Quantifying Dryland Niche Space and Spatial Connectivity Under Climate Forcing: A Hybrid AI–Mechanistic Ensemble Approach — Synthetic Simulation Code and Output Data (v1.0.0)

Abstract

This repository contains the complete simulation code and datasets supporting the manuscript "Quantifying Dryland Niche Space and Spatial Connectivity Under Climate Forcing: A Hybrid AI–Mechanistic Ensemble Approach." The package includes an AI-enhanced mechanistic ensemble biodiversity modelling framework that integrates ecological niche theory (40% weighting) with machine learning algorithms (Random Forest 30%, Gradient Boosting 30%) to project habitat suitability for 50 European dryland specialist species across four climate scenarios (RCP2.6, RCP4.5, RCP6.0, and RCP8.5), with climate trajectories spanning 1995 to 2024 and biodiversity prediction outputs spanning 1995 to 2020. All species traits, site characteristics, and climate trajectories in this repository are fully synthetic, generated internally by the deposited Python script through parametric random sampling under a fixed seed (BASE_SEED = 42). No external climate, occurrence, or trait databases were queried or ingested at any stage of this pipeline. Contents: [simulation_script_filename].py — complete Python implementation of the EnhancedBiodiversitySimulator class, including species/location/climate generation, mechanistic suitability calculation, Random Forest (100 trees, maximum depth 10) and Gradient Boosting (100 trees, maximum depth 6, learning rate 0.10) training, ensemble weight-sensitivity grid search, ten-repeat robustness checking, and hotspot identification. Nine simulation output datasets (CSV format): species_characteristics_[timestamp].csv — eco-physiological traits for 50 simulated species location_data_[timestamp].csv — geographic and environmental attributes for 100 simulated locations climate_projections_[timestamp].csv — climate trajectories for four RCP scenarios, 1995–2024 biodiversity_predictions_[timestamp].csv — habitat suitability, occurrence probability, and extinction risk predictions validation_metrics_[timestamp].csv — TSS, AUC-ROC, sensitivity, specificity, and Cohen's kappa by scenario and year biodiversity_hotspots_[timestamp].csv — conservation priority hotspots with species richness and protection status ensemble_weight_sensitivity_[timestamp].csv — single-run AUC-ROC, TSS, and composite index L across eight tested mechanistic-weight configurations (0.20 to 0.80, including the equal-weight node) weight_sensitivity_robustness_summary_[timestamp].csv — mean and standard deviation of AUC, TSS, and composite index L across 10 repeated resampling runs per weight configuration weight_sensitivity_robustness_all_runs_[timestamp].csv — complete per-repeat, per-configuration results underlying the robustness summary Documentation (README, requirements.txt, LICENSE) The mechanistic component integrates thermal performance functions, precipitation requirements, drought sensitivity, habitat fragmentation and land-use penalties, and a protected-area enhancement multiplier with the two machine learning components to generate ensemble habitat suitability, extinction risk, and conservation hotspot outputs. Model validation employed five-block spatial cross-validation achieving a mean True Skill Statistic of 0.578 ± 0.019 and a mean AUC-ROC of 0.881 ± 0.007 across 100 evaluation locations; the selected ensemble configuration achieved an overall AUC-ROC of 0.883. This code and data are fully reproducible. Running the Python script generates new simulation outputs with current timestamps; setting the fixed base seed reproduces the results reported in the associated manuscript exactly. Computational environment: Python 3.14.6, scikit-learn 1.4.1, NumPy 1.26.4, SciPy 1.12.0. Full simulation runtime is approximately 25 minutes on standard consumer hardware; individual component runtimes range from under 1 minute (mechanistic model) to approximately 18 minutes (standalone Gradient Boosting). For citation and usage instructions, see README.md. All code is released under the MIT License.