Ensemble machine learning for climate-smart cultivation zoning of Glycyrrhiza glabra L. under CMIP6 climate scenarios

Accurate habitat suitability projections under climate change are essential for climate-smart cultivation planning and sustainable production of medicinal plants. However, conventional species distribution modeling (SDM) approaches often overlook the combined effects of algorithmic uncertainty, edaphic–topographic heterogeneity, and dynamic habitat turnover, limiting their effectiveness as decision-support tools for agricultural planning. Here, we developed an ensemble machine learning framework integrating six complementary machine-learning and Bayesian algorithms (RF, XGBoost, LightGBM, MaxEnt, BART, and SVM), four CMIP6 global climate models (GCMs), four future time horizons, two Shared Socioeconomic Pathways (SSP245 and SSP585), and high-resolution (∼1 km) climatic, edaphic, and topographic predictors to identify suitable cultivation zones for Glycyrrhiza glabra L. across Iran. The performance-weighted ensemble achieved high predictive performance (AUC = 0.835; 95% CI: 0.720–0.951) while substantially reducing algorithmic uncertainty (mean SD = 0.020–0.028). In contrast, variability among GCMs represented the dominant source of projection uncertainty (SD = 0.045–0.072). Although the total suitable cultivation area remained relatively stable through 2100, decreasing only from 43.9% under current conditions to 37.6–38.9% under future scenarios (∼560,000–596,000 km²), range-dynamic analyses revealed substantial hidden habitat turnover characterized by extensive habitat loss and negligible habitat gain. Under SSP585 (2081–2100), suitable habitat loss reached 259,413 km², whereas habitat gain was nearly absent (0.7 km²). The suitability centroid shifted consistently southwestward (16–28 km) with minimal elevational change (+2 to +21 m), indicating persistence of core cultivation zones rather than large-scale geographical redistribution. Persistent climate refugia, representing the most reliable areas for long-term cultivation and germplasm conservation, occupied only 46,400 km² (3.03% of Iran). Annual mean temperature (Bio1), annual precipitation (Bio12), soil organic carbon, and clay content were identified as the dominant environmental drivers governing cultivation suitability. These findings demonstrate that apparently stable suitable area can conceal substantial habitat turnover and progressive range contraction, emphasizing the importance of integrating uncertainty and range dynamics into AI-assisted cultivation planning under climate change. The proposed ensemble machine learning framework provides a transferable decision-support tool for climate-smart cultivation zoning, agricultural land-use planning, sustainable medicinal plant production, and long-term management of high-value medicinal crops under future climate scenarios.