Predicting the Current Potential Suitable Habitat of Pinus tabuliformis in China Using Multi-Source Environmental Data and Ensemble Learning Models.

Pinus tabuliformis is an endemic tree species in China that supports windbreak and sand fixation, soil and water conservation, and ecosystem stability. However, some stands experience growth decline and degradation, especially in arid and semi-arid regions with constrained regeneration. Identifying key environmental drivers and suitable habitats is therefore important for restoration and management. Because species distribution model outputs vary among algorithms, we developed a spatially validated stacking framework integrating maximum entropy (MaxEnt), random forest (RF), and a generalized additive model (GAM) using multi-source environmental data. Predictors were screened using variance inflation factor (VIF < 5) and correlation threshold (|r| < 0.8). The results showed that: (1) Seasonal hydrothermal conditions and topography primarily shaped the current potential distribution of P. tabuliformis. BIO9, DEM, and BIO18 were the leading predictors, with BIO9 ranking first across all models and highest suitability occurring at -10°C to 5°C. (2) Under five-fold spatial block cross-validation, RF achieved the highest mean AUC, TSS, and width-normalized partial AUC values of 0.973 ± 0.031, 0.726 ± 0.140, and 0.574 ± 0.245, respectively. The stacking ensemble showed numerically similar AUC and TSS values but did not outperform RF. No pairwise performance differences remained statistically significant after Holm correction. (3) Highly suitable habitats were concentrated in low- to mid-elevation mountains and hills of northern and north-central China, especially the Loess Plateau and adjacent areas, with the Qinling Mountains forming a southern transition zone. The ensemble predicted a total suitable habitat area of 28.45 × 105 km2 (95% CI: 25.23-31.42 × 105 km2), with a suitability-weighted centroid at approximately 34.50° N and 110.72° E. Overall, stacking integrated structurally different algorithms into a synthesized spatial estimate supporting field validation, conservation, restoration, afforestation zoning, and sustainable management.