, and radiometrically corrected. And, while some geometric and radiometric distortions due to
terrain relief may persist, these processed and derived products are generally more appropriate for use
in mapping. Among the distortions that may persist is speckle. SAR speckle causes pixel-to-pixel
variation in intensities even over a homogenous area; this grainy ‘salt and pepper’ texture degrades the
quality of the image and complicates interpretation. Speckle can be reduced by averaging the
backscatter response within a pixel, though this can effectively reduce resolution, or by applying
smoothing filters.
Another relevant RADAR concept is SAR interferometry. As mentioned in the introduction, a
RADAR image contains information about the intensity of the signal and the phase. If two SAR images
have been acquired over the same area from very close antenna positions, the different path lengths
from these positions to the object on the earth’s surface cause the differences in phase. The path
difference is geometrically related to the distance between two antennas and the terrain height. Since
the antenna positions are precisely known, the observed phase differences can be used to infer three-
dimensional information about the terrain height. The technique is known as SAR Interferometry.
Finally, SAR polarimetry is a relevant RADAR concept. As previously discussed, more parameters can
be measured from polarimetric RADAR compared to single-channel RADAR. The different polarization
bands may contain unique and additional information about the surface object. For example, a signal
that reflects off a tree trunk to the ground surface is likely to show distinctive polarization shifts from
signals that return directly off the soil. Surface objects that scatter are vertically oriented and show high
backscatter in vertically polarized imagery and low backscatter in horizontally polarized imagery. Such
unique information is important for discriminating different land-cover types.
Applications of SAR
Applications on the use of SAR data for forest mapping, measuring and monitoring aboveground biomass
(AGB) and scaling-up ground-based AGB measurements are increasing. Multiple studies have tested the
potential of combined RADAR channels of different frequencies and polarization for deforestation
monitoring. For example, Saatchi et al. (1997) used C-SIR data to map land-cover types and monitor
deforestation in the tropics, with an emphasis on characterizing several clearing practices and forest
regeneration characteristics. They also mapped forest patches and fragmentation and found these data helpful
in delineating areas with different degrees of forest disturbance. Rignot et al. (1997) compared SIR-C data
with Landsat TM in a test site in Rondonia, Brazil; while the Landsat TM yielded a more accurate
deforestation extent classification, the combined use of both Landsat and RADAR imagery further improved
the mapping accuracy.
The recent systematic availability of fully polarimetric SAR data from the ALOS-PALSAR, ENVISAT, and
RADARSAT-2 has led to further land-cover classification research using SAR imagery. Walker et al. (2010)
assessed the ability of PALSAR and LANDSAT data to classify and map forest cover in the Xingu River
headwaters in southeastern Amazonia, producing overall accuracies of 92 and 94 percent with PALSAR and
Landsat, respectively, for the forest versus non-forest classifications. They also found a high degree of spatial
similarity among maps derived from PALSAR, Landsat, and existing data from Projeto De Estimativa De
Desflorestamento da Amazonia (PRODES), the Brazilian Amazon deforestation monitoring program.
In addition to polarimetric information, polarimetric interferometric SAR (PolInSAR) provides
interferometric information (see Box 5.5) related to the structure and complexity of the observed objects.
Substantial improvements in land-cover change classification can be achieved by combining polarimetric and
polarimetric interferometric information (Shimoni et al., 2009). In addition, the fusion of spatial and textural
information derived from various SAR polarizations has been shown to improve classification results
(Borghys et al., 2006).
SAR data are also being evaluated for scaling up ground-based AGB and monitoring changes over large scales
(Lu, 2006; Mitchard et al., 2009). These data are sensitive to the geometric properties of the forest and directly
related to measurements of AGB. However, this sensitivity appears to saturate at biomass levels of around
100 tons ha-1 (Imhoff et al., 2000) and approximately 81 percent of global forests are above this saturation
limit (Nelson et al., 2007).
Recently, data from PALSAR, the first long wavelength (L-band, 25 cm) SAR satellite with the capability to
collect cross-polarized responses, has yielded improved estimates of AGB with little or no saturation, up to
250-300 tons per hectare based on the sensor’s cross-polarized ability to exploit the strong response of three
dimensional objects, such as trees, compared to bare soil. Mitchard et al. (2011) used L-band synthetic
aperture RADAR backscatter data from 1996 and JERS-1 and PALSAR data from 2007 to produce biomass
maps of a forest–savanna ecotone region in central Cameroon, characterized by small scale deforestation and
degradation. They found that the RADAR data detected changes in a broad AGB class in forest–savanna
transition areas with an accuracy of 95 percent. Similarly, Ryan et al. (2012) generated biomass maps and
changes in carbon stocks with known uncertainties using PALSAR imagery in a region in central
Mozambique yielding maps with sufficient accuracy to enable changes in forest carbon stocks of as little as 12
tons per hectare over 3 years, with 95 percent confidence to be detected. Mitchard et al. (2012) used a
