ArticlePDF Available

Abstract and Figures

Cocoa growing is one of the main activities in humid West Africa, which is mainly grown in pure stands. It is the main driver of deforestation and encroachment in protected areas. Cocoa agroforestry systems which have been promoted to mitigate deforestation, needs to be accurately delineated to support a valid monitoring system. Therefore, the aim of this research is to model the spatial distribution of uncertainties in the classification cocoa agroforestry. The study was carried out in Côte d’Ivoire, close to the Taï National Park. The analysis followed three steps (i) image classification based on texture parameters and vegetation indices from Sentinel-1 and -2 data respectively, to train a random forest algorithm. A classified map with the associated probability maps was generated. (ii) Shannon entropy was calculated from the probability maps, to get the error maps at different thresholds (0.2, 0.3, 0.4 and 0.5). Then, (iii) the generated error maps were analysed using a Geographically Weighted Regression model to check for spatial autocorrelation. From the results, a producer accuracy (0.88) and a user’s accuracy (0.91) were obtained. A small threshold value overestimates the classification error, while a larger threshold will underestimate it. The optimal value was found to be between 0.3 and 0.4. There was no evidence of spatial autocorrelation except for a smaller threshold (0.2). The approach differentiated cocoa from other landcover and detected encroachment in forest. Even though some information was lost in the process, the method is effective for mapping cocoa plantations in Côte d’Ivoire.
This content is subject to copyright. Terms and conditions apply.
Vol.: (0123456789)
1 3
Agroforest Syst (2023) 97:109–119
Modelling thespatial distribution oftheclassification error
ofremote sensing data incocoa agroforestry systems
DanKanmegneTamga · HoomanLatifi·
TobiasUllmann· RolandBaumhauer·
MichaelThiel· JulesBayala
Received: 29 November 2021 / Accepted: 19 October 2022 / Published online: 28 October 2022
© The Author(s) 2022
vegetation indices from Sentinel-1 and -2 data respec-
tively, to train a random forest algorithm. A classified
map with the associated probability maps was gener-
ated. (ii) Shannon entropy was calculated from the
probability maps, to get the error maps at different
thresholds (0.2, 0.3, 0.4 and 0.5). Then, (iii) the gen-
erated error maps were analysed using a Geographi-
cally Weighted Regression model to check for spatial
autocorrelation. From the results, a producer accuracy
(0.88) and a user’s accuracy (0.91) were obtained. A
small threshold value overestimates the classification
error, while a larger threshold will underestimate it.
The optimal value was found to be between 0.3 and
0.4. There was no evidence of spatial autocorrelation
except for a smaller threshold (0.2). The approach dif-
ferentiated cocoa from other landcover and detected
encroachment in forest. Even though some informa-
tion was lost in the process, the method is effective
for mapping cocoa plantations in Côte d’Ivoire.
Keywords Cocoa mapping· Geographically
weighted regression· Sentinel-1· Sentinel-2·
Shannon entropy· Spatial error assessment
The Cocoa tree (Theobroma cocoa) is a major cash
crop in the forest region of West Africa. It is grown
for cocoa beans which are known to be used as input
in the agro-industrial sector (chocolate and beverage
Abstract Cocoa growing is one of the main activi-
ties in humid West Africa, which is mainly grown in
pure stands. It is the main driver of deforestation and
encroachment in protected areas. Cocoa agroforestry
systems which have been promoted to mitigate defor-
estation, needs to be accurately delineated to support
a valid monitoring system. Therefore, the aim of this
research is to model the spatial distribution of uncer-
tainties in the classification cocoa agroforestry. The
study was carried out in Côte d’Ivoire, close to the Taï
National Park. The analysis followed three steps (i)
image classification based on texture parameters and
D.KanmegneTamga(*)· H.Latifi· M.Thiel
Department ofRemote Sensing, Institute forGeography
andGeology, Julius-Maximilians-University ofWürzburg,
Oswald-Külpe-Weg 86, 97074Würzburg, Germany
Department ofPhotogrammetry andRemote Sensing,
Faculty ofGeodesy andGeomatics Engineering,
K. N. Toosi University ofTechnology, P.O. Box,
15433-19967Tehran, Iran
T.Ullmann· R.Baumhauer
Department ofPhysical Geography, Institute
forGeography andGeology, Julius-Maximilians-
University ofWürzburg, Am Hubland, 97074Würzburg,
Centre forInternational Forestry Research (CIFOR) -
World Agroforestry (ICRAF), CIFOR-ICRAF, Sahel 06,
BP 9478, Ouagadougou06, BurkinaFaso
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol:. (1234567890)
production), but also in cosmetic and pharmaceutical
products. Cocoa beans and cocoa food preparations
are the first exported commodities of the region. They
represent about 5% of the revenues from exports,
more than precious stones (3%) and their primary
destinations are the Americas (40%), Europe (28%),
and Asia (16%) (ECOWAS 2016). At a microeco-
nomic level, cocoa farming is the basis of the liveli-
hood of about two million farmers in the West Afri-
can cocoa belt (Boeckx etal. 2020). Despite the fact
that revenues from cocoa stricto sensu are not suf-
ficient to reach descent living income (only 33% of
farmers in Ghana has a sufficient income from cocoa
Boeckx etal. 2020), and the burning question of child
labour in the cocoa value chain (Busquet etal. 2021;
ILO 2017; International Anti-slavery 2004), the sus-
tainable intensification of cocoa production could be
a solution to poverty alleviation (Boeckx etal. 2020).
The demand for cocoa will increase since the global
cocoa beans market size is projected to expand by
7.3% in 2025, mainly due to the rapid growing of
chocolate industries in emerging economies such as
China and India (GVR 2019).
Côte d’Ivoire is the world largest producer of
cocoa beans producer, with over 2 million tons in
2020, which makes up for approximately 39% of the
world’s cocoa production (Shahbandeh 2021). For
that performance, 47 000 ha of forest were cleared
in 2020 (Mighty Earth 2021). Smallholder farmers
take advantage of soil fertility and land availability
to establish new cocoa farms at the expense of forest
(Ruf etal. 2014). Encroachments in protected areas
where reported, with alarming consequences such as
habitat degradation and loss of wildlife -especially for
primate species-, carbon emissions and soil degrada-
tion (Abu etal. 2021; Dean 2019; Bitty et al. 2015).
Cocoa farming is the main driver of deforestation,
responsible of the loss of more than 80% of the forest
cover in Côte d’Ivoire between 1961 and 2000 (Sabas
etal. 2020). Here the term cocoa and cocoa planta-
tion will be used alike to refer to cocoa farming.
Certification of cocoa, which includes environ-
mental standards, was introduced to limit deforesta-
tion. These standards include sustainable farming and
reduced deforestation, mainly to ensure that cocoa
farms are established outside protected area (ProLand
2020). Since its introduction in Côte d’Ivoire in 2005,
certification has not consistently prevented deforesta-
tion. One reason is that it is voluntary, but the main
reason is the gap in the geographic coverage. Weak-
ness in traceability was reported as the most critical
factors due to a lack of accurate maps of cocoa, which
results in poor auditing and difficulty for the enforce-
ment of standards (ProLand 2020). Efforts have been
made to limit cocoa extension, by improving old
cocoa farm productivity through the introduction of
forest trees in cocoa plantations; but also by advance-
ment of surveillance around the Comoé and Taï
National parks. (GIZ 2020; Pye-Smith etal. 2016).
