PNV

PNV maps potential natural vegetation by applying machine learning to estimate vegetation cover in equilibrium with climate in the absence of human influence.

Key Features:

• Machine Learning Integration: Implements and compares neural networks (nnet), random forest (ranger), gradient boosting (gmb), K-nearest neighbors (class), and cubist algorithms.
• Case Studies: Applied to three datasets: BIOME 6000 global biomes with 8,057 modern pollen-based site reconstructions, European forest tree taxa with 1,546,435 ground-occurrence records covering 73 species, and global monthly FAPAR from 30,301 randomly sampled points.
• Explanatory Variables: Uses a stack of 160 global maps representing atmospheric, climatic, relief, and lithologic variables as predictors.
• Performance Metrics: Random forest (ranger) showed the best overall performance; BIOME 6000 (20 classes) accuracy ranged from 33% with spatial cross-validation to 68% with simple random subsetting with key predictors including total annual precipitation, monthly temperatures, and bioclimatic layers; forest tree species mapping achieved 25% accuracy with major predictors monthly cloud fraction, mean annual and monthly temperatures, and elevation; FAPAR regression models attained R² ≈ 90% with total annual precipitation, monthly cloud fraction, CHELSA bioclimatic layers, and month of year as primary predictors.
• Spatial Resolution: Generates global maps at 1 km spatial resolution.

Scientific Applications:

• Land potential and degradation assessment: Estimates potential vegetation to inform assessments of land potential and land degradation under natural conditions.
• Ecological research and conservation planning: Supports studies of vegetation dynamics, conservation planning, and environmental policy analysis.
• Predictive studies and scenario analysis: Enables incorporation of future climate scenarios through dynamic modeling for predictive and scenario-based analyses.

Methodology:

Models were built using machine learning algorithms (nnet, ranger, gmb, class, cubist) trained on a stack of 160 global predictor maps and the three input datasets (BIOME 6000: 8,057 pollen-based reconstructions; European tree occurrences: 1,546,435 records; FAPAR: 30,301 points), evaluated with spatial cross-validation and simple random subsetting, and used to produce 1 km resolution maps of biomes, tree distributions, and FAPAR.