Tune, and evaluate, species distribution models

Usage

tune_sdm(
  prep,
  out_dir = FALSE,
  return_val = "path",
  algo = c("all", "maxnet", "envelope", "rf"),
  max_corr = list(maxnet = 0.7, envelope = 0.9, rf = 0.99),
  fc = "auto_feature",
  limit_p = FALSE,
  rm = seq(1, 6, 0.5),
  trees = c(999),
  mtry = TRUE,
  limit_spat_mtry = 4,
  nodesize = c(1, 2),
  keep_model = FALSE,
  best_run = FALSE,
  metrics_df = envSDM::sdm_metrics,
  use_metrics = c("auc_po", "CBI_rescale", "IMAE"),
  do_gc = FALSE,
  force_new = FALSE,
  ...
)

Arguments

prep: Character or named list. If character, the path to an existing prep.rds. Otherwise, the result of a call to prep_sdm with return_val = "object"
out_dir: FALSE or character. If FALSE the result of tune_sdm will be saved to a temporary folder. If character, a file 'tune.rds' will be created at the path defined by out_dir.
return_val: Character: "object" or "path". Both return a named list. In the case of "path" the named list is simply list(tune = out_dir). Will be set to "object" if out_dir is FALSE.
algo: Character. Name of algorithm to use.
max_corr: Named list. Names of list elements must match algorithms being used. For each pair of predictor variables correlated at or above max_corr one will be dropped using caret::findCorrelation().
fc: Character. Used to generate levels of classes argument to maxnet::maxnet() that are tuned.
limit_p: TRUE, FALSE or number of predictor variables above which to limit the use of p in the classes argument used in maxnet::maxnet(). Useful with many predictor variables when it becomes unwieldy to generate interactions for all predictors.
rm: Numeric. Used to generate levels of regmult argument to maxnet::maxnet() that are tuned.
trees: Used to generate the levels of ntree argument to randomForest::randomForest() that are tuned. TRUE (tune with default trees), FALSE (don't tune trees) or numeric (the trees values to tune with).
mtry: Used to generate the levels of mtry argument to randomForest::randomForest() that are tuned. TRUE (tune with sensible guesses for mtry), FALSE (only use default randomForest::randomForest() mtry) or numeric (the mtry values to tune with).
limit_spat_mtry: Numeric. If mtry is TRUE and if using spatial cross validation, the values of mtry to tune will be limited to less than or equal to limit_spat_mtry.
nodesize: Used to generate the levels of nodesize argument to randomForest::randomForest() that are tuned. TRUE (tune with default nodesize), FALSE (only use default randomForest::randomForest() nodesize) or numeric (the nodesize values to tune with).
keep_model: Logical. If TRUE the model results will be appended as a list column in the returned tibble (as column m)
best_run: Logical. If TRUE this alters the behaviour of the tune_sdm() by, well, not tuning. :). Sets all folds to the same value so no cross-validation.
metrics_df: Dataframe. Defines which metrics to use when deciding on 'good' SDMs.
use_metrics: Character. Vector of values in metrics_df$metric to use when finding the 'best' model.
do_gc: Logical. Run base::rm(list = ls) and base::gc() at end of function? Useful when running SDMs for many, many taxa, especially if done in parallel.
force_new: Logical. If outputs already exist, should they be remade?
...: Passed to evaluate_sdm(). e.g. thresholds for use in predicts::pa_evaluate() (as tr argument, although if used, the values of the thresholds element of the pa_ModelEvaluation object returned by predicts::pa_evaluate() will be limited to the values in tr).

