Detection of early warning signs of ecosystem transformations using remote sensing
data/ stores raster inputs, output/ holds intermediate artifacts and model results, and scripts/ contains the analysis workflow.scripts/01_load_rasters.R loads and aligns NDVI, soil moisture, drought, transition, fire, and static driver rasters that support later processing steps.scripts/02_extract_ews_and_drivers.R sources the raster loaders, computes early warning signals (autocorrelation, variance, NDVI slope), estimates signal emergence, and compiles pixel-level feature tables—including static drivers and fire frequency—before fitting an exploratory GAM for lead time.scripts/03_fit_bart_model.R trains a Bayesian Additive Regression Trees (BART) model on the engineered features to predict warning lead times and saves the fitted model.scripts/04_predict_lead_time_raster.R projects model predictions back into raster space for mapping.scripts/full_primary_workflow.R ties together metric extraction, GAM modeling, and ROC evaluation with fire vs. non-fire stratification for reproducing the core analysis.build_leadtime_bart_model.R, model_accuracy_fire.R, prediction_sup2.R, and bart_to_predict.R explore alternate feature sets, diagnostics, accuracy comparisons, and partial dependence visualization.terra package for raster I/O, alignment, and pixel-wise time series extraction.detect_ews_year and compute_ews.dbarts and mgcv to understand how Bayesian Additive Regression Trees and Generalized Additive Models are used to interpret environmental drivers and lead-time predictions.targets or drake) and replace placeholder raster paths with actual datasets for reproducible large-scale processing.