Research report: Ensemble species distribution modelling of climate change impacts on the distribution of Pinus krempfii in Vietnam from the Last Glacial Maximum to 2070

A climate-based ensemble model predicts that suitable habitat for Pinus krempfii, the flat-need pine restricted to Vietnam’s Central Highlands, could contract by 57.6–71.1% by the 2070s compared with the present climate. The projections combine seven statistical and machine-learning algorithms with five CMIP6 global climate models and identify the Lang Biang Plateau and Chu Yang Sin mountains as the species’ most persistent climatic refuges. These areas retain modelled suitable conditions from the Last Glacial Maximum to the late 21st century, but the species has little elevational room as its modelled climatic range moves upward.

The study reconstructed potential distribution under Last Glacial Maximum conditions, represented by the CCSM4 palaeoclimate simulation, present-day climate from 1970–2000, and two future emissions pathways for 2061–2080: the low-emissions SSP1-2.6 and high-emissions SSP3-7.0. The analysis used 109 occurrence records collected from field surveys, the Global Biodiversity Information Facility and published genetic studies. Spatially thinning those records to a minimum separation of 5 km left 25 presence locations.

The researchers used block cross-validation to account for spatial autocorrelation and provide a more realistic assessment of model transferability to unsampled regions and future climates. Environmental predictors came from 19 bioclimatic variables and elevation at 30 arc-second resolution. After screening for multicollinearity, the researchers retained eight bioclimatic variables with pairwise Spearman correlations below 0.7: mean diurnal temperature range, temperature seasonality, minimum temperature during the coldest month, mean temperature of the wettest quarter, annual precipitation, precipitation in the driest month, precipitation seasonality and precipitation in the warmest quarter.

Elevation was obtained separately. Models were calibrated in biomod2 using MAXNET, generalized linear models, generalized additive models, multivariate adaptive regression splines, random forests, artificial neural networks and XGBOOST, with three pseudo-absence replicates and three cross-validation runs for each algorithm. The modelling framework produced 84 individual models: 63 cross-validated models, comprising seven algorithms, three pseudo-absence replicates and three cross-validation runs, plus 21 full-data models.

Only models with cross-validated area under the receiver operating characteristic curve above 0.7 entered the weighted-mean ensemble used for the principal habitat maps. XGBOOST recorded the strongest mean performance, with an AUC of 0.894 ± 0.107, followed by MAXNET at 0.870 ± 0.184 and artificial neural networks at 0.841 ± 0.179. Generalized additive models had the lowest mean AUC, 0.684 ± 0.193, and were excluded from the ensemble.

Coefficient-of-variation maps were used to measure disagreement among models, while a multivariate environmental similarity analysis identified locations where future climates fell outside the environmental range represented by the calibration data. Temperature seasonality emerged as the strongest predictor, with a mean permutation importance of 57.7%. Minimum temperature of the coldest month ranked second at 49.2%, followed by mean temperature of the wettest quarter at 37.1%.

Precipitation of the warmest quarter, precipitation of the driest month and precipitation seasonality also contributed substantially, with mean importances of 35.9%, 33.3% and 31.9%, respectively. The ranking indicates that seasonal thermal conditions were the strongest overall determinants of the model’s climatic niche, while moisture variables also contributed substantially. The researchers interpret this combination as evidence that both intra-annual temperature variability and moisture availability help define where the species can persist.

The model estimated 13,549 km² of climatically suitable habitat around 21,000 years before present, when the range extended farther south and east into what are now the Dong Nai Biosphere Reserve and lowland areas of southern Vietnam. Suitable area fell to 4,313 km² under the 1970–2000 baseline, a decrease of approximately 68%. The present-day prediction was concentrated in the Lang Biang Plateau, Chu Yang Sin mountains and Da Lat highlands and broadly matched the species’ known geographic range.

The median elevation of suitable habitat was 933 m during the Last Glacial Maximum, with a modelled range of 755–1,152 m, compared with 1,304 m and a range of 1,088–1,499 m in the baseline climate. Under SSP1-2.6, the five global climate models produced a median 1,829 km² of suitable habitat for the 2070s, a 57.6% reduction from the baseline, with individual-model estimates spanning 1–2,427 km². Under SSP3-7.0, the median fell to 1,246 km², corresponding to a 71.1% reduction, while estimates ranged from less than 1 km² to 1,849 km².

GFDL-ESM4 under SSP1-2.6 and ACCESS-CM2 under SSP3-7.0 each projected near-complete climatic habitat loss, whereas MRI-ESM2-0 retained the largest area under both pathways. In all future projections, suitable conditions remained concentrated in the Lang Biang and Chu Yang Sin complexes, with approximately 42% of projected habitat falling within protected areas. The projected refuge area also moved uphill.

Its median elevation reached 1,490 m under SSP1-2.6 and 1,528 m under SSP3-7.0, about 190–225 m above the present median. The modelled elevational band of suitable habitat shifted from 1,088–1,499 m in the baseline climate to 1,241–1,712 m under SSP1-2.6 and 1,250–1,732 m under SSP3-7.0. Because the highest peaks in the species’ principal mountain range reach only about 2,000 m, continued warming could compress the remaining habitat toward summits.

Model disagreement was greatest around range margins and under the high-emissions pathway, whereas the refugial cores had higher prediction confidence and remained within the climatic domain represented by the calibration data. The authors describe P. krempfii as a narrow-range tropical conifer listed as Vulnerable in both the IUCN Red List and Vietnam Red Book. Approximately 39.5% of current modelled suitable habitat lies within existing protected areas, and the modelled overlap rises to about 42% under both future scenarios as the range contracts toward refugia.

The study suggests that protection should encompass all predicted habitat around Lang Biang and Chu Yang Sin, preserve elevational connectivity between populations, support long-term monitoring, and prepare ex situ contingencies if in situ persistence becomes unlikely. It also proposes testing assisted migration to suitable sites 200–400 m above the current range boundary, although that is a proposed management measure rather than an intervention tested by the models. These results come with important limitations.

The models are correlative and do not represent competition, facilitation, herbivory, dispersal, genetic adaptation or changes in land use. Coarse climatic layers may also smooth over cool, moist gullies and other local microrefugia, while forest conversion and agricultural expansion could make the climate-only projections optimistic. The binary maps therefore estimate potential climatic suitability rather than the number of surviving trees or the certainty of extinction.

Even with those qualifications, the agreement among the baseline map, the modelled historical contraction and the concentration of future habitat in two high-elevation refuges indicate that conserving connected mountain terrain will be central to the flat-need pine’s response to future warming.

Subject of Research: Climate-driven contraction and upslope movement of suitable habitat for the Vietnamese endemic Pinus krempfii

Article Title: Ensemble species distribution modelling of climate change impacts on the distribution of Pinus krempfii in Vietnam from the Last Glacial Maximum to 2070

Article References: Tuan, L. Q., Duy, V. D., Phuong, P. M., & Vladimirovich, P. V. (2026). Ensemble species distribution modelling of climate change impacts on the distribution of Pinus krempfii in Vietnam from the Last Glacial Maximum to 2070. Nature Conservation, 66, 19-37. https://doi.org/10.3897/natureconservation.66.206478

Image Credits: AI Generated

DOI: 10.3897/natureconservation.66.206478

Keywords: Pinus krempfii, Vietnam, climate change, species distribution models, habitat loss, CMIP6, Last Glacial Maximum, montane forests, climate refugia, conservation, Ensemble, species