Value

If return_val is "object" a named list. If return_val is "path" a path to the saved file. If out_dir is a valid path, the 'full result' (irrespective of return_val) is also saved to fs::path(out_dir, "prep.rds"). The 'full result' is a named list with elements:

Examples


  out_dir <- file.path(system.file(package = "envSDM"), "examples")

  data <- fs::path(system.file(package = "envSDM"), "examples") |>
    fs::dir_ls(regexp = "prep\\.rds$"
               , recurse = TRUE
               ) |>
    tibble::enframe(name = NULL, value = "prep") |>
    dplyr::mutate(taxa = gsub("\\.rds", "", basename(dirname(prep)))
                  , out_dir = fs::path(out_dir, taxa)
                  )

  purrr::map(data$out_dir
              , \(x) tune_sdm(prep = fs::path(x, "prep.rds")
                              , out_dir = x
                              , fc = "lq"
                              , rm = c(2, 3)
                              , trees = 500
                              , mtry = c(1:3)
                              , nodesize = c(1, 2, 3)
                              , limit_p = 3
                              , use_metrics = c("auc_po", "CBI_rescale", "IMAE")
                              #, force_new = TRUE
                              )
              )
#> [[1]]
#> /home/nwilloug/temp/RtmpO4BWL8/temp_libpathfdac86abd0f83/envSDM/examples/chg__0.3__1/tune.rds
#> 
#> [[2]]
#> /home/nwilloug/temp/RtmpO4BWL8/temp_libpathfdac86abd0f83/envSDM/examples/chg__0.3__5/tune.rds
#> 
#> [[3]]
#> /home/nwilloug/temp/RtmpO4BWL8/temp_libpathfdac86abd0f83/envSDM/examples/chg__0__1/tune.rds
#> 
#> [[4]]
#> /home/nwilloug/temp/RtmpO4BWL8/temp_libpathfdac86abd0f83/envSDM/examples/chg__0__5/tune.rds
#> 

  # which tune args were best for each taxa using 'combo'?
  data %>%
    dplyr::mutate(tune = fs::path(out_dir, "tune.rds")
                  , tune = purrr::map(tune, rio::import, trust = TRUE)
                  , tune_mean = purrr::map(tune, "tune_mean")
                  ) %>%
    tidyr::unnest(cols = c(tune_mean)) %>%
    dplyr::filter(best) %>% # used 'combo' to determine 'best' as default in tune_sdm
    dplyr::select(taxa, algo, tune_args, combo, auc_po, IMAE, CBI, max_spec_sens)
#> # A tibble: 4 × 8
#>   taxa        algo  tune_args             combo auc_po  IMAE   CBI max_spec_sens
#>   <chr>       <chr> <chr>                 <dbl>  <dbl> <dbl> <dbl>         <dbl>
#> 1 chg__0.3__1 rf    tr: 500. mt: 2. ns: 2 0.657  0.875 0.820 0.832         0.551
#> 2 chg__0.3__5 rf    tr: 500. mt: 1. ns: 3 0.532  0.785 0.805 0.684         0.281
#> 3 chg__0__1   rf    tr: 500. mt: 1. ns: 1 0.519  0.782 0.811 0.638         0.275
#> 4 chg__0__5   rf    tr: 500. mt: 1. ns: 2 0.510  0.792 0.814 0.584         0.263

  # or best tune args choosing on just auc_po?
  data %>%
    dplyr::mutate(tune = fs::path(out_dir, "tune.rds")
                  , tune = purrr::map(tune, rio::import, trust = TRUE)
                  , all = purrr::map(tune, "tune_mean")
                  ) %>%
    tidyr::unnest(cols = c(all)) %>%
    dplyr::group_by(taxa) %>%
    dplyr::filter(auc_po == max(auc_po)) %>%
    dplyr::ungroup() %>%
    dplyr::select(taxa, algo, tune_args, auc_po, IMAE, CBI, max_spec_sens)
#> # A tibble: 4 × 7
#>   taxa        algo  tune_args             auc_po  IMAE   CBI max_spec_sens
#>   <chr>       <chr> <chr>                  <dbl> <dbl> <dbl>         <dbl>
#> 1 chg__0.3__1 rf    tr: 500. mt: 1. ns: 1  0.877 0.817 0.788         0.606
#> 2 chg__0.3__5 rf    tr: 500. mt: 2. ns: 2  0.786 0.809 0.640         0.263
#> 3 chg__0__1   rf    tr: 500. mt: 2. ns: 3  0.788 0.811 0.581         0.272
#> 4 chg__0__5   rf    tr: 500. mt: 1. ns: 2  0.792 0.814 0.584         0.263