embarcadero: Species distribution modelling with Bayesian additive regression trees in R

Abstract

8 1. Classification and regression tree methods, like random forests (RF) or 9 boosted regression trees (BRT), are one of the most popular methods of 10 mapping species distributions. 11 12 2. Bayesian additive regression trees (BARTs) are a relatively new alterna-13 tive to other popular regression tree approaches. Whereas BRT iteratively 14 fits an ensemble of trees each explaining smaller fractions of the total vari-15 ance, BART starts by fitting a sum-of-trees model and then uses Bayesian 16 backfitting with an MCMC algorithm to create a posterior draw. So far, 17BARTs have yet to be applied to species distribution modeling. 18 19 3. embarcadero is an R package of convenience tools for researchers in-20 terested in species distribution modeling with BARTs. It includes function-21 ality for spatial prediction, an automated variable selection and importance 22 procedure, and other functionality for rapid implementation and data visu-23 alization. 24 25 4. To show how embarcadero can be used by ecologists, we re-map the distri-26 bution of Crimean-Congo haemorrhagic fever and a likely vector, Hyalomma 27 truncatum, in Africa.28 29 Keywords: Bayesian additive regression trees, species distribution model-30 ing, ecological niche modeling, Crimean Congo haemorrhagic fever 31 1 Introduction 32 In the last two decades, over a dozen statistical and machine learning methods have 33 been proposed for species distribution modeling (SDM). Over time, a handful of 34 methods have risen to predominance due to ease of implementation, computational 35 speed, and strong predictive performance in rigorous cross-validation. Some meth-36 ods are especially popular for specific applications, mostly because of disciplinary 37 tradition. For example, maximum entropy (MaxEnt) models are widely popular 38 for studies of global ecological responses to climate change. (VanDerWal et al., 39 2013; Warren et al., 2013) In disease ecology, boosted regression trees (BRTs) have 40 become the dominant tool for mapping vectors, reservoirs, and transmission risk of 41 infectious zoonoses and vector-borne diseases (Carlson et al., 2019; Pigott et al., 42 2014; Messina et al., 2016), largely due to an influential 2013 paper on dengue 43 virus. (Bhatt et al., 2013) 44 Boosted regression trees are easily implemented as an out-of-the-box machine 45 learning method, are powerful for ecological inference, and consistently perform 46 well in rigorous tests of SDM performance. The classification tree approach is also 47 more intuitive than the complex fitting procedures "under the hood" of MaxEnt 48 or Maxlike methods. However, BRTs also have downsides: they can be prone 49 to overfitting, and fitting procedures are largely handed down as anecdotal best 50 practices, with most studies choosing learning rates and tree depth based on the de-51 fault settings recommended by early work (Elith et al., 2008), with very few studies 52 selecting parameters from formal cross-validation with packages like caret. Fur-53 thermore, uncertainty is usually measured by ...

Associations Between Habitat Quality and Body Size in the Carpathian-Podolian Land Snail Vestia turgida (Gastropoda, Clausiliidae): Species Distribution Model Selection and Assessment of Performance

“…There exists a large suite of algorithms for modelling the distribution of species, but because there is no single 'best' algorithm some authors have reasonably concluded that niche or distribution modelling studies should begin by testing a suite of algorithms for predictive ability under the particular circumstances of the study and choose an algorithm for a particular challenge based on the results of those tests (Qiao et al, 2015). Accordingly, we assessed the relative performance of various categories of SDM algorithms: BIOCLIM (Busby, 1991;Booth et al, 2014), Generalized Linear Models (GLMs, Guisan et al, 2002), MaxLike (Royle, et al, 2012), Random forests (Breiman, 2001), Boosted Regression Trees (Elith et al, 2008), Support Vector Machines (SVMs; Vapnik, 1998), and Bayesian additive regression trees (BART, Carlson, 2020).…”

“…In computer science, BARTs are used for everything from medical diagnostics to self-driving car algorithms, however they have yet to fi nd widespread application in ecology and in predicting species distributions. Running SDMs with BARTs has recently been greatly facilitated by the development of an R package, 'embarcadero' (Carlson, 2020), including an automated variable selection procedure being highly eff ective at identifying informative subsets of predictors. Also the package includes methods for generating and plotting partial dependence curves.…”

Identifying Environmental Refuges (“Coldspots”) from Infection by Batrachochytrium Dendrobatidis of Amphibians in Eastern Europe

“…SDMs were generated by employing Bayesian additive regression trees (BART), a powerful machine learning approach. Running SDMs with BARTs has recently been greatly facilitated by the development of an R package, 'embarcadero' [13], including an automated variable selection procedure being highly effective at identifying informative subsets of predictors. Also the package includes methods for generating and plotting partial dependence curves and visualization called spatial partial dependence plots, which reclassifies predictor rasters based on their partial dependence plots, and show the relative suitability of different regions for an individual covariate.…”

Optimizing predictive models to prioritize viral discovery in zoonotic reservoirs

“…In particular, posterior width directly measures model uncertainty (rather than approximating it by permuting training data), and a single model can be run (instead of an ensemble trained on smaller subsets of training data), allowing the model to use the full training dataset all at once. 105…”

“…This often produces a much more reduced model without going through a stepwise variable selection process, which can be slow and very subject to stochasticity. 105…”