Bayesian Additive Regression Trees (BART) are a powerful machine learning technique with very promising applications in ecology and biogeography in general, and in species distribution modelling (SDM) in particular. BART can produce highly accurate predictions without overfitting to noise or to particular cases in the data. Notably, unlike most SDM methods, BART generally shows a well-balanced performance regarding both main aspects of predictive accuracy: discrimination (i.e., distinguishing presence from absence localities) and calibration (i.e., having predicted probabilities reflect gradual occurrence frequencies across space and environment). Moreover, the Bayesian framework inherently handles prediction uncertainty, and it has a built-in complexity penalty with very sensible defaults, freeing the user from arbitrary or intensively cross-validated parameter choices.
This workshop will take participants through a worked example, from essential data preparation to model output analysis, using sample data and annotated R scripts. These scripts can be adapted on-the-spot by participants to work on their own species presence-only or presence-(pseudo)absence data and predictor variables (but mind that computation time can be very large for large datasets). We’ll prepare species occurrence and environmental data, compute and evaluate BART distribution models, identify influential predictors, map prediction uncertainty, plot partial response curves with Bayesian credible intervals, and map relative presence probability regarding particular predictors. All sessions will include both theoretical lectures and hands-on practicals in R.
Depth of knowledge needed by attendees: This workshop requires basic experience using and modifying R scripts; Basic knowledge about species distribution (or ecological niche) models; A computer with recent versions of R, RStudio, and R packages ‘embarcadero’ and ‘terra’ already installed; A good enough internet connection for live video sessions, and preferably a webcam for enhanced interactivity.
Level of detail of workshop: Intermediate
Mode of organization: Online on Zoom, with a Slack space for workshop materials, questions and discussion
Length of workshop: 7.5 hours, divided into three 2.5-hour blocks (on three separate days)
Dates: 17 – 19 December
Time: 2:00 – 4:30 p.m. GMT