combination of PALSAR, space-borne LiDAR (ICESAT GLAS), and ground-based data to map AGB in
Gabon’s Lopé National Park.
While these results highlight the potential for space-borne imaging RADAR for estimating forest area and
biomass, several limitations exist and the history of SAR data usage for land-cover classification remains
relatively recent. One major limitation of SAR data utilization and analysis is the difficulty involved in
interpreting RADAR backscatter as compared to optical spectral data (Saatchi et al., 2000). The presence of
topographic effect and speckle complicates both visual and digital analysis of RADAR images, and complex
areas with a greater abundance of secondary forests may yield significantly lower accuracies. Additional
evaluations are also needed to assess the utility of newer RADAR data sources in mountainous areas.
Uncertainty related to the long-term data continuity of space-borne RADAR systems could also prove a
limiting factor for forest monitoring. PALSAR and ENVISAT, for example, which both provided fully
polarimetric L-band data, are no longer collecting. However, the Japanese Space Agency (JAXA) launched
ALOS-PALSAR-2 in May 2014, and this instrument again contains an L-band.
Finally, though the range of advanced SAR processing techniques capitalizing on the availability of multi-
polarimetric information is evolving, and classification methods based on polarimetric decomposition are
being developed, SAR data generally yield less robust results than Landsat data for forest/non-forest
classification in most studies.
Use of Field Sampling To Facilitate Remote Sensing Product Validation
Remote sensing data, in particular high resolution satellite imagery, RADAR, and various kinds of LiDAR
hold promise to more efficiently map biomass stocks. This is of particular interest in the often vast and
difficult-to-access landscapes contemplated under most national and jurisdictional REDD+ MRV. These data
likely will not provide more accurate biomass data for plot sites than those derived from field data, and any
RADAR or LiDAR-based estimates of biomass will need to be calibrated using field-plot data. However,
where significant relationships can be calibrated, these data can be valuable in extrapolating field-based
estimates over larger areas, especially remote areas that are costly to access. While these approaches hold
promise, they are still in the process of refinement.
The same field measurement and QA/QC procedures described in Chapter 4 are also relevant in the
collection of ground truth data. Sampling strategies differ however, as the objective is not to produce a field-
based estimate but instead to model predictions, over a wider spatial extent, from remotely-sensed data.
Therefore, ground-reference data used to build prediction models for remotely-sensed data do not need to be
strictly uniformly sampled. Although, ideally the sample should be well-distributed across the area to which
the model will be applied. The sample should achieve a roughly even distribution across the range of
conditions of the remotely-sensed data, such as the range of reflectance values. This can be achieved via
weighted random sampling, other unequal probability sampling approaches or a systematic sample approach.
Note that these preferred sample designs require that the remote sensing data already be in hand, providing
an equivalent range of spectral values within the area sampled via remote sensing, to optimize the design.
Large, fixed-area plots are best suited for generating ground-reference data to compare with remotely-sensed
data. Larger plots facilitate the alignment of ground samples and remote sensing imagery, reducing the impact
of GPS inaccuracy (Mascaro et al., 2011). Circular plot designs and a minimum plot size of 0.2 ha is
recommended for use with airborne LiDAR to reduce model errors (Zolkos et al., 2013). Field measurements
should also be collected as close as possible to the acquisition date of the remote sensing image.
Field measurement plots can calibrate and validate prediction models. In the latter case, a sample of plots are
measured and held in reserve, and after model construction and application, used to assess the accuracy of the
predicted values at different scales. This approach to assess model uncertainty is recommended because it is
straightforward and captures the overall result of many potential sources of error that would otherwise have
to be independently quantified and then propagated.
5.6 REFERENCES
Asner, G.P., D.E. Knapp, E.N. Broadbent, P.J.C. Oliveira, M. Keller and J.N. Silva. 2005. Selective logging in
the Brazilian Amazon. Science 310: 480–482.
Baccini, A., S. J. Goetz, W. S. Walker, N. T. Laporte, M. Sun, D. Sulla-Menashe, J. Hackler, P. S. A. Beck, R.
Dubayah, M. A. Friedl, S. Samanta and R. A. Houghton. 2012. Estimated carbon dioxide emissions from
tropical deforestation improved by carbon-density maps. Nature Climate Change 2: 182-185.
Borghys, D., Y. Yvinec, C. Perneela, A. Pizurica and W. Philips. 2006. Supervised feature-based classification
of multi-channel SAR images, Pattern Recognition Letters 27: 252–258.
Clark, M.L., D.B. Clark, and D.A. Roberts. 2004. Small-footprint lidar estimation of sub-canopy elevation and
tree height in a tropical rain forest landscape. Remote Sensing of Environment 91(1): 68-89.
Congalton, R.G. 1991. A Review of Assessing the Accuracy of Classifications of Remotely Sensed Data.
Remote Sensing of Environment 37: 35-46.
Evans, T.L., M. Costa, K. Telmer and T.S.F. Silva. 2010. Using ALOS/PALSAR and RADARSAT-2 to Map
Land-cover and Seasonal Inundation in the Brazilian Pantanal. IEEE Journal of Selected Topics in applied Earth
Observation and Remote Sensing 3: 560-570.
FAO. 2006. Choosing a Forest Definition for the Clean Development Mechanism. Forests and Climate Change
Working Paper 4. http://www.fao.org/forestry/media/11280/1/0/
