American Beech SDM DATA

This project presents a comprehensive spatial analysis and species distribution modeling (SDM) of American beech (Fagus grandifolia), integrating a range of geospatial and statistical methods to assess both current habitat suitability and future distribution projections under various climate scenarios. The workflow began by merging selected ecoregions to delineate the study area and aligning all spatial datasets to a common coordinate reference system. Raster datasets—including presence data and the target group background data—were processed by resampling to a coarser resolution (1km2) for ecological consistency and then masked to the study area, ensuring that only relevant data were analyzed.

Subsequent data processing involved extracting non‐NA pixel values, applying k‑means clustering and Fisher’s natural breaks classification to determine optimal thresholds for categorizing the presence and background raster data. These thresholds served as critical cut-offs for distinguishing between different levels of basal area—a key ecological parameter. The derived threshold values were then used to filter raster data, facilitating the extraction of presence (and background) points which were subsequently converted into spatial vector (sf) objects.

A pivotal aspect of the analysis was the measurement of spatial autocorrelation. This was accomplished through the computation of semivariograms for both the basal area and target group datasets. By fitting a spherical model to the semivariogram, the study was able to quantify the range of spatial autocorrelation, thereby providing insights into the spatial structure inherent in the ecological data. This analysis not only confirmed the degree of spatial clustering but also informed subsequent modeling efforts by highlighting the spatial dependency in the dataset.

Environmental variables were then carefully selected and processed, with multicollinearity among predictors being assessed using the variance inflation factor (VIF) and visualized through correlation matrices. The refined set of environmental predictors was integrated into the BIOMOD2 framework, where the modeling data were formatted to include both species presence–absence data and environmental layers. A diverse array of algorithms—including ANN, CTA, FDA, GAM, GBM, GLM, MARS, MAXENT, MAXNET, RF, SRE, and XGBOOST—was employed to develop individual models using a k‑fold cross-validation approach, ensuring robust model evaluation based on metrics such as TSS, ROC, Kappa, and Boyce.

Ensemble modeling strategies were also implemented. Selected models (GAM, GBM, GLM, MARS, MAXNET, and RF) were combined using both algorithm-specific ensemble approaches and an “all models” ensemble strategy, where predictions were weighted (with options to incorporate basal area weights) and then aggregated to produce ensemble forecasts. Variable importance metrics and response curves were generated to elucidate the influence of each predictor on species distribution.

Looking forward, the project incorporated future climate scenarios (SSP1, SSP2, and SSP3) for multiple time periods (2011–2040, 2041–2070, and 2071–2100). For each scenario, ensemble forecasting was performed to predict potential shifts in habitat suitability. In addition, a Multivariate Environmental Similarity Surface (MESS) analysis was conducted to assess the reliability of model predictions when extrapolating to novel future conditions. Finally, a range size analysis was carried out by comparing current and future model predictions, quantifying changes in the potential distribution area of American beech.

In summary, this study integrates advanced spatial data processing, semivariogram-based spatial autocorrelation analysis, multivariate statistical techniques, and ensemble species distribution modeling to provide a robust evaluation of American beech habitat suitability and projected distribution shifts in response to climate change.