You are on page 1of 12

Climate Services 16 (2019) 100123

Contents lists available at ScienceDirect

Climate Services
journal homepage: www.elsevier.com/locate/cliser

Original research article

Recommendation domains to scale out climate change adaptation in cocoa T


production in Ghana
Christian Bunna, , Läderach Peterb, Amos Quayec, Muilerman Sanderd, Martin R.A. Noponene,

Mark Lundya
a
International Center for Tropical Agriculture, Cali, Colombia
b
International Center for Tropical Agriculture, Rome, Italy
c
Cocoa Research Institute of Ghana, Tafo, Ghana
d
Wageningen University & Research
e
Rainforest Alliance, London, United Kingdom

ARTICLE INFO ABSTRACT

Keywords: Climate change is threatening cocoa production in West Africa and guidance towards site-specific adaptation is
Cocoa required. We developed recommendation domains with common degree of impact requiring incremental, sys-
Recommendation domains temic or incremental adaptation effort to provide decision support for interventions to scale out adaptive
Climate impacts practices. We used Random Forests to divide the cocoa production belt into four zones with distinct climatic
Ghana
features under current and future climate conditions. To make model results actionable we used an expert
West Africa
validation approach. Cocoa experts evaluated and verified cocoa occurrence data for model input, prioritized
Decision support
climate and soil variables for modeling use and confirmed the validity of the distribution of climate zones.
Climate change will reduce the available area for cocoa production in the north due to a shift of the northern
transition to the Savanna zone. The current area for cocoa in central Ashanti will remain suitable but will face
uncertain climatic conditions. Areas in the Western, Central and Eastern regions will likely become hotter and
wetter. Each of these projected impacts will require site-specific adaptation strategies matching the degree of
impacts. Failing to prepare may subject rural communities to high risks of losing their livelihoods. Our re-
commendation domains can support impact specific preparation so that the majority of Ghana’s cocoa pro-
duction area may be sustained despite adverse climatic changes. Institutional and private actors can use our
work to scale out locally conceived interventions to alleviate impacts from drought, heat and erratic rainfall.

understandable decision support. Scalable climate smart tech-


Practical implications nologies exist, but robust processes for large scale interventions to
support farmers need to be developed.
In this study we developed recommendation domains to adapt
Cocoa provides a livelihood to millions of people in West Africa. cocoa production in Ghana to climate change. The re-
It is oftentimes the main cash source for smallholder families that commendation domains are based on a degree of impact concept
struggle daily to make a better life for their families and them- that supports stakeholders to design and roll out locally adequate
selves. Adapting these farmers to climate change is at the same adaptation measures. The domains are easier to interpret for
time trivial and highly complex: trivial, because most good stakeholders than previous impact assessments, but are equally
agricultural practices, correctly applied, can provide resilience methodologically robust. To develop the domains we combined a
against shocks and sustainability for long term changes. However, data driven method with a participatory validation process. The
developing individualized solutions for all cocoa households resulting maps were in line with state of the art climate impact
currently appears unfeasible and smallholders commonly don’t assessments but additionally relate to local perceptions of cocoa
have the means to develop forward looking solutions themselves. ecological zones and production systems. As a result, the domains
As a result, the risk of maladaptation or inefficient use of re- are framed within local knowledge and are more acceptable and
sources is high. The complexity is thus in the need to translate the accessible than conventional work.
crop- and site specificity of climatic changes into actionable and We differentiated zones by the degree of necessary adaptation


Corresponding author at: International Center for Tropical Agriculture, Km 17 Recta Cali-Palmira CP 763537, Apartado Aéreo 6713, Cali, Colombia.
E-mail address: cbunn@cgiar.org (C. Bunn).

https://doi.org/10.1016/j.cliser.2019.100123
Received 18 January 2019; Received in revised form 28 June 2019; Accepted 10 September 2019
Available online 20 September 2019
2405-8807/ © 2019 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/BY-NC-ND/4.0/).
C. Bunn, et al. Climate Services 16 (2019) 100123

effort. At low change adaptation can be incremental, using rela- currently occupies the entire climatic range in which it can be culti-
tively minor interventions such as integrated pest management, vated. For Ghana this assumption appears reasonable because of its
adjustment of shade levels and more generally, good agricultural widespread production in distinct ecological zones and a history of
practices. A higher degree of change requires a systemic adapta- devastating bush fires outside of the cocoa areas (Asante et al., 2017).
tion using more shade trees, potentially drip irrigation, or a
These studies projected reductions of suitability for cocoa mostly
change of cocoa variety. In some zones, the climate of the future
towards the West of Ghana. Possible increases of suitability were found
was found to be unsuitable at current cocoa locations and a
transformation to other crops may be considered or production for the Eastern region towards Lake Volta (Läderach et al., 2013). These
systems with a high degree of functional shade and income di- effects were likely caused by increases in dry season temperatures
versification. Additionally, we point out zones where global cli- (Schroth et al., 2016) or increased potential evapotranspiration during
mate models don’t exhibit the necessary degree of agreement to the short dry season (Läderach et al., 2013). Multiple regression ana-
support specific technological packages and emphasis should be lysis of time series from the Tafo cocoa region also suggested negative
put on increasing the resilience of producers. total production effects in Ghana (Anim-Kwapong and Frimpong,
Our recommendation domains can be used by public, private 2005). The latter study linked a regression model with results of global
and non-governmental actors to roll out climate adaptation in- climate models (GCMs) that projected reduced annual precipitation and
terventions at scale. Training activities can be differentiated by
higher mean temperatures to conclude that climate change will have
the likely degree of impact to provide assistance that responds to
negative impacts on cocoa production in Ghana.
local needs. For example, farmers can be made aware of di-
versification options where needed, or encouraged to intensify Thus, without adaptive action, a geographical shift of cocoa pro-
production where adequate. Other stakeholders may use our duction in Ghana caused by climate change appears likely and stake-
work to focus deforestation prevention where most needed. holders increasingly demand decision support to direct adaptation.
Certification agencies can adjust compliance requirements across Available work has limited value to develop adaptation strategies that
geographies in a feasible manner. Therefore our approach and account for local climatic limits to production (Sloan et al., 2019).
concept should be considered for other regions and crops as a Läderach et al. (2013) presented absolute changes of suitability scores
valuable step towards mainstreaming climate smart agriculture. between current and 2050s climate. Similar absolute changes, however,
may result in different degrees of impact depending on the current
suitability for production. Anim-Kwapong and Frimpong (2005) report
their regression model for a single district without accounting for dif-
1. Introduction ferences in local environmental limits so that their findings may not be
used to scale-out adaptation strategies. On the other spatial extreme,
Tools and processes need to be developed to close the gap between Schroth et al. (2016) did evaluate the potential degree of impact for the
climate change impact evaluation and implementation of adaptive ac- entire West African cocoa belt so that no region specific potential ha-
tion (Campbell et al., 2016). Recommendation domains may support zards could be described. It was therefore difficult to develop crop and
decision makers to assess the scaling potential of adaptation portfolios site-specific adaptation strategies for Ghana based on these previous
and direct their efficient out-scaling (Notenbaert et al., 2017). Such studies.
domains have to respond to the need of stakeholders for granular cli- The objective of our work was to differentiate recommendation
mate risk data to assess the extent and nature of risk for their sourcing domains with common degrees of projected impacts to inform stake-
regions (Sloan et al., 2019). Furthermore, adaptation to climate change holders in Ghanaian cocoa production. Because sustainable action is
will be specific to the local context in which the crop is produced and is most likely if risk assessments appeal to decision frameworks that are
not only driven by yield considerations (Eitzinger et al., 2018). Impact already in use (Weber, 2010), we included expert consultations during
based prioritization of adaptation approaches depends on the degree of the model validation process.. Our degree of impact concept follows the
impact: for low degrees of projected impacts commonly accepted sus- framework by Vermeulen et al. (2013) in that it is defined by the ne-
tainable agricultural practices may be sufficient adaptation strategies, cessary adaptation effort. At low climate impacts, the climate change
whereas high degrees of impacts may require a shift to alternative crops signal is indistinguishable from the observed variability and incre-
(Vermeulen et al., 2013). mental adjustments to the farming systems are sufficient for adaptation.
Cocoa production in Ghana has been projected to be challenged by Higher degrees of impact require systemic adaptation measures, or even
uncertain precipitation patterns and increasing temperatures (Anim- a transformational adaptation to other crops. We added a category of
Kwapong and Frimpong, 2005; Läderach et al., 2013; Schroth et al., systemic resilience where adaptation efforts face a high degree of un-
2016). Cocoa is produced all over the southern part of Ghana, whereas certainty from GCM disagreement. In addition, we sought to provide
in the Northern Savanna the lack of suitable precipitation prohibits its information about projected hazards from changes in mean climate to
production. Cocoa production is an integral part of the rural economy support anticipatory adaptation strategies.
in Ghana. An estimated 800,000 families cultivate and market cocoa on
1.45 million ha in plots of 0.4 to 4 ha (Asamoah and Baah, 2003). Gross 2. Methods and data
Domestic Product (GDP) generated from cocoa accounted for 16% of
the total GDP, and 68% of the GDP in the primary sector (GSS, 2017). Our description of recommendation domains for climate change
Negative impacts of climate change on cocoa would have repercussions adaptation and associated threats followed three steps: We first devel-
for the Ghanaian economy and especially for rural development. oped a multiclass suitability distribution model that distinguished dif-
Previous studies have investigated the potential impacts of climate ferent agro-climatic zones with distinct climatic properties. Second,
change on cocoa production for Côte d’Ivoire and Ghana (Läderach these zones were validated and aligned with conceptions of ecological
et al., 2013), West-Africa (Schroth et al., 2016), or for Tafo in Ghana zones in Ghana in expert workshops. Third, understanding the fate of
(Anim-Kwapong and Frimpong, 2005). The former studies used the each of the agro-climatic zones in future climate conditions is the basis
machine learning based species distribution modeling approach Maxent to enable the link to site specific adaptation strategies.
(Phillips et al., 2006) to model the impacts of climate change on the Agro-climatic zones for cocoa production in Ghana were defined
geographic distribution of suitable climate conditions for cocoa pro- using a combination of interpolated climate data from WorldClim
duction. Such methodological approaches assume that a location is (Hijmans et al., 2005) and a database of occurrences of cocoa produc-
suitable if its combination of climate variables is more similar to known tion. The Random Forest algorithm (RF) (Breiman, 2001) was used to
occurrences than to a usually random sample drawn from the general identify climate clusters within the occurrence data. RFs have the
environment. The implicit assumption of these studies was that cocoa capability of predicting several classes and may be used to extrapolate