Foody, G.M. 2002. Status of land-cover classification accuracy assessment, Remote Sensing Environment 80: 185-
201.
Friedl, M.A., C.E. Brodley, and A.H. Strahler. 1999. Maximizing Land Cover Classification Accuracies
Produced by Decision Trees at Continental to Global Scales, Transaction on Geosciences and Earth Observation
37(2): 969-977.
Friedl, M.A., D. Sulla-Menashe, B. Tan, A. Schneider, N. Ramankutty, A. Sibley, and X. Huang. 2010.
MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets, Remote
Sensing of Environment 114: 168-182.
Gislason, P.O., J.A. Benediktsson and J.R. Sveinsson. 2006. Random Forests for land cover classification,
Pattern Recognition Letters 27: 294-300.
GFOI. 2013. Integrating remote-sensing and ground-based observations for estimation of emissions and
removals of greenhouse gases in forests: Methods and Guidance from the Global Forest Observations
Initiative. Group on Earth Observations. Geneva, Switzerland.
GOFC-GOLD. 2013. A sourcebook of methods and procedures for monitoring and reporting anthropogenic
greenhouse gas emissions and removals associated with deforestation, gains and losses of carbon stocks in
forests remaining forests, and forestation. GOFC-GOLD Report version COP19-2 (GOFC-GOLD Land Cover
Project Office, Wageningen University, The Netherlands). Available at
http://www.gofcgold.wur.nl/redd/index.php.
Hansen M.C., R.S DeFries, J.R.G. Townshend and R. Sohlberg. 2000. Global land cover classification at 1 km
spatial resolution using a classification tree approach, International Journal of Remote Sensing 21(6): 1331-1364.
Hansen M.C., D.P. Roy, E.J. Lindquist, B. Adusei, C.O. Justice and A. Altstatt. 2008a. A method for
integrating MODIS and Landsat data for systematic monitoring of forest cover and change in the Congo
Basin. Remote Sensing Environment 112: 2495–251.
Hansen, M.C., S.V. Stehman, P.V. Potapov, T.R. Loveland, J.R.G. Townshend, R.S. DeFries, K.W. Pittman,
F. Stolle, M.K. Steininger, M. Carroll, and C. Dimiceli. 2008b. Humid tropical forest clearing from 2000 to
2005 quantified using multi-temporal and multi-resolution remotely sensed data, PNAS 105: 9439-9444.
Harper, G.J., M.K. Steininger, C.J. Tucker, D.J. and F. Hawkins. 2007. Fifty years of deforestation and forest
fragmentation in Madagascar. Environmental Conservation 34: 325-333.
Hirata Y., G. Takao, T. Sato, J. Toriyama, eds. 2012. REDD-plus Cookbook. REDD Research and
Development Center, Forestry and Forest Products Research Institute Japan, 156pp.
ISBN 978-4-905304-15-9.
Imhoff, M.L., P. Johnson, W. Holford, J. Hyer, L. May, W. Lawrence, P. Harcombe. 2000. BioSAR™: an
inexpensive airborne VHF multiband SAR system for vegetation biomass measurement. Transaction on
Geosciences and Earth Observation 38: 1458–1463.
IPCC. 2003. Definitions and methodological options to inventory emissions from direct human-induced
degradation of forests and devegetation of other vegetation types. In: Penman, J., M. Gytarsky, T. Krug, D.
Kruger, R. Pipatti, L. Buendia, K. Miwa, T. Ngara, K. Tanabe and F. Wagner. IPCC-IGES, Kanagawa (eds.).
IPCC GPG-LULUCF. 2003. Good Practice Guidance for Land Use, Land-Use Change and Forestry.
Published: IGES. Japan.
IPCC. 2006. 2006 IPCC Guidelines for National Greenhouse Gas Inventories, Prepared by the National
Greenhouse Gas Inventories Programme. Eggleston, H.S., L. Buendia K. Miwa, T. Ngara and K. Tanabe,
eds. Published: IGES Japan.
Jiang, D., Y. Huang, D. Zhuang, Y. Zhu, X. Xu and H. Ren. 2012. A Simple Semi-Automatic Approach for
Land Cover Classification from Multispectral Remote Sensing Imagery. PLoS ONE 7(9): e45889.
doi:10.1371/journal.pone.0045889
Lambin, E.F., H.J. Geist, and E. Lepers. 2003. Dynamics of land-use and land-cover change in tropical
regions. Annual Review Environment Resource 28: 205–41.
Lefsky, M.A., WB. Cohen, G.G. Parker and D.J. Harding. 2002. Lidar Remote Sensing for Ecosystem
Studies. BioScience 52(1) 19-30.
Lindquist E.J., M.C. Hansen, D.P. Roy and C.O. Justice. 2008. The suitability of decadal image data sets for
mapping tropical forest cover change in the Democratic Republic of Congo: implications for the global land
survey. International Journal of Remote Sensing 29: 7269-7275.
Lim, K., P. Treitz, M. Wulder, B. St-Onge and M. Flood. 2003. LiDAR remote sensing of forest structure.
Progress In Physical Geography 27(1): 88-106.
Longépé, N., P. Rakwatin, O. Isoguchi, M. Shimada, Y. Uryu, and K. Yulianto. 2011. Assessment of ALOS
PALSAR 50 m orthorectified FBD Data for Regional Land-cover Classification by Support Vector Machines.
IEEE Transaction on Geosciences and Earth Observation 49: 2135-2050.
Lu, D. S. 2006. The potential and challenge of remote sensing-based biomass estimation. International Journal of
Remote Sensing 27: 1297-1328.
Mascaro, J., M. Detto, G.P. Asner. And H.C. Muller-Landau. 2011. Evaluating uncertainty in mapping forest
carbon with airborne LiDAR. Remote Sensing Environment 115: 3770–3774.
Masek, J. G., Huang, C., Wolfe, R., Cohen, W., Hall, F., Kutler, J., and Nelson, P. (2008). North American
Forest Disturbance Mapped from a Decadal Landsat Record. Remote Sensing Environment 112: 2914–2926.
Mather, P. 1999. Computer Processing of Remotely-Sensed Images. Wiley & Sons. Chicester, England.
Mitchard, E. T. A., S. S. Saatchi, I.H. Woodhouse, G. Nangendo, N.S. Ribeiro, M. Williams, C.M. Ryan, S.L.
Lewis, T.R. Feldpausch and P. Meir. 2009. Using satellite RADAR backscatter to predict aboveground woody
biomass: A consistent relationship across four different African landscapes. Geophysical Research Letters 36.
Mitchard, E. T. A., S.S. Saatchi, I. Woodhouse, T. Feldpausch, S. Lewis, B. Sonké, C. Rowland, and P. Meir.
2011. Measuring biomass changes due to woody encroachment and deforestation/degradation in a forest-
savanna boundary region of central Africa using multi-temporal L-band RADAR backscatter. Remote Sensing
