Species distributions are an important EBV in the ‘species populations’ class. Knowing where species are is essential for understanding biodiversity patterns and informing conservation efforts. However, less than 10% of the world is well sampled, and even the longest running and well-sampled biodiversity observation networks have substantial data gaps. Information on species occurrences is often sparse and heavily spatially and taxonomically biased, necessitating the need for species distribution models (SDMs) to fill these data gaps and provide a better, less biased idea of where species are. SDM outputs be used as key base layers for a wide variety of purposes including: creating maps for sampling prioritization, quantifying the impact of environmental stressors on species, mapping habitat suitability for at-risk species, mapping biodiversity hotspots across the landscape, identifying the locations of conservation priorities and protected area expansion, identifying sampling gaps and the needed locations of future sampling, and calculating a range of biodiversity indicators including the Species Habitat Index (SHI), the Species Protection Index (SPI)
Methods:
SDMs predict where species are likely to occur based on a suite of environmental variables that are associated with known occurrences (Peterson, 2001; Elith and Leathwick, 2009). The MaxEnt pipeline pulls occurrences of the species of interest from GBIF and environmental raster layers from the GEO BON STAC catalog. Then, the pipeline cleans the GBIF data by only including one occurrence per pixel and removes collinearity between the environmental layers. Third, the pipeline creates a set of pseudo-absences (background points) and combines this with presences and the environmental predictors to create a dataset that is ready to be input into the SDM model. The pipeline runs the SDM on this data using the MaxEnt algorithm using the ENMeval R package (Kass et al. 2021). The MaxEnt SDM is run by 1) partitioning occurrence and background points into subsets for training and evaluation, 2) building the model with different algorithmic settings (model tuning), and 3) evaluating their performance (see package vignette). Lastly, the pipeline computes the 95% confidence interval using bootstrapping and cross validation techniques.
BON in a Box pipeline:
The BON in a Box pipeline allows you to run an SDM for a specific region and species (or multiple species) of interest. The pipeline has the following inputs:
The pipeline creates the following outputs:
Contributors:
Citations:
Elith, J., & Leathwick, J. R. (2009). Species Distribution Models: Ecological Explanation and Prediction Across Space and Time. Annual Review of Ecology, Evolution, and Systematics, 40(Volume 40, 2009), 677–697. https://doi.org/10.1146/annurev.ecolsys.110308.120159
Kass JM, Muscarella R, Galante PJ, Bohl CL, Pinilla-Buitrago GE, Boria RA, Soley-Guardia M, Anderson RP (2021). “ENMeval 2.0: Redesigned for customizable and reproducible modeling of species’ niches and distributions.” Methods in Ecology and Evolution, 12(9), 1602-1608. https://doi.org/10.1111/2041-210X.13628.
Peterson, A. T. (2001). Predicting Species’ Geographic Distributions Based on Ecological Niche Modeling. The Condor, 103(3), 599–605. https://doi.org/10.1093/condor/103.3.599
This document describes the methodology behind the BON-in-a-Box (BiaB) pipeline for using Boosted Regression Trees (BRTs) for species distribution modeling.
Summary
This pipeline builds a model to predict the distribution of a species (a type of essential biodiversity variable), by using occurrence data from the Global Biodiversity Information Facility (GBIF), and environmental predictors from an arbitrary STAC Catalogue.
In particular, this pipeline uses a specific model called a Boosted Regression
Tree (BRT), a machine-learning model which tends to work well with spatial data. The
details of how a BRT works are in the description of the key script in the
pipeline, fitBRT.jl.
Inputs:
Outputs
See an example pipeline output here
[!IMPORTANT]
Using BRTs to fit a species distribution model requires absence data. For the majority of species where no absence data is available, there are various methods to generate pseudoabsences (PAs) based on heuristics about species occurrence. However, the performance characteristics of an SDM fit using PAs can be widely variable depending on the method and parameters used to generate PAs. This means the results of BRT should be explicitly considered as a function of how PAs were generated, and sensitivity analysis to different PAs is highly encouraged.
Pipeline Steps
flowchart LR
a{input species} --> b[Load GBIF Occurrences]
c{input bounding box} --> b
d{input layers} --> e[Load Layers from STAC]
c --> e
b --> f[Clean presences]
f --> g[Generate Pseudoabsences]
c --> g
g --> h[Fit BRT]
e --> h
h --> i(predicted sdm)
h --> j(uncertainty map)
c --> k[create water mask]
h --> l[model fit statistics]
h --> m[diagnostic plots]
k --> hMethods: The species distribution modeling method provided in the package ewlgcpSDM (Effort-Weighted Log-Gaussian Cox Process) is based on spatial point processes and presence-only observations. It implements the method proposed by Simpson et al. (2016) to estimate log-Gaussian Cox processes using INLA (Rue et al. 2009) and the SPDE approach (Lindgren et al. 2009). The model relies on a discrete grid (the mesh) of arbitrary resolution to approximate the spatial component of the model. The method proposed in ewlgcpSDM contains three key aspects for species distribution modeling, namely:
The current version of the pipeline does not make use of the spatial component yet as some more work is needed to allow the adjustments necessary for the spatial component to work properly. The current version of the pipeline thus corresponds to an effort-weighted inhomogeneous Poisson point process.
BON in a Box pipeline: The pipeline is used to run an SDM for a set of species in a specific region and using a set of environmental predictors. Some inputs are yet to be added to the list of inputs required by the user. Currently, the pipeline mostly reuses the same inputs as the MaxEnt pipeline, namely:
The pipeline creates the following outputs:
See an example pipeline output here
Contributors:
Citations: Lindgren, F., Rue, H., and Lindström, J. 2011. An explicit link between Gaussian fields and Gaussian Markov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society Series B: Statistical Methodology, 73(4): 423-498.
Phillips, S. J., Dudík, M., Elith, J., Graham, C. H., Lehmann, A., Leathwick, J. and Ferrier, S. 2009. Sample selection bias and presence-only distribution models: implications for background and pseudo-absence data. Ecological Applications, 19(1): 181-197, https://doi.org/10.1890/07-2153.1
Rue, H., Martino, S. and Chopin, N. 2009. Approximate Bayesian Inference for Latent Gaussian models by using Integrated Nested Laplace Approximations, Journal of the Royal Statistical Society Series B: Statistical Methodology, 71(2): 319–392, https://doi.org/10.1111/j.1467-9868.2008.00700.x
Simpson, D., Illian, J. B., Lindgren, F., Sørbye, S. H. and H. Rue. 2016. Going off grid: computationally efficient inference for log-Gaussian Cox processes, Biometrika 103(1): 49–70, https://doi.org/10.1093/biomet/asv064