tidysdm overview
SDMs with tidymodels
Species Distribution Modelling relies on several algorithms, many of
which have a number of hyperparameters that require turning. The
tidymodels universe includes a number of packages
specifically design to fit, tune and validate models. The advantage of
tidymodels is that the models syntax and the results
returned to the users are standardised, thus providing a coherent
interface to modelling. Given the variety of models required for SDM,
tidymodels is an ideal framework. tidysdm
provides a number of wrappers and specialised functions to facilitate
the fitting of SDM with tidymodels.
This article provides an overview of the how tidysdm
facilitates fitting SDMs. Further articles, detailing how to use the
package for palaeodata, fitting more complex models and how to
troubleshoot models can be found on the tidisdm
website. As tidysdm relies on tidymodels,
users are advised to familiarise themselves with the introductory
tutorials on the tidymodels
website.
When we load tidysdm, it automatically loads
tidymodels and all associated packages necessary to fit
models:
Preparing your data
We start by reading in a set of presences for a species of lizard
that inhabits the Iberian peninsula, Lacerta schreiberi.
data(lacerta)
lacerta
#> # A tibble: 1,297 × 3
#> ID latitude longitude
#> <dbl> <dbl> <dbl>
#> 1 858029749 42.6 -7.09
#> 2 858029738 42.6 -7.09
#> 3 614631090 41.4 -7.90
#> 4 614631085 41.3 -7.81
#> 5 614631083 41.3 -7.81
#> 6 614631080 41.4 -7.83
#> 7 614631072 41.4 -7.81
#> 8 614559731 40.3 -7.70
#> 9 614559728 40.4 -7.70
#> 10 614559657 40.4 -7.56
#> # ℹ 1,287 more rows
First, let us visualise our presences by plotting on a map.
tidysdm works with sf objects to represent
locations, so we will cast our coordinates into an sf
object, and set its projections to standard lonlat (crs = 4326).
library(sf)
lacerta <- st_as_sf(lacerta, coords = c("longitude", "latitude"))
st_crs(lacerta) <- 4326
It is usually advisable to plot the locations directly on the raster
that will be used to extract climatic variables, to see how the
locations fall within the discrete space of the raster. For this
vignette, we will use WorldClim as our source of climatic information.
We will access the WorldClim data via the library pastclim;
even though this library, as the name suggests, is mostly designed to
handle palaeoclimatic reconstructions, it also provides convenient
functions to access present day reconstructions and future projections.
pastclim has a handy function to get the land mask for the
available datasets, which we can use as background for our locations. We
will cut the raster to the Iberian peninsula, where our lizard lives.
For this simply illustration, we will not bother to project the raster,
but an equal area projection would be desirable…
library(pastclim)
land_mask <-
get_land_mask(time_ce = 1985, dataset = "WorldClim_2.1_10m")
# Iberia peninsula extension
iberia_poly <-
terra::vect(
"POLYGON((-9.8 43.3,-7.8 44.1,-2.0 43.7,3.6 42.5,3.8 41.5,1.3 40.8,0.3 39.5,
0.9 38.6,-0.4 37.5,-1.6 36.7,-2.3 36.3,-4.1 36.4,-4.5 36.4,-5.0 36.1,
-5.6 36.0,-6.3 36.0,-7.1 36.9,-9.5 36.6,-9.4 38.0,-10.6 38.9,-9.5 40.8,
-9.8 43.3))"
)
crs(iberia_poly) <- "lonlat"
# crop the extent
land_mask <- crop(land_mask, iberia_poly)
# and mask to the polygon
land_mask <- mask(land_mask, iberia_poly)
For plotting, we will take advantage of tidyterra, which
makes handling of terra rasters with ggplot a
breeze.
library(tidyterra)
library(ggplot2)
ggplot() +
geom_spatraster(data = land_mask, aes(fill = land_mask_1985)) +
geom_sf(data = lacerta)
Thinning step
Now, we thin the observations to have one per cell in the raster (it
would be better if we had an equal area projection…):
set.seed(1234567)
lacerta <- thin_by_cell(lacerta, raster = land_mask)
nrow(lacerta)
#> [1] 231
ggplot() +
geom_spatraster(data = land_mask, aes(fill = land_mask_1985)) +
geom_sf(data = lacerta)
Now, we thin further to remove points that are closer than 20km.