Environment 115: 2861–2873.
Mitchard, E. T. A., S.S. Saatchi, L.J.T. White, K.A. Abernethy, K.J. Jeffery, S.T. Lewis, M. Collins, M.A
Lefsky, M.E. Leal, I.H. Woodhouse P. and Meir. 2012. Mapping tropical forest biomass with radar and
spaceborne LiDAR in Lopé National Park, Gabon: overcoming problems of high biomass and persistent
cloud. Biogeoscience 9: 179-191.
Nelson R.F., P. Hyde, P. Johnson, B. Emessiene, M.L. Imhoff, R. Campbell, and W. Edwards 2007.
Investigating RADAR-LiDAR synergy in a North Carolina pine forest. Remote Sensing Environment 110: 98–108
Nelson, R., J. Boudreau, T.G. Gregoire, H. Margolis, E. Naesset, T. Gobakken and G. Stahl. 2009.
Estimating Quebec provincial forest resources using ICESat/GLAS, Canadian Journal of Forestry Research 39:
862-881.
Nepstad, D. C., A. Verissimo, A. Alencar, C. Nobre, E. Lima, P. Lefebvre, P. Schlesinger, C. Potter, P.
Moutinho, E. Mendoza, M. Cochrane and V. Brooks. 1999. Large-scale impoverishment of Amazonian
forests by logging and fire. Nature 398: 505-508.
Olofsson, P., G.M. Foody, S.V. Stehman and C.E. Woodcock 2013. Making better use of accuracy data in
land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimate. Remote
Sensing Environment 129: 122-131.
Olofsson, P., G.M. Foody, M. Herold, S.V. Stehman, C.E. Woodcock, M. Wulder 2014. Good practices for
estimating area and assessing accuracy of land change. Remote Sensing Environment 148: 45-57.
Pal, M. and P.M. Mather. 2003. An assessment of the effectiveness of decision tree methods for land cover
classification. Remote Sensing Environment 86: 554-565.
Pal, M. 2005. Random forest classifier for remote sensing classification. International Journal of Remote Sensing
26(1): 217-222.
Puyravaud, J.P. 2003. Standardizing the calculation of the annual rate of deforestation. Forest Ecology and
Management 177: 593-596.
Rignot, E., W.A. Salas and D.L. Skole. 1997. Mapping deforestation and secondary growth in Rondonia,
Brazil, using imaging RADAR and thematic mapper data. Remote Sensing Environment 59: 167-179.
Rodriguez-Galiano, V.F., B. Ghimire, E. Pardo-Iguzquiza, M. Chica-Olmo, and R.G. Congalton. 2012.
Incorporating the Downscaled Landsat TM Thermal Band in Land-cover Classification using Random
Forest. Photogrammetric Engineering & Remote Sensing 78(2): 129-137.
Rogan, J., J. Miller, D. Stow, J. Franklin, L. Levian and C. Fischer 2003. Land-Cover Change Monitoring with
Classification Trees Using Landsat TM and Ancillary Data, Photogrammetric Engineering & Remote Sensing 69(7):
793-804.
Ryan, S.M, T. Hill, E. Woollen, C. Ghee, E. Mitchard, G. Cassells, J. Grace, I.H. Woodhouse and M.
Williams. 2012. Quantifying small-scale deforestation and forest degradation in African woodlands using
RADAR imagery. Global Change Biology 18(1): 243-257.
Saatchi, S., J.V. Soares and D.S. Alves. 1997. Mapping deforestation and land use in Amazon rainforest by
using SIR-C imagery. Remote Sensing Environment 59: 191–202.
Saatchi, S. and M. Moghaddam. 2000. Estimation of crown and stem water content and biomass of boreal
forest using polarimetric SAR imagery. IEEE Transaction on Geosciences and Earth Observation 38(2): 697-709.
Saatchi, S., D. Agosti, K. Alger, J. Delabie, and J. Musinsky. 2001. Examining Fragmentation and Loss of
Primary Forest in the Southern Bahian Atlantic Forest with RADAR Imagery. Conservation Biology 15(4): 867-
875.
Saatchi, S., N.L. Harris, S. Brown, M. Lefsky, E.T.A. Mitchard, W. Salas, B.R. Zutta, W. Buermann, S.L.
Lewis, S. Hagen, S. Petrova, L. White, M. Silman and A. Morel. 2011. Benchmark map of forest carbon
stocks in tropical regions across three continents. PNAS 108(24): 9899-9904.
Shimoni, M., D. Borghys, R. Heremans, C. Perneel and M. Acheroy. 2009. Fusion of Pol SAR and PolInSAR
data for land-cover classification. International Journal of Applied Earth observation and Geoinformation 11: 169-180.
Song, C., C.E. Woodcock, K.C. Seto, M.P. Lenney and S.A. Macomber. 2001. Classification and Change
Detection Using Landsat TM Data: When and How to Correct Atmospheric Effects? Remote Sensing
Environment 75: 230-244.
Souza, C. M., Jr. and D. Roberts. 2005. Mapping forest degradation in the Amazon region with Ikonos
images. International Journal of Remote Sensing 26(3): 425-429.
Souza, C., D.A. Roberts and M.A. Cochrane. 2005. Combining spectral and spatial information to map
canopy damages from selective logging and forest fires. Remote Sensing Environment 98: 329-343.
Stehman, S., P. Olofsson, C. Woodcock, M. Friedl, A. Sibley, J. Newell, D. Sulla-Menashe and M. Herold
2010. Designing a reference validation database for accuracy assessment of land cover. Accuracy 2010
Symposium, July 20-23, Leicester, UK.
Stickler, C.M., D.C. Nepstad, M.T. Coe, D.G. Mcgrath, H.O . Rodrigues, W.S. Walker, B.S. Soares-Filho and
E. Davidson. 2009. The potential ecological costs and cobenefits of REDD: a critical review and case study
from the Amazon region. Global Change Biology 15: 2803–2824.
Walker, W.S., M.S. Claudia, J.M. Kellndorfer, K.M. Kirsch and D.C. Nepstad. 2010. Large-Area Classification
and Mapping of Forest and Land-cover in the Brazilian Amazon: A Comparative Analysis of
ALOS/PALSAR and Landsat Data Sources. IEEE Journal of Selected Topics in Applied Earth Observations and
Remote Sensing 3: 594-604.
UN-REDD Programme, National Forest Monitoring Systems: Monitoring and Measurement, Reporting and
Verification (M & MRV) in the context of REDD+ Activities, 26-27 October 2012, Brazzaville, Republic of
Congo.
Zolkos, S. G., S.J. Goetz and R. Dubayah. 2013. A meta-analysis of terrestrial aboveground biomass
estimation using lidar remote sensing. Remote Sensing Environment 128: 289-298.
5.7 COMMON SATELLITE DATA SOURCES FOR LAND-USE MONITORING
Sensor Satellite – Swath Resolution Repeat Systematic Operational Monitoring