2
C. Bunn, et al. Climate Services 16 (2019) 100123

their spatial distribution under current and future climate (Bunn et al., months with less than 100 mm precipitation (Table 1). A short dry
2015; Moriondo et al., 2013). Ghanaian experts of the Cocoa Research season induces uniform flowering, dry periods longer than three
Institute of Ghana (CRIG) evaluated and verified cocoa occurrence data months are considered harmful for the trees (Lass and Wood, 1985).
for model input, suggested soil variables for modeling use, aligned agro- For climate conditions in the period 2040 to 2069 (2050s) we used
climatic zone descriptions with their conceptions and confirmed the data from 19 global circulation models (GCMs) (Table 2) from the Fifth
validity of the resulting model. Assessment Report (AR5) of the Intergovernmental Panel on Climate
Change (IPCC) (Stocker et al., 2013). The data was downscaled using
2.1. Database of locations of cocoa the delta method (Ramirez and Jarvis, 2010). This approach computes
the difference between model outputs for current and future conditions,
We collected data of the current distribution of cocoa to define the interpolates the data to 0.5 arc-minutes resolution and the resulting
climates that are suitable for cultivation. The data came from four delta is added to the WorldClim current climate data. The bioclimatic
sources: variables can then be calculated for all future GCM outputs.
We chose the representative concentration pathway (RCP) 6.0, an
1. Läderach et al. (2013) compiled 294 occurrence locations from ex- intermediate scenario in which radiative forcing continues to increase
isting databases, maps, expert knowledge and field missions where until the end of the century (Fujino et al., 2006), because with current
coordinates were gathered using a Global Positioning System (GPS). emissions trajectories few scenarios appear possible that may result in
2. Schroth et al. (2016) determined cocoa growing areas in West Africa extreme radiative forcing pathways (RCP 8.5 and 2.6) (Riahi et al.,
as reported in the Atlas on Regional Integration in West Africa 2017) so that an intermediate RCP is an appropriate choice for adap-
(ECOWAS, 2007). From within these areas 558 occurrence locations tation guidance. For adaptation of a perennial crop mid-century is a
were sampled on a 0.3 degree grid. relevant time scale, at which time less difference can be found between
3. GPS coordinates of Rainforest Alliance project locations and training the potential impact of different mitigation pathways than between
sites. global climate models (Stocker et al., 2013) so that our choice covers
4. GPS coordinates of CRIG research project sites. most uncertainty from climate projections.

The raw database contained 2914 geo-referenced locations growing 2.3. Agro-climatic zones for cocoa production
cocoa in Ghana. Cocoa farms are often small and the occurrence data
may be highly clustered, which may affect model performance nega- We used Random Forests (RF) (Breiman, 2001) to estimate the re-
tively. Spatial filtering has been shown to be an effective method to lationship between climate, soil and cocoa distribution. RFs are ma-
reduce sampling bias (Kramer-Schadt et al., 2013). We reduced the chine-learning classifiers that are formed by ensembles of classification
database to unique pixel cells on a five arc-minute grid, which we call trees. We used the random Forest package (Liaw and Wiener, 2002) in
“occurrence pixels”. In addition, we removed as outliers sites for which the statistical software R (R Core Team, 2014) that implements the RF
one or more environmental variables exceeded 3.5 standard deviations approach. We used RFs in three distinct applications: (1) we used RF in
from the mean. From the resulting data we further removed several an unsupervised variation to produce a dissimilarity measure to group
sites at the margins of the area defined by Schroth et al. (2016) because occurrence locations into clusters with similar agro-climatic char-
they were unfeasible according to CRIG experts. 259 occurrence pixels acteristics, (2) using the resulting groups from (1) as response variable
remained (Fig. 1) that were confirmed correct by CRIG experts. we used RF to classify layers of climate and soil data of current and
future conditions, (3) finally, the RF classifier was used to evaluate
projected climatic changes in the different impact zones.
2.2. Climate data
2.4. Definition of climate types for cocoa production
For the current climate (1950–2000), we used the WorldClim data
sets at 5 and 0.5 arc-minute resolution (Hijmans et al., 2005). World- Agro-ecological zoning (AEZ) is a common tool to identify analo-
Clim provides data of monthly precipitation, mean monthly minimum gous environments, to prioritize interventions, and can be useful to
and maximum temperatures, and 19 bioclimatic variables derived from investigate impact of climate change (Bunn et al., 2015; Fischer et al.,
these data. For our analysis, we used 19 bioclimatic variables and 2002). In our case, we developed agro-climatic zones by identifying
complemented these with another derived variable of consecutive groups with common climate characteristics in the occurrence data
using RF clustering (Shi and Horvath, 2006). The unsupervised problem
of suitability type definition from ungrouped occurrence locations can
be turned into a supervised one by using artificial class labels to dis-
tinguish the original unlabeled data from synthetic data drawn from a
reference distribution (Shi and Horvath, 2006). RF produces a measure
of dissimilarity that when used as input for partitioning around medoid
(PAM) clustering instead of common Euclidean distances results in
biologically meaningful groupings (Shi and Horvath, 2006).
We used bioclimatic variables isothermality, mean maximum tem-
perature of the warmest month, mean temperature, precipitation sea-
sonality, precipitation of the driest and coldest quarters (Bioclim 3, 5, 9,
15, 17, and 19) for clustering. These variables were found to be phy-
siologically meaningful for cocoa production by Ghanaian experts. RF
prediction rules are useful even for correlated variables (Boulesteix
et al., 2012). The number of cluster centers for PAM was determined
using the Calinski and Harabasz (Caliński and Harabasz, 1974) index
and evaluation of a cluster dendrogram. Agro-climatic zone descrip-
tions were based on the differences of the group means from the grand
Fig. 1. Distribution of validated occurrence pixels. Each dot represents a unique mean for a subset of variables and their confidence intervals as calcu-
georeference of confirmed cocoa production. lated by R’s multcomp package (Bretz et al., 2010).