However, note that the standard map units for a ‘lonlat’ projection are
meters. tidysdm provides a convening conversion function,
km2m(), to avoid having to write lots of zeroes):
set.seed(1234567)
lacerta_thin <- thin_by_dist(lacerta, dist_min = km2m(20))
nrow(lacerta_thin)
#> [1] 113
Let’s see what we have left of our points:
ggplot() +
geom_spatraster(data = land_mask, aes(fill = land_mask_1985)) +
geom_sf(data = lacerta_thin)
Now we sample pseudo-absences (we will constrain them to be at least
50km away from any presences), selecting as 3 times as many points as
presences:
set.seed(1234567)
lacerta_thin <- sample_pseudoabs(lacerta_thin,
n = 3 * nrow(lacerta_thin),
raster = land_mask,
method = c("dist_min", km2m(50))
)
Let’s see our presences and absences:
ggplot() +
geom_spatraster(data = land_mask, aes(fill = land_mask_1985)) +
geom_sf(data = lacerta_thin, aes(col = class))
Generally, we can use pastclim to check what variables
are available for the WorldClim dataset:
climate_vars <- get_vars_for_dataset("WorldClim_2.1_10m")
We first download the dataset at the right resolution (here 10
arc-minutes):
download_dataset("WorldClim_2.1_10m")
And then create a terra SpatRaster object. The dataset
covers the period 1970-2000, so pastclim dates it as 1985
(the midpoint). We can directly crop to the Iberian peninsula:
climate_present <- pastclim::region_slice(
time_ce = 1985,
bio_variables = climate_vars,
data = "WorldClim_2.1_10m",
crop = iberia_poly
)
We start by selecting variables for which presences are markedly
different from the underlying background. First, let’s extract climate
for all presences and pseudo-absences:
lacerta_thin <- lacerta_thin %>%
bind_cols(terra::extract(climate_present, lacerta_thin, ID = FALSE))
We can use violin plots to compare the distribution of the climatic
variables for presences and pseudo-absences:
lacerta_thin %>% plot_pres_vs_bg(class)
We
want to choose variables for which presences use values different from
the background (pseudo-absences). We can qualitatively look at the
plots, or use a quantitative approach that ranks them based on the
overlap of the respective density plots:
lacerta_thin %>% dist_pres_vs_bg(class)
#> bio19 bio12 bio16 bio02 bio13 bio07 bio05 bio04 bio09 bio10 bio17
#> 0.6935026 0.6929654 0.6885097 0.6776561 0.6679894 0.6542063 0.6243068 0.6071527 0.5506741 0.4855064 0.4318872
#> bio15 bio18 bio06 bio08 bio14 bio11 bio03 bio01 altitude
#> 0.4241999 0.3600587 0.3454245 0.3108560 0.2954351 0.2317079 0.2173934 0.2045136 0.1554400
As a first step, we can focus on variables that have at least 30% of
non-overlapping distribution between presences and pseudo-absences:
vars_to_keep <- lacerta_thin %>% dist_pres_vs_bg(class)
vars_to_keep <- names(vars_to_keep[vars_to_keep > 0.30])
lacerta_thin <- lacerta_thin %>% select(all_of(c(vars_to_keep, "class")))
vars_to_keep
#> [1] "bio19" "bio12" "bio16" "bio02" "bio13" "bio07" "bio05" "bio04" "bio09" "bio10" "bio17" "bio15" "bio18"
#> [14] "bio06" "bio08"
These are still a lot of variables. Among them, there are several
that have been argued in the past to be important for the distribution
of the species of interest. Based on this paper (https://doi.org/10.1007/s10531-010-9865-2), we are
interested in:
suggested_vars <- c("bio05", "bio06", "bio13", "bio14", "bio15")
Environmental variables are often highly correlated, and collinearity
is an issue for several types of models. We can inspect the correlation
among variables with:
pairs(climate_present[[suggested_vars]])
We can see that some variables have rather high correlation
(e.g. bio05 vs bio14). We can subset to variables below a certain
threshold correlation (e.g. 0.7) with:
climate_present <- climate_present[[suggested_vars]]
vars_uncor <- filter_high_cor(climate_present, cutoff = 0.7)
vars_uncor
#> [1] "bio15" "bio05" "bio13" "bio06"
#> attr(,"to_remove")
#> [1] "bio14"
So, removing bio14 leaves us with a set of uncorrelated
variables.
