3. Overview of algorithms

\(~\)

Introduction

In this exercise, you will:

• Gain familiarity and experience working with a variety of different algorithms used for SDMs/ENMs

Please spend a short time to look through some of the materials in the Overview of algorithms section. But there is a lot of information there and impossible to absorb in a short period. When you feel motivated, move on to the R Practical section to work through some examples.

\(~\)

Overview of algorithms

There is a wide range of different modeling algorithms used for correlative SDMs/ENMs (mechanistic SDMs are not covered here). For presence-only or presence-absence data, these range from envelope (or ‘profile’) methods (e.g., BIOCLIM), to GLMs (e.g., logistic regression), machine-learning algorithms (e.g., maxent, Random Forest), and complex methods that can be used to model multiple species at once (‘joint SDMs’).

The Biodiversity and Climate Change Virtual Laboratory has compiled this useful site that provides a detailed description of different model algorithms but does not include R examples.

Some of the most commonly used R packages for SDMs, based on the general type of algorithm are provided in the list below:

• Presence-absence algorithms
  • GLMs (Generalized linear models; base R stats packae): D. Zurell has a nice example of fitting a simple SDM using GLMs, and here is another example from R Spatial.
  • GAMs (Generalized additive models; gam R package): Here is a basic tutorial put together by Phoebe Zarnetske’s lab.
• Presence-only algorithms
  • BIOCLIM, as implemented in the dismo R package.
  • maxent is available as a Java program, which can be called in R with the dismo package.
  • maxnet is a new and slightly different implementation of maxent. See code on Github.
  • biomod2 is an ensemble approach where results from multiple algorithms are combined.
  • embarcadero uses Bayesian additive regression trees (BARTs) as an alternative to popular classification tree methods. There is an extended tutorial here.
  • randomForest is another machine learning method, see an R tutorial here.
  • brt (‘boosted regression trees’) is another machine learning method. See an R example here, paper by Elith et al.
• Joint (multi-species) SDMs
  • hSDM
  • HMSC, publication, Github

Although these are beyond the scope of this course, an even wider variety of approaches are available when additional data on abundance or repeat surveys or demographic rates are available (e.g., occupancy models, integral projection models). Here are some links for your own interest:

• Occupancy models
  • A lecture and R exercise by Res Altwegg and Tad Fukami has a nice R tutorial
  • MacKenzie et al.’s book “Occupancy Estimation and Modeling: Inferring Patterns and Dynamics of Species Occurrence” is an authoritative guide on these models.
• Integral projection models (IPMs)
  • Merow et al.’s excellent (2013) paper provides a thorough R tutorial in the appendix.
  • Cory Merow also has this nice presentation on IPMs.

\(~\)

R Practical

For this part of the exercise, we will work through this tutorial by R Spatial.

For additional examples, I also highly recommend Damaris Zurell’s “SDM algorithms” exercise.