3
C. Bunn, et al. Climate Services 16 (2019) 100123

Table 1
Bioclimatic variables used and mean values at current occurrence locations in Ghana under current, and 2050s conditions; the values for 2050s were calculated as
mean impact across 19 GCMs.
Type Bioclimatic variable Description Current Mean 2050s Mean Unit

Temperature BIO 1 Annual Mean Temperature 26,2 27,7 °C


BIO 2 Mean Diurnal Range (Mean of monthly (max temp - min temp)) 9,1 8,6 °C
BIO 3 Isothermality (BIO2/BIO7) (*100) 72 70 –
BIO 4 Temperature Seasonality (standard deviation *100) 103,7 109,8
BIO 5 Max Temperature of Warmest Month 33,1 34,5 °C
BIO 6 Min Temperature of Coldest Month 20,6 22,3 °C
BIO 7 Temperature Annual Range (BIO5-BIO6) 12,5 12,2 °C
BIO 8 Mean Temperature of Wettest Quarter 26,5 27,6 °C
BIO 9 Mean Temperature of Driest Quarter 26,6 28,2 °C
BIO 10 Mean Temperature of Warmest Quarter 27,4 29,1 °C
BIO 11 Mean Temperature of Coldest Quarter 24,7 26,2 °C

Precipitation BIO 12 Annual Precipitation 1453 1476 mm


BIO 13 Precipitation of Wettest Month 234 235 mm
BIO 14 Precipitation of Driest Month 22 21 mm
BIO 15 Precipitation Seasonality (Coefficient of Variation) 53 55 –
BIO 16 Precipitation of Wettest Quarter 570 575 mm
BIO 17 Precipitation of Driest Quarter 117 113 mm
BIO 18 Precipitation of Warmest Quarter 335 329 mm
BIO 19 Precipitation of Coldest Quarter 361 382 mm
BIO 20 Number of Consecutive Months < 100 mm precipitation 3,63 3,63 –

Table 2 the occurrence pixels to exclude unfeasible locations (Barbet-Massin


List of global climate models used to model future climate conditions. et al., 2012). We used all 20 bioclimatic variables as independent
bcc_csm1_1 gfdl_cm3 ipsl_cm5a_lr mri_cgcm3 variables and the SRC variable as presented above.
bcc_csm1_1_m gfdl_esm2g miroc_esm ncar_ccsm4 For each RF model, we grew 1000 decision trees with seven vari-
cesm1_cam5 gfdl_esm2m miroc_esm_chem ncc_noresm1_m ables selected at each node. Trees were grown to full size. We followed
csiro_mk3_6_0 giss_e2_h miroc_miroc5 nimr_hadgem2_ao the recommendation to increase the number of trees until the predic-
fio_esm giss_e2_r mohc_hadgem2_es
tion error of independent runs stabilizes (Boulesteix et al., 2012). To
make the process more robust we trained 25 RF classifiers. We drew
2.5. Agro-climatic zone distribution five random samples from the entire population of occurrence pixels,
from each of which we drew five random subsamples of 80% of the
RF was applied to classify the climate data for Ghana into the agro- sample.
climatic zones for cocoa. The suitability of a site for cocoa production is We extrapolated the classifiers onto maps of the bioclimatic vari-
not only defined by its climate, but also by soil attributes. Cocoa experts ables and determined the modal value across all 25 model results. We
from CRIG therefore suggested including such information in the agro- obtained maps of each cocoa agro-climatic zone plus class “0″ back-
climatic zone mapping. Our classification model therefore used the 20 ground sites where cocoa is unlikely to be cultivated for current and
bioclimatic variables (Table 1). In addition, we included a soil variable future climate data. The most likely future agro-climatic zone was de-
from several soil attributes using RF classification focusing on soil at- termined for each pixel by the mode across the results for the 19 GCMs.
tributes that provide resilience against adverse climatic events and la-
belled the variable Soil Resilience Capacity (SRC). Such an approach 2.6. Identification of climatic threats
has previously been shown to improve model skill (Piikki et al., 2017).
The SRC variable was developed by implementing a supervised binary In addition to the recommendation domains of common degree of
classification approach based on the assumption that cocoa is cultivated impact, we evaluated potential hazards which may cause the projected
at locations with appropriate soils. As input we obtained 122 soil impacts. We reviewed the climate risks that farmer innovations try to
variables from the African Soils Information Service (Leenaars et al., manage (Adjewodah, 2016) and matched them with changes in bio-
2015) and cocoa experts prioritized a subset of 44 variables that are climatic variables. We found that innovations address problems to
meaningful for cocoa Supplementary materials). cocoa production from low/high rainfall, drought, high temperatures
For the agro-climatic zoning, we firsttrained RF with random sam- and prolonged dry season. We used the following indicator variables:
ples from each climate type of occurrence pixels and a random back- low/high rainfall- Bio 12 (total precipitation); drought – Bio 17 (pre-
ground sample of pixels within Ghana that did not have cocoa. From the cipitation in driest quarter), Bio 9 (mean temperature of driest quarter),
occurrence pixels in each agro-climatic zone, we selected samples of the and Bio 20 (consecutive months with less than 100 mm precipitation);
same size of the smallest suitability group. Such reduction of ecological high temperatures – Bio 1 (annual mean temperature), Bio 5 (max
sampling bias has been shown to improve the capacity of niche based temperature of the warmest month), Bio 10 (mean temperature of
approaches to correctly predict species distributions (Varela et al., warmest quarter); and prolonged dry season - Bio 20 (consecutive
2014). months with less than 100 mm precipitation).
The background sample had a size twice that of the smallest climate For each of these threats we used RF to evaluate whether changes in
group. For binary classification problems a 1:1 sampling ratio is re- indicator variables will be positive or negative for cocoa production.
commended to avoid the preferential prediction of the majority class We trained RF to estimate the probability whether a location is part of
(Barbet-Massin et al., 2012). The sampling ratio we chose accounted for the occurrence or background group of samples, using the indicator
the trade-off between the multi-class agro-climatic zone classification, variables as independents. In 25 repeats, we trained 100 trees on 50%
as well as the binary classification in suitable classes and unsuitable of the entire population. To avoid overconfident classification despite
background locations. We constrained the background samples to the small number of variables we limited the number of nodes to four.
contain only pixels with annual mean temperatures within the range as We then applied the RF classifiers to current and future climate data as