Agency Width Cycle Acquisitions Status Applications
Landsat 7 ETM+ Landsat 7 - 165km 15m panchromatic 16-21 days Yes Yes, with SLC-off deforestation,
NASA 30m multispectral gaps encroachment,
60m thermal roads, log ponds
Landsat 8 OLI LDCM - 185km 15m panchromatic 16-21 days Yes yes deforestation,
NASA 30m multispectral encroachment,
100m thermal roads, log ponds
ASTER Terra - NASA 60km 15m multispectral Varies No Partial (no SWIR deforestation,
channels) encroachment,
roads, log ponds
SPOT-5 CNES 60km 20m multispectral Varies No Yes deforestation,
5m panchromatic (2.5m encroachment,
interpolated) roads, log ponds
SPOT-6 Airbus Defense & 60km 6m multispectral Varies No Yes deforestation,
Space 1.5m panchromatic encroachment,
roads, log ponds
CCD CBERS-2B – INPE 113km 20m multispectral 26 days Yes Yes deforestation,
encroachment,
HRC CBERS-2B – INPE 27km 2.7m panchromatic 26 days Yes Yes skid trails, illegal fishing
vessels
MODIS Terra / Aqua – 2330km 250m visible 4 times per Yes Yes fires, large-scale
NASA 500m multispectral day deforestation
1km thermal (diurnal/
nocturnal)
AWIFS Resource Sat-1 730km 56m 5 days Yes Yes large-scale deforestation
IKONOS IKONOS – 11.3km @ 4m multispectral (4) Varies No Yes skid trails, canopy gaps,
GeoEye nadir 1m panchromatic (1) illegal fishing vessels /
logging vehicles
Pleiades 1a and Airbus Defense & 20km 50cm panchromatic 2m Varies No Yes Skid trails, canopy gaps,
1b Space multispectral logging roads
GeoEye-1 GeoEye-1 - Varies 0.4m panchromatic (1) Varies No Yes skid trails, canopy gaps,
GeoEye (resampled to 0.5m) illegal fishing vessels /
1.65m multispectral (4) logging vehicles
QuickBird QuickBird – Varies 0.6m (1) panchromatic Varies No Yes skid trails, canopy gaps,
DigitalGlobe 2.4m multispectral (4) illegal fishing vessels /
logging vehicles
RapidEye BlackBridge 77km 6.5m multispectral 1-5 days No Yes Deforestation,
Constellation encroachment, roads,
skid trails
WorldView-2 WorldView-2 Varies 0.5m panchromatic Varies No Yes skid trails, canopy gaps,
DigitalGlobe 1.8m multispectral illegal fishing vessels /
logging vehicles
Radarsat-2 CSA varies 8 m quad pol fine 24 days varies Yes deforestation,
25 m quad pol standard roads, log ponds
100 m wide
ASAR ENVISAT-ESA varies 30 m polarization mode 36 days varies No deforestation
150 m Wide Swath mode
1 km Global Monitoring
mode
PALSAR PALSAR-JAXA varies 9 m Single pol 45 days Yes- all mode No deforestation,
19 m dual pol No- quad pol roads, log ponds
30 m quad pol
100 m Scan SAR
WorldView-3 WorldView – 13.1km 0.31m panchromatic (1) < 1 day No Yes skid trails, canopy gaps,
DigitalGlobe 1.24m multispectral (8) illegal fishing vessels /
3.7m SWIR (8) logging vehicles
PALSAR-2 PALSAR-JAXA varies 1-3 m Spotlight 15 days Varies Yes deforestation,
3-10 m Stripmap encroachment,
100 m Scan SAR roads, log ponds
Future Missions
Sensor Satellite – Swath Resolution Repeat Cycle Systematic Operational Monitoring