lacerta_thin <- lacerta_thin %>% select(all_of(c(vars_uncor, "class")))
climate_present <- climate_present[[vars_uncor]]
Fit the model by cross-validation
Next, we need to set up a recipe to define how to handle
our dataset. We don’t want to do anything to our data in terms of
transformations, so we just need to define the formula (class
is the outcome, all other variables are
predictors; note that, for sf objects,
geometry is automatically replaced by X and
Y columns which are assigned a role of coords,
and thus not used as predictors):
lacerta_rec <- recipe(lacerta_thin, formula = class ~ .)
lacerta_rec
#>
#> ── Recipe ──────────────────────────────────────────────────────────────────────────────────────────────────────
#>
#> ── Inputs
#> Number of variables by role
#> outcome: 1
#> predictor: 4
#> coords: 2
In classification models for tidymodels, the assumption
is that the level of interest for the response (in our case, presences)
is the reference level. We can confirm that we have the data correctly
formatted with:
lacerta_thin %>% check_sdm_presence(class)
#> [1] TRUE
We now build a workflow_set of different models,
defining which hyperparameters we want to tune. We will use
glm, random forest, boosted_trees and
maxent as our models (for more details on how to use
workflow_sets, see this
tutorial). The latter three models have tunable hyperparameters. For
the most commonly used models, tidysdm automatically
chooses the most important parameters, but it is possible to fully
customise model specifications (e.g. see the help for
sdm_spec_rf).
lacerta_models <-
# create the workflow_set
workflow_set(
preproc = list(default = lacerta_rec),
models = list(
# the standard glm specs
glm = sdm_spec_glm(),
# rf specs with tuning
rf = sdm_spec_rf(),
# boosted tree model (gbm) specs with tuning
gbm = sdm_spec_boost_tree(),
# maxent specs with tuning
maxent = sdm_spec_maxent()
),
# make all combinations of preproc and models,
cross = TRUE
) %>%
# tweak controls to store information needed later to create the ensemble
option_add(control = control_ensemble_grid())
We now want to set up a spatial block cross-validation scheme to tune
and assess our models. We will do an 80:20 split, by creating 5 folds.
We use the spatial_block_cv function from the package
spatialsample. spatialsample offers a number
of sampling approaches for spatial data; it is also possible to convert
objects created with blockCV (which offers further features
for spatial sampling) into an rsample object suitable to
tisysdm with the function blockcv2rsample.
library(tidysdm)
set.seed(100)
lacerta_cv <- spatial_block_cv(lacerta_thin, v = 5)
autoplot(lacerta_cv)
We can now use the block CV folds to tune and assess the models (to
keep computations fast, we will only explore 3 combination of
hyperparameters per model; this is far too little in real life!):
set.seed(1234567)
lacerta_models <-
lacerta_models %>%
workflow_map("tune_grid",
resamples = lacerta_cv, grid = 3,
metrics = sdm_metric_set(), verbose = TRUE
)
#> i No tuning parameters. `fit_resamples()` will be attempted
#> i 1 of 4 resampling: default_glm
#> ✔ 1 of 4 resampling: default_glm (881ms)
#> i 2 of 4 tuning: default_rf
#> i Creating pre-processing data to finalize unknown parameter: mtry
#> ✔ 2 of 4 tuning: default_rf (1.5s)
#> i 3 of 4 tuning: default_gbm
#> i Creating pre-processing data to finalize unknown parameter: mtry
#> ✔ 3 of 4 tuning: default_gbm (7.5s)
#> i 4 of 4 tuning: default_maxent
#> ✔ 4 of 4 tuning: default_maxent (2.9s)
Note that workflow_set correctly detects that we have no
tuning parameters for glm. We can have a look at the
performance of our models with:
Now let’s create an ensemble, selecting the best set of parameters
for each model (this is really only relevant for the random forest, as
there were not hype-parameters to tune for the glm and gam). We will use
the Boyce continuous index as our metric to choose the best random
forest and boosted tree. When adding members to an ensemble, they are
automatically fitted to the full training dataset, and so ready to make
predictions.
lacerta_ensemble <- simple_ensemble() %>%
add_member(lacerta_models, metric = "boyce_cont")
lacerta_ensemble
#> A simple_ensemble of models
#>
#> Members:
#> • default_glm
#> • default_rf
#> • default_gbm
#> • default_maxent
#>
#> Available metrics:
#> • boyce_cont
#> • roc_auc
#> • tss_max
#>
#> Metric used to tune workflows:
#> • boyce_cont
And visualise it
autoplot(lacerta_ensemble)
Projecting to the present
We can now make predictions with this ensemble (using the default
option of taking the mean of the predictions from each model).
