mapspamc includes two types of models: (1) Minimization
of cross-entropy, which is used by SPAM (You and
Wood 2006; You, Wood, and Wood-Sichra 2009; You et al. 2014; Yu et al.
2020) and the maximization of fitness score, which was developed
by Dijk et al. (2022). The sections below
describe the basic structure of the two models. Additional technical
information is presented in the appendix.
The cross-entropy approach (min_entropy) is defined as a
non-linear optimization problem in which the error between prior
information on the location of crop area shares (\(\pi_{ij}\)) and model allocated area shares
(\(s_{ijl}\)) is minimized subject to a
number of constraints.1 More precisely the cross-entropy objective
function (\(CE\)) is specified as
follows:
\[\begin{equation} \min\ CE(s_{ijl}, \pi_{ijl})= \sum_{i} \sum_{j} \sum_{l}s_{ijl} \ lns_{ijl}- \sum_{i} \sum_{j} \sum_{l}s_{ijl} \ ln\pi_{ijl} \end{equation}\]
In the maximization of fitness score approach
(max_score) the spatial allocation model is defined as a
linear optimization problem in which the weighted ‘fitness score’ over
all grid cells, crops and production systems is maximized. The fitness
score, which ranges from 0 to 1, is a composite indicator of the
economic and biophysical suitability for each crop and production system
combination at the grid cell level. The solution to the model implies
that on average all crops will be allocated to locations, which are
considered as most appropriate from both an economic and biophysical
viewpoint. The objective function (\(S\)) is defined as follows:
\[\begin{equation} \max\ S(s_{ijl}, score_{ijl})= \sum_{i} \sum_{j} \sum_{l}s_{ijl} \times score_{ijl} \end{equation}\]
where, \(s_{ijl}\) is the share of total physical area of crop \(j\) and production system \(l\) allocated to grid cell \(i\), and \(score_{ijl}\) is the fitness score for all \(j\) and \(l\) combinations.
Cross entropy minimization is better suited for maps at a resolution of 5 arc minute as it results in a more fragmented allocation per grid cell. This is desirable when the resolution is relatively coarse (e.g. 5 arc minutes) and the number of crop and production systems that can be observed a grid cell is relatively large. In contrast, maximizing the fitness score results in a much more concentrated crop area allocation, where only a few crops and production systems are allocated to a grid cell. This makes sense for relatively small grid cells (e.g. 30 arc seconds) where one only expects to observed a small number of crop and production systems, but is not realistic for large grid cells. Hence, we recommend to use the cross-entropy model for 5 arc minute resolution and the fitness score model for 30 arc second resolution crop distribution maps.
To solve the models in mapspamc, either fitness scores
or prior area information is needed to allocate the statistics to each
grid cell. The fitness scores and priors are informed by a combination
of economic (e.g. market access and population density) and biophysical
factors (e.g growing conditions determined by soil, temperature and
precipitation) that jointly determine the location of crops. As the
impact of these factors will be different for production system, we
define separate fitness scores and priors for each of them.
For subsistence farmers, who mainly grow crops for their own consumption, we assume that rural population density can be used as proxy for the location of subsistence production systems. We use min-max normalization to create an index of rural population density between 0 and 1:
\[\begin{equation} score_{ijl} = \frac{rpd_{ijl}-min(rpd_{ijl})}{max(rpd_{ijl})-min(rpd_{ijl})} \quad l = subsistence \quad \forall i\ \forall j \end{equation}\]
where \(rpd_{ijl}\) is the rural population density defined for each grid cell \(i\), crop \(j\) and production system \(l\), which is converted to the \(score_{ijl}\) between 0 and 1. In contrast to the other three systems, the allocation of subsistence farmers is modeled by means of a constraint instead of through the objective function (see next section). In this way, it is ensured that the total subsistence crop area is allocated proportional to the rural population density. Distributing subsistence production systems by means of maximizing the fitness score, might lead to a solution in which a large share of subsistence crop area will be allocated to grid cells with the highest rural population density. Such a concentrated allocation is not realistic for subsistence farming.
Low-input rainfed systems are characterized by smallholder farms that mostly sell their products on local markets. We assume that farmers will specialize in growing crops with the highest biophysical suitability, while other factors such as market access, plays a minor role for this production system. The fitness score is defined as follows:
\[\begin{equation} score_{ijl} = \frac{suit\_index_{ijl}-min(suit\_index_{ijl})}{max(suit\_index_{ijl})-min(suit\_index_{ijl})} \quad l = low\ input \quad \forall i\ \forall j \end{equation}\]
where \(score_{ijl}\) and \(suit\_index_{ijl}\) are the fitness score and the biophysical suitability score for grid cell \(i\), crop \(j\) and production system \(l\), respectively. Similar to the fitness score for subsistence farming, we apply the min-max normalization approach to create an index in the range of 0 and 1.
