Protected Areas Grasslands and Savannas Latin America
Description
Protected Areas (PAs) are important for maintaining ecosystem functioning and for harboring habitat for biodiversity, key aspects of these conservation strategies, especially for threatened and degraded ecosystems such as grasslands and savannas in Latin American landscapes. This study was developed for three lowland Latin American regions: the Casanare department in Colombia, the flooding Pampa in Argentina, and the ecoregions of Humid Chaco, Pantanal, Aquidabán, Litoral Central, and Selva Central in Paraguay. The data published here are summary tables of multitemporal averages of the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI) for trend comparisons between vegetation activity in grasslands inside and outside PAs across these three Latin American countries. This information was processed and exported from Google Earth Engine (GEE) using the MODIS/Terra Vegetation Indices 16-Day Global 250m product (MOD13Q1, Version 6.1; Didan, 2021), which is publicly available via https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MOD13Q1. Alongside this, there is a series of rasters of predicted species richness developed from Species Distribution Models (SDMs) for plants and birds in the study areas. Notably, based on this information, PAs show higher predicted plant and bird richness than in unprotected areas. The information in the predicted richness raster's pixels represents the number of species that could coexist in each pixel, making it valuable for future efforts to understand potential biodiversity distribution under different scenarios. The data on bird and plant occurrences that were downloaded from the Global Biodiversity Information Facility (GBIF) and used for SDMs training are also available, along with the R scripts and Python code used to run the analysis in GEE via Google Colab.
Files
Steps to reproduce
We identified stable grassland areas using LULC products from MapBiomas Colombia, Argentina, and Paraguay (1994–most recent year available per region), filtering grassy biomes (savannas, wetlands, pastures) that did not convert to woodland or plantations. Monocrops, oil extraction areas, forests, water bodies, and urban areas were excluded, assuming grasslands outside protected areas (PAs) are managed under extensive livestock and rotational grazing systems. This was implemented in Google Earth Engine (GEE) via Python's "geemap" package in Google Colab, then vectorized in R 4.60 (R Core Team, 2024) to produce reference polygons for stable grassland/savanna areas. Using these mapped areas, we generated 1000 random points inside PAs and 2000 outside them, extracting NDVI and EVI values from MODIS MOD13Q1 V6.1 (Didan, 2021) from 2000 onward via GEE as ANPP proxies to compare grassland productivity trends. Significant differences between categories over time were assessed via smooth differences—year-by-year differences between fitted GAM trajectories for protected vs. unprotected areas. To assess potential species richness, we ran Species Distribution Models (SDMs) using plant and bird occurrences from GBIF (2000–2025). Species with fewer than 10 observations were excluded; those with over 200 occurrences were subsampled to 200. Following Barbet-Massin et al. (2012), we used five-fold cross-validation, splitting occurrences across folds and sampling pseudo-absences (10:1 ratio) from areas >5 km from recorded occurrences. Models were built using the "maxent" R package with predictors from four sources: (i) 19 WorldClim v2.1 bioclimatic variables (Fick & Hijmans, 2017), downscaled to 250 m; (ii) mean and SD of NDVI from 2023 Sentinel-2 imagery; (iii) topographic variables (elevation, slope, aspect) from SRTM at 30 m (NASA, 2013); and (iv) grassland probability layers (natural, semi-natural, cultivated) from Global Pasture Watch (Parente et al., 2024). Spatial block cross-validation was used to compute accuracy metrics (ROC-AUC, AUPRC, normalized AUPRC) and determine optimal thresholds for binarization. After generating continuous suitability rasters (0–1), thresholds were averaged across cross-validation folds per species to create binary SDMs (bS-SDM), which were summed to produce predicted species-richness maps per taxonomic group and region. Biodiversity values were then extracted at randomly selected points within and outside PAs (Fig. 4).
Institutions
- Leibniz Centre for Agricultural Landscape ResearchBrandenburg, Müncheberg