4
C. Bunn, et al. Climate Services 16 (2019) 100123

before. For each recommendation domain, we evaluated whether pro- series of meetings over the course of a week, the experts were presented
jected changes were positive, unchanged, negative, or beyond lower with preliminary input data and intermediate model results. For this
current limits. The current lower limit was defined by the 10th per- purpose map data was complemented with village locations to provide
centile of probabilities at occurrence locations. orientation. Following the experts’ suggestions input data and model
parameters were adapted to improve the alignment of model projec-
2.7. Model uncertainty tions and expert knowledge.
In addition, following a suggestion by the expert panel we compared
Model uncertainty was evaluated by estimating the model agree- our modeled distribution of agro-climatic zones with a map of historic
ment across all independent classification repeats and GCMs. If agree- canopy types (Hall and Swaine, 1981). Commonly the distribution of
ment on the majority class was below 60% the zone was labeled cocoa production in Ghana is described by the natural vegetation dis-
“mixed”. For current conditions, model agreement was assessed across tribution and canopy type (Hall and Swaine, 1981). Especially the re-
the 25 individual repeats of model training. Therefore, for a pixel cell to gions with moist semi-deciduous or moist evergreen canopy are major
be classified into any of the suitability types at least 15 classification cocoa production zones. As natural vegetation is to some part shaped by
steps had to agree on the same class. For future conditions, we repeated climatic conditions, this model has previously been used to distinguish
the classification for the outputs of 19 GCMs. Model agreement was agro-ecological zones of cocoa production.
therefore assessed across 19*25 = 475 individual classification out- Finally, the current distribution was validated using the multiclass
comes of which at least 285 had to agree in order to indicate a agro- area under receiver operating characteristic curve (AUC) (Hand and
climatic zone. If however, the modal classification was any one of the Till, 2001) as implemented in the R package “pROC” (Robin et al.,
agro-climatic zones, rather than of the “unsuitable” class, but less than 2011). The AUC assumes values 0 – 1. An AUC of 0.5 indicates that the
60% of repeats agreed to which zone, the respective pixel cell was still performance was no better than random sampling, while 1.0 is perfect
classified as suitable, and labelled “uncertain”. classification. This definition of the AUC measure can be extended to
Where the modal classification outcome was of the unsuitable class, multiclass problems by averaging all pairwise AUC comparisons to a
but less than 95% of the models agreed, the suitability class “limita- multiclass AUC (Hand and Till, 2001). We used this measure to evaluate
tions” was assigned. Thus, grid cells of the suitability class “limitations” the discrimination of agro-climatic zones by the RF classifier.
are most likely unsuitable, but at least two model repeats, or up to 12
model repeats, disagreed and classified these cells as suitable. For fu- 2.9. Recommendation domains by degree of impact
ture conditions, at least 24 of the 475 models had to disagree with the
“unsuitable” classification, or up to 237. Finally, the recommendation domains were defined by the quality
To demonstrate the effect of uncertainty from different GCM pro- of change between agro-climatic zones under current conditions, and
jections we picked a representative subset of six GCMs. We determined future conditions in each of the 19 GCM projections. We defined the
annual mean temperature and annual precipitation across Ghana for domain “incremental adaptation” where the climate zone was not
current and future conditions and ranked GCMs by the degree of projected to change; “systemic adaptation” where one zone changed to
change. The six GCMs chosen represent extreme changes in precipita- another agro-climatic zone, “transformational adaptation” where a
tion and mean temperature (Table 3). previously suitable zone likely becomes unsuitable for cocoa, and
“systemic resilience” where GCM uncertainty was high. Along this
2.8. Validation gradient, we provided an assessment of potential climate stressors to
cocoa production.
Evaluation of species distribution models from machine learning
algorithms is commonly done by using the area under the receiver 3. Results
operating characteristic curve (AUC). However, this measure is often
misleading (Lobo et al., 2008) and not meaningful where a potential 3.1. Definition of climate suitability types
distribution is evaluated (Merow et al., 2013). When presented with
different models biological knowledge of the species should be taken Metrics to guide the number of climate clusters in the occurrence
into account during model evaluation (Mainali et al., 2015). An alter- data suggested four distinct agro-climatic clusters. We named the agro-
native method is to use regional expert knowledge to evaluate model climatic zones in a North-South order from 1 to 4 after extrapolation.
performance (Uusitalo et al., 2015). This concept follows the idea of a For expert-selected climate variables, we compared the confidence in-
participatory research methodology (Bergold and Thomas, 2012; tervals for each cluster resulting from RF classification to the grand
Cornwall and Jewkes, 1995; Pain and Kindon, 2007) where research is mean (Fig. 2). Differences to the grand mean of all occurrence locations
carried out by inclusion of local experts. These local experts are sta- were used to describe the climate types as follows:
keholders in Ghanaian cocoa production and provide local knowledge
and experience. • Cluster 1 had near average conditions except for elevated tem-
All input data, intermediate and final results were subjected to peratures during the driest quarter and an even distribution of
evaluation of a panel of regional cocoa experts from the Cocoa Research precipitation throughout the year (low precipitation seasonality).
Institute of Ghana (CRIG). Experts were invited because of their This cluster belongs to the agro-climatic zone3, typical for the
knowledge on the ecophysiology of cocoa production in Ghana. In a Eastern region.

Table 3
Representative set of global climate models and projected climatic changes in Ghana.
GCM Annual mean temperature change in °C Annual total precipitation change in mm Temp. rank Temp. label Prec. rank Prec. label

miroc_esm 1.1 46 1/19 Cool 12/19 Interm.


csiro_mk3_6_0 1.9 −80 15/19 Hot 1/19 Dry
giss_e2_h 1.4 123 3/19 Cool 18/19 Wet
ipsl_cm5a_lr 2.6 14 19/19 Hot 6/19 Interm.
miroc_miroc5 1.6 154 6/19 Interm. 19/19 Wet
gfdl_cm3 2.3 93 17/19 Hot 16/19 Wet

5
C. Bunn, et al. Climate Services 16 (2019) 100123

Fig. 2. Contrast plots of group confidence intervals to grand mean; confidence intervals for each cluster are shown relative to the grand mean (dotted line), the
average value is given below each plot, the x-axis indicates the deviation; number of cluster groups is not identical with suitability types.

• Cluster 2 had average temperatures but low annual precipitation production in Ghana. The type 1 zone overlapped largely with the moist
and a strong dry season (very low precipitation in the driest quarter semi-deciduous north western zone. The moist semi-deciduous south
and the highest number of dry months). This cluster belongs to the eastern type zone is largely covered by the type 3 agro-climatic zone.
agro-climatic zone1, typical for Southern Brong Ahafo. Between the type 1 and type 3 zones, the type 2 zone spans parts of both
• Cluster 3 had the lowest diurnal temperature range, a high annual semi-deciduous zones. The fourth agro-climatic zone corresponded with
average temperature and a weak dry season (short, with compara- the moist evergreen zone. Both the dry semi-deciduous zones and the
tively high precipitation in the driest quarter) with high annual wet evergreen zone were classified to have strong limitations for cocoa
precipitation. This is the agro-climatic zone4, which spans the South production (Fig. 3A).
of the cocoa region in Ghana. By the 2050s in the RCP 6.0 emissions scenario, nearly the entire
• Cluster 4 had comparatively low temperatures and average pre- type 1 agro-climatic zone will become unsuitable for cocoa production
cipitation, but the dry season is quite long. This is the agro-climatic or will show strong limitations. The Ashanti region that is currently
zone2, which can be found in the Ashanti region. relatively cool will increasingly be of a mixed climate type. The type 4
climate zone in the Western and Central region expands into the Eastern
region (Fig. 3B) which under current conditions was a type 3 zone. The
3.2. Distribution of agro-climatic zones for cocoa under current and future area with limitations along the margins of the main agro-climatic zones
conditions that under current conditions overlapped with the dry semi-deciduous
zones becomes unsuitable. The area with limitations in the wet ever-
Extrapolation of the RF classifiers on current climate data resulted green zone remains equally challenged.
in maps of the spatial distribution of the four cocoa agro-climatic zones The average ability of the RF algorithm to discriminate agro-cli-
(Fig. 3A). The northern type 1 zone has low annual precipitation and a matic zones was satisfactorily high on the consensus map, with the
strong dry season. Central Ashanti is of the second agro-climatic zone multiclass AUC averaging 0.82 which is much better than chance. The
with cooler temperatures and a rather long dry season. The Eastern conventional AUC measure averaged 0.75, which demonstrates the
region is characterized by the third agro-climatic zone with higher robustness of the algorithm to discriminate between suitable and un-
temperatures and reliable precipitation. The south (Western and Cen- suitable pixel cells.
tral region) can be described with the fourth agro-climatic zone with
high temperatures and precipitation. Between agro-climatic zones,
patches of the “mixed” climate zone could be observed where the 3.3. Expert workshop results
suitability classification was uncertain. At the margins of these agro-
climatic zones area with limitations for cocoa production could be All input data, intermediate and final results were subjected to
found. In these regions, a majority of model repeats suggested that this evaluation of a panel of regional cocoa experts from the Cocoa Research
area is unsuitable, but some model repeats disagreed and classified this Institute of Ghana (CRIG). According to feedback from the panel, un-
area as suitable (Fig. 3A). feasible locations were eliminated from the occurrence data and addi-
This model of distribution of climate zones overlapped partially tional data from CRIG was added to the final dataset (Fig. 1). The cli-
with the distribution of natural vegetation (Hall and Swaine, 1981) that mate variables to describe agro-climatic zones were defined based on
was previously used to describe agro-ecological zones of cocoa the recommendations by the panel. In addition, a soils suitability

