To analyze and understand potential impacts of climate change on crop production on a regional scale, we applied the following steps:
-
(a)
We selected a range of model treatments that represent farming practices of drybeans in the target countries.
-
(b)
We identified high-impact spots where climate change will impact drybean production in the first planting season of the year (in Central America called primera).
-
(c)
We compared simulated impacts on drybean yields for the second (called postrera) and the third (called apante) seasons on selected sites within the different types of hotspots.
-
(d)
We used data from multiple global circulation models (GCMs) for selected sites to assess uncertainty.
We used WorldClim (Hijmans et al. 2005) and downscaled GCMs (Ramirez-Villegas and Jarvis 2010) to provide monthly climate data for the climate baseline (current climate) and future climates. We generated daily climate from the monthly data, which we then used in the DSSAT drybean model (Fig. 1). We describe each procedure in more detail below.
Climate data
We used monthly total precipitation and mean monthly minimum and maximum temperature data as input to the MarkSim weather generator. For the climate baseline, we used the WorldClim database (Hijmans et al. 2005), which interpolated between observed data from more than 47,000 weather stations worldwide for the period 1950–2000 (Hutchinson 1995).
For future climates, we used the GCMs that the Intergovernmental Panel on Climate Change (IPCC) used for its “Fourth Assessment Report (AR4)” (Intergovernmental Panel on Climate Change (IPCC) 2007). We selected the GCMs’ outputs for the A2 scenario from the IPCC’s Special Report on Emissions Scenarios (SRES) (Intergovernmental Panel on Climate Change (IPCC) 2000). The A2 scenario describes a very heterogeneous world with high population growth, slow economic development, and slow technological change. It is commonly called the “business as usual scenario,” and 13 years after publication of the SRES, it reflects the current situation.
The spatial resolution of the GCMs (1–2°) is inappropriate for analyzing the impacts on agriculture (Jarvis et al. 2010), which therefore needs downscaling to provide higher resolution surfaces. We used the delta method of downscaling (Ramirez-Villegas and Jarvis 2010), which is based on the sum of interpolated anomalies to 30″ monthly climate surfaces of WorldClim. The method assumes that changes in climates are only relevant at coarse scales and that relationships between variables will be maintained in the future.
We used downscaled data from all 19 GCMs from IPCC’s AR4 for two different 30-year running-mean periods, 2010–2039 [2020s] and 2040–2069 [2050s]. We took means of the 30″ data (Ramirez-Villegas and Jarvis 2010) to produce 2.5′ and 5′ spatial resolution (roughly 5 and 10 km) for Nicaragua, Honduras, El Salvador, and Guatemala. We used the monthly data for each 2.5′ pixel as input to the MarkSim climate generator to produce daily weather data (Jones and Thornton 1993, 2000; Jones et al. 2002).
MarkSim uses a third-order Markov function to generate daily weather data that reflects the synoptic control of rainfall in the tropics by convection cells. It generates daily data of maximum and minimum temperatures, rainfall, and solar radiation for as many years as the user requires. We generated 99 replicate years of daily weather data for the climate baseline and for each of the 19 GCM models for the 2020s and 2050s for each pixel in the four countries. We automated this step by using the MarkSim 1.0 code compiled as an executable and running it from a command line under the control of a master FORTRAN procedure. In this way, we were able to run the process unattended as the run of MarkSim for each site was independent. The master procedure logged any failures of MarkSim but continued with the next site, which was not possible using MarkSim’s shell routine in batch mode.
Crop modeling
DSSAT is a widely tested series of simulation models (Jones et al. 2003; Hoogenboom et al. 2010). It incorporates detailed understanding of crop physiology, biochemistry, agronomy, and soil science to simulate performance of the main food crops, as well as pastures and fallows. It simulates crop water balance, photosynthesis, growth, and development on a daily time step. DSSAT requires input of the soil water characteristics and genetic coefficients of the crop cultivar, plus any relevant agronomic inputs such as fertilizer and irrigation. It is driven by daily maximum and minimum temperatures, rainfall, and solar radiation.
BEANGRO is a simulation model for drybeans (P. vulgaris L.) that was integrated into the crop simulation module component of DSSAT (Hoogenboom et al. 1994). It simulates vegetative growth, reproductive development, and yield. It has been validated many times (see, for example, Oliveira et al. 2012) and accurately reflects the phenological development and yield of drybean cultivars (see, for example, Oliveira et al. 2012 and Fig. 2). Here, we used it to examine the difference between yields using the climate data described above.
We prepared DSSAT management files (FILEX) that included initial conditions at planting, cultivar selection, planting data, and row and plant spacing, among others. We consulted experts from the CIAT bean program and from national bean programs in the four countries in Central America on the appropriate management to apply.
We assessed final impact using the mean of the treatments and calculated the anomalies across sites (pixels) of future and baseline yields.
