Spatially evaluated statistical reconstruction of Quaternary vegetation with PaleoVeg

Statistical vegetation–climate models translate palaeoclimate simulations into vegetation estimates for reconstructing Quaternary environments. We present PaleoVeg, which builds on the probabilistic approach of Shao et al. (2018) using a Random Forest and WorldClim 2.1 bioclimatic variables. It predicts probabilities for 11 vegetation and bare-ground classes and option- ally applies class-specific CO2 response factors. We compare configurations using annual mean temperature and precipitation alone or 17 bioclimatic pre- dictors. Both use identical five-fold spatial cross-validation and a separate test comprising 132 withheld five-degree tiles. Expanding the predictor set in- creases mean cross-validation accuracy from 0.695 to 0.790 and macro F1 from 0.468 to 0.622. The full model achieves accuracy of 0.838 and macro F1 of 0.652 on the separate spatial test, although performance varies among classes and regions. In a matched Last Glacial Maximum experiment, PaleoVeg and archived Shao reconstructions show the same direction of CO2 -induced area change for all 11 classes. A complementary European application by Kern et al. (accepted for publication) finds similar broad temporal changes in Pale- oVeg, pollen-based reconstructions, and JSBACH over 70–20 ka BP, alongside regional disagreements and differences in absolute forest estimates.