Professional Documents
Culture Documents
March 2017
2
CONTENTS
1. Contents ........................................................................................ 2
3. Introduction .................................................................................. 8
4. Scale .............................................................................................. 9
8.1. Ethiopia’s Land Use and Land Cover Map 2013 ................................ 21
Image analysis .................................................................................................21
14. References................................................................................. 63
4
List of Tables
Table 4: Inputs needed for the statistical formula that calculates the number of
overall samples for the accuracy assessment .................................................. 36
Table 8: Matrix which gives correct classification, omission (under detection) and
commission (over detection) errors expressed in corresponding map area (ha).... 38
Table 11: Description of NFI strata and number of sampling units located in each
stratum ..................................................................................................... 43
Table 12: Tree and other vegetation measurements and observations in NFI. ...... 45
Table 13: IPCC ratios for Below Ground Biomass (2006) ................................... 47
Table 14: Number of SUs accessible and SUs accessible with cover sections
classified as forest per region ....................................................................... 48
Table 15: Number of NFI SU analyzed per Stratum and per Biome ..................... 49
Table 16: Number of SU classified as forest strata included in the EF analysis. ..... 49
Table 17: distribution of the plots per Biome and Ecological Zones: .................... 54
Table 18: IPCC 2006 default AGB values comparison with the Biomes estimates .. 56
5
List of Figures
Figure 1: Extent of the forest cover under the current Ethiopia’s forest definition 19
Figure 2: Land-uses replacing forest over the period 2000-2013 (as % of the total
forest loss over this period) .......................................................................... 21
Figure 5: High-resolution imagery from Google Earth and user interface from Open
Foris Collect Earth to classify sample data ...................................................... 34
Figure 11: Carbon estimated per strata per all the land uses ............................. 53
Figure 12: The average AGB (tC/ha) with their confidence intervals for forest in the
4 biomes is compared between primary (NFI) and secondary (literature and local
studies) ..................................................................................................... 57
Figure 13: NFI results for the average AGB carbon stock per biome in forest ....... 60
Figure 14: NFI results for the average BGB carbon stock per biome in forest ....... 61
Figure 15: NFI results for deadwood carbon stock per biome in forest ................ 61
6
Acronyms
AD Activity Data
EF Emission Factor
Ht height
RS remote sensing
SU sampling unit
WD wood density
7
Ethiopia’s FRL is in the context of receiving results based payments for REDD+
implementation. The FRL includes deforestation and afforestation, AGB, BGB,
deadwood and CO2 emissions; it is national and based on a historical average of
emissions and removals between 2000 and 2013. The Forest Reference Emission
Level for deforestation is: 17,978,735 tCO2 /year; the Forest Reference Level for
afforestation is: 4,789,935 tCO2 /year. The choice of construction approach and
historical period is provisional and may change in the future following appropriate
and comprehensive assessment and national circumstances. In this revised version
of the FRL the estimates have been improved by estimating the emission factors
using countrywide data from the National Forest Inventory and taking into account
the revision of the first submission.
8
INTRODUCTION
Ethiopia welcomes the invitation to submit a Forest Reference Emission Level (FRL)
on a voluntary basis expressed in Decision 12/CP.17, paragraph 13. This FRL
submission is in the context of results-based payments for the implementation of
reducing emissions from deforestation and forest degradation, and the role of
conservation, sustainable management of forests and enhancement of forest carbon
stocks (REDD+) under the United Nations Framework Convention on Climate
Change (UNFCCC).
Ethiopia has followed the guidance provided by the UNFCCC through the decisions
taken at the Conference of the Parties (CP), notably the modalities for forest
reference emission levels and forest reference level in Decision 12/CP.17 and the
guidelines for submissions of information on reference levels in the Annex of
Decision 12/CP.17. The original FRL submission was made in December 2015. This
revised version includes revisions that were already foreseen in the original (e.g.
national-level emission factors from recently completed NFI) and some clarifications
as a consequence of the technical assessment process over the last 8 months. This
submission does not prejudge or modify any of Ethiopia’s Nationally Determined
Contributions or Nationally Appropriate Mitigation Actions pursuant to the Bali
Action Plan.
SCALE
The Forest Reference Emission Level will cover the national territory of Ethiopia.
Ground inventory data has been collected in the past three years through the
Ethiopia’s National Forest Inventory and wall-to-wall and sample based remote
sensing have been carried out at the national scale. The current submission
therefore uses national activity data and emission factors.
10
Forest degradation is defined as the loss of carbon stock in forest land remaining
forest land, and forest enhancement is defined as the enrichment of carbon stock in
forest land remaining forest land (or the opposite of forest degradation). These
activities are not accounted for in the FRL.
With regard to the soil, this may constitute a very large carbon pool in Ethiopian
forests however, little is known about emissions from soil after forest conversion
and data collection in soils is very costly and needs monitoring over an extended
period. For this reason, the soil carbon pool is not included in the FRL.
Ethiopia does not have a recent soil map, a study (Tekalign et al., 1991) indicates
the Organic Carbon (OC) fraction to be between 0 to 3%. Currently there are no
other extensive and complete studies.