Modeling steps
Simulations of bean management for the primera season
We defined a sowing window between 15 April and 30 June (the primera planting season) with a sowing trigger of 50 % available soil water in the top 30-cm layer of soil. The simulations started with available soil water (ASW) set at the lower limit (−1.5 MPa water potential) 60 days before the start of the sowing window to allow early season rain to accumulate in the soil. In consultation with experts, we selected one cultivar (black-seeded ICTA OSTUA) and one breeding line (red-seeded BAT1289) representative of cultivars commonly used in Central America. Because we could not obtain spatially distributed soil data, we used representative generic medium sandy loam and medium silt loam soils from the DSSAT package. We simulated two levels of fertilizer applications, 64 kg/ha 12-30-06 and 128 kg/ha 18-46-00 (N-P-K) at sowing, both with a side dressing of 30 kg N/ha as urea 22 days after sowing. The design was therefore a factorial arrangement of two cultivars, two soils, and two levels of fertilizer:
$$ \left\{\begin{array}{c}\hfill \mathrm{ICTA}\ \mathrm{OSTUA}\hfill \\ {}\hfill \mathrm{BAT}1289\hfill \end{array}\right\}\times \left\{\begin{array}{c}\hfill \mathrm{generic}\ \mathrm{medium}\ \mathrm{silty}\ \mathrm{loam}\hfill \\ {}\hfill \mathrm{generic}\ \mathrm{medium}\ \mathrm{sandy}\ \mathrm{loam}\hfill \end{array}\right\}\times \left\{\begin{array}{c}\hfill 64\;\mathrm{kg}/\mathrm{ha}\ 12\hbox{-} 30\hbox{-} 06(F1)\hfill \\ {}\hfill 128\;\mathrm{kg}/\mathrm{ha}\ 18\hbox{-} 46\hbox{-} 00(F2)\hfill \end{array}\right\} $$
Equation 1: experimental design used in DSSAT
The lower level of fertilizer represents a typical farmer’s management in Central America. A more advanced farmer might use the higher level, which also gives an estimate of the potential yields of the selected cultivars.
We used the averaged climate for the 19 GCMs as input data in a first step at 5′ resolution. After identifying high-impact spots (see below), we ran the simulations at 2.5′ using all 19 GCMs in step 4 (see Section 2.3.4).
Identify future high-impact spots
We calculated the yield change (future-baseline) from the yield outputs of the simulations (the mean of the eight treatments in step 1). We used the climate baseline and the ensemble of future climate data from the GCMs. We used distance statistics (Getis and Ord 1992) to identify the significant outliers and the high-impact spots (HISs).
Distance statistics analyze spatial association by measuring the degree of association within a population of weighted points. Spatial association is when the deviation of the variable of interest with respect to the mean (z-value) is greater than some specified level of significance. Here, we used a robust version of the root mean square (Darrouzet-Nardi and Bowman 2011) to scale the data and identify points that lie outside positive and negative cutoffs.
We used the HISs to identify priorities for diversification, adaptation, or conservation strategies. The three types of HISs were:
-
(a)
Adaptation spots: We identified pixels whose negative z-values of spatial association were equal to or greater than one standard deviation of the mean (68 %). We expect that yields of drybean in the primera season in these pixels will decrease in the 2020s and even more in the 2050s.
-
(b)
Hotspots: Pixels whose negative z-values were greater than two standard deviations of the mean (95 %). Yields will be so low that it will probably not be economic for farmers to continue to grow drybeans.
-
(c)
Pressure spots: Pixels whose positive z-values are greater than one standard deviation of the mean are where the future climate will favor drybeans. Most of these pixels lie outside the current zone of bean production, but we did map them (see Section 3).
We identified hotspots and adaptation spots within the areas that currently grow drybeans. We overlaid the pixels on maps from the Bean Atlas for the Americas (Mejia et al. 2001) using a kernel density analysis (Silverman 1986). By this means, we also identified the pressure spots outside the areas that currently grow drybeans.
Comparison of different growing seasons for selected sites and estimation of fertilizer responses
Changing planting dates would be an adaptation option, if alternate growing seasons were to give a yield advantage in future climates. We therefore ran the same set of treatments for the postrera and apante planting seasons and compared results with those for the primera. Then, within the identified hot- and adaptation spots, we selected 15 communities within municipalities that produce drybeans, distributed across all four countries. We selected pixels that lay within 15 km of the selected communities that intersected with the bean-growing areas identified in the Bean Atlas. We constrained the selection to those pixels whose elevation lay within 100 m of the elevation of the selected community. In total, we selected 171 points for the comparison between seasons (Table 3).
Table 3 Selected municipalities and points used from the Bean Atlas for simulating different planting seasons
We defined the planting date windows 15 April–30 June for the primera season, 20 August–30 September for the postrera, and 25 October–5 December for the apante. We also ran the model without simulating nutrient options to assess the fertilizer response on each site. We estimated the yield with no fertilizer increase by disabling the fertilizer application in the simulation control options. We did this for the 15 selected sites using current and climate input data for climate baseline and GCM ensembles for the 2020s and 2050s.
Run data from multiple GCMs on selected sites for the primera season to assess the prediction uncertainty
Uncertainty in climate projections raises doubts as to their applicability in crop models (Asseng 2013). Acknowledging that uncertainty exists is the first step towards being able to quantify it (Challinor et al. 2009; Ramirez-Villegas et al. 2013). We used data from all 19 GCMs on the 171 points selected in the previous step to generate daily data for the 2020s and calculated the change of yield for each GCM. For each point, we estimated the uncertainty of the simulated yields for the predicted future climates:
-
(a)
The yield change of the GCM ensemble mean
-
(b)
The standard deviation (SD) of the yield change
-
(c)
The agreement among the model simulations using the 19 GCMs’ climate projections, calculated as the percentage of the model outputs predicting changes in the same direction