However, these efforts have not been geographically
evaluated due to the lack of a robust method to gener-
ate accurate maps.
Freely-available remote sensing imagery from the
Sentinel mission, provided by the European Space
Agency (ESA) has been used for the detection and
mapping of cocoa plantations in West Africa. Abu
et al. (2021) detected cocoa and its impact on pro-
tected areas in Côte d’Ivoire. With the same data-
set processed with deep learning, it was possible
to separate full sun cocoa and cocoa agroforestry
(cocoa associated with forest trees) in the same region
(Ashiagbor et al. 2020). However, cocoa farms are
difficult to map, especially in the regions where they
share the landscape with other spectrally overlap-
ping crops such as rubber and palm oil. This create
a heterogenous landscape also referred to as complex
landscape where the classified map will have a dis-
proportionate number of misclassified pixels (Fillela
2018). To ensure image classification of good qual-
ity and with high resolution imagery could be used
in combination with open-source data. This solution
was used to successfully delineate cocoa agroforests
in Cameroon, and to classify different agroforestry
systems (AFS) in Mali (Aguilar etal. 2018; Numbisi
etal. 2019). Such high-quality map comes at a cost
in terms of the price, geographic restriction, and the
difficulty to integrate it in a near-real time monitoring
Another solution is to use freely available imagery
for the classification of cocoa farms while present-
ing a spatial assessment of the classification error
alongside the classification. It is acknowledged that
the confusion matrix is not informative enough to
evaluate a classification, since it does not provide
any spatial information on the error. Therefore, we
introduced a spatial error assessment workflow as a
supplement for the evaluation of a classified map
(Comber etal. 2020; Narayan etal. 2021; Roodposhti
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol.: (0123456789)
et al. 2019). The spatial evaluation of the classifi-
cation is a post classification method that is used
to identify pixels with a high probability of being
misclassified. It is based on the analysis of the het-
erogeneity at each pixel based on the prediction prob-
abilities (Roodposhti et al. 2019). It also evaluates
the spatial correlation of the error to identify local
conditions that affect the classfication (Comber etal.
2020). The application of this classification workflow
is still missing in West Africa. Therefore, the aim of
this study is to model the spatial distribution of error
on the classified map to provide reliable cocoa maps
in Côte d’Ivoire. The study neither assesses the per-
formance of classification techniques nor evaluates
the type of input data, as it rather aims at detecting
and mapping misclassified pixels. The hypotheses
are: (i) the analysis of the heterogeneity of each pixel
could reveal misclassified pixels; and (ii) spatial error
assessment could reduce false detection of cocoa
encroachment from non-protected into protected area;
(iii) the distribution of the classification error is spa-
tially correlated and could be explained by satellite
derived indices.
Study area
The study was conducted in the southern part of Côte
d’Ivoire. The area was a forest region associated
with the transitional equatorial climate. Agriculture
was the main activity to support the economic life
in the region. Within the region, available lands are
promptly converted into plantations of cash crops
such as cocoa (Theobroma cocoa), rubber (Hevea
brasiliensis) and oil palm (Elaeis guineensis), but
also for food crops like yam (Dioscorea alata), rice
(Oriza sativa), and cassava (Manihot esculenta).
Agricultural plantations are located next to two pro-
tected forests in the study area, including the northern
part of Taï National Park and a portion of the wildlife
reserve of N’Zo. These are fully protected areas, with
no human intervention.
A field campaign was conducted in March 2020,
in the administrative subdivisions of Gueyo and Buyo
(Fig. 1). The data acquisition consisted of collect-
ing georeferenced points and polygons in AFS, using
sampling plots of 25 × 25m, which were used in the
image classification workflow. Four AFS were con-
sidered: (i) cocoa plantations (full sun cocoa or cocoa
agroforestry) with at least two production cycles; (ii)
Rubber and (iii) oil palm plantations as monoculture
planted in rows, often established on formal agricul-
tural plots; and (iv) agricultural farms dedicated to
food production including cotton (Gossypium sp).
In total, 139 referenced polygons were collected: 30
plots of cocoa, 31 plots of rubber, 16 plots of palm
and 62 agricultural farms.
Overview of methodology
The data processing was organized in three steps:
(1) supervised image classification from which two
outputs were generated: a classification map and
the probability maps of each class. (2) the Shannon
entropy at pixel level was calculated using the prob-
ability maps, and different thresholds were specified
to model the spatial distribution of the classification
error (Roodposhti etal. 2019). (3) A geographically
weighted regression was used to evaluate the spatial
autocorrelation of the classification error (Comber
etal. 2020).
Data pre-processing
The Sentinel-1 (S1) C-band Synthetic Aperture Radar
data provided by the ESA was used in combination
with optical data to deal with clouds in the region of
interest. The selected acquisition mode was the Inter-
ferometric Wide swath (IW) with VV and VH polari-
sations, which is the main acquisition mode over land
and satisfies most current service requirements such
as preserving revisit performance (Copernicus 2022).
With Google Earth Engine (GEE), a composite image
(median pixel value) was created between 2017 and
2020, which is corresponding to the period of expo-
nential expansion of cash crops in the area (Yao etal.
2020). The median image was used to be able to com-
bine S1 with Sentinel-2 (S2). A mosaic image of the
study area was finally downloaded at the spatial reso-
lution of 10m.
The texture information was retrieved from the VV
and VH bands using the Grey Level Co-occurrence
Matrix (GLCM) (Table1). Texture parameters from
GLCM provide valuable information on the spatial
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol:. (1234567890)
relationship between pixels on the image, and also
significantly improve the classification accuracy,
especially for heterogenous landscapes with open
canopy like cocoa plantations (Mishra et al. 2019).
To generate the GLCM texture parameters, a window
size of 5 by 5 pixels, and 32 number of grey levels
were used as suggested in the literature (Mishra etal.
2019; Numbisi etal. 2019).
The S2 archive provided by ESA was used in combi-
nation with S1. Five bands were used: red (R), green
(G), blue (B), near-infra-red (NIR) and red-edge (RE)
at the spatial resolution of 10 m. The images were
acquired using GEE, which uses the Sen2Cor module
for the preprocessing of S2 images. Since the area is
cloudy throughout the year, the cloud percentage per
pixel was set to 5%, and for each band to get a use-
ful image. The composite image (median image) was
generated between 2017 and 2020. Seven vegetations
VARI, see Table 2) were generated to provide
Fig. 1 Study area in Southern Côte d’Ivoire. Land cover mask represents the area corresponding to water forest, and urban areas.