FOREST DEFINITION
In February 2015 Ethiopia adopted a new forest definition (MEF 2015) as follows:
'Land spanning more than 0.5 ha covered by trees (including bamboo) (with a
minimum width of 20m or not more than two‐thirds of its length) attaining a height
of more than 2m and a canopy cover of more than 20% or trees with the potential
to reach these thresholds in situ in due course (Minutes of Forest sector
management, MEFCC, Feb. 2015). Ethiopia is in the process of approving this as its
national legal definition.
This forest definition differs from the definition used for international reporting to
the Global Forest Resources Assessment (FRA) and from the forest definition used
in the NFI which both applied the FAO forest definition with the thresholds of 10%
canopy cover, a 0.5 ha area and a 5 m height.
The reason for changing the national forest definition is to better capture the
natural primary state of Ethiopia’s forest vegetation. Specifically, the reason for
lowering the tree height from 5 meters to 2 meters is to capture natural forest
vegetation types like the dryland forests, which of trees reaching a height of around
13
2-3 m. The proposed change in forest definition results in the inclusion of what
previously was classified as Ethiopia’s dense woodlands that have a wider
distribution through the country (see Figure 1). Commercial agriculture is
expanding mainly on dense woodlands and Ethiopia desires to enable REDD+
incentives for its conservation.
Figure 1: Extent of the forest cover under the current Ethiopia’s forest
The reason for increasing the canopy cover threshold from 10% to 20% is to avoid
acceptance of highly degraded forest lands into the forest definition and in this way
provide incentives for protecting quality forest.
This forest definition also differs from the definition used for reporting greenhouse
gas (GHG) emissions and removals from the forestry sector within the framework of
the Clean Development Mechanism (CDM) that was submitted to the UNFCCC
14
Since Ethiopia has improved the data quality on the forest area change assessment
and changed the forest definition, some inconsistencies currently exist between the
emissions and removals from forestry in the FRL and the GHG inventory. However,
the future GHG inventory reporting in the Biennial Update Report (BUR) will use the
improved data and new forest definition and full consistency will be sought when
reporting results in the technical annex to the BUR.
Free grazing affects the plains and lowland woodlands to the largest extent. The
large-scale investment agricultural schemes – both private ones and state owned
ones - have been significant drivers in Gambella, Benishangul-Gumuz and Afar
Regional States. In Ethiopian Somali and Afar Regional States charcoal is produced
by almost all rural households as one of the core livelihood income sources.
The findings of this study are confirmed by map comparison of land-use replacing
forest after deforestation assessed by Ethiopia’s National Forest Monitoring System,
described under section 6.2.
1 https://cdm.unfccc.int/DNA/cdf/index.html
15
Land uses replacing forest land
Bareland
Shrubland
Grassland
Agriculture
Open Woodland
Figure 2: Land-uses replacing forest over the period 2000-2013 (as % of the total forest
loss over this period)
16
HISTORICAL PERIOD
The historical period for which forest area change has been assessed is 2000-2013.
This period was selected to assess change because:
The year 2000 is well covered by standard products (Global Land Surveys
(GLS) and most global products) making it easier to connect or compare to
existing imagery and products
data for this period is available from the Global Forest Change (GFC) product,
(Hansen et al., 2013), which is used as a comparison and as a base for the
change analysis.
The period and dates are in line with requirements by FCPF Carbon Fund
Methodological Framework (criterion 11, indicator 11.1 and 11.2)
However, Ethiopia is still exploring whether the emissions and removals over this
period are representative of emissions and removals expected in absence of REDD+
implementation (and thus whether this historical period is appropriate as a
benchmark against which to assess performance). Tree cover loss estimates from
the Global Forest Change (GFC) product in 2013 suggest an upwards trend in tree
cover loss in Ethiopia. Ethiopia is still exploring and evaluating which FRL
methodology and/or choice of historical period best reflects emissions expected in
the near future in absence of REDD+ implementation.
17
A potential vegetation map of Ethiopia has been released by Friis and Sebsebe
(2009) and Friis, Demissew and van Breugel (2010). This map divided the Ethiopian
vegetation into 12 major types, which have been aggregated into 5 biomes.
The vegetation types are based on the information related to previous literature,
field experience of the authors, as well as on an analysis of about 1300 species of
woody plants in the flora of Ethiopia and Eritrea. The map is based on broad field
surveys, mainly along the country roads, and on a set of classification criteria
defining the altitudinal and rainfall limits for each of the vegetation types. The data
on altitude was obtained from 90x90 meter resolution digital elevation model
provided by the CGIAR-CSI (2008). The monthly total rainfall data with 30 arc
seconds resolution is from WorldClim. The Global Lakes and Wetlands Database
(GLWD) was used to delineate wetlands and lakes, and the AEON river database
(average stream separation of 15 km) was used to define the boundaries of water
bodies and related vegetation types.
18
The potential vegetation map has been also used as an input to create the strata
map that is the base layer used to design the national Forest and Landscape
Inventory of Ethiopia.
