Maxent species distribution models, in C++

sdm_maxent() fits Maxent (Phillips, Anderson and Schapire 2006) as an infinitely-weighted logistic regression with a lasso penalty (Fithian and Hastie 2013; Phillips et al. 2017), the formulation of the maxnet package, with the same feature classes and default regularisation. The lasso path is solved by coordinate descent in C++, so no Java and no glmnet are needed. Predictions are on the cloglog scale by default (Phillips et al. 2017), clamped to the training range.

Usage

sdm_maxent(
  presence,
  data,
  features = "default",
  regmult = 1,
  add_samples_to_background = TRUE,
  standardize = TRUE
)

# S3 method for class 'tmr_maxent'
predict(
  object,
  newdata,
  type = c("cloglog", "logistic", "exponential", "link"),
  clamp = TRUE,
  ...
)

Arguments

presence: 0/1 vector: 1 for presences, 0 for background points.
data: Data frame of predictors for the same rows; numeric columns are continuous, factors are categorical.
features: Feature classes, any of "l" (linear), "q" (quadratic), "p" (product), "h" (hinge), "t" (threshold) as one string such as "lqh", or "default", which uses maxnet's rule (l below 10 presences, lq below 15, lqh below 80, else lqph).
regmult: Regularisation multiplier: larger gives smoother models.
add_samples_to_background: Add presences to the background, as Maxent does?
standardize: Rescale every continuous predictor to 0-1 over the training data before building features (default TRUE), as the original Java Maxent does. The maxnet package builds linear, quadratic and product features on the raw scale and fits them unstandardised, so a predictor in metres (elevation) and one in millimetres (rainfall) produce squares and products that differ by many orders of magnitude. The penalties then do not compare across features and the lasso path can stall. On the 0-1 scale every feature class is comparable, the default regularisation means the same thing for every predictor, and the fit is stable. Use FALSE to reproduce maxnet exactly.
object: A model from sdm_maxent().
newdata: Data frame or SpatRaster of predictors.
type: "cloglog" (default), "logistic", "exponential" or "link".
clamp: Clamp predictors and features to the training range?
...: Not used.

Value

sdm_maxent(): an object of class tmr_maxent with the non-zero coefficients (betas), alpha, entropy and the feature definitions. predict(): a numeric vector, or a SpatRaster for raster input.

References

Phillips SJ, Anderson RP, Schapire RE (2006) Maximum entropy modeling of species geographic distributions. Ecological Modelling 190:231-259.

Fithian W, Hastie T (2013) Finite-sample equivalence in statistical models for presence-only data. Annals of Applied Statistics 7:1917-1939.

Phillips SJ, Anderson RP, Dudik M, Schapire RE, Blair ME (2017) Opening the black box: an open-source release of Maxent. Ecography 40:887-893.

Examples