Gueyo and Buyo are the administrative subdivisions where the field work was carried out
Table 1 GLCM texture parameters (Hall-Beyer 2017)
Texture measures Abbreviation (for
VV polarization)
Entropy VV_entro ∑−ln(Pij)Pij
Contrast VV_contr ∑Pij(i−j)2
Variance VV_var ∑Pij(i−µi)2
Correlation VV_cor ∑Pij[(i−µi)(j−µi)/σ2]
Mean VV_mean ∑Pij/N
Homogeneity VV_homo ∑Pij/(1 + (i−j))2
Dissymmetry VV_diss ∑(∑(|i−j|Pij))
Second moment VV_sec ∑(∑(Pij)2)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol.: (0123456789)
additional information on the biophysical properties
of the vegetation in the study area.
Data analysis
Random forest classification
Random Forest (RF) is a commonly used machine
learning algorithm for classification, which combines
the output of multiple decision trees to reach a single
classification result. It could predict a feature class,
or the probability of belonging to a class. RF reduces
the risk of overfitting, provides flexibility (as it is not
affected by data distribution or missing values) and
gives information on the feature importance.
To classify the image features from S1 and S2, the
data were split in train (75%) and test (25%) sets. A
10-folds cross validation with 10 repetitions was used
to train the model using the train set, and the optimal
variables were randomly sampled at each split. The
Overall Accuracy (OA), the Producer and User Accu-
racies (PA and UA), and the area under the curve
were generated as evaluation metrics for the classifi-
cation (Abu etal. 2021). The trained model was used
to predict AFS classes on the image, and to generate
the probability map of each feature class. To reduce
the salt and pepper effect on the map, a probability
pass filter of 3 by 3 pixels was applied on the clas-
sified map. The spatial resolution was evaluated and
integrated in the area analysis.
Spatial error assessment
The probability map of each feature class, generated
from RF algorithm, was used to generate an entropy
map using the Shannon entropy (Eq.1). The entropy
values range from 0 to 1, where high entropy values
are associated with high probability of classification
error (Roodposhti etal. 2019).
where H(x) is the entropy value for pixel (x), and Pi
is the probability value of the pixel (x) for the AFS
class i.
The pixels of the entropy map were grouped into
two classes (error and good classification) using pre-
defined entropy threshold’s value. This was referred
to as error map. Different threshold values were
evaluated: (1) a single entropy threshold value for the
entire map (Roodposhti etal. 2019). Here four values
were evaluated (0.2; 0.3, 0.4; 0.5). (2) Entropy thresh-
old value corresponding to each feature class. To get
the value for each class, field polygons were used to
extract entropy pixel values for each class. Under
normal distribution, the mean value was selected as
threshold for the given class. The extracted values
were cocoa = 0.32; palm = 0.52; farm = 0.3 and r ub-
ber = 0.43. The error maps were then used to remove
error pixels from the classified map. Since the error
map from (2) was based on field plots, it was used as
reference with which other error maps were compared
to evaluate under- and overestimation for each class.
Geographically weighted regression (GWR)
A GWR model (2) was implemented to evaluate
non-stationarity on the error maps. Non-stationarity
occurs when a global model such as ordinary least
square model fails to capture the spatial relationship
Table 2 Vegetation indices and formulae
Bands: (B blue, G green, R red, RE red-edge, NIR near-infrared)
Vegetation indices Formula References
1. Normalized difference vegetation index (NDVI) (NIR–R)/(NIR + R) Rouse etal. (1973)
2. Green leaf index (GLI) (2 * G–R–B)/(2 * G + R + B) Gobron etal. (2000)
3. Enhanced vegetation index (EVI) 2.5 * (NIR–R)/(NIR + 6 * R–7.5 * B + 1) Huete (1999)
4. Soil adjusted vegetation index (SAVI) (1 + L) * (NIR–R)/(NIR + R + L), L = 0.5 Huete (1988)
5. Modified soil adjusted vegetation index (MSAVI) 0.5 x (2 * NIR + 1–sqrt((2 * NIR + 1)2–8 * (NIR–R)) Qi etal. (1994)
6. Transformed chlorophyll absorption in reflec-
tance index (TCARI)
3 * ((RE–R)- 0.2 * (RE-G) * (RE/R)) Haboudane etal. (2002)
7. Visible atmospherically resistance Index (VARI) (G–R)/(G + R–B) Gitelson etal. (2002)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol:. (1234567890)
between variables (Narayan Mishra et al. 2021).
GWR is described as a spatial and statistical tech-
nique for modelling heterogeneous processes when
the relationship between the dependent and the inde-
pendent variables varies as a function of spatial loca-
tion on the map (Brunsdon etal. 1996).
where yi is the classification (error of good classifica-
tion) at location i; βi0 is the intercept variable at loca-
tion i; βik: represents the local regression coefficient
for variable k at location i. and Xik the independent
variable k at location i and εj represents the residual
error at location i.
To implement GWR, 500 points were randomly
distributed over the study area to extract pixel values
from the independent variables (satellite data) and
the dependent variable (error maps). The relevant
features were selected based on the Akaike Informa-
tion Criterion (AIC) in stepwise logistic regression,
and the residuals were evaluated for spatial autocor-
relation (Comber et al. 2020). No evidence of spa-
tial autocorrelation would suggest a stationary map,
therefore a GWR model is not applicable. However,
if there is evidence of spatial autocorrelation, a Mul-
tiscale-GWR model (MS-GWR) would be applied
to determine the spatial scale at which each predic-
tor varied. If the residuals from the MS-GWR model
are spatially autocorrelated, then a specific variant of
GWR (MS-GWR or Mixed-GWR) will be applied
to calculate the coefficient of the variable across the
map (Comber etal. 2020). This analysis was done in
R using the packages Raster, GWmodel and ggplot2.
Image classification
Texture parameters and vegetation indices generated
from S1 and S2 data, respectively, were used to train
a RF model for the classification of four AFS in the
region. The classification suggested a high perfor-
mance of OA = 0.94 (Kappa = 0.92), with the small-
est producer’s accuracy for cocoa (PA = 0.88 and
UA = 0.91). The vegetation over the protected area
was classified mainly as cocoa and rubber planta-
tions (Fig. 2). When looking at the field plots, the
𝛃𝐢𝐤𝐗𝐢𝐤 +𝛆
classification showed some confusion between fea-
tures classes. In Fig.1a, rubber and farm pixels were
falsely classified as cocoa. Moreover, pixels cor-
responding to cocoa are falsely classified as palm
or rubber (Fig. 1c, d respectively). As suggested by
the AUC, cocoa and rubber plantations were more
difficult to distinguish (AUC = 0.95) compared to
farm and palm plantation (AUC = 0.98 and 0.99
Spatial error assessment
To model the spatial distribution of the error, an
entropy map was generated based on the probabil-
ity of each feature class, where different thresholds
were used to produce the error maps (Fig. 3). The
figure indicated that higher entropy threshold values
(thr = 0.5 for example) resulted in smaller area of the
classification error (38.5% compared to 58.3% for
thr = 0.2), which could result in underestimation of
the classification error for larger threshold values or
overestimation for smaller values. For each threshold,
the protected areas were labelled as error, with some
points corresponding to an AFS within their bound-
ary. The area of each AFS at different thresholds
was evaluated against the reference, which showed
that there was an overestimation of the error with
thr = 0.2 and thr = 0.3, corresponding to ca. 1 865.5
km2 and 645.3 km2, respectively. In addition, thr = 0.4
and thr = 0.5 showed an underestimation of the error
(376.3 km2 and 1 141.5 km2, respectively). The dif-
ference in the cocoa area (134.5 km2) between the
error map with thr = 0.3 and the reference error map
(based on field plots) was small compared to the other
The spatial distribution of the classification error
(reference error map) was used to remove errone-
ous pixels from the classified AFS map (Fig.4). The
result showed that the method was efficient to deal
with the classification error as illustrated by Fig.4c.