As the Forest and Landscape Inventory is not focused only on forest strata, during
2015 a new aggregation map was proposed to better represent the reliable carbon
stock estimates. Using the 12 vegetation types as input, these have been
aggregated into 5 biomes following expert judgment by Ethiopian botanical
scientists. Based on their knowledge of the vegetation types and their physiology
they have suggested the following aggregation into four biomes with expected
homogenous carbon contents (Figure 4).
19
This new stratification has been adopted to estimate the carbon content for the FRL
purposes.
20
The activity data for deforestation and afforestation was assessed as average of
annual forest loss and forest gain in hectares in the period 2000-2013. Satellite
imagery was used in combination with sample data from Google Earth observations
to determine the spatial extent of forest and forest area change. The creation of the
forest and forest area change map is explained under section 8.1, 8.2, 8.3 and 8.4
(step 1 and 2). Errors in the map are identified through verification of the map
classification by visual interpretation (section 8.4, step 3). Accordingly the map
area estimates are corrected for systematic error (bias) in the map and confidence
intervals are calculated around the bias-corrected area estimates as a measure of
the random error remaining in the estimate after removal of the bias (8.4, step 3).
The results are then compared against the Global Forest Change product and a
possible explanation for differences is provided (section 8.5).
Image analysis
Landsat 8 imagery was acquired from www.glovis.usgs.gov for images with less
than 3% cloud cover data acquisition dates from March 2013 to December 2013.
The land use/land classes were identified and based on past mapping experiences
of the Woody Biomass Inventory Strategic Planning Project (WBISPP, 2004) and the
prevailing ground situation of the country from the forest inventory. Representative
areas of interests (AOIs) were collected for each of the LULC classes using Google
Earth.
Geometric and radiometric corrections were applied to the Landsat images. The
AOIs served as training data and were used to classify the satellite data using the
Maximum Likelihood algorithm. The libraries of radiometric signatures for the
22
Landsat scenes were iteratively edited to harmonize the scenes. The classified
scene maps were mosaicked to form the thematic land cover/land use map for
Ethiopia.
23
The proposed approach by Ethiopia follows the GFOI guiding principle 1 for remote
sensing (GFOI, 2014): ‘When mapping forest change, it is generally more accurate
to find change by comparing images as opposed to comparing maps estimated from
images’. The approaches tested for change detection include purely automatic
spectral methods (e.g. IMAD algorithm) and supervised change detection using
change training points/loss and gain (Tewksebury et al. 2015). Post classification
change detection is not a suitable option for Ethiopia because historical land cover
maps do not have sufficient accuracy to derive change. Studies examining post
classification change have shown two forest/non-forest maps can be highly accurate
with user's accuracies of about 95%, the user's accuracy of the deforestation class
in the change map is likely to be much lower, indicating that the forest change
obtained by post-classification is inaccurate (Olofsson, et al., 2013).
26
The method chosen was a supervised classification of imagery, in which the user
identifies representative spectral samples for each class in the digital image. The
representative spectral samples are used as a dictionary and the classification
algorithm uses this dictionary to classify all objects/pixels depending on what their
spectral signature resembles most in the dictionary. In the case of change
detection, the object to be classified is a multi-temporal stack of imagery, and the
classes are change (loss and gain) or stable.
The imagery chosen for the exercise is Landsat data2, adapted to forest land cover
detection (FAO & JRC, 2012). The process assessed two mosaics for the year 2000
and 2013, to assess the change occurred in this period. For each year, all available
pixels covering a specified area and date range are collected and corrected for sun-
sensor-target anomalies3. A target day is fixed in order to get the maximum
vegetation cover and least cloud cover as possible. The proximity to this target day,
the pixel temperature and pixel wetness are computed to create a best-pixel
mosaic. The two best pixel mosaics (one for each time period) are finally stacked
into a consistent multi-temporal object. All the data collection, correction and
composition are implemented within Google Earth Engine (GEE) API.
The Ethiopian FRL considered two types of land use changes; non-forest land use
change to forest as gain, and forest land use change to the non-forest land use as
loss. Forest area change was corrected for bias using a stratified random sample,
using visually interpreted high temporal and spatial resolution satellite imagery. The
change detection process included:
2 The choice is justified by the fact that the Landsat images covers the entire time period, correspond to
the technical requirements, are consistent with the data used for the forest area map and are freely available and
likely to be sustainably available in the future (GFOI, 2014).
3 http://landsat.usgs.gov/Landsat8_Using_Product.php
27
Points for loss and gains were carefully assessed through a visual assessment using
a time series of Landsat images and vegetation indices and very high resolution
imagery available in the Google Earth, Bing Maps, and Here maps, all assessed
using the Open Foris Collect Earth interface. Points with a high level of classification
confidence, assigned by the technicians, were used for as the training dataset for
the supervised classification.
28
Figure 5: High-resolution imagery from Google Earth and user interface from Open Foris
Collect Earth to classify sample data
Sets of criteria that assist for the filtering and acquisition of Landsat images such as
date of acquisition, date and year range, month of image acquisition and target day
that controls phenology and cloud cover were set by national experts and
incorporated into Google Earth Engine API (table 3). Based on the selection and
filtering criteria, the collection of relevant Landsat archives were filtered through to
composite a best pixel mosaic.
