poEvaluationMetrics

R package for evaluating species distribution models (SDMs): discrimination metrics (AUC, correlation, PRG), calibration-style error (MAE, bias), Boyce-type Spearman indices (SBI_* via smoothed GAMs), and threshold-based scores (TSS, Kappa, omission, F1-style Fbp, SEDI, ORSS, etc.).

Install

Dependencies (beyond CRAN defaults) include terra, sf, CAST, predicts, mecofun, PresenceAbsence, prg, and Metrics. Some must be installed from GitHub / Git:

install.packages(c("remotes", "devtools"))

# PRG package (GitHub)
remotes::install_github("meeliskull/prg/R_package/prg")

# mecofun (GitLab)
remotes::install_git("https://gitup.uni-potsdam.de/macroecology/mecofun.git")

Then install poEvaluationMetrics (install any other missing packages from CRAN with install.packages as needed):

remotes::install_github("envima/poEvaluationMetrics")

Vignette: after installation, see vignette("poEvaluationMetrics", package = "poEvaluationMetrics").

Citation

After installing the package:

citation("poEvaluationMetrics")

Or cite the inst/CITATION entry when reporting methods.

Usage

High-level: run one or more evaluation designs (presence–absence, presence–background with sampled background, presence–artificial-absence via AOA) with optional replication for PBG/PAA. presence is required (no default).

library(poEvaluationMetrics)

result <- performanceEstimation(
  prediction = prediction_raster,           # terra::SpatRaster
  presence = presence_sf,                   # sf — required
  absence = absence_sf,                     # or FALSE to skip PA
  background = TRUE,                        # or FALSE, or an sf object
  aa = TRUE,                                # PAA via CAST::aoa; FALSE to skip
  environmentalVariables = env_rasters,    # needed if background or aa is TRUE
  noPointsTesting = 1000,                   # defaults to nrow(presence)
  replicates = 50                           # mean over replicates for PBG/PAA
)

# One row per scenario; use the `scenario` column (PA / PAA / PBG)
result[result$scenario == "PBG", ]

Single map + one set of absence/background points:

df_metrics <- calculateMetrics(prediction_raster, presence_sf, background_or_absence_sf)

Tabular data only (continuous predictions + binary observed, plus a raster for Boyce/SBI sampling):

tab <- data.frame(predicted = c(0.2, 0.8, 0.1), observed = c(0L, 1L, 0L))
indexCalculation(tab, prediction_raster)

Standalone helpers (binary classifications for most): Fbp(), sedi(), orss(), omission(), brier_score(), fowlkes_mallows(), mcc(), tcr(). See ?Fbp and other help pages.

Documentation website (pkgdown)

Configure the site URL in _pkgdown.yml if it differs from the default. Build locally:

install.packages("pkgdown") # if needed
pkgdown::build_site()

Development

devtools::document()
devtools::check()

License

Not decided yet—there is no License field in DESCRIPTION. When you pick one (e.g. MIT or GPL-3), add it there and any matching LICENSE file so others know how they may use the code.

Note: R CMD check (and devtools::check(as_cran = TRUE)) expects a License entry, so those checks will report an error until you add one. You can still develop with load_all(), tests, and pkgdown in the meantime.