Historical forest coverage (HFC) dataset

We reconstructed a historical forest cover dataset for China from 1900 to 2020 at five-year intervals and 1 km spatial resolution (Fig. 1). Evaluation of the Mixed-Cell Cellular Automata (MCCA) model indicated an overall accuracy of 0.97, F1-score of 0.91, and Kappa coefficient of 0.89 for the 2000s (SI Appendix, Table S1). Subsequently, we used historical topographic maps surveyed in the 1950s to further validate the dataset, which yielded an overall accuracy of 0.97, F1-score of 0.82, and Kappa coefficient of 0.81 (SI Appendix, Fig. S1, Table S2). To assess historical accuracy, we used the locations of 283,707 georeferenced old trees, each over 100 years old, as indicators of early 20th-century forest fragments (SI Appendix, Fig. S2). To validate the forest cover maps, we identified a 1 km2 pixel as a ‘hit’ if it contained both at least one old tree and a forested area ≥1 hectare (i.e., ≥1% forest cover). This threshold provides a conservative yet ecologically justified benchmark for forest presence, capturing the minimal extent of a forest patch likely to contain ancient trees, given their average density (~0.36 trees/km²) (39), while reducing misclassification risks from isolated pixels or geolocation uncertainty. Based on this criterion, 93.86% (n = 266,294) of old trees in 2020 and 78.66% (n = 223,152) of those estimated to be present in 1900 fell within qualifying pixels. Additionally, when using a stricter 10-hectare threshold, 84.21% (n = 238,912) in 2020 and 70.05% (n = 198,742) in 1900 still met the criterion. These high concordance rates across spatial thresholds and time periods underscore the reliability of our forest reconstruction and the robustness of the multi-source validation strategy.