Predicting Global Pine Distributions Under Climate Change
A machine learning framework that predicts how climate change will affect the global distribution and richness of pines (genus Pinus) by 2070.
Problem
Species distribution models learn where a species can live from the climate at the places where it occurs, and then project that relationship onto future climates. Three things make those projections unreliable if handled naively: the choice of algorithm changes the predictions, the climate models and emission scenarios disagree with each other, and nearby occurrence records inflate performance estimates through spatial autocorrelation. Future climates also include conditions that do not exist today, where a model trained only on present data has nothing to learn from.
Approach
I combined global occurrence records for all pine species collected across the globe with climate and soil layers (GIS), and reconstructed the past evolution of each species' climatic niche from the pine phylogeny. This reconstructed niche information was used where model predictions were most uncertain.
For every species I fitted an ensemble of random forests, GAMs and GLMs across multiple cross-validation partitions, so the uncertainty from algorithm choice and data partitioning carries into the predictions. Each model was then projected under multiple CMIP5 climate models and emission scenarios (RCPs) for 2070, yielding hundreds of thousands of projections (species × algorithm × partition × climate model × scenario), run on an HPC cluster with Slurm and Apptainer.
To get an honest estimate of performance, the models were also evaluated on an independent dataset from regions on other continents that were never used in training.
Results
- High predictive performance: AUC ≥ 0.98, Kappa ≥ 0.75 and TSS ≥ 0.90
- On the independent dataset, performance was significantly better than random expectation, and adding the reconstructed past niches improved it further
- Projected changes in the global distribution and richness of pines by 2070, published in Diversity and Distributions (2024)
Solution Overview
Occurrence records (Pinus species) + climate & soil layers (GIS)
→ Past climatic niche reconstruction (pine phylogeny)
→ Ensemble per species: random forest / GAM / GLM × CV partitions
→ Projections: CMIP5 climate models × RCP scenarios (2070)
→ hundreds of thousands of projections on Slurm (Apptainer)
→ Validation on independent data from regions not used in training
→ Global maps of future distribution and richness
Tech Stack
- R, Bash
- Random forests, GAMs, GLMs
- GIS (raster and vector processing)
- Frequentist and Bayesian statistics, Monte Carlo simulation
- Slurm, Apptainer
-
Let's have a virtual coffee together!
Want to see if we're a match? Let's have a chat and find out. Schedule a free 30-minute session to discuss your data challenges and explore how we can work together.