However, removal of the erroneous pixels also partly
removed classified pixels (Fig. 4d). The area sta-
tistics indicated that 72.8% (7 425 ha) of the area
is allocated to cocoa plantation, 15.3% to rubber (1
555ha), 7.9% to agriculture (804ha) and 4.09% for
to palm (417ha). Based on the results, there was evi-
dence of encroachment in the protected areas, where
the detected feature classes included cocoa (8.1%:
197ha), rubber (8.1%: 16ha) and farm (1ha).
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol.: (0123456789)
Geographically weighted regression (GWR)
The stepwise regression indicated that 12 features
were significant to model the distribution of the
classification error across the map (R2 = 0.78). The
residual analysis of the regression showed that there
was no evidence of spatial autocorrelation (p = 0.66).
However, there was evidence of spatial autocorre-
lation in the scenario with the overestimated error
(thr = 0.2), as indicated in Table3.
Cocoa farms are established in small scale planta-
tions, which resulted in a salt and pepper effects
that are difficult to interpret. Therefore, a smooth
filter was used to correct this effect. The analysis
detected 88% of the cocoa in the region, of which
91% were accurately classified. This performance
was superior to a comparable study in Côte d’Ivoire
(PA = 82.9 and UA = 62.2) carried out by Abu etal.
(2021). This could be explained by the size of the
study area: Abu etal. (2021) worked on the entire
country level, whereas our study area was limited
to the cocoa production zone. In central Africa,
an accuracy of 89.76% was obtained in Cameroon
(Numbisi etal. 2019). This study used higher reso-
lution images to delineate cocoa agroforestry from
secondary forest. In Côte d’Ivoire, cocoa planta-
tions are established as full sun (with few to no
forest trees), thus their structure looks like rubber
plantations, whereas in Cameroon, cocoa is grown
under trees and are infeasible to be separated from
secondary forest. This difference in context justifies
the difference in methodology and in the results. An
accuracy of 97.6% was obtained for the classifica-
tion of a cocoa plantation in Brazil using S1 & S2
and deep learning (Filella 2018). The study reported
a high level of prediction error in the final map,
leading to the conclusion that the spatial and spec-
tral resolutions of S2 is inappropriate to map cocoa
agroforests. In fact, classifying cocoa agroforests
requires high resolution imagery, which is not the
Fig. 2 Classification of agroforestry systems in the south of Côte d’Ivoire. Zoom on the field plots is presented in the right and left
side of the classification. White dots on the image represented the areas that have been masked out
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol:. (1234567890)
case for full sun cocoa, for which S1 and S2 have
been successfully used (Abu etal. 2021).
To reduce the error in the classification, Shan-
non entropy was added to the classification work-
flow to model the classification error. Here, entropy
refers to the likelihood of a pixel to be a particular
feature class. It is used in classification workflow as
a weighted uncertainty metric to compare different
classification models, or to compute the uncertainty
of classification at pixel level (Numbisi et al. 2019;
Roodposhti etal. 2019). A threshold value of 0.4 was
used by Numbissi etal. (2019) for the classification
of cocoa agroforestry in central Africa, where optimal
threshold value was found between 0.3 and 0.4 with
0.4 slightly underestimating the map error. Rood-
poshti etal. (2019) found that the optimal threshold
value was 0.2 for homogenous farming landscapes
like corn, grass, and soybean. Thus, depending on
final use of the map, a threshold of 0.3 or 0.4 could
be used. The value at 0.3 would be a rigorous choice
to limit probability of false detection for monitoring
cocoa extension around protected areas. The spectral
confusion between cocoa, rubber, palm, and forest
was used to detect encroachment. Consequently, high
entropy values were associated with pixels in the pro-
tected areas, so that only footprints of encroachment
could be detected.
The expansion of cocoa plantation is acknowl-
edged as the main driver of deforestation in Côte
d’Ivoire, and evidence of encroachment in pro-
tected areas has been reported (Abu etal. 2021). We
detected encroachment of cocoa in the Taï National
Park. The distribution of the plantations within the
park were located around the limit of the study area
with a pattern that was like artefacts. This could be
because a composite image was used as input layer
since the use of single layers was infeasible due to
presence of clouds in the region throughout the year.
Therefore, validation would only be possible via
fieldwork, interpretation of aerial photography or
Fig. 3 Error maps for different entropy thresholds of agroforestry systems in south Côte d’Ivoire
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol.: (0123456789)
UAV data. Evaluation of the spatial autocorrelation
on the image was necessary to identify local drivers
of the error on the map. The study reported that the
spatial distribution of the classification error was not
related to any local condition. However, signs of local
conditions associated with errors revealed when the
error was highly overestimated (thr = 0.2).
The proposed methodology provides a frame-
work for more reliable and accurate classification,
particularly for AFS established within complex
landscapes. It shows a pathway to model the spa-
tial distribution of the classification error on the
map and improve the classification by showing pix-
els which are accurately classified. This is neces-
sary for the monitoring of cocoa expansion around
protected areas to limit the risk of false detection.
In addition, an accurate map is required to deliver
accurate estimation of carbon stocks in each land-
scape for carbon estimation studies.
Fig. 4 Improved classification of agroforestry systems in southern Côte d’Ivoire using entropy value threshold derived from field
Table 3 Summary of the GWR
Threshold values Number of
Moran’ I (p value) Conclusion GWR Moran’I (p
Appropriate GWR
Thr = 0.2 13 0.8 6.09 E−6 Evidence of Spatial
0.005 MS-GWR
Thr = 0.3 12 0.79 0.33 No spatial autocor-
Not applicable
Thr = 0.4 13 0.77 0.61
Thr = 0.5 11 0.77 0.63
Thr = field plot 12 0.78 0.66
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol:. (1234567890)
This study aimed at modelling the spatial distribu-
tion of the classification error in cocoa AFS in Côte
d’Ivoire. The proposed model integrates the Shan-
non entropy in the classification framework to cap-
ture the classification error on the map. It was found
that the entropy threshold value should be selected
between 0.3 and 0.4 for the classification of full sun
cocoa or cocoa agroforestry, respectively. Addition-
ally, the model was able to detect encroachment in
protected areas and could be used for monitoring
the expansion of cocoa. This proposed methodol-
ogy could also be used for image classification in
presence of classes that are difficult to be separated.
