AI Ensemble Maps Where Licorice Can Survive a Warming Century

Licorice is one of those plants most people know only as a flavoring, yet its roots carry glycyrrhizin, a triterpenoid saponin prized for anti-inflammatory, antiviral, antioxidant, and immunomodulatory properties. The species behind it, Glycyrrhiza glabra, is a perennial legume scattered across the arid and semi-arid belt of Eurasia, and Iran holds one of the world’s largest reservoirs of wild genetic resources for it. That makes the plant strategically important for pharmaceutical supply chains, biodiversity conservation, and rural livelihoods alike. A new study published in Smart Agricultural Technology by Emran Dastres and Hassan Esmaeili of Shahid Beheshti University now asks a deceptively simple question with profound consequences: where, across Iran’s 1.53 million square kilometers, will licorice still find suitable conditions as the twenty-first century unfolds?

Answering that question required more than a single predictive model. The researchers built an ensemble machine learning framework that combines six complementary algorithms: Random Forest, XGBoost, LightGBM, MaxEnt, Bayesian Additive Regression Trees, and a Support Vector Machine. Each algorithm was calibrated independently with carefully tuned hyperparameters, from Bayesian-optimized boosting settings to regularization multipliers in MaxEnt, and each was evaluated with spatially stratified five-fold cross-validation using 50-kilometer blocks. This blocking matters because species occurrence records are spatially autocorrelated; if training and testing points sit too close together, models earn inflated performance scores that collapse when extrapolating to new climates. The final ensemble weighted each algorithm by its cross-validated True Skill Statistic, so better-performing models contributed more to the consensus map.

The occurrence data itself was assembled from three sources: systematic field surveys conducted between 2020 and 2025 with handheld GPS units accurate to under five meters, digitized herbarium specimens, and records from the Global Biodiversity Information Facility. After removing duplicates, discarding points with coordinate uncertainty above 2.5 kilometers, and applying a five-kilometer spatial thinning rule, the team retained 304 spatially independent occurrence records. Because true absences are unavailable in such studies, 10,000 pseudo-absence points were generated using a target-group background approach, restricted to cells where other medicinal plants from the same surveys had been recorded, ensuring that sampling bias affected presences and backgrounds similarly.

What distinguishes this framework from many earlier species distribution studies is the breadth of environmental information fed into the models. Alongside the standard 19 bioclimatic variables from WorldClim at roughly one-kilometer resolution, the researchers incorporated six soil properties from SoilGrids 2.0, including pH, electrical conductivity, texture fractions, and soil organic carbon, averaged across the 0 to 100 centimeter profile that represents licorice’s root zone. Topographic predictors were derived from the 12.5-meter ALOS PALSAR digital elevation model, including slope, aspect, the Topographic Wetness Index, and the Multi-Resolution Valley Bottom Flatness index, which helps identify valley-bottom micro-refugia where groundwater access may buffer warming. After a two-stage multicollinearity filter using Spearman correlations and variance inflation factors, 25 of 35 candidate predictors were retained. For future projections, the team drew on four CMIP6 global climate models, EC-Earth3-Veg, MIROC6, MPI-ESM1-2-HR, and UKESM1-0-LL, under two Shared Socioeconomic Pathways: the intermediate SSP245 and the high-emission SSP585, across four consecutive 20-year periods spanning 2021 to 2100.

The ensemble performed well. Random Forest led the individual algorithms with an AUC of 0.833, but the weighted ensemble edged ahead at 0.835, with a True Skill Statistic of 0.64 and a Continuous Boyce Index of 0.73. Under baseline conditions, the consensus model estimated roughly 673,000 square kilometers of suitable habitat, about 44 percent of Iran, concentrated in three zones: the western and northwestern Zagros, the northern Alborz foothills and Caspian lowlands, and the northeast near Khorasan. All models agreed on the unsuitability of the hyper-arid Dasht-e Kavir and Dasht-e Lut basins, the southern coastal plains, and peaks above 2,500 meters. Suitable habitat peaked between 800 and 1,600 meters in elevation, with the highest frequencies between 1,100 and 1,300 meters.