Agency Width Acquisitions Status Applications
SPOT-7 Airbus 60km 1.5m No Launched June 2014, deforestation,

Defense & panchromatic (1) in calibration encroachment,
Space 6.0m roads, log ponds
multispectral (4)
Sentinel-2 ESA 290km 10/20m/60m 5 day (once both 2A Anticipated launch deforestation
A/B & 2A are launched) summer 2014 (2A),
then (2B)
VIIRS NPV – NASA 3000km 750m 2 times per day Yes recently launched, in fires
calibration
5.8 SELECTED RESOURCES
Online guides and other materials
United Nations Space Science and Technology:
http://www.unoosa.org/oosa/en/SAP/centres/index.html
Systems for World Surveillance, Inc.
http://www.rsat.com/tutorials.html
Biodiversity Informatics Facility
http://biodiversityinformatics.amnh.org/index.php?section_id=17
European Space Agency Earthnet
http://earth.eo.esa.int/download/eoedu/Earthnet-website-material/to-access-from-Earthnet/
NASA Earth Observatory
http://earthobservatory.nasa.gov/
Software
EXELIS: ENVI
http://www.exelisvis.com/ProductsServices/ENVI.aspx
INTERGRAPH: ERDAS Imagine
http://geospatial.intergraph.com/products/ERDASIMAGINE/ERDASIMAGINE/Details.aspx
PCI Geomatics
http://www.pcigeomatics.com/
American Museum of Natural History (AMNH) Biodiversity Informatics Facility, Open source GIS and
remote sensing software
http://biodiversityinformatics.amnh.org/index.php?section_id=33&content_id=138
GRASS GIS
http://grass.fbk.eu
IDRISI GIS and Image Processing Software
http://www.clarklabs.org/products/idrisi.cfm
Quantum GIS
http://www.qgis.org/en/site/
Random forests Software
http://www.stat.berkeley.edu/~breiman/RandomForests/cc_software.htm
Rulequest data mining tools; See5 classification software
http://www.rulequest.com/see5-info.html
R statistical language
http://www.r-project.org/
Open source
Alaska Satellite Facility - Map Ready, SAR Tool Kit
http://www.asf.alaska.edu/downloads/software_tools
ESA – polsarpro (Polarmetric SAR Data Processing and Educational Tool)
http://earth.eo.esa.int/polsarpro/
NEST – Next ESA SAR toolbox
http://nest.array.ca/web/nest/release-4B-1.1
RAT – Radar Tools
http://radartools.berlios.de/
Data access
USGS Earth Resources Observation and Science Center (EROS)
http://glovis.usgs.gov/
(Landsat Archive, Global Land Survey (GLS) data, as well as various ASTER and MODIS products)
USGS LandsatLook Viewer
http://landsatlook.usgs.gov/
(Enables searching of both LandsatLook images & Level1 Landsat data)
National Research Institute (INPE) of Brazil
http://www.dgi.inpe.br/CDSR/
(Range of Landsat and CBERS imagery, as well as various MODIS products)
Global Land Survey (GLS) 2005 products: Global Land-cover Facility
http://www.land-cover.org/data/
(Range of data sources including the Landsat archive and selected imagery for a range of instruments
including ASTER, Ikonos, Quickbird, Orbview, and MODIS)
RapidEye Catalog
http://eyefind.rapideye.com
(Access to the BlackBridge RapidEye image archive)
SPOT Catalog
http://catalog.spotimage.com/PageSearch.aspx?language=UK
(Access to the SPOT satellite archive)
Earth Remote Sensing Data Analysis Center (ERSDAC)
http://imsweb.aster.ersdac.or.jp/ims/cgi-bin/dprSearchMapByMenu.pl
(Access to the ASTER imagery archive)
Global tree-cover loss and biomass data

