Machine Learning Maps Plant Species Shifts Under Climate Change

In a striking example of how artificial intelligence is reshaping conservation science, a team of researchers in Iran has used machine learning to forecast the fate of a medicinal and industrially valuable shrub under climate change, with results that paint a sobering picture. The plant, Vitex pseudo-negundo, is a close relative of the well-known chaste tree and produces essential oils with documented pharmacological and biopesticide properties. Yet according to a new study published in Natural Resources Research, the suitable habitat for this species could collapse dramatically by the end of the century, shrinking to a fraction of its current extent under high-emission climate scenarios.

The research, led by Musa Neyestani of the Department of Natural Resources and Environmental Engineering at Shiraz University, together with Atiyeh Amindin, Soroor Rahmanian, Gholamabbas Ghanbarian, Roja Safaeian and Hamid Reza Pourghasemi, employed species distribution modeling to map where the plant thrives today and where it might survive tomorrow. Species distribution models are a cornerstone of modern biogeography: they relate known occurrence records of a species to environmental conditions at those locations, then use those relationships to predict suitability across unsampled areas and future times. What distinguishes this study is its head-to-head comparison of several machine learning algorithms applied to a single medicinal species, combined with projections under the latest generation of climate scenarios.

Four algorithms were tested: random forest (RF), support vector machine (SVM), generalized linear model (GLM) and multivariate adaptive regression splines (MARS). Each approaches the prediction problem differently. The generalized linear model, a classical statistical technique dating back to the foundational work of Nelder and Wedderburn, fits a mathematically explicit relationship between environmental predictors and the probability of species presence. Support vector machines, by contrast, construct decision boundaries in a high-dimensional feature space, using kernel functions to separate suitable from unsuitable conditions with maximal margin. Multivariate adaptive regression splines build piecewise linear functions that can capture nonlinear thresholds and interactions, automatically detecting breakpoints where, for example, a small change in temperature produces a large change in habitat quality.

Random forest, the eventual winner, is an ensemble method that grows hundreds of decision trees, each trained on a random bootstrap sample of the occurrence data and a random subset of predictor variables. Individual trees are noisy, but their aggregate vote — averaging predictions across the forest — is remarkably stable and resistant to overfitting. Originally introduced by Leo Breiman in 2001, random forest has become a workhorse of ecological modeling precisely because it handles nonlinear interactions, mixed variable types and relatively small datasets with grace. In this study, it achieved an area under the curve (AUC) of 0.995, a near-perfect discrimination score on the standard receiver operating characteristic evaluation. An AUC of 1.0 means flawless separation of presences from absences; values above 0.9 are generally considered excellent. The random forest outperformed the other three algorithms, establishing it as the most reliable engine for the species’ habitat projections.

With the modeling framework settled, the team turned to the question of what actually governs where the plant grows. The analysis of variable importance yielded a result with clear ecological logic: the single most influential factor was BIO9, the mean temperature of the driest quarter — a bioclimatic variable capturing thermal conditions during the most water-stressed part of the year. This makes intuitive sense for a Mediterranean and Irano-Turanian shrub whose physiology must contend with the combined stress of summer heat and drought. The second most important factor was soil electrical conductivity, a proxy for salinity. Soil chemistry is often neglected in species distribution studies, which default to climate layers, but a growing body of literature shows that edaphic variables can rival or exceed climate in explaining plant distributions. The finding that conductivity matters so strongly for V. pseudo-negundo underscores its tolerance for, or dependence on, particular soil conditions — and highlights a vulnerability, since salinization patterns may shift independently of temperature and rainfall.

Equally revealing were the variables that barely mattered. Topographic descriptors such as the topographic wetness index, aspect and plan curvature — measures of how landscape position influences water accumulation, solar exposure and slope shape — had the least influence on the model. For a species whose distribution is apparently governed by regional climate and soil chemistry rather than fine-scale terrain, this simplifies conservation targeting: coarse-resolution climate and soil data may suffice for identifying priority areas, at least at the scale of the study region.