6
C. Bunn, et al. Climate Services 16 (2019) 100123

Fig. 3. Distribution of agro-climatic zones for cocoa


production in Ghana in the A) current climate, B)
across 19GCMs in the RCP 6.0 by the 2050s. Colored
patches represent the different climate types. Light
yellow represents area of classification uncertainty,
grey areas were found to be unsuitable but with high
uncertainty. Thick black lines in (A) show the bor-
ders of canopy zones adapted from Hall and Swaine
(1981) with the following abbreviations: WE = Wet
Evergreen; ME = Moist Evergreen; MS = Moist
Semi-deciduous (NW = North-west subtype;
SE = South-east subtype); DS = Dry Semi-decid-
uous. (For interpretation of the references to color in
this figure legend, the reader is referred to the web
version of this article.)

variable was added to the climate variables because experts argued that Western region and most of Ashanti also showed very high model un-
suitable soils may provide resilience against adverse climate conditions. certainty (F REF _Ref12651057 \h Fig. 4). On the other hand, model
The relevant soil attributes for this purpose were picked by CRIG ex- agreement was high for Southern Brong Ahafo and parts of the Eastern
perts. Finally, the resulting agro-climatic zoning for current climate region.
conditions was discussed with the expert panel. The experts agreed that A comparison of the climate zone classification of a representative
it correctly described the current production zones of cocoa in Ghana set of GCMs showed that most uncertainty was caused by disagreement
and suggested the overlap with the canopy type based ecological between GCMs (Fig. 4). All GCM projections agreed that the type 4
zoning. agro-climatic zone with its high temperatures and high precipitation
will be the dominant agro-climatic zone in the future: it is likely to
3.4. Model uncertainty retain area in the Western region, and the agro-climatic zone 3 was
projected to change to zone 4. Equally high agreement of GCMs pro-
We differentiated two sources of uncertainty: the agreement be- jected losses of suitable area in Brong Ahafo. The exception was the
tween individually trained classifiers and the agreement between global giss_e2_h GCM, a wet scenario with moderate temperature increases
climate models. We made the agreement between classifiers explicit by and a high precipitation increase. The largest uncertainty from GCMs
including the “mixed” and the “limitations” climate zones. could be observed in the Ashanti region. Most scenarios projected that
Nevertheless, it was insightful to compare the model agreement for this region remains suitable, but disagree on the agro-climatic zone in
current and future climate conditions. For current conditions, between future decades. In a similar pattern, the future of the Volta region re-
patches of high model agreement areas with low or no model agree- mained uncertain (Fig. 4).
ment could be found. Towards the margins model agreement was also Thus, we found that the projected loss of suitable area in Brong
low (Fig. 4). This area was reported as “limitations” before. The patches Ahafo was highly certain. The change from type 3 to type 4 climate in
of model disagreement within the cocoa zone were “mixed” climate the west of the Eastern region was also highly certain. The change of the
zones. Areas with high model agreement areas were reported as the extent of the type 4 agro-climatic zone in the Western and Central re-
climate zones. gion was of intermediate certainty. The future climate in the North of
This interpretation changed with future conditions. For the 2050s, the Western region and Ashanti was highly uncertain. GCM uncertainty
we found model agreement to be lower. For most of the suitable climate in this region did not permit an unambiguous classification.
zones, no clear classification outcome can be shown. The North of the

7
C. Bunn, et al. Climate Services 16 (2019) 100123

Fig. 4. Top: classification uncertainty for the current and future time slices. Bottom: distribution of suitability zones in the 2050s (RCP 6.0) by global climate model.

3.5. Gradient of climate change impacts region and the Southern margin will require systemic change because a
change of one climate zone to another was projected (Fig. 5).
We derived a gradient of impacts from these projected changes and For other parts of the country, the climate change signal was found
GCM model agreement. The gradient showed that Southern Brong to be less significant. The Southern stretch of the current production
Ahafo, Northern Ashanti and the North and South of Volta will become zone was projected to remain in the same climate zone so that no
transformation zones. These areas were classified as no longer suitable fundamental changes in agronomic practices will be necessary due to
by the RF classifier. The Northern half of the central cocoa production climate change. Incremental adaptation is the dominant strategy. Only
zone (Northern Ashanti, Northern Western region and most of the Volta few sites were projected to become suitable because of climate change.
region) was classified as a zone with little certainty of climate projec- Some locations at the margins of the currently suitable area may be-
tions. The main adaptation strategy would therefore require an increase come suitable but with strong limitations.
of resilience by systemic change. In Southern Ashanti, the Eastern

8
C. Bunn, et al. Climate Services 16 (2019) 100123

Fig. 5. Recommendation domains to scale adaptation strategies by degree of impact in the consensus across 19 GCMs in the RCP 6.0 scenario.

3.6. Climate threats by recommendation domain certain: For most recommendation domains, GCMs projected drought
risk that would be considered unsuitable for cocoa under current con-
For each recommendation domain we evaluated changes in biocli- ditions. The risk of high temperatures will also increase in all scenarios
matic variables and their potential to result in the projected degrees of beyond current suitable values.
impact. Specifically, we considered increasing or decreasing risk of
low/high rainfall, drought, high temperatures or prolonged dry season. 4. Discussion
We assigned probability scores for whether a location will have suitable
or unsuitable risk conditions. The 10th percentile at presence locations Climate change impact assessments should be able to support sta-
was taken as threshold below which conditions are risky. Table 4 keholders in the development of adequate adaptation strategies. We
summarizes the scores. demonstrated how a machine-learning modeling framework can be
The projections for precipitation variables were mixed between used to develop impact based recommendation domains. We employed
GCMs. For low/high rainfall and prolonged dry season some GCMs an agro-climatic zoning approach that was validated by local experts to
point to improved conditions, while others project riskier conditions. frame the impact assessment in local climate heuristics. Together with
The transformation zone was characterized by suitability scores near the identification of projected hazards, our recommendation domains
the lower end. Therefore, any reduction in precipitation or extension of will improve the ability of stakeholders to scale out site-specific adap-
the dry season will result in a critical change. These projections were tion strategies and dissemination of farmer innovation. The impact
different for the incremental or systemic change zones. Projected assessment itself is in line with previous projections in that it showed a
changes were not found to be below current suitable values, but posi- potential reduction in suitable area towards the Savanna in the North of
tive changes may occur in about half of GCM projections. The systemic the cocoa belt (Läderach et al., 2013; Schroth et al., 2016), but also that
resilience zone was projected with a slightly higher likelihood of ne- with targeted adaptation measures Ghana may remain a major cocoa
gative changes. producer into the future. However, our modeling approach went be-
When taking into account temperature variables the result was more yond previous binary suitability classifications that were hard to

Table 4
Percent of GCMs that project increasing or decreasing climate threats for each recommendation domain.
GCM projection

Drought High temperatures

Adaptation zone Unsuitable Reduced Unchanged Improved Unsuitable Reduced Unchanged Improved

Transform 100 100 0 0 100 100 0 0


Resilience 100 100 0 0 100 100 0 0
Change 95 100 0 0 100 100 0 0
Incremental 89 100 0 0 100 100 0 0
Opportunity 100 100 0 0 100 100 0 0

GCM projection

Prolonged dry season Total precipitation

Adaptation zone Unsuitable Reduced Unchanged Improved Unsuitable Reduced Unchanged Improved

Transform 95 79 0 21 79 42 0 58
Resilience 26 47 5 47 11 42 11 47
Change 0 37 11 53 0 47 5 47
Incremental 0 5 37 58 0 63 37 0
Opportunity 16 21 74 5 5 26 5 68