prediction_present <- predict_raster(lacerta_ensemble, climate_present)
ggplot() +
geom_spatraster(data = prediction_present, aes(fill = mean)) +
scale_fill_terrain_c() +
# plot presences used in the model
geom_sf(data = lacerta_thin %>% filter(class == "presence"))
We can subset the ensemble to only use the best models, based on the
Boyce continuous index, by setting a minimum threshold of 0.8 for that
metric. We will also take the median of the available model predictions
(instead of the mean, which is the default). The plot does not change
much (the models are quite consistent).
prediction_present_boyce <- predict_raster(lacerta_ensemble, climate_present,
metric_thresh = c("boyce_cont", 0.8),
fun = "median"
)
ggplot() +
geom_spatraster(data = prediction_present_boyce, aes(fill = median)) +
scale_fill_terrain_c() +
geom_sf(data = lacerta_thin %>% filter(class == "presence"))
Sometimes, it is desirable to have binary predictions (presence vs
absence), rather than the probability of occurrence. To do so, we first
need to calibrate the threshold used to convert probabilities into
classes (in this case, we optimise the TSS):
lacerta_ensemble <- calib_class_thresh(lacerta_ensemble,
class_thresh = "tss_max"
)
And now we can predict for the whole continent:
prediction_present_binary <- predict_raster(lacerta_ensemble,
climate_present,
type = "class",
class_thresh = c("tss_max")
)
ggplot() +
geom_spatraster(data = prediction_present_binary, aes(fill = binary_mean)) +
geom_sf(data = lacerta_thin %>% filter(class == "presence"))
Projecting to the future
WorldClim has a wide selection of projections for the future based on
different models and Shared Socio-economic Pathways (SSP). Type
help("WorldClim_2.1") for a full list. We will use
predictions based on “HadGEM3-GC31-LL” model for SSP 245 (intermediate
green house gas emissions) at the same resolution as the present day
data (10 arc-minutes). We first download the data:
download_dataset("WorldClim_2.1_HadGEM3-GC31-LL_ssp245_10m")
Let’s see what times are available:
get_time_ce_steps("WorldClim_2.1_HadGEM3-GC31-LL_ssp245_10m")
#> [1] 2030 2050 2070 2090
We will predict for 2090, the further prediction in the future that
is available.
Let’s now check the available variables:
get_vars_for_dataset("WorldClim_2.1_HadGEM3-GC31-LL_ssp245_10m")
#> [1] "bio01" "bio02" "bio03" "bio04" "bio05" "bio06" "bio07" "bio08" "bio09" "bio10" "bio11" "bio12" "bio13"
#> [14] "bio14" "bio15" "bio16" "bio17" "bio18" "bio19"
Note that future predictions do not include altitude (as
that does not change with time), so if we needed it, we would have to
copy it over from the present. However, it is not in our set of
uncorrelated variables that we used earlier, so we don’t need to worry
about it.
climate_future <- pastclim::region_slice(
time_ce = 2090,
bio_variables = vars_uncor,
data = "WorldClim_2.1_HadGEM3-GC31-LL_ssp245_10m",
crop = iberia_poly
)
And predict using the ensemble:
prediction_future <- predict_raster(lacerta_ensemble, climate_future)
ggplot() +
geom_spatraster(data = prediction_future, aes(fill = mean)) +
scale_fill_terrain_c()
Visualising the contribution of individual variables
It is sometimes of interest to understand the relative contribution
of individual variables to the prediction. This is a complex task,
especially if there are interactions among variables. For simpler linear
models, it is possible to obtain marginal response curves (which show
the effect of a variable whilst keeping all other variables to their
mean) using step_profile() from the recipes
package. We use step_profile() to define a new recipe which
we can then bake to generate the appropriate dataset to make the
marginal prediction. We can then plot the predictions against the values
of the variable of interest. For example, to investigate the
contribution of bio05, we would:
bio05_prof <- lacerta_rec %>%
step_profile(-bio05, profile = vars(bio05)) %>%
prep(training = lacerta_thin)
bio05_data <- bake(bio05_prof, new_data = NULL)
bio05_data <- bio05_data %>%
mutate(
pred = predict(lacerta_ensemble, bio05_data)$mean
)
ggplot(bio05_data, aes(x = bio05, y = pred)) +
geom_point(alpha = .5, cex = 1)
It is also possible to use DALEX,to explore
tidysdm models; see more details in the tidymodels
additions article.