29
No Criteria Decisions
1 Path and Row Landsat paths between 162 to 172 and Row between 50 and 58
2 Extent of dates for each Start date for initial period in the year of 1997-01-01 to 2003-
time period ( 2000 to 12-31 and range for the time period two start date 2010-01-01
2013) and 2015-03-31
3 Target day , the time with Start Julian day 330th day of the year/November to end of 60th
least clouds and most day of the year /February
vegetation
To use specific bands of Landsat images and normalize different band naming
between Landsat 5, 7 and 8 normalization function was implemented under GEE-
API. A function to correct for latitudinal components of sun-sensor of target
geometry was used from landsat.usgs.gov/Landsat8_Using_Product.php. This
function further produces weighted pixel based on the user specified day of the
year, reflectance values of band ratios, pixel temperature and wetness. To correct
longitude and latitudes, solar elevation angel and radiometric characteristics of each
pixel was implemented using the information obtained from metadata of acquired
images. Using the quality function, each band of images from multiple temporal
initial and final time of change assessment were mosaicked. Single images for each
time periods were created using the same script and composited into a multi-
temporal stack for detection of change between the two time periods (figure 5).
The classification process consists of compiling the spectral signature for all the
training points, creating a model from this spectral library and applying the model
to the entire imagery. Two classifiers have been tested, the CART algorithm
(Breiman et al, 1984) and the RandomForest algorithm (Breiman, 2001).
30
The validated training points were used at the national scale to train and directly
classify loss and gain from the mosaicked Landsat image. After a first run of
classification is completed, the training dataset can be improved by visually
assessing zones of false change (i.e stable areas classified as change) and missed
changes (i.e change classified as stable). Examples of potentially incorrect
classifications include agricultural area with strong greenness variations or shadows
due to elevation, which could be mistaken for false change and areas with known
deforestation classified as stable. The training sites are added on the misclassified
locations for the correct class. The new sites were entered in the spectral library
with appropriate labeling. Classification was iteratively improved by adding more
points of gain, loss and stable (figure 6). The outputs from the two algorithms were
visually compared and the RF classification was considered superior and used as the
result. The processing chain, from classification of the change, iterative
improvement of the training data, and export of the results was performed in the
Google Earth Engine API, with the following script (a GEE Trusted Tester account is
needed to open this link)
Figure 6: Number of available Landsat archives queried to create a best pixel mosaic in
Google Earth Engine.
31
classification, the results can be exported. The exported result from the GEE-API
was filtered to only include loss or gains in clumps of 5 pixels to match the national
minimum mapping unit of forest, 0.5 hectares. The 2013 land cover map was used
to filter false loss by eliminating loss detected on forest and false gain by
eliminating gain detected on non-forest. Manual editing was implemented to
improve the change detection product by delineating false and missed changes
using the Landsat composite image of 2000 and high-resolution imagery for 2013
from Google Earth. An accuracy assessment of the change was finally produced to
estimate the reliability of the change measured (FAO, 2015), produce corrected
estimates of change and confidence intervals around those estimates.
The required overall number of samples which was calculated using the Olofsson et
al. methodology (2014) adopting the following assumptions, using conservative
user inputs to have a cost effective and robust sample size:
Table 4: Inputs needed for the statistical formula that calculates the number of overall
samples for the accuracy assessment
The overall number of samples needed for the accuracy assessment are distributed
over the four map classes as follows: a minimum of 50 samples were allocated to
the “rare” map classes (classes covering a relatively small area) of forest loss and
forest gain. The remaining samples are distributed proportionally based on the
extent of the classes stable forest and stable non-forest.
Table 5: Samples distributed in the map classes for verification (collection of reference data)
Forest loss 84
Forest gain 97
Ethiopia has collected reference sample data to check whether the map
classification was correct at the location of the sample point. The sample data is
34
collected by visual interpretation of the point using a time series of very high
resolution imagery and Landsat imagery. The visual interpretation of sample points
is referred to as “reference data”. Only points where the users had high confidence
of the classification were used, therefore the total number of reference points used
in the area estimation is 1984 samples.
An error matrix was used to compare the map classes against the reference data.
The overall accuracy of the map is calculated by dividing the samples where the
map and reference data agree (the bold sample counts in diagonal) divided by the
total number of samples (sum of sample counts in all cells in the matrix).
Reference
2000-2013 Classes User’s
FL FG SF SNF Total
accuracy
Forest Loss (FL) 20 2 12 50 84 24%
Map data
To convert the sample counts into areas by which the map area per class can be
corrected, we first need to convert the absolute sample counts into proportions of
the total amount of samples per map class. E.g. for forest loss 20 out of 84 samples
are in agreement which equals to a proportion of 0.24, 2/84 = 0.02 was forest
gain, 12/84 = 0.14 was stable forest and 50/84 = 0.60 was stable non-forest.
Reference
2000- 2013 classes
FL FG SF SNF
FL= Forest Loss , FG= Forest Gain, SF= Stable Forest, SNF= Stable Non Forest
In order to obtain the bias-corrected area estimate, first the map area was
calculated multiplying proportional error matrix of the corresponding rows in the
matrix (Table 8) multiplied by the map areas per class. The result of this
multiplication is shown in Table 10
Table 8: Matrix which gives correct classification, omission (under detection) and
commission (over detection) errors expressed in corresponding map area (ha)
The bias-corrected area estimates are obtained by the map area minus the over-
detections (commission errors) plus the under-detections (omission errors). E.g.