9
C. Bunn, et al. Climate Services 16 (2019) 100123

interpret by stakeholders by demonstrating how different climate zones which is not the case. Changes of precipitation and higher temperatures
of cocoa production will be affected. Describing the climatic properties will require confronting changed pest and disease patterns, inefficient
of the zones and their fate in future scenarios makes it easier for sta- agronomic practices or changed variability. On the other extreme, the
keholders to understand the differential effects of otherwise similar transformation domain will require expansive adaptation efforts to
GCM projections. sustain cocoa, or to facilitate a non-catastrophic transition to other li-
The threats from climate change by recommendation domain that velihoods. However, the chosen “red” may imply an abandonment of
we identified concurred with previously described threats to cocoa affected farmers by their customary relationships in the cocoa.
production. For example, dry season temperatures were hypothesized Another question is to which degree uncertainty should be disclosed
to negatively affect changes of crop suitability near the Savanna tran- when presenting climate model results. We included two types of un-
sition zone in Western Africa (Schroth et al., 2016). Our observation certainty explicitly in the map outputs: from the classification algorithm
that dry season conditions will be less suitable for cocoa production in and the GCM projections. Both sources of uncertainty have different
the future in all zones is similar to the projected loss of suitability across implications for adaptation. The increasing inability of the algorithm to
the cocoa belt by Läderach et al. (2013). While a dry season induces clearly identify a specific agro-climatic zone under future conditions
uniform flowering, overall yields are higher without such a period can be explained by the emergence of novel climates in the region. Such
(Zuidema et al., 2005). In addition, we here point out the potential novel climates will predominantly appear in the core cocoa zone and to
effects of rising temperatures. High temperatures induce faster leaf a lesser extent in the Savanna zone (Radeloff et al., 2015). Thus, even
growth, ageing and abscission possibly caused by high maintenance where the algorithm concluded that such climates are likely similar to
respiration rates (Almeida and Valle, 2007). High temperatures may suitable locations, the lack of historic analogues impedes the develop-
also further reduce transpiration rates (Amorim and Valle, 1992) when ment of adapted production systems. This difficulty increases ex-
in the dry season net carbon assimilation is already reduced (Carr and ponentially where GCM uncertainty makes statements about future
Lockwood, 2011). The increased temperatures could thus aggravate the precipitation impossible. We therefore think that showing uncertainty
effects even of an otherwise unchanged dry season. made the resulting maps harder to interpret, but ultimately will be
In the core cocoa zone, projected annual total precipitation was more credible to stakeholders (Burke et al., 2015).
equally likely to decrease or increase in our analysis. For cocoa this In summary, our analysis showed that in the incremental adaptation
makes it difficult to evaluate the benefits of adaptation strategies to domain known production systems may be adapted through minor
address potential hazards because the seasonal distribution and total changes to practices that respond to increases in temperature and
availability of precipitation are the dominant yield determining factors precipitation which may harbor additional risk and changed pests and
(Zuidema et al., 2005). Every GCM should be considered an equally diseases. In the systemic adaptation domain increasing precipitation
likely scenario (Burke et al., 2015) and we stated this uncertainty in our may make traditional knowledge about production practices obsolete
analysis in a coherent way. Other authors pointed to a slight increase of and farmers will have to make comprehensive adjustments to their
annual total precipitation when averaging over GCMs, to argue that dry systems. In the domain where climate models showed high uncertainty,
season temperatures (for which GCMs agree on the direction of change) we suggest to increase the resilience of the farming system to respond to
will be more important for future suitability (Schroth et al., 2016). an unknown future. Such adaptation may include a diversification of
While this may be true for the Western African cocoa belt as a whole, crops, or comprehensive measures to respond to increasing drought.
our analysis showed that for the high impact domain towards the Sa- Finally, our study and others have shown that towards the transition
vanna margins in Ghana precipitation projection agreement was better. zone with the Savanna a transformation of cocoa systems may be ne-
A majority of GCM scenarios agreed that the future seasonality of cessary. Such transformation may include the development of alter-
precipitation will be more pronounced, and that dry season character- native value chains or novel cocoa systems that are viable under con-
istics may pose unsustainable stress to the cocoa tree. ditions that would in the past be considered hostile for cocoa, for
Additionally, our validation process ensured that the re- example by breeding new varieties, efficient survival irrigation, and the
commendation domains are meaningful to stakeholders and biologi- establishment of fire breaks.
cally valid. Our motivation to use expert validation was to improve our The climate data that we used compared thirty-year average cli-
ability to interpret impacts within the local context (Corner et al., matologies. However, smallholders often have immediate challenges to
2018). Describing the agro-climatic zones in terms of climate variables their food and income security. The changing likelihood of extreme
that are used by experts within Ghana will make the zones relatable to precipitation and temperature events can therefore be more concerning
their specific agronomic knowledge. This facilitates the development of in this setting than long-term developments. This is exemplified by
adaptation strategies for the recommendation domains using existing farmers expanding their cocoa area in the transformation domain
cultivation practices. Furthermore, test statistics are harder to interpret knowingly accepting low yields and high seedling mortalities, because
than a confident value suggests (Lobo et al., 2008) and biological of a lack of income alternatives (Asante et al., 2017). Similarly, also
knowledge needs to be considered in model evaluation (Mainali et al., some downstream stakeholders don’t consider long term climate change
2015). Our finding that the spatial distribution of agro-climatic zones in relevant to their business performance because they are more pre-
our model reproduced canopy-type zones (Hall and Swaine, 1981) that occupied with securing short term supply (Sloan et al., 2019). Our work
were previously used to define ecological zones for cocoa production is is therefore most meaningful to actors for which the long-term per-
therefore a validation approach that was credible to local experts and spective is relevant, for example research institutes, traders with long
immediately demonstrated the biophysical relevance of the classifica- established relationships with producers or companies that build their
tion. Similar approaches should be considered where the communica- brand value around the sustainability concept.
tion of projected impacts is of interest, because doing so will make it Our work provided the necessary background information to de-
easier to relate abstract modeling exercises to concrete decision heur- velop potential pathways for sustainable cocoa production, but addi-
istics used in the local context, as is recommended for effective com- tional research into recommendation domains for changing climate
munication (Corner et al., 2018). variability would be complementary to our work. Whereas sustain-
We chose a simple color coding for the recommendation domain ability is the capacity of a system to function in the future, resilience to
maps: green for low adaptation needs, yellow for systemic adaptation, shocks is considered the means to achieve this (Tendall et al., 2015).
orange for high uncertainty zones and red where transformative Interventions that increase the system resilience by increasing the ca-
adaptation may be needed. This scheme was highly suggestive, and may pacity to avoid damages from hazards, and its flexibility to recover will
not be optimal in all cases. For example, the green of the incremental depend on a thorough analysis of variability. Ideally, such research
adaptation domain might suggest that no adaptation will be required, would also consider market effects, as has been done for barley (Xie