S1 and S2 could be used to the near-real time moni-
toring of cocoa plantation in Côte d’Ivoire, and
the classification on the map could be corrected
using other techniques like spectral-spatial super
resolution image reconstitution, in combination
with advanced classification approaches like deep
Acknowledgements Authors acknowledge the support from
the German Federal Ministry for Education and Research
(BMBF) via the project carrier at the German Aerospace
Center (DLR Projektträger)through the research project:WAS-
CAL-DE-Coop (FKZ: 01LG1808A).
Funding Open Access funding enabled and organized by
Projekt DEAL.
Open Access This article is licensed under a Creative Com-
mons Attribution 4.0 International License, which permits
use, sharing, adaptation, distribution and reproduction in any
medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Crea-
tive Commons licence, and indicate if changes were made. The
images or other third party material in this article are included
in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not
included in the article’s Creative Commons licence and your
intended use is not permitted by statutory regulation or exceeds
the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this licence, visit
http:// creat iveco mmons. org/ licen ses/ by/4. 0/.
Abu I, Szantoi Z, Brink A, Robuchon M, Thiel M (2021)
Detecting cocoa plantations in Côte d’Ivoire and Ghana
and their implications on protected areas. Ecol Ind. https://
doi. org/ 10. 1016/J. ECOLI ND. 2021. 107863
Aguilar R, Zurita-Milla R, Izquierdo-Verdiguier E, de By RA
(2018) A cloud-based multi-temporal ensemble classifier
to map smallholder farming systems. Remote Sens 10:5.
https:// doi. org/ 10. 3390/ rs100 50729
Dean A (2019) Deforestation and Climate Change | Climate
Council. Retrieved Sep 1 2021 from https:// www. clima
tecou ncil. org. au/ defor estat ion/
Ashiagbor G, Forkuo EK, Asante WA, Acheampong E, Quaye-
Ballard JA, Boamah P, Foli E (2020) Pixel-based and
object-oriented approaches in segregating cocoa from for-
est in the Juabeso-Bia landscape of Ghana. Remote Sens
Appl Soc Environ 19:100349. https:// doi. org/ 10. 1016/J.
RSASE. 2020. 100349
Bitty AE, Gonedele SB, Koffi Bene JC, Kouass PQ, Mcgraw
WS (2015) Tropical conservation science | ISSN 1940–
0829 | Cite this paper
as. Mongabay.Com Open Access J-Trop Conserv Sci
8(1), 95–113. http:// creat iveco mmons. org/ licen ses/ by/3.
0/ us/. The: 95- 113. Avail ableo nline: www. tropi calco nserv
ation scien ce. org
Boeckx P, Bauters M, Dewettinck K (2020) Poverty and cli-
mate change challenges for sustainable intensification of
cocoa systems. Curr Opin Environ Sustain 47:106–111.
https:// doi. org/ 10. 1016/J. COSUST. 2020. 10. 012
Brunsdon C, Fotheringham AS, Charlton ME (1996) Geo-
graphically weighted regression: a method for explor-
ing spatial nonstationarity. Geogr Anal 28(4):281–298.
https:// doi. org/ 10. 1111/J. 1538- 4632. 1996. TB009 36.X
Busquet M, Bosma N, Hummels H (2021) A multidimen-
sional perspective on child labor in the value chain: the
case of the cocoa value chain in West Africa. World Dev
146:105601. https:// doi. org/ 10. 1016/J. WORLD DEV.
2021. 105601
Comber A, Brunsdon C, Charlton M, Dong G, Harris R,
Lu B, Harris P (2020) The GWR route map: a guide to
the informed application of Geographically Weighted
Regression, 1–34. http:// arxiv. org/ abs/ 2004. 06070
Copernicus (2022) User guides—Sentinel-1 SAR—acqui-
sition modes—Sentinel Online—Sentinel Online.
Accessed 3 June 2022 https:// senti nels. coper nicus.
eu/ web/ senti nel/ user- guides/ senti nel-1- sar/ acqui
sition- modes
Earth M (2021) (4) Mighty Earth ? (@StandMighty)/Twitter.
Accessed 15 June 2022 https:// twitt er. com/ Stand Mighty?
ref_ src= twsrc% 5Etfw% 7Ctwc amp% 5Etwe etemb ed%
7Ctwt erm% 5E136 24512 52659 986433% 7Ctwgr% 5E%
7Ctwc on% 5Es1_ & ref_ url= https% 3A% 2F% 2Fwww. agenc
eecofi n. com% 2Fcac ao% 2F1902- 85365- la- cote-d- ivoire- a-
perdu- 47- 000- hecta res- de- forets- au- profit-d
ECOWAS (2016) Import and export | Economic Community
of West African States(ECOWAS). Accessed 31 August
2021 https:// www. ecowas. int/ doing- busin ess- in- ecowas/
import- and- export/
Filella GB (2018) Cocoa segmentation in Satellite images with
deep learning. ETH Zurich, Zürich
GIZ (2020) Strengthening governance and sustainable manage-
ment of natural resources in the Comoé and Taï regions.
Accessed 1 September 2021 https:// www. giz. de/ en/ world
wide/ 30013. html
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Agroforest Syst (2023) 97:109–119
1 3
Vol.: (0123456789)
GVR (2019) Cocoa Beans Market Size, Analysis | Global
Industry Report, 2019–2025. Accessed 31 August 2021
https:// www. grand viewr esear ch. com/ indus try- analy sis/
cocoa- beans- market
Hall-Beyer M (2017) GLCM texture: a tutorial v.3.0 March
2017. http:// www. ucalg ary. ca/ UofC/ nasdev/ mhall bey/
resea rch. htm
ILO (2017) Global estimates of child labour: results and
trends, 2012–2016 International Labour Office (ILO),
Geneva, 2017 ISBN: 978–92–2–130152–3 (print) ISBN:
International Anti-slavery (2004) The cocoa industry in West
Africa: a history of exploitation. Thomas Clarkson House.
https:// doi. org/ 10. 1038/ 16430 6b0
Mishra VN, Prasad R, Rai PK, Vishwakarma AK, Arora A
(2019) Performance evaluation of textural features in
improving land use/land cover classification accuracy of
heterogeneous landscape using multi-sensor remote sens-
ing data. Earth Sci Inf 12(1):71–86. https:// doi. org/ 10.
1007/ S12145- 018- 0369-Z/ TABLES/ 10
Narayan Mishra V, Kumar V, Prasad R, Punia M (2021)
Geographically weighted method integrated with logis-
tic regression for analyzing spatially varying accu-
racy measures of remote sensing image classification.