36
The 95% confidence interval for the bias-corrected area estimates are calculated by
multiplying 1.96 by the standard error of the area estimate. The standard error of
the area estimate is calculated using formula 11 in Olofsson et al 2014. The results
are displayed in the table below:
The results from the accuracy assessment for forest loss for the period 2000 to
2013 is 1,193 thousand ha +/- 579 thousand ha and for forest gain is 246 thousand
ha +/- 216 thousand ha. The annual forest loss is approximately 92 thousand
37
The forest change map has also been overlaid with a map of biomes to divide the
forest loss, gain and forest cover estimates by biome. Forest loss and gain per
biome is used for the calculation of emissions and removals. The gain and loss per
biome is obtained by multiplying the loss/gain/forest areas in each biome by the
overall correction in the map, e.g. forest loss per biome is multiplied by a factor
1,192,55/372,18= 3.18. The relatively high annual forest area gain in the Dry
Afromontane biome gives some evidence that Ethiopia is already implementing
several mitigating actions that aim to restore forest resources. The on-going
mitigation actions reducing emissions are watershed management, agricultural
intensification, trees on farm for fuelwood, declining livestock (due to stall-feeding,
diseases, lack of own fodder and livestock raids), non-wood and alternative energy
sources, and controlled migration.
According to the importance of the carbon stock in the strata types, the sampling
distances were determined and the plot coordinates were generated using grid
generator. Accordingly, within the grid resolution of 1/4 x 1/4 degree square and
triangular combination grids plots coordinates were generated in the Stratum I,
while a ½ x ½ degree square and triangular combination grids were used for Strata
II and IV, ½ x ½ degree square grid for Stratum III, and a 1x1 degree square grid
in Stratum V, resulting in a total of 631 Sampling Units (SU).
Table 11: Description of NFI strata and number of sampling units located in each stratum
40
In the NFI, data was collected in the field through observations and measurements
at different levels: within the limits of the sampling units (SU) and in smaller
subunits within each SU, and Land Use/Cover Sections (LUCS). A sampling unit
consists of four subunits (sample plots) that can be divided into LULC. Trees and
stumps in the entire plot area have been recorded and small trees (in forest) and
saplings were recorded in smaller subplots (see the table below).
42
None None
Table 12: Tree and other vegetation measurements and observations in NFI.
43
The following parameters are needed to calculate above ground biomass in carbon
stock: diameter at breast height (dbh), tree height, a wood density factor and a
carbon fraction. The dbh and height parameters are measured in the field. A carbon
fraction of 0.47 has been applied, which is the default value for wood in the tropical
and subtropical domain (IPCC 2006).
AGB = 0.0673 WD DBH 2
H
0.976
Where:
In order to express the AGB pool in carbon stock, the AGB is multiplied by a carbon
fraction of 0.47 (kg C/kg dry matter).
According to Chave et al. (2014) the inclusion of country specific wood density in
the equation significantly improves biomass estimation. Therefore, MEFCC
conducted an extensive study to determine the most appropriate wood density
estimate for the country and basic wood density of 421 indigenous and exotic tree
species growing in Ethiopia (see table in annex). The overall average wood density
44
for the species is 0.612 g cm-3. This overall average wood density is comparable
with the global average value and that of tropical Africa (Chave et al. 2009; Reyes
et al 1992; Brown and Lugo 1984, IPCC 2006). The minimum value of wood density
was 0.262 for Moringa species, and the maximum was 1.040 g cm-3 for Dodonaea
angustifolia.
The BGB carbon pool default conversion factors proposed by IPCC (2006) were
applied using the ecological zone and ratios in Table 13.
R
Ecological Zone Above Ground Biomass [tonne root d.m. (tonne
shoot d.m,)-1
For fallen deadwood, the De Vries’ formula (De Vries, 1986) was applied, estimating
log volume in m3 ha−1. This formula requires the length of the transect (L) and the
log diameter (d) at the point of intersection
Where:
Two decomposition classes were recorded for deadwood particles: sound and
rotten. Missing data in the decomposition classes were assumed to be sound
deadwood piece. Because a rotten wood contains less biomass than a sound wood,
the wood density of dead wood is scaled down using lower wood densities than for
standing trees, as follows:
The default wood density for the species is 0.612 g cm-3, similarly as for trees.
The NFI data analysis is based on the analysis of 539 accessible and surveyed
sample units from the NFI out of the 631 sample units from the original sampling
design. The surveyed sample units are from the different regions (Table 14).
The calculations for the EF are based on the land use sections of each sample unit
classified as forest.