Global Landsat-based estimates of tree-cover loss from 2000 to 2012, produced by the University of
Maryland and distributed by Google.
http://earthenginepartners.appspot.com/science-2013-global-forest
Two maps of global forest biomass, with a 1-km resolution, based on a suite of satellite data inputs, calibrated
with plot data. These may be useful for national stratification of field sampling of biomass in a MRV system.
A. Baccini, S J. Goetz, W.S. Walker, N. T. Laporte, M. Sun, D. Sulla-Menashe, J. Hackler, P.S.A. Beck, R.
Dubayah, M.A. Friedl, S. Samanta and R. A. Houghton. 2012. Estimated carbon dioxide emissions from
tropical deforestation improved by carbon-density maps. Nature Climate Change
http://dx.doi.org/10.1038/NCLIMATE1354
Pantropical National Level Carbon Stock Dataset
http://www.whrc.org/mapping/pantropical/carbon_dataset.html
Saatchi S, N.L. Harris, S. Brown, M. Lefsky, E.T. Mitchard, W. Salas, B.R. Zutta, W. Buermann, S.L. Lewis, S.
Hagen, S. Petrova, L. White, M. Silman and A. Morel. 2011. Benchmark map of forest carbon stocks in
tropical regions across three continents. Proceedings of the National Academy of Sciences 14;108(24):9899-904.
http://carbon.jpl.nasa.gov/
Tutorials
The remote sensing tutorial: Federation of American Scientists (FAS)
http://www.fas.org/irp/imint/docs/rst/Front/tofc.html
General remote sensing: Canada Centre of Remote Sensing
http://www.nrcan.gc.ca/earth-sciences/about/organization/organization-structure/canada-centre-for-
remote-sensing/11740
USGS Change-tracking tool
http://pubs.usgs.gov/gip/133/
NASA fundamentals of remote sensing
http://gcmd.nasa.gov/records/remote_sensing_tutorial-00.html
Introduction to remote sensing - Virtual Hawaii
http://satftp.soest.hawaii.edu/space/hawaii/vfts/oahu/rem_sens_ex/rsex.spectral.1.html
NOAA’s Satellite and Information Service: Learning About Satellites and Remote Sensing
http://noaasis.noaa.gov/NOAASIS/ml/education.html
An introduction to remote sensing
CSIRO
http://www.cmis.csiro.au/rsm/intro/
An introduction to radar remote sensing: Canada Centre of Remote Sensing
http://www.nrcan.gc.ca/earth-sciences/geography-boundary/remote-sensing/radar-remote/2122
Radar polarimetry: Canada Centre of Remote Sensing
http://www.nrcan.gc.ca/earth-sciences/geography-boundary/remote-sensing/radar/1893
ESA’s RADAR Tutorial
http://earth.esa.int/applications/data_util/SARDOCS/spaceborne/Radar_Courses/
ESA’s Synthetic Aperture radar: Land applications tutorial
http://earth.eo.esa.int/download/eoedu/Earthnet-website-material/to-access-from-Earthnet/2008_Bilko-
SAR-Land-Applications-Tutorial/sar_land_apps_1_theory.pdf
U.S. Agency for International Development
1300 Pennsylvania Avenue, NW
Washington, DC 20523
Tel: (202) 712-0000
Fax: (202) 216-3524
www.usaid.gov