10
C. Bunn, et al. Climate Services 16 (2019) 100123

et al., 2018). Cocoa markets are highly volatile because of the inelastic approach in motion. Hist. Soc. Res. Sozialforschung 191–222.
supply and the high geographical concentration of production (Oomes Boulesteix, A.-L., Janitza, S., Kruppa, J., König, I.R., 2012. Overview of random forest
methodology and practical guidance with emphasis on computational biology and
et al., 2016). Changing climate variability in West Africa is therefore bioinformatics. Wiley Interdiscip. Rev. Data Min. Knowl. Discov. 2, 493–507.
likely to amplify uncertainty on cocoa markets globally, and as a result Breiman, L., 2001. Random Forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/
would have repercussions for all stakeholders. A:1010933404324.
Bretz, F., Hothorn, T., Westfall, P., 2010. Multiple Comparisons using R. CRC Press.
We conclude that the use of multi-class RF classification of agro- Bunn, C., Läderach, P., Pérez Jimenez, J.G., Montagnon, C., Schilling, T., 2015. Multiclass
climatic zones in combination with an expert validation process can classification of agro-ecological zones for Arabica coffee: an improved understanding
provide actionable recommendation domains to guide adaptation of the impacts of climate change. PLoS One 10, e0140490. https://doi.org/10.1371/
journal.pone.0140490.
strategies at scale. The approach combined a valid impact assessment Burke, M., Dykema, J., Lobell, D.B., Miguel, E., Satyanath, S., 2015. Incorporating climate
that required little input data, allowed for the explicit statement of uncertainty into estimates of climate change impacts. Rev. Econ. Stat. 97, 461–471.
uncertainty but is interpretable for a wide audience of stakeholders. https://doi.org/10.1162/REST_a_00478.
Caliński, T., Harabasz, J., 1974. A dendrite method for cluster analysis. Commun. Stat.
Our approach would be relatively easy to replicate in other regions and
Theor. Meth. 3, 1–27.
for other crops, where climate data is hard to interpret for stakeholders. Campbell, B.M., Vermeulen, S.J., Aggarwal, P.K., Corner-Dolloff, C., Girvetz, E.,
Involving experts facilitates the stakeholder dialogue and results in Loboguerrero, A.M., Ramirez-Villegas, J., Rosenstock, T., Sebastian, L., Thornton, P.,
more relatable impact assessments, which are a precondition for ef- Wollenberg, E., 2016. Reducing risks to food security from climate change. Global
Food Secur. https://doi.org/10.1016/j.gfs.2016.06.002.
fective adaptation at scale. Carr, M.K.V., Lockwood, G., 2011. The water relations and irrigation requirements of
cocoa (Theobroma Cacao L.): a review. Exp. Agric. 47, 653–676. https://doi.org/10.
Declaration of Competing Interest 1017/S0014479711000421.
Core Team, R., 2014. R: A Language and Environment for Statistical Computing. R
Foundation for Statistical Computing, Vienna, Austria.
The authors declare that they have no known competing financial Corner, A., Shaw, C., Clarke, J., 2018. Principles for Effective Communication and Public
interests or personal relationships that could have appeared to influ- Engagement on Climate Change: A Handbook for IPCC Authors. Climate Outreach,
Oxford.
ence the work reported in this paper. Cornwall, A., Jewkes, R., 1995. What is participatory research? Soc. Sci. Med. 41,
1667–1676. https://doi.org/10.1016/0277-9536(95)00127-S.
Acknowledgements ECOWAS, 2007. Atlas on Regional Integration in West Africa. ECOWAS-SWAC/OECD,
Abuja.
Eitzinger, A., Binder, C.R., Meyer, M.A., 2018. Risk perception and decision-making: do
This work was implemented as part of the CGIAR Research Program farmers consider risks from climate change? Clim. Change. https://doi.org/10.1007/
on Climate Change, Agriculture and Food Security (CCAFS), which is s10584-018-2320-1.
Fischer, G., Shah, M.M., Van Velthuizen, H.T., 2002. Climate change and agricultural
carried out with support from CGIAR Fund Donors and through bi-
vulnerability.
lateral funding agreements. For details please visit https://ccafs.cgiar. Fujino, J., Nair, R., Kainuma, M., Masui, T., Matsuoka, Y., 2006. Multi-gas mitigation
org/donors. The views expressed in this document cannot be taken to analysis on stabilization scenarios using AIM global model. Energy J. 343–354.
reflect the official opinions of these organizations. GSS, 2017. Annual GDP Estimate 2006 to 2013 [WWW Document]. Ghana Stat. Serv. URL
http://www.statsghana.gov.gh/docfiles/GDP/GDP2015/2015_Annual_GDP_
We benefited greatly from the substantial input by our colleagues at September_2015_Edition.pdf. (Accessed 2.22.17).
the Cocoa Research Institute of Ghana (CRIG) under the lead of Dr. Hall, J.B., Swaine, M.D., 1981. Distribution and Ecology of Vascular Plants in a Tropical
Amos Quaye. We are very grateful to Dr. Gilbert J. Anim-Kwapong in Rain Forest. Springer, Netherlands, Dordrecht doi: 0.1007/978-94-009-8650-3.
Hand, D.J., Till, R.J., 2001. A simple generalisation of the area under the ROC curve for
his former position as executive Director of CRIG for his support. We are multiple class classification problems. Mach. Learn. 45, 171–186.
also thankful for the contributions from COCOBOD, represented by Dr. Hijmans, R.J., Cameron, S.E., Parra, J.L., Jones, P.G., Jarvis, A., 2005. Very high re-
E. Tei Quartey and Eric Amengor. solution interpolated climate surfaces for global land areas. Int. J. Climatol. 25,
1965–1978.
The discussions with the panel of experts of Ghanaian cocoa pro- Kramer-Schadt, S., Niedballa, J., Pilgrim, J.D., Schröder, B., Lindenborn, J., Reinfelder,
duction that Dr. Amos Quaye put together were of very high value for V., Stillfried, M., Heckmann, I., Scharf, A.K., Augeri, D.M., 2013. The importance of
this project. This study would not have been possible without the ex- correcting for sampling bias in MaxEnt species distribution models. Divers. Distrib.
19, 1366–1379.
pertise of CRIG Scientists and we would like to acknowledge the kind
Läderach, P., Martinez-Valle, A., Schroth, G., Castro, N., 2013. Predicting the future cli-
support that was given to us by these CRIG Scientists, including A. matic suitability for cocoa farming of the world’s leading producer countries, Ghana
Afrifa, Fredrick Amon-Armah, G. Awudzi, Jerome Dogbatse, Esther and Côte d’Ivoire. Clim. Change 119, 841–854.
Lass, R.A., Wood, G.A.R., 1985. Cocoa Production, World Bank Technical Paper.
Gyan,and Sampson Kolan who were part of the core group. Other key
Washington, D.C., U.S.A.
experts who were part of the core working group included Emmanuel Leenaars, J.G.B., Hengl, T., González, M.R., de Jesus, J.M., Heuvelink, G.B.M., Wolf, J.,
Amamoo-Otchere (GISD), Robert Asugre (IFDC), Williams Atakorah van Bussel, L.G.J., Claessens, L., Yang, H., Cassman, K.G., 2015. Root Zone Plant-
(SARI), Rose Birago-Sefa (Tropenbos International) and Reuben Ottou Available Water Holding Capacity of the Sub-Saharan Africa soil, version 1.0.
Gridded Funct. Soil Inf. Dataset RZ-PAWHC SSA V 10 ISRIC Rep. 2.
(SNV). Liaw, A., Wiener, M., 2002. Classification and regression by random forest. R News 2,
18–22.
References Lobo, J.M., Jiménez-Valverde, A., Real, R., 2008. AUC: a misleading measure of the
performance of predictive distribution models. Global Ecol. Biogeogr. 17, 145–151.
https://doi.org/10.1111/j.1466-8238.2007.00358.x.
Adjewodah, P., 2016. CSA recommendations for cocoa in Ghana. Mainali, K.P., Warren, D.L., Dhileepan, K., McConnachie, A., Strathie, L., Hassan, G.,
Almeida, A.-A.F. de, Valle, R.R., 2007. Ecophysiology of the cacao tree. Br. J. Plant Karki, D., Shrestha, B.B., Parmesan, C., 2015. Projecting future expansion of invasive
Physiol. 19, 425–448. species: comparing and improving methodologies for species distribution modeling.
Amorim, S.M.C. de, Valle, R.R., 1992. Efeito da temperatura radicular na resistência ao Global Change Biol. 21, 4464–4480. https://doi.org/10.1111/gcb.13038.
movimento de água no cacaueiro (Theobroma cacao L.). Acta Bot. Bras. 6, 55–64. Merow, C., Smith, M.J., Silander, J.A., 2013. A practical guide to MaxEnt for modeling
https://doi.org/10.1590/S0102-33061992000100004. species’ distributions: what it does, and why inputs and settings matter. Ecography
Anim-Kwapong, G.J., Frimpong, E.B., 2005. Vulnerability of agriculture to climate No-No. https://doi.org/10.1111/j.1600-0587.2013.07872.x.
change-impact of climate change on cocoa production. Final Rep. Submitt. Neth. Moriondo, M., Jones, G.V., Bois, B., Dibari, C., Ferrise, R., Trombi, G., Bindi, M., 2013.
Clim. Change Stud. Assist. Programme NCAP 2. Projected shifts of wine regions in response to climate change. Clim. Change 119,
Asamoah, M., Baah, F., 2003. Improving research-farmer linkages: The role of CRIG, in: 825–839. https://doi.org/10.1007/s10584-013-0739-y.
4th International Seminar on Cocoa-Pests and Diseases (INCOPED), Accra. pp. 19–21. Notenbaert, A., Pfeifer, C., Silvestri, S., Herrero, M., 2017. Targeting, out-scaling and
Asante, W.A., Acheampong, E., Kyereh, E., Kyereh, B., 2017. Farmers’ perspectives on prioritising climate-smart interventions in agricultural systems: lessons from applying
climate change manifestations in smallholder cocoa farms and shifts in cropping a generic framework to the livestock sector in sub-Saharan Africa. Agric. Syst. 151,
systems in the forest-savannah transitional zone of Ghana. Land Use Policy 66, 153–162. https://doi.org/10.1016/j.agsy.2016.05.017.
374–381. Oomes, N., Tieben, B., Laven, A., Ammerlaan, T., Appelman, R., Biesenbeek, C., Buunk,
Barbet-Massin, M., Jiguet, F., Albert, C.H., Thuiller, W., 2012. Selecting pseudo-absences E., 2016. Market Concentration and Price Formation in the Global Cocoa Value
for species distribution models: how, where and how many? Meth. Ecol. Evol. 3, Chain.
327–338. https://doi.org/10.1111/j.2041-210X.2011.00172.x. Pain, R., Kindon, S., 2007. Participatory Geographies. SAGE Publications Sage UK,
Bergold, J., Thomas, S., 2012. Participatory research methods: a methodological London, England.