Table 14: Number of SUs accessible and SUs accessible with cover sections
classified as forest per region
N of accessible sampling N of accessible sampling units
Region
units with forest
Tigray 31 5
Afar 23
Amhara 94 12
Oromia 200 68
Somali 80 1
Beneshangul Gumu 28 12
SNNPR 68 25
Gambela 15 5
Total 539 128
46
Table 15: Number of NFI SU analyzed per Stratum and per Biome
Biome
13 18 59 95
5
Stratum I
107 107
Stratum II
1 93 6 1 101
Stratum III
36 38 114 29 1 218
Stratum IV
15 2 1 18
Stratum V
164 146 132 94 3 539
Total
9 12 50 72
1
Stratum I
2 2
Stratum II
14 2 16
Stratum III
1 7 17 12 37
Stratum IV
1 1
Stratum V
5 30 29 64 128
Total
The NFI sample plot design implies a different sampling probability for trees in the
plots and small trees in the sub-plots. Thus, two different areal weighting methods
47
for tree and sapling data were applied. All results were first computed at the LUCS
level by plots, and then aggregated up to strata level by regions.
Figure 11: Carbon estimated per strata per all the land uses
The first step involves using a ratio estimator, defined by the ratio between the
variable y (tons of carbon) and the variable x (hectares of forest). Both y and x
have been measured for each plot. The index d, which in statistical literature is
referred to as a “domain”, is the biomes.
48
The sample selected (the 631 plots from the sample design) is called s. sd is the
portion of the sample that fall into biome d (so, for example, for the NFI strata 1, sd
correspond to the 5 plots that have fallen into the biome 1).
The NFI design is stratified, therefore each stratum has a different sampling
intensity defined by the inclusion probability ( ) of each plot. The has been
computed by dividing the number of hectares sampled in each strata by the total
area of the strata (when sampling intensity is higher, inclusion probability is
higher). Because the total area of the strata is much bigger than the area actually
inventoried, the value can be very small. In the same domain, the plots can
come from different strata and therefore they can have different inclusion
probabilities.
The domain ratio estimator has been calculated using the following equation:
Where:
inclusion probability ( ): ;
The domain ratio estimator provides the total domain. The average has been
calculated dividing by the size of the domain as follows:
49
This calculation has been used to compute the estimates per hectare for each
biome.
where:
probability:
and:
When the plots k and l come from different strata, their selection has been
independent:
and is therefore 0.
When the plots k and l come from the same stratum then:
50
While
As k and l are from the same plot and from the same stratum .
To summarize, the
In order to correctly apply the BGB ratios from the IPCC (2006), the ecological zone
layers have been applied. Tables 15, 16 and 17 shows that the distribution of SU by
biomes and ecological zones differs from the distribution by biomes and NFI strata.
Table 17: distribution of the all the SUs per Biome and Ecological Zones:
Ecological Zones
Tropical
Tropical Tropical
Tropical moist Tropical Total
dry mountain Water
desert deciduous shrubland Result
forest system
Biomes forest
Acacia-Commiphora 18 0 0 54 123 0 195
Combretum-
Terminalia 0 43 13 92 31 0 179
Dry Afromontaine 0 1 148 2 0 151
Moist Afromontaine 0 1 7 85 0 0 93
Others 1 1 0 2 4 1 9
Total 19 46 20 381 160 1 627
The difference in the distribution is why there is not direct correspondence between
AGB and BGB in some estimates.
Figure 12: The average AGB (tC/ha) with their confidence intervals for forest in the 4
biomes is compared between primary (NFI) and secondary (literature and local studies)
The use of primary data greatly reduces the confidence intervals of the NFI analysis
when compared to the secondary data analysis. When comparing the results from
secondary and primary data analysis, there is a large difference in AGB estimates
for Dry and Moist Afromontane forests, where the secondary sources suggest much
higher carbon contents (220% and 62% higher for Dry and Moist Afromontane
forest, respectively). This difference is believed to be attributed to the sample
design in the secondary data, which most likely targeted primary and dense forest
patches. Therefore, the NFI data is thought to be more representative for
estimating emissions and removals from country-wide forest area changes.
A further comparison has been done with the IPCC default values, showing a
substantial concordance with the NFI estimates (cfr. table 18).
53
Table 18: IPCC 2006 default AGB values comparison with the Biomes estimates.The
asterisk identifies the most represented ecological zones in Ethiopia
Correspondent
Above-ground
above-ground
Ecological zones biomass
biomass range for
(t ha-1)
Biomes (t ha-1)
Tropical moist deciduous
260 (160-430) 65-200
forest
Above-ground
Biomes biomass
(t ha-1)
Acacia-Commiphora 55
Combretum-Terminalia 65
Commiphora biome are not considered reliable as some very large diameters
strongly influence the results. It is expected that this data will become available
once the national level data has been collected in the course of 2016.
Figure 13: NFI results for the average AGB carbon stock per biome in forest
55
Figure 14: NFI results for the average BGB carbon stock per biome in forest
Figure 15: NFI results for deadwood carbon stock per biome in forest
56
Ethiopia assumes total oxidation of AGB, BGB and deadwood after forest
conversion, therefore emission factors are approximated by the full carbon stock in
AGB, BGB and deadwood for forest in the different biomes. The removal factor for
forest gain is estimated as the inverse of the emission factor therefore assuming
full average carbon stock for each hectare of gain detected. As such, Ethiopia does
not take into account the age structure in the forest, which would introduce too
much complexity (for the time being). Assuming the full carbon stock is removed
from the atmosphere at the time gain is detected may over-estimate the removals
corresponding to the early years of forest growth. However, this may be
compensated by the fact that gain is generally detected by remote sensing in a
later stage of growth (therefore removals already preceded the time of detection).