The most striking result is what the future projections did not show. Total suitable area remained surprisingly stable through 2100, fluctuating between roughly 557,000 and 596,000 square kilometers under both scenarios. Yet beneath that apparent stability, the range-dynamics analysis revealed a hidden crisis of turnover. Under SSP585 in the final decades of the century, habitat loss reached 259,413 square kilometers while habitat gain was essentially zero, a fraction of a square kilometer. Under the moderate SSP245 pathway, losses ranged from about 44,000 to 114,000 square kilometers with similarly negligible gains. In other words, licorice is not projected to lose its overall footprint so much as to watch its existing habitats erode without any compensatory colonization elsewhere, a pattern the authors attribute to dispersal limitations, fragmentation, and landscape barriers that prevent species from tracking shifting climates.

The geographic reorganization was gradual rather than dramatic. The suitability centroid, initially located in central Iran’s Isfahan province at about 33.07 degrees north, drifted consistently southwestward by 13 to 27.6 kilometers depending on scenario and period, equivalent to a mean velocity of roughly 0.35 kilometers per year. Elevational shifts were minor, staying within a range of minus 2 to plus 21 meters relative to the baseline mean of 1,262 meters. The northeastern Khorasan region showed remarkable persistence under both scenarios across all time horizons, suggesting it may function as a terminal refugium even under the most extreme trajectory. Crucially, only 3.03 percent of Iran, about 46,400 square kilometers, remained suitable across every climate model, scenario, and period. These persistent core refugia are, in the authors’ framing, the most reliable targets for long-term cultivation investment, germplasm conservation, and climate-smart agricultural planning.

Perhaps the most methodologically consequential finding concerns uncertainty. By running all six algorithms against all four climate models, the team could decompose prediction variability into algorithmic, climate-model, and temporal components. Algorithmic uncertainty proved tiny, with mean standard deviations between 0.020 and 0.028, and high-uncertainty areas never exceeding 0.5 percent of the country. Climate-model uncertainty was consistently larger, with standard deviations rising to 0.072 under late-century SSP585, concentrated in central Iran and the southern Zagros where the GCMs diverged most on precipitation. The lesson is clear: the choice of machine learning algorithm matters far less than the choice of climate model, reinforcing the argument that ecological forecasts relying on a single GCM are systematically overconfident. A Multivariate Environmental Similarity Surface analysis added a further safeguard, showing that under SSP245 no novel climates emerged at all, while under SSP585 strict extrapolation affected at most about 4.44 percent of the territory by 2081 to 2100, confined to low-elevation arid zones.

Variable importance rankings, independently confirmed by Random Forest and XGBoost metrics, identified annual precipitation as the dominant driver of suitability, followed by annual mean temperature, clay content, soil organic carbon, and elevation. The prominence of the soil variables is no statistical accident. Because licorice is harvested for its roots, properties governing water retention, aeration, nutrient availability, and penetration resistance translate directly into root biomass and glycyrrhizin accumulation, making edaphic data agronomically meaningful rather than merely decorative. The authors argue that climate-only models systematically oversimplify the niche of root-harvested medicinal plants, and that integrating SoilGrids and terrain derivatives is essential for credible cultivation zoning.

The framework’s real significance may lie beyond licorice. By fusing ensemble machine learning, multi-model CMIP6 projections, heterogeneous environmental layers, multi-temporal analysis, and explicit uncertainty decomposition into a single workflow, the study sketches what AI-assisted decision support for climate-smart agriculture can look like: probabilistic maps that tell growers not just where conditions may be suitable, but how confident to be and where refugia will persist regardless of which future arrives. The authors acknowledge limitations, including the assumption of niche stability, the treatment of soils as static, and the omission of dispersal, biotic interactions, and land-use change, and they point toward explainable AI, deep learning, and dynamic soil modeling as next steps. But the core message stands: apparent stability in total suitable area can conceal severe internal contraction, and adaptation strategies that protect today’s productive zones and the handful of persistent refugia will matter far more than hoping new frontiers of suitability will open on their own.

Subject of Research: Ensemble machine learning projection of climate-driven cultivation suitability shifts for Glycyrrhiza glabra in Iran under CMIP6 scenarios

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

Article References: Dastres, E., & Esmaeili, H. (2026). Ensemble machine learning for climate-smart cultivation zoning of Glycyrrhiza glabra L. under CMIP6 climate scenarios. Smart Agricultural Technology, 15, Article 102633. https://doi.org/10.1016/j.atech.2026.102633

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102633

Keywords: licorice, Glycyrrhiza glabra, species distribution modeling, ensemble machine learning, CMIP6, climate change, climate refugia, Iran, medicinal plants, soil variables, uncertainty quantification, climate-smart agriculture