J Indian Soc Remote Sens. https:// doi. org/ 10. 1007/
s12524- 020- 01286-2
Numbisi FN, Van Coillie FMB, De Wulf R (2019) Delinea-
tion of cocoa agroforests using multiseason sentinel-1
SAR images: a low grey level range reduces uncertainties
in GLCM texture-based mapping. ISPRS Int J Geo-Inf.
https:// doi. org/ 10. 3390/ ijgi8 040179
ProLand U (2020) The role of governments in making certi-
fication effective: a synthesis of the evidence and a case
study of cocoa in Côte d’Ivoire. Retrieved from www. tetra
techi ntdev. com
Pye-Smith K, Toledano (2016) A brighter future for cocoa
farmers: how the vision for change programme is raising
productivity and improving rural livelihoods. ICRAF
Trees for Change no. 13. Nairobi: World Agroforestry
Rinkesh (2020) Various causes, effects and surprising solutions
to soil degradation—conserve energy future. Accessed
1 September 2021 https:// www. conse rve- energy- future.
com/ causes- effec ts- solut ions- soil- degra dation. php
Roodposhti MS, Aryal J, Lucieer A, Bryan BA (2019) Uncer-
tainty assessment of hyperspectral image classification:
deep learning vs. random forest. Entropy 21(1):1–15.
https:// doi. org/ 10. 3390/ e2101 0078
Ruf F, Schroth G, Doffangui K (2014) Climate change, cocoa
migrations and deforestation in West Africa: what does
the past tell us about the future? Sustain Sci 10(1):101–
111. https:// doi. org/ 10. 1007/ S11625- 014- 0282-4
Sabas BYS, Danmo KG, Madeleine KAT, Jan B (2020) Cocoa
production and forest dynamics in ivory coast from 1985
to 2019. Land 9(12):524. https:// doi. org/ 10. 3390/ LAND9
Shahbandeh M (2021) Cocoa production by country 2019/2020
| Statista. Accessed 28 May 2021 https:// www. stati sta.
com/ stati stics/ 263855/ cocoa- bean- produ ction- world
wide- by- region/
Yao B, Sabas S, Danmo KG, Akoua K, Madeleine T, Jan B, Ci
KATM (2020) Cocoa production and forest dynamics in
ivory coast from 1985 to 2019. Land. https:// doi. org/ 10.
3390/ land9 120524
Publisher’s Note Springer Nature remains neutral with regard
to jurisdictional claims in published maps and institutional
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”),
for small-scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are
maintained. By accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use
(“Terms”). For these purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or
a personal subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or
a personal subscription (to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the
Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data
internally within ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking,
analysis and reporting. We will not otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of
companies unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that
Users may not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to
circumvent access control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil
liability, or is otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by
Springer Nature in writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer
Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates
revenue, royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain.
Springer Nature journal content cannot be used for inter-library loans and librarians may not upload Springer Nature journal
content on a large scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any
information or content on this website and may remove it or features or functionality at our sole discretion, at any time with or
without notice. Springer Nature may revoke this licence to you at any time and remove access to any copies of the Springer Nature
journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express
or implied with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or
warranties imposed by law, including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be
licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other
manner not expressly permitted by these Terms, please contact Springer Nature at
... Moreover, the applicability of RS in AFS is challenging in West Africa due to the spectral similarities among AFS categories. In a recent study, Kanmegne Tamga et al. reported that RS data were unable to effectively separate AFS in West Africa, returning a high level of mixed pixels across the map (more than one AFS in a pixel) [23]. Therefore, RS-based analysis resulted in maps and estimations associated with higher level of uncertainties that should be addressed. ...
... The model with the best score (higher R 2 and lower RMSE) was used to generate the AGB map of the study area. In addition, the spatial distribution of the prediction uncertainties was modelled [23]. For that, the field data reference AGB estimation was used to assess the AGB prediction, and the difference was used to generate the RMSE for each plot. ...
... The AGB map in the Guineo-Congolian region was generated ( Figure 11A) with the corresponding uncertainty map (Figure 11B), where the RMSE value of the model (3.82 was used as threshold to classify the error [23]. The error level was high above the thresh old, while the error level below the threshold was low. ...
Full-text available
Agroforestry systems (AFS) offer viable solutions for climate change because of the above-ground biomass (AGB) that is maintained by the tree component. Therefore, spatially explicit estimation of their AGB is crucial for reporting emission reduction efforts, which can be enabled using remote sensing (RS) data and methods. However, multiple factors including the spatial distributions within the AFS, their structure, their composition, and their variable extents hinder an accurate RS-assisted estimation of the AGB across AFS. The aim of this study is to (i) evaluate the potential of spaceborne optical, SAR and LiDAR data for AGB estimations in AFS and (ii) estimate the AGB of different AFS in various climatic regions. The study was carried out in three climatic regions covering Côte d'Ivoire and Burkina Faso. Two AGB reference data sources were assessed: (i) AGB estimations derived from field measurements using allometric equations and (ii) AGB predictions from the GEDI level 4A (L4A) product. Vegetation indices and texture parameters were generated from optical (Sentinel-2) and SAR data (Sentinel-1 and ALOS-2) respectively and were used as predictors. Machine learning regression models were trained and evaluated by means of the coefficient of determination (R 2) and the RMSE. It was found that the prediction error was reduced by 31.2% after the stratification based on the climatic conditions. For the AGB prediction, the combination of random forest algorithm and Sentinel-1 and-2 data returned the best score. The GEDI L4A product was applicable only in the Guineo-Congolian region, but the prediction error was approx. nine times higher than the ground truth. Moreover, the AGB level varied across AFS including cocoa (7.51 ± 0.6 Mg ha −1) and rubber (7.33 ± 0.33 Mg ha −1) in the Guineo-Congolian region, cashew (13.78 ± 0.98 Mg ha −1) and mango (12.82 ± 0.65 Mg ha −1) in the Guinean region. The AFS farms in the Sudanian region showed the highest AGB level (6.59 to 82.11 Mg ha −1). AGB in an AFS was mainly determined by the diameter (R 2 = 0.45), the height (R 2 = 0.13) and the tree density (R 2 = 0.10). Nevertheless, RS-based estimation of AGB remain challenging because of the spectral similarities between AFS. Therefore, spatial assessment of the prediction uncertainties should complement AGB maps in AFS.
... This agroforestry and intercropping system aim to raise the standard of living in the village communities surrounding the forest by giving smallholder farmers or rural residents the chance to cultivate food crops. In addition, villagers close to the forest are required to actively participate in efforts to preserve and stop forest and land destruction in this way (Mayrowani and Ashari 2011;Tamga et al. 2022). ...
Full-text available
Natsir M, Ulya Z, Fitriani R. 2022. Mangrove forest utilization policies reconceptualized with a view to improving the regional economy in Aceh Tamiang District, Indonesia. Biodiversitas 23: 6570-6578. One of the Aceh's (Indonesia) districts with mangrove forests is Aceh Tamiang District. Manyak Payeud, Bendahara, Seureuwey, and Banda Mulia are the four sub-district in the Aceh Tamiang region that have the most potential tourist sites. However, due to the widespread illegal logging of mangroves carried out by the community to suit their daily requirements, the preservation of mangrove forests in the four sub-districts above is seriously damaged. As a result of illegal logging, mangroves are becoming less common in Aceh Tamiang District. Therefore, a unique policy, which can incorporate both community needs and economic development, is needed to manage the mangroves of Aceh Taminga District. The expectation is that mangrove management will enable Aceh Tamiang District to develop tourist attractions in this study. The normative juridical research methodology with SWOT analysis was applied in this study to understand the potential of mangrove management in improving the regional Economy in Aceh Tamiang District. The findings of this study are based on a SWOT analysis of four subdistricts in Aceh Tamiang District. The result suggests that the notion of forestry policy can support both local community interests and regional economic development. Additionally, it can be carried out through the Village Qanun to raise community awareness of the need to maintain and manage mangroves to enable the implementation of regional plans. To harmonize the understanding of the significance of mangroves as one of the regional tourist destinations, institutional coordination between the Aceh Tamiang district government and the provincial and central governments can also be used to realize the development of Aceh Tamiang tourism policies.