57
The strategic directions of the forest sector in GTP-2 are enabling the community to
actively participate in environmental protection and forest development activities,
and implementing the green economy strategy at all administration levels and
embarking on environmental protection and forest development at a scale. In the
GTP-2, the sector has thus set goals mainly in relation to building climate resilient
green economy, environmental protection and forest development. This will be
applied mainly in priority sectors identified by the CRGE strategy. In addition, a goal
of the sector is mobilizing resources that can enable full implementation the CRGE
strategy. In terms of forest development, the aim is to increase the share of the
forest sector in the overall economy. The strategy also aims to increase the forest
coverage through research-based forest development. A target set during the GTP-
2, is to reduce deforestation by half.
58
UPDATING FREQUENCY
In order to ensure the accuracy of the FRL with updated socio-economic conditions
and in order to incorporate new or improved data that may be available, the FRL
will be revised periodically. Ethiopia proposed this FRL to be valid at least for 5
years, yet it may be improved and updated more frequently. Ethiopia is also
evaluating the possibility of undertaking a new region based forest inventory, with a
5 year cycle.
61
FUTURE IMPROVEMENTS
REFERENCES
1. Breiman, L., Freidman, JH, Olshen, RA, and Stonem CJ. 1984. Classification
and regression trees. Wadsworth, Belmont CA.
2. Breiman. 2001. Random Forests. Machine Learning. 45, 2-32.
https://www.stat.berkeley.edu/~breiman/RandomForests/cc_home.htm
3. CGIAR-CSI SRTM 90m DEM Digital Elevation Database, versión 4. CGIAR
Consortium for Spatial Information (CGIAR-CSI). URL:
http://srtm.csi.cgiar.org/Index.asp
4. Bodart, C., Brink A., Donnay F., Lupi, A., Mayaux, P., & Achard, F. (2013).
Continental estimates of forest cover and forest cover changes in the dry
ecosystems of Africa between 1990 and 2000. Journal of biogeography,
40(6), 1036-1047.
5. Brown, S. (1997). Estimating biomass and biomass change of tropical
forests: a Primer. FAO Forestry Paper 134:120-130.
6. Brown, S. (2002). Measuring carbon in forests: current status and future
challenges. Environmental Pollution 116: 363-372.
7. Brown, S. and A. E. Lugo (1990). Tropical secondary forests. Journal of
Tropical Ecology 6:1-32.
8. Brown, S. and Lugo, A. (1982). The storage and production of organic matter
in tropical forests and their role in the global carbon cycle. Biotropica 14:
161-187.
9. Brown, S.A.J., Gillespie, J.R. and Lugo, A.E.(1989). Biomass estimation
methods for tropical forests with application to forest inventory data. Forest
Science 35: 110-115.
10.Chave J, Coomes D, Jansen S. 2009. Towards a worldwide wood economics
spectrum. Ecology Letters, 12, 351–366.
11.Chave, J., Andalo, C., Brown, S., Cairns, M.A., Chambers, J.Q., Eamus, D.,
Fölster, H.,Fromard, F., Higuchi, N., Kira, T., Lescure, J.-P., Nelson, B., Ogawa,
H., Puig, H.,Riéra, B., Yamakura, T., 2005. Tree allometry and improved
estimation of carbon stocks and balance in tropical forests. Oecologia 145,
87–99.
12.Chave J, et al. 2014. Improved allometric models to estimate the
aboveground biomass of tropical trees. Global Change Biology 20, pp. 3177–
90.
13.Cornelissen, J. H. C., S. Lavorel, E. Garnier, S. Diaz, N. Buchmann, D. E.
Gurvich, P. B. Reich, H. ter Steege, H. D. Morgan, M. G. A. van der Heijden,
J. G. Pausas, and H. Poorter. 2003. A handbook of protocols for standardized
and December 2006 WOOD DENSITY IN NEOTROPICAL FORESTS 2365 easy
measurement of plant functional traits worldwide. Australian Journal of
Botany 51:335–380.
14.CSA (2007). Population and Housing Census Report-Country – 2007
(http://www.csa.gov.et/index.php/2013-02-20-14-51-51/2013-04-01-11-53-
00/census-2007)
15.Demel Teketay, Mulugeta Lemenih, Tesfaye Bekele, Yonas Yemshaw, Sisay
Feleke, Wubalem Tadesse, Yitebitu Moges, Tesfaye Hunde & Demeke Nigussie
63
30.Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., &
Wulder, M. A. 2014. Good practices for estimating area and assessing
accuracy of land change. Remote Sensing of Environment, 148, 42-57.
31.Pearcy R. W., Ehleringer J. R., Mooney H. A., Rundel P. W. . 1989. Plant
physiological ecology. Chapman and Hall, New York, New York, USA.
