To select low-correlated bioclimatic variables for species distribution modeling (SDM), you can use the usdm R package to calculate variance inflation factor (VIF) and remove multicollinear variables. This improves model interpretability and avoids redundancy in predictors.

one would emphasize geospatial reasoning, multivariate data structure, and landscape evolution context. Below is a detailed explanation framed with a geoscientific rigor, integrating recent research insights: 🔷 1. Contextualizing Bioclimatic Variables in Landscape and Tectonic Systems As structural geologists, we understand that species distributions are constrained not only by climatic variables (temperature, precipitation, seasonality) but also by tectonically controlled topography, lithological heterogeneity, and geomorphic evolution. These climatic variables often covary due to orographic effects, latitude, or climatic gradients inherited from past tectonic uplift (e.g., the Andes or Himalayas). Therefore, before any modeling, it is critical to understand the covariance structure of the bioclimatic variables—PCA helps isolate orthogonal climatic gradients that may control species niches. 🔷 2. Why Use PCA? Bioclimatic layers from datasets like WorldClim 2.1 or CHELSA are highly collinear. For instance: Bio1 (Mean Annual Temperature) and Bio10 (Mean Temp of Warmest Quarter) are almost always strongly correlated. Bio12 (Annual Precipitation) and Bio16/17 (Precipitation of Wettest/Driest Quarter) show similar redundancy. Multicollinearity: Inflates variance in SDM models. Obscures which variables are truly driving the species’ response. Increases model complexity and reduces transferability across space/time. Thus, PCA allows us to: Derive independent environmental axes (Principal Components). Avoid arbitrary variable selection. Retain ecologically meaningful gradients. 📌 Key Reference: Zurell, D., et al. (2020). "A standard protocol for reporting species distribution models." Ecography. DOI:10.1111/ecog.04960 This paper recommends PCA as a robust method to address multicollinearity prior to SDM calibration. 🔷 3. Procedure for Correlating Bioclimatic Variables via PCA a. Extract Bioclimatic Data Extract from raster stacks at species presence points or across the entire study region to understand background environmental variation. In regions with high relief (e.g., fold-thrust belts), include topographic derivatives like elevation, slope, or ruggedness to test their correlation with bioclim variables. b. Standardize and Run PCA Perform centered and scaled PCA (prcomp() in R or PCA() in Python’s scikit-learn) to ensure variables on different scales are treated equally. Interpret the loadings matrix: Variables that cluster together on PC1/PC2 indicate strong correlation. PC1 may capture temperature gradient, PC2 may capture precipitation seasonality. c. Use PCA for Correlation Structure There are two workflows from here: Variable Selection Based on PCA Loadings Retain one variable per group of strongly loading variables. Drop collinear ones. (e.g., keep Bio1, drop Bio5 & Bio6) Use Principal Components as Predictors in SDM Especially useful when projecting into novel climates (e.g., paleo-SDM or future climate projections). PCs are linearly uncorrelated, which simplifies model interpretation. 📌 Key Reference: Braunisch, V., Coppes, J., et al. (2013). "Selecting from correlated climate variables: a major source of uncertainty for predicting species distributions under climate change." Ecography, 36(9), 971–983. Highlights PCA as a solution to climate variable correlation issues in mountainous landscapes. 🔷 4. Geologically Informed Considerations In tectonically active or paleoclimatically dynamic regions, species’ niches may not be in equilibrium with current climate. Thus, you may: Compare PC gradients with tectonostratigraphic zones to explore controls on endemism. Integrate PCA-derived climatic space with substrate constraints (e.g., carbonate vs. siliciclastic terrains). Validate PCA axes with modern ecoregions or biogeographical provinces. 🔷 5. Example: Alpine Biodiversity Gradient In the Alps, temperature lapse rates, seasonal snow cover, and aspect—all heavily driven by structural relief—co-vary with Bio1, Bio4, Bio10, etc. PCA in SDM studies here (e.g., Mateo et al., 2016) often shows PC1 as a latitudinal temperature axis, PC2 as precipitation seasonality, reflecting glacial refugia patterns and postglacial dispersal. 📌 Reference: Mateo, R.G., et al. (2016). "Integrating species distribution modelling and landscape connectivity analysis for habitat conservation and restoration." Diversity and Distributions, 22(7), 703–717.

Indeed, before modeling, it is recommended to preselect bioclimatic or abiotic variables to avoid multicollinearity noise. It has been proven that the variance inflation factor (VIF) will help you with this selection, and PCA also helps you see their influence. This will aid in this initial selection and avoid multicollinearity noise.

Tristan Millet I really recommend this article for working with SDM: Article chooseGCM: A Toolkit to Select General Circulation Models in R

I'd like to clarify two points. Using the Pearson correlation matrix and the VIF index, taking into account the species' ecology, is sufficient to select variables while avoiding collinearity. PCA can be useful for reducing dimensionality and establishing some relationships between variables.