Full-text available
Due to little adoption of the agroforestry practices, land degradation has become a serious pressing problem in various parts of the world in general and in study area in particular. Therefore, these studies aim to assess the determents of the agroforestry practices in the study area. Therefore, based on multi-stage sampling techniques 184 randomly selected sample households’ heads was determined. Descriptive and inferential statistics such as T-test, chi-square tests, and binary logit model was used to compare the mean difference between adopters and non-adopters households. The finding the study shows Senegal (L. Britton), home garden, and boundary planting are the most common agroforestry practices in the study area. The mean annual income for adopter farm households' heads was 1148743.00 birr, and for non-adopters, was 138675.00 birr. This implies that agroforestry practices make a significant contribution to the adopter's income. However, lack of farmland (27.7%), pests (16.3%), and low market access (15.2%) are major constraints that determine the adoption of the agroforestry practices in the study area. The results of the binary logistic model specify that age, farm size, and distance negatively affect the adoption of agroforestry practices in the study area. While perception were positively and significantly influence adoption of agroforestry practices in the study area, Therefore, the governmental & non-governmental organizations should have to develop new land policies to remove barriers to land access, tree tenure & an array of exotic tree species should be required to resist pests and drought conditions in study area.
Full-text available
Côte d'Ivoire and Ghana are the largest producers of cocoa in the world. In recent decades the cultivation of this crop has led to the loss of vast tracts of forest areas in both countries. Efficient and accurate methods for remotely identifying cocoa plantations are essential to the implementation of sustainable cocoa practices and for the periodic and effective monitoring of forests. In this study, a method for cocoa plantation identification was developed based on a multi-temporal stack of Sentinel-1 and Sentinel-2 images and a multi-feature Random Forest (RF) algorithm. The Normalized Difference Vegetation Index (NDVI) and second-order texture features were assessed for their importance in an RF classification, and their optimal combination was used as input variables for the RF model to identify cocoa plantations in both countries. The RF model-based cocoa map achieved 82.89% producer's and 62.22% user's accuracy, detecting 3.69 million hectares (Mha) and 2.15 Mha of cocoa plantations for Côte d'Ivoire and Ghana, respectively. The results demonstrate that a combination of an RF model and multi-feature classification can distinguish cocoa plantations from other land cover/use, effectively reducing feature dimensions and improving classification efficiency. The results also highlight that cocoa farms largely encroach into protected areas (PAs), as 20% of the detected cocoa plantation area is located in PAs and almost 70% of the PAs in the study area house cocoa plantations.
Full-text available
The cocoa economy of Ivory Coast started in the eastern part of the country in the 1970s and spread to the central-western and then south-western regions. For nearly a decade, it has been in the West of Ivory Coast with a population increase caused by large waves of migration. This study aims to determine different factors explaining dynamics of the cocoa economy from the East to West of Ivory Coast. The method adopted consisted of processing Landsat images from 1985–2018 and an individual survey of 278 heads of households. The results obtained showed that the development of the cocoa economy led forest cover degradation with a total loss estimated at 60.80%, 46.39%, 20.76% and 51.18% of forest area in the East, Centre-West, South-West and West, respectively. The creation of new cocoa farms in the West of Ivory Coast is governed by non-native people (51.13%) settled between 2010 and 2018. About 41% of these producers come mainly from the Centre-West (25%) and the South-West (16%). In addition, 29% of producers come from the West of Ivory Coast. Despite the abiotic characteristics being considered unfavourable, the west of Ivory Coast is in the process of becoming the country’s new zone of high cocoa production.
Full-text available
Geographically Weighted Regression (GWR) is increasingly used in spatial analyses of social and environmental data. It allows spatial heterogeneities in processes and relationships to be investigated through a series of local regression models rather than a global one. Standard GWR assumes that the relationships between the response and predictor variables operate at the same spatial scale, which is frequently not the case. To address this, several GWR variants have been proposed. This paper describes a route map to inform the choice of whether to use a GWR model or not, and if so which of three core variants to apply: a standard GWR, a mixed GWR or a multiscale GWR (MS-GWR). The route map comprises primary steps: a basic linear regression, a MS-GWR, and investigations of the results of these. The paper provides guidance for deciding whether to use a GWR approach, and if so for determining the appropriate GWR variant. It describes the importance of investigating a number of secondary issues at global and local scales including collinearity, the influence of outliers, and dependent error terms. Code and data for the case study used to illustrate the route map are provided, and further considerations are described in an extensive Appendix.
Full-text available
Delineating the cropping area of cocoa agroforests is a major challenge in quantifying the contribution of land use expansion to tropical deforestation. Discriminating cocoa agroforests from tropical transition forests using multispectral optical images is difficult due to the similarity of the spectral characteristics of their canopies. Moreover, the frequent cloud cover in the tropics greatly impedes optical sensors. This study evaluated the potential of multiseason Sentinel-1 C-band synthetic aperture radar (SAR) imagery to discriminate cocoa agroforests from transition forests in a heterogeneous landscape in central Cameroon. We used an ensemble classifier, Random Forest (RF), to average the SAR image texture features of a grey level co-occurrence matrix (GLCM) across seasons. We then compared the classification performance with results from RapidEye optical data. Moreover, we assessed the performance of GLCM texture feature extraction at four different grey levels of quantization: 32 bits, 8 bits, 6 bits, and 4 bits. The classification’s overall accuracy (OA) from texture-based maps outperformed that from an optical image. The highest OA (88.8%) was recorded at the 6 bits grey level. This quantization level, in comparison to the initial 32 bits in the SAR images, reduced the class prediction error by 2.9%. The texture-based classification achieved an acceptable accuracy and revealed that cocoa agroforests have considerably fragmented the remnant transition forest patches. The Shannon entropy (H) or uncertainty provided a reliable validation of the class predictions and enabled inferences about discriminating inherently heterogeneous vegetation categories.
Full-text available
Uncertainty assessment techniques have been extensively applied as an estimate of accuracy to compensate for weaknesses with traditional approaches. Traditional approaches to mapping accuracy assessment have been based on a confusion matrix, and hence are not only dependent on the availability of test data but also incapable of capturing the spatial variation in classification error. Here, we apply and compare two uncertainty assessment techniques that do not rely on test data availability and enable the spatial characterisation of classification accuracy before the validation phase, promoting the assessment of error propagation within the classified imagery products. We compared the performance of emerging deep neural network (DNN) with the popular random forest (RF) technique. Uncertainty assessment was implemented by calculating the Shannon entropy of class probabilities predicted by DNN and RF for every pixel. The classification uncertainties of DNN and RF were quantified for two different hyperspectral image datasets—Salinas and Indian Pines. We then compared the uncertainty against the classification accuracy of the techniques represented by a modified root mean square error (RMSE). The results indicate that considering modified RMSE values for various sample sizes of both datasets, the derived entropy based on the DNN algorithm is a better estimate of classification accuracy and hence provides a superior uncertainty estimate at the pixel level.