11
C. Bunn, et al. Climate Services 16 (2019) 100123

Phillips, S.J., Anderson, R.P., Schapire, R.E., 2006. Maximum entropy modeling of species Agriculture Papers. Springer International Publishing, Cham, pp. 227–233. https://
geographic distributions. Ecol. Model. 190, 231–259. https://doi.org/10.1016/j. doi.org/10.1007/978-3-319-92798-5_19.
ecolmodel.2005.03.026. Stocker, T.F., Qin, D., Plattner, G.-K., Tignor, M., Allen, S.K., Boschung, J., Nauels, A., Xia,
Piikki, K., Winowiecki, L., Vågen, T.-G., Ramirez-Villegas, J., Söderström, M., 2017. Y., Bex, V., Midgley, P.M., 2013. Climate Change 2013. The Physical Science Basis.
Improvement of spatial modelling of crop suitability using a new digital soil map of Working Group I Contribution to the Fifth Assessment Report of the
Tanzania. South Afr. J. Plant Soil 34, 243–254. https://doi.org/10.1080/02571862. Intergovernmental Panel on Climate Change-Abstract for decision-makers. Groupe
2017.1281447. d’experts intergouvernemental sur l’evolution du climat/Intergovernmental Panel on
Radeloff, Volker C., Williams, John W., Bateman, Brooke L., Burke, Kevin D., Carter, Climate Change-IPCC, C/O World Meteorological Organization, 7bis Avenue de la
Sarah K., Childress, Evan S., Cromwell, Kara J., et al., 2015. The rise of novelty in Paix, CP 2300 CH-1211 Geneva 2 (Switzerland).
ecosystems. Ecol. Appl. 25 (8), 2051–2068. https://doi.org/10.1890/14-1781.1. Tendall, D.M., Joerin, J., Kopainsky, B., Edwards, P., Shreck, A., Le, Q.B., Kruetli, P.,
Ramirez, J., Jarvis, A., 2010. Disaggregation of Global Circulation Model Outputs. Int. Grant, M., Six, J., 2015. Food system resilience: defining the concept. Glob. Food
Cent. Trop. Agric, CIAT Cali Colomb. Secur. 6, 17–23. https://doi.org/10.1016/j.gfs.2015.08.001.
Riahi, K., van Vuuren, D.P., Kriegler, E., Edmonds, J., O’Neill, B.C., Fujimori, S., Bauer, Uusitalo, L., Lehikoinen, A., Helle, I., Myrberg, K., 2015. An overview of methods to
N., Calvin, K., Dellink, R., Fricko, O., Lutz, W., Popp, A., Cuaresma, J.C., Kc, S., evaluate uncertainty of deterministic models in decision support. Environ. Model.
Leimbach, M., Jiang, L., Kram, T., Rao, S., Emmerling, J., Ebi, K., Hasegawa, T., Softw. 63, 24–31. https://doi.org/10.1016/j.envsoft.2014.09.017.
Havlik, P., Humpenöder, F., Da Silva, L.A., Smith, S., Stehfest, E., Bosetti, V., Eom, J., Varela, S., Anderson, R.P., García-Valdés, R., Fernández-González, F., 2014.
Gernaat, D., Masui, T., Rogelj, J., Strefler, J., Drouet, L., Krey, V., Luderer, G., Environmental filters reduce the effects of sampling bias and improve predictions of
Harmsen, M., Takahashi, K., Baumstark, L., Doelman, J.C., Kainuma, M., Klimont, Z., ecological niche models. Ecography 37, 1084–1091. https://doi.org/10.1111/j.1600-
Marangoni, G., Lotze-Campen, H., Obersteiner, M., Tabeau, A., Tavoni, M., 2017. The 0587.2013.00441.x.
Shared Socioeconomic Pathways and their energy, land use, and greenhouse gas Vermeulen, S.J., Challinor, A.J., Thornton, P.K., Campbell, B.M., Eriyagama, N., Vervoort,
emissions implications: an overview. Global Environ. Change 42, 153–168. https:// J.M., Kinyangi, J., Jarvis, A., Läderach, P., Ramirez-Villegas, J., Nicklin, K.J.,
doi.org/10.1016/j.gloenvcha.2016.05.009. Hawkins, E., Smith, D.R., 2013. Addressing uncertainty in adaptation planning for
Robin, X., Turck, N., Hainard, A., Tiberti, N., Lisacek, F., Sanchez, J.-C., Müller, M., 2011. agriculture. Proc. Natl. Acad. Sci. 110, 8357–8362. https://doi.org/10.1073/pnas.
pROC: an open-source package for R and S+ to analyze and compare ROC curves. 1219441110.
BMC Bioinf. 12, 77. Weber, E.U., 2010. What shapes perceptions of climate change?: what shapes perceptions
Schroth, G., Läderach, P., Martinez-Valle, A.I., Bunn, C., Jassogne, L., 2016. Vulnerability of climate change? Wiley Interdiscip. Rev. Clim. Change 1, 332–342. https://doi.org/
to climate change of cocoa in West Africa: patterns, opportunities and limits to 10.1002/wcc.41.
adaptation. Sci. Total Environ. 556, 231–241. https://doi.org/10.1016/j.scitotenv. Xie, W., Xiong, W., Pan, J., Ali, T., Cui, Q., Guan, D., Meng, J., Mueller, N.D., Lin, E.,
2016.03.024. Davis, S.J., 2018. Decreases in global beer supply due to extreme drought and heat.
Shi, T., Horvath, S., 2006. Unsupervised learning with random forest predictors. J. Nat. Plants 4, 964–973. https://doi.org/10.1038/s41477-018-0263-1.
Comput. Graph. Stat. 15, 118–138. https://doi.org/10.1198/106186006X94072. Zuidema, P.A., Leffelaar, P.A., Gerritsma, W., Mommer, L., Anten, N.P., 2005. A phy-
Sloan, K., Teague, E., Talsma, T., Daniels, S., Bunn, C., Jassogne, L., Lundy, M., 2019. One siological production model for cocoa (Theobroma cacao): model presentation, va-
size does not fit all: private-sector perspectives on climate change, agriculture and lidation and application. Agric. Syst. 84, 195–225.
adaptation. In: Rosenstock, T.S., Nowak, A., Girvetz, E. (Eds.), The Climate-Smart

12

You might also like