Mountain JSDMs

Joint Species Distribution Models for alpine vegetation along elevational gradients.

Calanda JSDM

Code for the manuscript:

Simultaneous disentangling processes behind species ranges and community assembly along an elevational gradient

Billur Bektas, Maximilian Pichler, Florian Hartig, Mikko Tiusanen, Camille Brioschi, Jake M. Alexander, Janneke Hille Ris Lambers

Model framework

Species distributions and community assembly are modeled using the sjSDM package (Pichler & Hartig, 2021), a GPU-accelerated joint species distribution model that decomposes variance into environment, spatial, and species association components.

All model fitting and experiments are run on Google Colab (GPU: A100, default RAM) via the notebook:

• Calanda_JSDM/R/02_model/06_sjsdm_experiments_colab.ipynb

The Colab notebook runs 6 experiments: (1) full grid search over spatial form, alpha, and lambda; (2) decoupled lambda sensitivity; (3) anova sampling saturation; (4) model fit sampling saturation; (5) drop-one environmental variable; (6) 10-fold cross-validation with species AUC and site log-loss evaluation.

Anova bug fix

The sjSDM anova function (v1.0.6) has a bug on line 346 of anova.R: inherits(object, "spatial ") has a trailing space, causing spatial models to never receive spatial_formula = ~0 in the null model. This is patched at runtime in the Colab notebook (cell 13). See: TheoreticalEcology/s-jSDM#172

Data dimensions

Stage	Species	Sites
Raw vegetation data	723 binomial	624 plots
After 5% prevalence + 70% coverage + env matching (Y matrix)	168	529
After AUC >= 0.7 filter	158	--
After logloss <= log(2) + ci_width < 0.1 filter	--	503
Environmental gradient analysis (script 09)	158 species, 503 sites
Functional trait analysis (script 10)	115 species, 374 sites

---

Directory structure

Calanda_JSDM/
├── R/
│   ├── 00_setup/
│   │   ├── 00_workflow.R              # Coordinator with TRUE/FALSE toggles
│   │   └── functions_calanda.R        # Shared helper functions
│   │
│   ├── 01_data_prep/
│   │   ├── 01_prepare_TRY_traits.R
│   │   ├── 02_prepare_field_trait_data.R
│   │   ├── 03_merge_and_assess_traits.R
│   │   └── 04_prepare_climate_vegetation_data.R
│   │
│   ├── 02_model/
│   │   ├── 05_test_cpu_jsdm.R                    # Local CPU test
│   │   └── 06_sjsdm_experiments_colab.ipynb       # Colab GPU fitting + experiments
│   │
│   ├── 03_analysis/
│   │   ├── 07.1–07.7_post_exp*.R                 # Model experiment post-analysis
│   │   ├── 07_modeling_experiments.R              # Consolidated experiment figures
│   │   ├── 08_variance_partitioning.R             # Venn diagram, VP violins, scatter plots
│   │   ├── 09_environmental_gradient_analysis.R   # VP vs env predictors and species betas
│   │   └── 10_functional_post_analysis.R          # VP vs species traits and community traits
│   │
│   ├── 04_visualization/
│   │   └── 11_map.R                               # Study area map with RGB climate triangle
│   │
│   ├── 05_validation/
│   │   ├── validation_checks.R        # Pipeline integrity checks
│   │   ├── pipeline_attrition.R       # Species/site attrition tracking
│   │   └── vp_diagnostic_figures.R    # Discard vs proportional VP diagnostics
│   │
│   └── archive/                       # Deprecated scripts
│
├── plot/                              # All figures (PDFs)
├── output/                            # Data outputs (gitignored)
├── data/                              # Raw data (gitignored)
└── docs/                              # Documentation (gitignored)

---

Pipeline steps

Step 1: Data preparation

Script	What it does
01_prepare_TRY_traits.R	Fetches trait data from TRY database and ecological indicators/dispersal from floraveg.eu
02_prepare_field_trait_data.R	Processes field-collected trait measurements (leaf area, LMA, LDMC, CN)
03_merge_and_assess_traits.R	Merges TRY + field traits, computes species medians and kCV, imputes missing values
04_prepare_climate_vegetation_data.R	Builds X and Y matrices from remote sensing and vegetation surveys; computes community traits and functional distinctiveness

Step 2: Model fitting

Script	What it does
05_test_cpu_jsdm.R	Local CPU test of model structure
06_sjsdm_experiments_colab.ipynb	Full model fitting + 6 experiments on Colab GPU (A100)

Step 3: Model experiment post-analysis

Script	What it does
07.1--07.7	Individual post-analysis for each experiment
07_modeling_experiments.R	Consolidated publication-ready figures

Step 4: Post-model analysis

Script	What it does
08_variance_partitioning.R	Aggregates VP across 10-fold CV, Venn diagram with mean R² values, species/site violins and scatter plots
09_environmental_gradient_analysis.R	AIC stepwise regressions: community VP vs 12 env predictors (unweighted); species VP vs species betas (weighted)
10_functional_post_analysis.R	AIC stepwise regressions: species VP vs trait medians, kCV, distinctiveness (weighted); community VP vs trait means and variances (unweighted)

Step 5: Visualization

Script	What it does
11_map.R	Study area map: greyscale DEM, RGB climate triangle (summer temp / ET / FDD), open/closed habitat shapes, species richness sizes

Step 6: Validation (optional)

Script	What it does
validation_checks.R	Data alignment, join correctness, VIF, residual normality, heteroscedasticity checks
pipeline_attrition.R	Species and site attrition from raw data through all filters, with reasons
vp_diagnostic_figures.R	Discard vs proportional allocation comparison; raw shared fraction distributions

---

Execution

Run 00_workflow.R to execute the full pipeline. Toggle steps at the top:

run_data_prep     = TRUE   # Step 1
run_model         = FALSE  # Step 2 (GPU, use Colab)
run_experiments   = FALSE  # Step 3
run_analysis      = TRUE   # Step 4
run_visualization = TRUE   # Step 5
run_validation    = FALSE  # Step 6

---

Last updated: 2026-03-24