Study the species distribution model

I am trying species distribution model recently and generating some nice figure, so I think my blog is a good place to store these pretty things.

Platform

Most ecologists use R and there is a curated list of packages¹².

Main Steps

• Presence-absence data & Environmental Data
• Fit model (which could use different algorithms)
• Project model to past/present/future environmental data

More than data-model-projection

• Variable selection: remove highly correlated and low-importance variable/feature
• Data partition: split data into train data (to feed the model); validation data (for model tuning); test data (an independent examination dataset for the result of train+validation). Cross validation is to use different methods (randomly or spatially, or environmentally) to partition the data and fit/evaluate data.
• Hyper-parameter of model: to tune a range of parameters in each specific model algorithm
• Model evaluation: use AUC/TSS/R2/RMSE etc.
• Ensemble model: train models using different data partition strategies/model algorithms/hyperparameters and pick the best/mean output

Planktic foraminifera case study

I use core-top data and convert them into presence/absence data. Then I model two species (G. ruber white/G. bulloides) using SST, SSS, NPP, Chl, phytoplankton biomass, ice cover, pH (both use Maxent model). Both models achieve AUC score at around 0.9, and SST are the most important variable for both species. So this is basically a just-for-fun version of Waterson et al. (2017) reconstruction³. Here is the project result in the modern ocean: clearly G. ruber (white) lives in the (sub)tropical gyres and G. bulloides lives in the subpolar ocean and some productive upwelling regions (California current system, Humboldt current system, Benguela current system, and Arabian Sea).

References