MRV Manual Chapter5

Uploaded by

Copyright:

Available Formats

MRV Manual Chapter5

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MRV Manual Chapter5

Uploaded by

Copyright:

Available Formats

REDD+ MEASUREMENT,

Tetra Tech Contacts:

Forest Carbon, Markets and Communities (FCMC) Program

Stephen Kelleher, Chief of Party

Olaf Zerbock, USAID Contracting Officer’s Representative

Please cite this report as:

REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 i

FOREST CARBON, MARKETS AND COMMUNITIES

REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 ii

REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 iv

REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 v

REDD+ MEASUREMENT, REPORTING AND VERIFICATION (MRV) MANUAL, VERSION 2.0 vi

5.2 LAND USES AND CATEGORIES IN THE UNFCCC

Box 5.1: Approaches

Approach 3: Tracking of land-use conversion on a mapped manner

5.2.1 Definition of national forest and other classes

Box 5.2: Definition of forest criteria

24 Examples of national forest definitions are available at http://cdm.unfccc.int/DNA/cdf/index.html

a) Forest Degraded Non-forest Sum T1 b) Forest Degraded Non- %

Forest 9,940 40 20 10,000 Forest 99.4 0.4 0.2 100

Degraded 5 1,970 25 2,000 Degraded 0.3 98.5 1.3 100

Non- 4,000 4,000 Non- 100.0 100

Sum T2 9,945 2,010 4,045 % T2 99.7 98.9 101.5

5.3 OVERALL STEPS AND NEEDS

1) What categories are most important to monitor?

3) What methodological aspects should be considered?

2) What type of classification approach should be used?

6) What post-classification processing steps should be applied?

5.4 REMOTE SENSING OVERVIEW

5.4.1 Types and characteristics of remote sensing data

Passive remote sensing

Active remote sensing

Resolution and other considerations

Box 5.3. Examples of the use of high resolution imagery in NFMSs

In summary, key data characteristics to consider are:

5.4.2 Image pre-processing, analysis, and post-processing

Geometric registration and co-registration

Replicability and analyst interaction versus automation

Calculating change rates

Forest Degraded Forest Non-forest Map total

Reference total 9,960 2,040 4,000 16,000

Producer's accuracy (%) User's accuracy (%)

Forest 99.2 Forest 98.8

Box 5.5. Additional characteristics of RADAR

Sensor Satellite – Swath Resolution Repeat Systematic Operational Monitoring

Sensor Satellite – Swath Resolution Repeat Cycle Systematic Operational Monitoring

SPOT-7 Airbus 60km 1.5m No Launched June 2014, deforestation,

Global tree-cover loss and biomass data

You might also like