Full-text available
Smallholder farmers cultivate more than 80% of the cropland area available in Africa. The intrinsic characteristics of such farms include complex crop-planting patterns, and small fields that are vaguely delineated. These characteristics pose challenges to mapping crops and fields from space. In this study, we evaluate the use of a cloud-based multi-temporal ensemble classifier to map smallholder farming systems in a case study for southern Mali. The ensemble combines a selection of spatial and spectral features derived from multi-spectral Worldview-2 images, field data, and five machine learning classifiers to produce a map of the most prevalent crops in our study area. Different ensemble sizes were evaluated using two combination rules, namely majority voting and weighted majority voting. Both strategies outperform any of the tested single classifiers. The ensemble based on the weighted majority voting strategy obtained the higher overall accuracy (75.9%). This means an accuracy improvement of 4.65% in comparison with the average overall accuracy of the best individual classifier tested in this study. The maximum ensemble accuracy is reached with 75 classifiers in the ensemble. This indicates that the addition of more classifiers does not help to continuously improve classification results. Our results demonstrate the potential of ensemble classifiers to map crops grown by West African smallholders. The use of ensembles demands high computational capability, but the increasing availability of cloud computing solutions allows their efficient implementation and even opens the door to the data processing needs of local organizations.
The International Labor Organization (ILO) states that most agricultural work carried out by children occurs within the family unit, is generally unpaid and often hazardous in its nature and/or in the circumstances in which it is carried out. At the same time, some scholars nuance this view by positing that children who work in agriculture in the spheres of their own household are not necessarily exploited. Making progress in addressing (worst forms of) child labor by value chain actors necessitates unpacking the complex dynamics, context and interlinkages that connect firms and farms at the local community level. This study responds to this call by proposing a new multidimensional perspective on child labor based on comparing and contrasting Global Value Chain (GVC) literature and the Sustainable Livelihood Approach (SLA). Adopting such a perspective allows for an explanation of both vertical dynamics, including global inter-firm linkages and power distribution, as well as horizontal dynamics, such as local norms and values, access to capitals and livelihood trajectories that contribute to the occurrence of child labor. This framework is illustrated by a case study on child labor in the cocoa value chain in Ghana and Côte d’Ivoire, based on information obtained from a variety of sources, including 38 key informant interviews, 12 focus group discussions and structural observations. This study shows that children are not only factors of production, but are socially embedded in family structures and local communities. Children participate in a wider range of rural and agricultural activities as part of rural upbringing and learning a livelihood, in which not only harms but also benefits can occur. These findings advance the discussion by moving away from a dichotomy on child labor as a good or bad practice and putting the development opportunities of children center stage.
The accuracy of thematic information extracted from remote sensing image is assessed recurrently using the confusion matrix method. But the accuracies have been criticized as a consequence of its aspatial nature. The work presented here describes a geographically weighted method combined with logistic regression for producing and visualizing the spatially distributed accuracy measures across the landscape. The outcomes compare the standard confusion matrix-based accuracy measures with those that have been permitted to differ locally. Furthermore, statistical parameters, i.e. Akaike information criterion, adjusted squared correlation coefficient (R2) and residual sum of squares (RSS) were employed to compare the performance of geographically weighted logistic regression (GWLR) with global ordinary least square regression technique. The GWLR technique was found to provide more reliable performance in estimating spatially varying accuracy measures. The results demonstrated that the geographically weighted approach offers additional and valuable insights for examining spatial variation in the context of landscape mapping accuracy.
Sustainable intensification of cocoa systems should embrace, among others, poverty alleviation and climate change policies. Using data from 710 households in Ghana, we showed that only 33 and 15% of cocoa producing families acquire sufficient income from cocoa to reach ‘living income’ according to World Bank assumptions with respectively actual yields with a farmgate price of 3435 $ t⁻¹ and double yields with a farmgate price of 1150 $ t⁻¹. Besides price and yield incentives, agroforestry for climate change adaptation and preservation of other ecosystem services (e.g. greenhouse gas mitigation) should be promoted, while recognizing that farmer’s willingness to invest in tree planting links with informal rights for trees and customary tenure systems. Finally, extension services and communal measures that mediate overall wellbeing should be complementary measures. In conclusion, we advocate for a interdisciplinary science-based approach that respects specific contextual socio-political systems to intensify cocoa production.
A wall-to-wall Earth observation (EO) data is required, as recommended by the Intergovernmental Panel on Climate Change, private sector organizations and major development partners, to allow for the implementation of forest monitoring commitments, and also to monitor commodity-led deforestation. However, a major limitation associated with the use of optical EO data in the high forest zone of Ghana is the presence of persistent cloud cover and the spectral limitations of segregating agroforestry cocoa (AFC) from open canopy forest (OCF) cover. The aim of the study was to investigate the synergistic use of Sentinel-1 (S1) and Sentinel-2 (S2) EO data to produce a Land Use Land Cover map which shows AFC and OCF as different land cover classes. It was hypothesized that, a hybrid method of spectral, radar and image objects will accurately segregate the different cocoa system from forest and other land use classes. The research was conducted in the Bia-Juaboso REDD + Hotspot Intervention Area in the cocoa-forest mosaic landscape within the High forest zone of Ghana. The S1 and S2 datasets were freely acquired for the periods January to March 2018. The S1 datasets were pre-processed from backscatter intensity values to the VV and VH bands. The S2 datasets were corrected for atmospheric effects, and cloud pixels were masked and filled using a temporal gap-filling method. Six vegetation indices (VIs) were extracted, and the Multiresolution Segmentation algorithm was used to derive image objects (IOs). The S2 bands, the six VIs, the S1 VV + VH data, and the IOs were stacked into 3 different multi-layer image dataset denoted with D (i.e. D1 = S2 + VIs; D2 = S2 + VIs + S1; and D3 = S2+VIs + S1+IOs). The three datasets were classified using Random Forest and 1228 training points. Overall accuracy (OA) and kappa (k) were calculated for the classification outcome using 615 independent validation points. McNemar's test (χ) was used to access the statistically significant difference between D1, D2 and D3. The results of the study show that, D3 significantly improved the overall classification output (OA = 89.76%, k = 0.877) compared to D1 (OA = 79.02%, k = 0.748; χ = 5.56, p-value = 0.018) and D2 (OA = 80.49%, k = 0.765; χ = 5.50, p-value = 0.019). Combining spectral pixels with image objects increases overall classification accuracy and specifically, the accuracy of isolating AFC from OCF. This research is significant because it will provide an improved decision support to government-led monitoring and the private sector's commitment to halt cocoa-driven deforestation. Furthermore, and most importantly, the map shows agroforestry cocoa separated from monoculture cocoa, which provides a tremendous boost to monitoring landscape-level improvements associated with the promotion and adoption of agroforestry in cocoa landscapes, as a climate-smart practice and also for monitoring various off-reserve landscape forest restoration activities.