Repeated ensembles
The steps of thinning and sampling pseudo-absences can have a bit
impact on the performance of SDMs. As these steps are stochastic, it is
good practice to explore their effect by repeating them, and then
creating ensembles of models over these repeats. In
tidysdm, it is possible to create
repeat_ensembles. We start by creating a list of
simple_ensembles, by looping through the SDM pipeline. We
will just use two fast models to speed up the process.
# empty object to store the simple ensembles that we will create
ensemble_list <- list()
set.seed(123) # make sure you set the seed OUTSIDE the loop
for (i_repeat in 1:3) {
# thin the data
lacerta_thin_rep <- thin_by_cell(lacerta, raster = climate_present)
lacerta_thin_rep <- thin_by_dist(lacerta_thin_rep, dist_min = 20000)
# sample pseudo-absences
lacerta_thin_rep <- sample_pseudoabs(lacerta_thin_rep,
n = 3 * nrow(lacerta_thin_rep),
raster = climate_present,
method = c("dist_min", 50000)
)
# get climate
lacerta_thin_rep <- lacerta_thin_rep %>%
bind_cols(terra::extract(climate_present, lacerta_thin_rep, ID = FALSE))
# create folds
lacerta_thin_rep_cv <- spatial_block_cv(lacerta_thin_rep, v = 5)
# create a recipe
lacerta_thin_rep_rec <- recipe(lacerta_thin_rep, formula = class ~ .)
# create a workflow_set
lacerta_thin_rep_models <-
# create the workflow_set
workflow_set(
preproc = list(default = lacerta_thin_rep_rec),
models = list(
# the standard glm specs
glm = sdm_spec_glm(),
# maxent specs with tuning
maxent = sdm_spec_maxent()
),
# make all combinations of preproc and models,
cross = TRUE
) %>%
# tweak controls to store information needed later to create the ensemble
option_add(control = control_ensemble_grid())
# train the model
lacerta_thin_rep_models <-
lacerta_thin_rep_models %>%
workflow_map("tune_grid",
resamples = lacerta_thin_rep_cv, grid = 10,
metrics = sdm_metric_set(), verbose = TRUE
)
# make an simple ensemble and add it to the list
ensemble_list[[i_repeat]] <- simple_ensemble() %>%
add_member(lacerta_thin_rep_models, metric = "boyce_cont")
}
#> i No tuning parameters. `fit_resamples()` will be attempted
#> i 1 of 2 resampling: default_glm
#> ✔ 1 of 2 resampling: default_glm (408ms)
#> i 2 of 2 tuning: default_maxent
#> ✔ 2 of 2 tuning: default_maxent (9.5s)
#> i No tuning parameters. `fit_resamples()` will be attempted
#> i 1 of 2 resampling: default_glm
#> ✔ 1 of 2 resampling: default_glm (400ms)
#> i 2 of 2 tuning: default_maxent
#> ✔ 2 of 2 tuning: default_maxent (9.2s)
#> i No tuning parameters. `fit_resamples()` will be attempted
#> i 1 of 2 resampling: default_glm
#> ✔ 1 of 2 resampling: default_glm (410ms)
#> i 2 of 2 tuning: default_maxent
#> ✔ 2 of 2 tuning: default_maxent (9.3s)
Now we can create a repeat_ensemble from the list:
lacerta_thin_rep_ens <- repeat_ensemble() %>% add_repeat(ensemble_list)
lacerta_thin_rep_ens
#> A repeat_ensemble of models
#>
#> Number of repeats:
#> • 3
#>
#> Members:
#> • default_glm
#> • default_maxent
#>
#> Available metrics:
#> • boyce_cont
#> • roc_auc
#> • tss_max
#>
#> Metric used to tune workflows:
#> • boyce_cont
We can then predict in the usual way (we will take the mean and
median of all models):
lacerta_thin_rep_ens <- predict_raster(lacerta_thin_rep_ens, climate_present,
fun = c("mean", "median")
)
ggplot() +
geom_spatraster(data = lacerta_thin_rep_ens, aes(fill = median)) +
scale_fill_terrain_c()