Under current conditions, the model classified 47.98 percent of the study area as having low habitat suitability, with only 8.07 percent falling into the highly suitable class. That modest baseline already frames the species as a habitat specialist. The projections into the future, however, are where the study delivers its most consequential message. Using the CMIP6 framework of shared socioeconomic pathways — specifically the low-emission SSP1-2.6 scenario and the high-emission SSP5-8.5 scenario — the researchers projected habitat suitability decades ahead, through 2090. The two scenarios bracket the plausible range of futures: SSP1-2.6 assumes strong mitigation consistent with the Paris Agreement’s ambitions, while SSP5-8.5 assumes continued fossil-fuel-intensive development and represents a worst-case trajectory of warming.

The geographic pattern of change is as important as its magnitude. The models project significant habitat loss in the central portion of the study region, where conditions are expected to become increasingly inhospitable, while western areas may actually gain suitability — a classic range shift signature, with the species’ climatic envelope migrating away from its present-day core. Such shifts have been documented worldwide as one of the most consistent ecological responses to modern warming, but they pose acute problems for sedentary organisms and for species whose dispersal cannot keep pace with the velocity of climate change. By 2090 under SSP5-8.5, unsuitable habitat is projected to expand to 88.12 percent of the study area. Perhaps most alarmingly, only about 3 percent of the study area is projected to remain suitable under both scenarios — a narrow sliver of climate refugia where conservation efforts would offer the highest return on investment.

Why does this matter beyond the study region? V. pseudo-negundo occupies a significant niche in both traditional and applied contexts. Its essential oils vary in composition across ecotypes and plant organs, and laboratory studies have demonstrated antifungal and antibiofilm activity against pathogenic fungal strains, as well as phytotoxic properties that position the plant as a candidate for natural herbicide development — a “biopesticide” role of growing commercial interest as agriculture seeks alternatives to synthetic chemicals. Relatives within the Vitex genus have a long history in herbal medicine, and phytochemical analyses of V. pseudo-negundo have revealed antioxidant-rich phenolic compounds. Losing the genetic diversity embedded in wild populations would mean losing raw material for future drug discovery and agrochemical innovation, not to mention the shrub’s role in agroforestry systems and ecosystem functioning.

The study’s methodology also carries lessons for the field at large. The researchers used the sdm platform in R, a reproducible and extensible environment for species distribution modeling that allows multiple algorithms to be fitted and compared within a unified framework. Ensemble and comparative approaches of this kind are increasingly regarded as best practice, because different algorithms impose different assumptions and can disagree in their projections; understanding which models perform best for a given species, and quantifying the uncertainty across them, strengthens the credibility of conservation recommendations. The authors also cite the ongoing debate over scenario plausibility and internal climate variability, acknowledging that projections of species distributions inherit uncertainty from the climate models that feed them. The convergence of the SSP1-2.6 and SSP5-8.5 results on a common refugial core of roughly 3 percent is, in this light, a robust finding: it suggests that even under optimistic emissions pathways, the species faces a contracted future.

For conservation planners, the implications are concrete. Priority should be given to protecting the western areas where suitability is projected to persist or expand, and to the central refugia identified under both scenarios. Assisted migration — deliberately relocating populations or propagules to projected future habitat — emerges as a plausible, if debated, tool. Seed collection and ex situ conservation of genetically diverse populations, particularly from ecotypes with distinctive essential oil chemistry, would preserve options that in situ measures alone cannot guarantee. The authors argue that integrating predictive modeling into conservation planning is no longer optional but essential for the long-term survival of species like this one.

The broader takeaway extends well beyond a single shrub in a semi-arid landscape. As machine learning tools become standard equipment in ecology, studies like this one demonstrate their power to convert sparse field observations into actionable maps of future risk. They also serve as a warning: plants that anchor traditional medicine, emerging biopesticide industries and agroforestry livelihoods are often habitat specialists, precisely the species most vulnerable to a rapidly shifting climate. Whether the refugia identified by this study’s algorithms are safeguarded in time will test whether conservation practice can keep pace with the predictive science that now precedes it.