32.Reusing, M. (2000). Change detection of natural high forests in Ethiopia
using remote sensing and GIS techniques. International Archives of
Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000
33.Reyes, G., S. Brown, J. Chapman, and A. E. Lugo. 1992. Wood densities of
tropical tree species. General Technical Report SO-88. USDA Forest Service,
Southern Forest Experiment Station, New Orleans, Louisiana, USA
34.Sarndal, B. Swensson, J. Wretman: Model Assisted Survey Sampling,
Springer, New York, 1992
35.Tewkesbury, A. P., Comber, A. J., Tate, N. J., Lamb, A., & Fisher, P. F. 2015. A
critical synthesis of remotely sensed optical image change detection
techniques. Remote Sensing of Environment, 160, 1-14.
36.De Vries P.G., 1986. Sampling theory for forest inventory. A teach-yourself
course, Springer-Verlag, Berlin, 420 p.
37.WBISPP (2004). A strategic plan for the sustainable development,
conservation and management of the woody biomass resources: Final report.
Federal Democratic Republic of Ethiopia, Ministry of Agriculture. 60 pp.
38.Westerman R. L. 1990. Soil testing and plant analysis, 3rd ed. Soil Science
Society of America, Madison, Wisconsin, USA.
39.de Wit, M. & Stankiewicz, J. 2006. Changes in Surface Water Supply across
Africa with Predicted Climate Change. Science 311: 1917-1921.
65
ANNEX
ANNEX I
SU area Samples
Strata SU Strata Area (Ha) Sampling Intensity
(Ha) distribution (%)
SU area Samples
Biome SU Biome Area (Ha) Sampling Intensity
(Ha) distribution (%)
Table 19: sampling distribution and sampling intensity comparisons between NFI strata and biomes.
NOTE:The difference between the number of SU (631 and 627) is due to the fact that 4 SU during the sampling design are accidentally fall
outside the national boundaries currently adopted.
66
ANNEX II
Table 20: Basic Wood Density of Indigenous and Exotic Tree Species in Ethiopia
Basic Density Data
No Scientific name (g/cm3) Reference quality* Remark
1 Acacia abyssinica 0.826 average of genus (ICRAF database) M
http://www.worldagroforestry.org/regions/southeast
2 Acacia albida 0.562 _asia/resources/db/wd M
3 Acacia asak 0.769 average of genus (ICRAF database) M
4 Acacia brevispica 0.769 >> M
5 Acacia bussei 0.769 >> M
Acacia decurrens Air dry
6 0.816 Getachew Desalegn et al., 2012 H density
7 Acacia dolichocephala 0.769 average of genus M
8 Acacia drepanolobium 0.769 >> M
9 Acacia etbaica 0.590 Vreugdenhil et al., 2012 H
Tropical Africa: global database (Zanne et al.,
10 Acacia gerrardii 0.775 2009) M
11 Acacia goetzei 0.883 >> M
12 Acacia hokii 0.769 average of genus (ICRAF database) M
13 Acacia Iahai 0.769 >> M
14 Acacia macrothyrsa 0.769 >> M
15 Acacia mellifera 0.482 Vreugdenhil et al., 2012 H
16 Acacia mollis 0.482 Vreugdenhil et al., 2012 H
17 Acacia nilotica 0.723 Vreugdenhil et al., 2012 H
18 Acacia oerofota 0.769 average of genus M
19 Acacia pentagon 0.826 average of genus (ICRAF database) M
20 Acacia polyacantha 0.769 average of genus M
21 Acacia reficiens 0.769 >> M
22 Acacia robusta 0.769 >> M
23 Acacia senegal 0.741 Vreugdenhil et al., 2012 H
24 Acacia seyal 0.497 Vreugdenhil et al., 2012 H
25 Acacia sieberiana 0.769 average of genus M
26 Acacia tortilis 0.590 Vreugdenhil et al., 2012 H
27 Acacia xiphocarpa 0.769 average of genus M
28 Acacia Zanzibarica 0.769 average of genus M
29 Acallypha acrogyna 0.300 A. Cauturus (Zanne et al.; global database), L
30 Acanthus sp. 0.592 Global database (Zanne et al., 2009) M
31 Acokanthera schimperi 0.784 Acokanthera oppositifolia (from Global database) L
Acrocarpus air dry
32 fraxinifolius 0.610 Getachew Desalegn et al., 2012 H density
33 Adansonia digitata 0.590 Vreugdenhil et al., 2012 H
same species from wood density for trees of
34 Adathada schimperiana 0.640 Uganda L
http://db.worldagroforestry.org//wd/species/Alangi
35 Alangium chinense 0.420 um_chinense M
36 Alangium Chinese 0.408 >> M
37 Albizia aylemeri 0.579 Genus average M
http://db.worldagroforestry.org//wd/species/Alangi
38 Albizia glaberiima 0.555 um_chinense M
39 Albizia grandibracteata 0.534 Albizia gummifera L
Albizia gummifera air dry
40 0.580 Getachew Desalegn et al., 2012 H density
http://db.worldagroforestry.org//wd/species/Alangi
41 Albizia lebbeck 0.596 um_chinense M
42 Albizia lophantha 0.579 Genus average M
67