Data and code for "Losing the Amazon twice: hybrid land-use modelling and causal matching reveal cascading threats to forests under weakened protection" (Amazon basin, 1 km, 1986–2023 and 2030–2090 projections)

This record contains the spatial data, model outputs and analysis code underlying the article "Losing the Amazon twice: hybrid land-use modelling and causal matching reveal cascading threats to forests under weakened protection" (do Couto, Durán-Díaz & de Vries; submitted to Applied Geography, 2026). The study projects land-use change across the entire Amazon basin (RAISG Cuencas boundary; ~7.0 million km²; ~1 km resolution) with a hybrid Markov chain–Random Forest–Cellular Automata framework calibrated with 38 annual MapBiomas Amazônia maps (1986–2023), compares a full-protection scenario (S0) with a no-protection scenario (S1) for 2030, validates the effect of protected areas and Indigenous Territories with propensity score matching on observed deforestation (2000–2020), and quantifies projected forest loss within floodplain and riparian zones. Contents of Losing_the_Amazon_Twice_Data_and_Code.zip 01_LULC_rasters/ — 38 annual land-use/land-cover maps (1986–2023, 8 classes, derived from MapBiomas Amazônia Collection 6.0), persistence and change-frequency layers, the hindcast map (2020 projected from 2010; FoM = 0.88, accuracy = 0.94), projections for 2030, 2050, 2070 and 2090 under current protection, and the 2030 counterfactual maps S0 (full protection) and S1 (no protection). The partial-protection variants S2–S5 are included for completeness; they share the demand of S1 and only redistribute deforestation spatially (they are not used to rank protection categories in the article). 02_Zones_and_protection/ — basin mask, GFPLAIN250m floodplain mask, riparian masks (HydroRIVERS buffered at 500, 1000 and 2000 m), presence and distance rasters for protected areas and Indigenous Territories (RAISG 2020: any protection, PAs, ITs, strict-use, sustainable-use) and zone pixel counts. 03_Susceptibility_surfaces/ — Random Forest susceptibility surfaces (0–1) for the forest-source transitions (Forest Formation → Flooded Forest, → Farming, → Other Natural) from the baseline model (covariates D1–D7), the counterfactual model (D1–D8) and the S0/S1 scenario surfaces; the remaining transitions are reproducible with the code. 04_Markov_and_bootstrap/ — zone-specific Markov transition matrices (protected, unprotected, recent-weighted and constrained versions; NPZ and JSON), the 25 bootstrap ensemble members and the uncertainty summary (2030–2090). 05_Diagnostics_and_source_data/ — diagnostic reports of pipeline steps 1–5 and of the matching analysis (step 7), figure source data (JSON) and the aquatic-zone results. 06_Supplementary_Data/ — the nine supplementary data files submitted with the article (source data workbook, counterfactual areas, Markov demand, matching results, feature importance and cross-validation AUC, validation metrics, RF parameters, 2020 baseline areas, pipeline configuration). 07_Code/ — the complete pipeline (Python 3.9+; numpy, scipy, scikit-learn, rasterio, joblib): data preparation, Markov matrices, Random Forest susceptibility, Cellular Automata projection, counterfactual scenarios, bootstrap uncertainty, matching analysis, aggregation and figures; configuration file and batch runner. Grid and formats. All rasters: EPSG:4326 (WGS84), 3762 × 2875 pixels, ~0.00898° (~1 km), LZW-compressed GeoTIFF; nodata = 0 (LULC), 255 (masks), −1 (susceptibility). LULC codes: 1 Forest Formation, 2 Flooded Forest, 3 Wetlands, 4 Water, 6 Farming, 7 Urban, 8 Other Anthropic, 9 Other Natural ("forest" in the article = classes 1 + 2). Sources. MapBiomas Amazônia Collection 6.0; RAISG PA_TI_RAISG_2020; HydroRIVERS v1.0 (Lehner & Grill, 2013); GFPLAIN250m (Nardi et al., 2019); CHELSA v2.1; FABDEM; SoilGrids; WorldPop; MODIS NDVI. Please cite the article and this dataset when using these data.