Measured and modelled rates of natural forest expansion in the European Alps (1800–2020)

General DescriptionThis repository contains spatial datasets supporting the manuscript Land abandonment and global warming drove two centuries of natural forest expansion in the Alps, published in Global Change Biology.The data provide both empirically measured and modelled estimates of natural forest expansion (NFE) across the European Alps over the past ~200 years in the form of rate of change (RoC; % forest change year-1), together with predictor variables used for statistical modelling.All spatial data are provided in ETRS89-extended / LAEA Europe (EPSG:3035) at 1-km resolution unless otherwise specified.Repository Content (openData/)Data include:Measured annual rate of change (RoC) of the compiled local cases as GeoTiff and CSV;Wall-to-wall spatial predictions of RoC with associated uncertainty levels (standard deviation of ensemble models and area of applicability) as GeoTiff;Reconstructed tree cover density (TCD) for 1800, 1870, 1950, 1980, and 2018 as GeoTiff.Spatial footprints of compiled case studies (Spatial Extents) as gpkg.Documented R code.1) ForestChange_predictors.csvSpreadsheet containing the forest RoC and predictors for the 1-km cells:ID_cell = the ID of the 1-km cellRefPeriod = reference period (e.g., 1800-2020)FORCHANGE_Absol = absolute value of forest RoC (% year-1)x, y = the x and y coordinates of the cell centroid in EPSG:3035CNTR_CODE = countryPark_pres = presence of a protected areaSOCIO_, CLIM_, VEG_, SOIL_, TOPO_, FORCOVER_INIT_ = socioeconomic, climate, vegetation, soil, topographic predictors and initial forest coverNotes:RoC represents normalized annual change in percent forest cover per 1-km grid cell.Elevation range restricted to 1000–3000 m a.s.l.No percentile-based outlier trimming was applied.This dataset was used for GLM inference and machine learning ensemble calibration.2) Measured_RoC/Folder with gridded measured RoC layers for compiled case studies (GeoTiffs). These rasters correspond to the harmonized 1-km INSPIRE grid and represent observed data used to calibrate models.3) spatPrediction/Folder with wall-to-wall gridded predictions (GeoTiffs) of annual forest cover change (RoC) across the entire European Alps at 1-km resolution. Cell values represent predicted weighted average of annual rate of forest cover change (% year⁻¹) derived from a two-stage ensemble machine learning framework (Random Forest, BRT, Lasso, GAM, BART). We provide:Average_Spat_Predict_* = weighted average of the ensemble predictions;StDev_Spat_Predict_* = pixel-wise standard deviation across the base learners, an indicator of model uncertainty;AoA_*_mask = area of applicability, a binary mask of the areas where environmental conditions are within the multivariate domain of the calibration data; predictions outside the AoA should be interpreted with caution.A README.md file is provided to support replicability.4) TCD/Folder with gridded reconstructed tree cover density (TCD) (% forest cover) at 1-km resolution for the years 1800, 1870, 1950, and 1980 + Copernicus TCD for 2018 (available at https://doi.org/10.2909/4dc35722-09ce-427f-9a1b-775a8640da27). Historical TCD layers were reconstructed by backward propagation of predicted RoC values from the 2018 Copernicus Tree Cover Density product. These layers represent estimated forest proportion (%) per grid cell and provide spatially explicit forest cover reconstructions across two centuries.5) VIMP-ALE_ML/ Outputs of the driver analysis (random forest per period): permutation variable importance (VIMP_*.csv), accumulated local effects (ALE2_*.csv) and cross-validated model performance (performance_*.csv).6) timeline_forestCover_bycountry.csv Relative forest cover per country at 1800, 1950, 1980 and 2020 in the compiled case studies (Total_Cells, Forest_Cells, Relative_Cover).7) boundaries/ Reference layers used by the code: Alpine Convention perimeter (alp_conv_perimeter/), NUTS level 2 boundaries (NUTS_lv2.gpkg), 1-km INSPIRE reference grids from ESDAC (inspire_1km/; see reference.txt) and the spatial footprints of the compiled case studies (CaseStudyExtents_1800/: CaseStudyExtents_1800/1950/1980/2020.gpkg, one polygon per case study, attribute: case_study). 8) inspire_1km_grid_alps.tif 1-km INSPIRE grid template for the Alps (EPSG:3035). 9) external/ dynamics.xlsx = population and temperature series used in the timeline of the synthesis figure; socioeconomic_data_LAU.gpkg = municipality-level socioeconomic statistics (Eurostat) with the population dynamics class (DIN_POP). 10) RCode/ R scripts and README to reproduce the analyses and figures (run order 00-13): RoC calculation, predictor table, spatial autocorrelation range, GLM (glmmTMB), machine-learning drivers and interpretation, spatial ensemble, TCD back-cast, treeline analysis, comparison with HILDA+, synthesis figure. The README states which scripts can be run from this repository and which public datasets must be downloaded for the others.Methodological SummaryForest cover maps derived from historical cadastral maps, aerial photography, and satellite imagery were harmonized to a binary forest / non-forest classification and aggregated to a standardized 1-km INSPIRE grid. Annual rate of change (RoC) was calculated as the difference between final and initial forest cover of a certain period divided by the temporal extent of that period (number of years). Spatial predictions were generated using a two-stage ensemble machine learning framework implemented in the mlr3 ecosystem, including spatial autocorrelation as explicit covariate and spatial block cross-validation for model evaluation. Full methodological details are provided in the associated publication.Data Interpretation NotesRoC values represent normalized annual rates and do not imply linear trajectories within each period.Forest definition harmonized across source datasets; slight differences in historical forest definitions may persist.Predictions should be interpreted within the Area of Applicability (AoA).Model uncertainty increases toward environmental margins and sparsely sampled regions.Code and Data AvailabilityThe R code is provided in RCode/. Land-cover maps used to derive the RoC rasters (harmonization step) are not shared, so scripts 01, 02 and 07 document the derivation of the deposited data but cannot be re-run from this repository. Predictor source data are openly available: Copernicus Tree Cover Density 2018 (https://doi.org/10.2909/4dc35722-09ce-427f-9a1b-775a8640da27), CHELSA TraCE21k (https://chelsa-climate.org/chelsa-trace21k/), CHELSA v2.1 BIOCLIM+ (https://chelsa-climate.org/exchelsa-extended-bioclim/), CHELSA v2.1 monthly time series (https://chelsa-climate.org/timeseries/), SoilGrids v2.0 (https://soilgrids.org/), SRTM GL1 (https://www.earthdata.nasa.gov/data/catalog/lpcloud-srtmgl1-003), Eurostat municipal statistics and agricultural census (https://ec.europa.eu/eurostat), Copernicus GHS-BUILT-S (https://ghsl.jrc.ec.europa.eu), OpenStreetMap (https://www.openstreetmap.org), World Database on Protected Areas (https://www.protectedplanet.net). HILDA+ and the Pecher et al. (2011) treeline data, used in the comparison and treeline analyses, must be downloaded from their original sources.