The smop package is a small R package (it has four main
functions) designed to perform the Mobility-Oriented Parity analysis
(MOP) both in fast fashion and taking care of memory usage. To do this
it uses optimized C++ code via the Rcpp and
RcppArmadillo
packages.
The MOP test was proposed in Owens et al. 2013 to estimate regions where extrapolation occurs (Figure 1).
Example of use
First we estimate zones where strict extrapolation occurs. To show this, we read environmental data from the calibration and the transfer zones.
m_path <- system.file("extdata/M_layers", package = "smop") |>
list.files(full.names=TRUE)
g_path <- system.file("extdata/G_layers", package = "smop") |>
list.files(full.names=TRUE)
M_calibra <- terra::rast(m_path)
G_transfer <- terra::rast(g_path)Calibration area
terra::plot(M_calibra)
Fig. 2. Calibration area in Central America.
Model transfer area
terra::plot(G_transfer)
Fig. 3. Transfer area in South America
Zones where strict extrapolation occurs.
extr_zones <- smop::extrapolation_zones(M_calibra = M_calibra,
G_transfer = G_transfer,
as_vec =FALSE)
terra::plot(extr_zones)
Zones of strict extrapolation. The green zones represent areas where strict extrapolation occurs
The MOP analysis
Now we perform the mop analysis
# Example 1 compute mop distance in serial
mop_test <- smop::mop(M_calibra = M_calibra, G_transfer = G_transfer,
percent = 10,comp_each = NULL,
normalized = TRUE, standardize_vars=TRUE)
terra::plot(mop_test)
MOP analysis. In this map zero values (in gray) represent zones of strict extrapolation and values above it are zones with analogus climates.
Parallel MOP
The MOP analysis can be performed in parallel by evoking functions
from the furrr
R package. Here, we use the comp_each parameter to specify the
number of pixels to be processed in each CPU core. The number of pixels
to be processed in each CPU core depends on the user’s RAM memory; if
the user has a computer with few RAM, it is recommended to process by
500 pixels, otherwise, it can be as high as 10,000 pixels. An
improvement from other versions of the mop function such as the
ntbox
and the kuenm, is
that the performance of the function is less affected by the
parametrization of the comp_each parameter.
# Example 2: Run the mop function in parallel
future::plan("future::multisession",workers = 2)
mop_test_parallel <- smop::mop(M_calibra = M_calibra,
G_transfer = G_transfer,
percent = 10,comp_each = 500,
normalized = TRUE, standardize_vars=TRUE)
future::plan("future::sequential")
terra::plot(mop_test_parallel)In fact, the better management of memory and the optimization of the code allows users to run MOP analyses on huge extents such as the whole world. Here, we show how to run the MOP using the bioclimatic variables for the current time at 10 arc minutes and a climate change scenario.
First, we downloaded the climatic layer for whole world for the current time at 10 arc minutes of spatial resolution. Then, we downloaded the climate change layers. Finally, we run the MOP analysis in a parallel fashion using 10 cores.
bio10_pre <- geodata::worldclim_global("bioc",res=10,path = tempdir())
bio10_fut <- geodata::cmip6_world("CNRM-CM6-1", "585", "2061-2080",
var="bioc", res=10, path=tempdir())
mop_time <- system.time({
future::plan("future::multisession",workers = 10)
mop_basic_res1 <- smop::mop(M_calibra = bio10_pre[[c(1,5,6,12,15)]],
G_transfer = bio10_fut[[c(1,5,6,12,15)]],
comp_each = 10000,
percent = 50,standardize_vars = TRUE,
normalized = TRUE)
future::plan("future::sequential")
})
mop_time
#> user system elapsed
#> 47.362 5.575 748.391 This process took about 12 minutes which seems to be a lot of time but note that this analysis was done for the whole world! Moreover, this kind of MOP analysis wont run on other packages.
Lets plot the result
terra::plot(mop_basic_res1)