We assume that high-input rainfed and irrigated production systems mostly consist of medium to large commercial farms, which location is largely determined by profit maximization. To simulate this, the fitness core is defined as the square root of potential revenue and market access:
\[\begin{equation} score_{ijl} = \frac{\sqrt{access_{i} \times rev_{ijl}}-min(\sqrt{access_{i} \times rev_{ijl}})}{max(\sqrt{access_{i} \times rev_{ijl}})-min(\sqrt{access_{i} \times rev_{ijl}})} \quad l \in high\ input,\ irrigated \quad \forall i\ \forall j \end{equation}\]
where \(access_{ijl}\) is market access approximated by the inverse of travel time to large cities and \(rev_{ijl}\) is potential revenue, calculated as the product of potential yield (\(yield_{ijl}\)) for grid cell \(i\), crop \(j\) and system \(l\) and national level crop price (\(price_j\)):
\[\begin{equation} rev_{ijl} = pot\_yield_{ijl} \times price_{i} \end{equation}\]
Again, we use the min-max normalization to create an index in the range of 0-1.
We use the fitness scores as a basis to estimate the priors. To priors for subsistence farming are calculated as follows:
\[\begin{equation} prior_{ijl} =\frac{score_{ijl}}{\sum_{j} \sum_{l}{score_{ijl}}} \times crop\_area_{jl} \quad l = subsistence \quad \forall i\ \forall j \end{equation}\]
where \(prior_{ijl}\) is the prior crop area in ha for subsistence farming defined for each grid cell \(i\) and crop \(j\) and \(crop\_area_{jl}\) is the total physical area per crop for the subsistence production system taken from the statistics.
The prior areas for the other three systems are estimated by distributing the residual grid cell cropland (i.e. after subtracting the prior area for subsistence production systems) using the scores as weights:
\[\begin{equation} prior_{ijl} = \frac{score_{ijl}}{\sum_{j} \sum_{l} score_{ijl}} \times residual\_cropland_{i} \quad l \in low\ input,\ high\ input,\ irrigated \quad \forall i\ \forall j \end{equation}\]
where \(residual\_cropland_{i}\) is the grid cell cropland after subtracting the prior for subsistence farming. In case the prior area for subsistence farming is larger than than the cropland in the grid cell all the priors for the other production systems are set to zero.
Finally, the priors are converted in to crop area shares (\(\pi_{ij}\)), which are an input into the cross-entropy objective function:
\[\begin{equation} \pi_{ij} = \frac{prior_{ijl}}{\sum_{i}{prior_{ijl}}} \end{equation}\]
The allocation of crop area is determined by either minimizing the cross-entropy function or maximizing the fitness scores subject to a set of constraints:
\[\begin{equation} 0 \leq s_{ijl} \leq 1 \end{equation}\]
\[\begin{equation} \sum_{i}s_{ij}=1 \qquad \forall j\ \forall l \end{equation}\]
\[\begin{equation} \sum_{j} \sum_{l}crop\_area_{jl} \times s_{ijl} \leq avail_{i} \qquad \forall i \end{equation}\]
where \(crop\_area_{jl}\) is total national-level physical area for crop \(j\) and production system \(j\).
\[\begin{equation} \sum_{i \in k} \sum_{l}crop\_area_{jl} \times s_{ijl} = sub\_crop\_area_{jk} \qquad \forall j \in J\ \forall k \end{equation}\]
\[\begin{equation} \sum_{j \in L}crop\_area_{jl} \times s_{ijl} \leq irr\_area_{i} \qquad \forall i \end{equation}\]
max_score model), that
allocates the subsistence production system crops proportional to rural
population density. In cases, where the biophysical conditions are
considered not suitable for a certain crop, we assume a zero
allocation:\[\begin{equation} \bar{s}_{ijl} = \frac{rur\_pop_i}{\sum_{i}{rur\_pop_i}} \quad l = subsistence \quad \forall i\ \forall j \end{equation}\]
where \(\bar{s}_{ijl}\) is the share of physical area allocated to grid cell \(i\) for crop \(j\) and production system \(l\) and \(pop_i\) is the rural population density in grid cell \(i\).
More precisely, SPAM uses a relative entropy framework because it minimizes the difference between the cross-entropy of probability distribution \(P\) with a reference probability distribution \(Q\) and the cross entropy of \(P\) with itself. This is also sometimes referred to as the the Kullback–Leibler divergence (Kullback and Leibler 1951)↩︎