🌲🌉 Species distribution models with BART 🌉 🌲
Colin J. Carlson (Georgetown University)
Last updated February 2020
How do I install?
# install.packages('devtools')
devtools::install_github('cjcarlson/embarcadero')Some current FAQs
Q: Where did the vignette go?
A: The NLMR package appears to be having some maintenance issues with upstream dependencies, so alas, the vignette is gone. It's a shame the R-spatial world is gradually falling apart due to dependency issues right now.
Q: Where did bigstack() go?
A: embarcadero no longer requires the velox package, which unfortunately has fallen into disrepair, and the bigstack() function is therefore gone by popular demand. One of the best things anyone could do for the R spatial community would be to take over velox maintenance or create a new clone of it.
Q: My predictions are all generating as 0.5. What should I do?
A: This is an issue you can find an answer to in the documentation of the dbarts package. Essentially, you need to "touch" part of the model object before you save it. For a bart() model x, just run invisible(x$fit$state) (or for rbart_vi(), run invisible(x$fit[[1]]$state)) before you save the model. If you do this, you should be okay making predictions with a saved model (though due to stochasticity, they may be slightly different precise values).
Other stuff
What's BART?
Bayesian additive regression trees (BARTs) are an exciting alternative to other popular classification tree methods being used in ecology, like random forests or boosted regression trees. Whereas boosted regression trees fit an ensemble of trees each explaining smaller fractions of variance, BART starts by fitting a sum-of-trees model and then uses Bayesian "backfitting" with an MCMC algorithm to create a posterior draw.
Why BART?
BART does well with model-free variable selection and handles irrelevant predictors well; the Bayesian aspect also comes with perks, like posterior distributions on predictions (without having to bootstrap, like you do with BRTs) and automated confidence intervals on partial dependency plots. So far, it also seems to avoid the problem BRTs have for niche modeling: with no automated selection of tree depth or tree complexity, BRTs tend to overfit especially with randomly-generated pseudoabsences and ensembling over subsets. In previous "bake-offs," BART commonly outperforms other methods, including comparable classification tree methods like random forests and boosted regression trees.
Are BARTs new to ecology?
No, there is nothing new under the sun. Yen et al. used BARTs to examine habitat selection of birds back in 2011. But so far, no, no one is using them for species distribution modeling (as far as I can tell).
What does embarcadero do?
This package is basically a wrapper around 'dbarts' with a few tools
- basic model summary statistics and diagnostics
- spatial prediction with raster data
- credible interval draws from the posterior distribution
- visualization of how posterior draws learn over time
- variable importance measures and plots
- stepwise variable elimination
- automatic Nice Plots for partials, including multiple ways to visualize posterior draws
- spatial projection of partials ("spartials")
- compatibility with random intercept BART models (riBART)
- plots for random intercepts
In future versions I'd hope to include compatibility with:
- explicitly-spatial adaptations of BART (spatial priors)
- compatibility with smoothed BART models (softBART) and sparse BART models with Dirichlet priors (DART)
How do I learn to use embarcadero?
The paper at Methods in Ecology and Evolution includes a bunch of code, and there's a new vignette included in the package that runs through all of the basic functionality using a virtual species. I've migrated the older, more advanced vignette - which is over 30 pages and takes a good 3 hours to run on an older machine - to another repo (cjcarlson/pier39) and am in the process of updating it to be much clearer.
How do I cite embarcadero?
The embarcadero package is published at Methods in Ecology and Evolution!
Can I help?
Please! Reach out to colin.carlson@georgetown.edu if you want to help with development or otherwise are interested in being one of the first users.
Update (February 2023): Hi all! Thanks for your notes over the past three years. I've been out of the office soaking in my tub, but I'm hoping to make some progress updating embarcadero to current best practices in the coming weeks. In the long run, I'd be interested in potentially handing the package off to someone else to curate, as my availability to do so is very limited - if you think that person should be you, maybe drop me a note.
Why is it called embarcadero?
Because of Embarcadero BART station, and because I'm homesick for Humphry Slocombe.