The Mediterranean Basin is recognised as a global biodiversity hotspot, characterised by a Mediterranean climate with hot, dry summers and mild, wet winters (Médail and Quézel, 1999). The region encompasses diverse landscapes, including coastal areas, mountain ranges, and semi-arid environments. This study adopts the Mediterranean region boundary defined by Olson et al. (2001), with a slight modification. A small part of north-west Spain was included, and the Canary Islands were excluded because the region has a subtropical climate (Luque Söllheim et al., 2024). The study area map was generated in QGIS v3.40.8 (EPSG:4326) (Fig. 1).

A total of 1283 presence records for O. europaea var. sylvestris were obtained from the Global Biodiversity Information Facility (GBIF, 2025). Only herbarium-preserved specimens were included to ensure data quality, as these records are based on physical plant material that has been collected and preserved, such as specimens pressed and affixed to herbarium sheets. Observational records lacking physical specimens were excluded. To ensure comprehensive coverage, the search in GBIF included several taxonomic synonyms listed in Plants of the World Online (POWO, 2025), an authoritative global database of accepted plant names and synonyms. The search terms included O. sylvestris Mill., O. europaea var. sylvestris (Mill.) Lehr, O. europaea subsp. sylvestris (Mill.) Rouy, O. europaea subsp. sylvestris (Mill.) Rouy ex Hegi, and O. europaea subsp. europaea (GBIF.org, 2025). To complement GBIF data, 170 herbaria listed in the Index Herbariorum (Sweetgum, 2025) and located within the Mediterranean Basin were contacted by email, coordinated through the Herbarium of the University of Málaga (MGC). From the herbaria that replied, 13 provided usable records for O. europaea var. sylvestris, yielding 422 records, of which 79 were retained after cleaning.

Each record was reviewed to confirm that it represented wild, non-cultivated individuals, using contextual information such as location, habitat type, and collection notes. Records from botanical gardens or clearly cultivated settings were excluded. Complementary data provided by herbaria were checked for duplication against GBIF data, and duplicates were removed.

Occurrence records were not spatially thinned because the final cleaned dataset contained a limited number of unique, verified presence points; applying even a moderate thinning distance would have substantially reduced the sample size and weakened the model's ability to characterise the species' environmental niche. As a consequence, the model remains sensitive to spatial sampling bias, with a risk of overfitting and overestimated suitability near densely sampled areas such as the Iberian Peninsula (see Sect. 4.5). Some records lacked precise coordinates; therefore, when possible, coordinates were estimated from locality descriptions. All occurrences were georeferenced, checked for spatial accuracy in QGIS version 3.40.8 (QGIS Association, 2025), and verified against the Mediterranean Basin boundary map (Fig. 1). After data cleaning, 479 of the 1283 GBIF records and 79 of the 422 herbarium records were retained for modelling, resulting in a total of 558 occurrence records.

We obtained 19 bioclimatic variables from WorldClim v2.1 at 30 arcsec (∼1km2; Fick and Hijmans, 2017) resolution, representing historical averages from 1970 to 2000. Descriptions of these variables (bio1–bio19) are provided in Table A1.

Future projections for 2041–2060 and 2061–2080 were derived from the CMIP6 model EC-Earth3-Veg (Döscher et al., 2022) under the SSP2–4.5 and SSP5–8.5 scenarios (O'Neill et al., 2016). This model was selected for its demonstrated accuracy in simulating European climates (Filahi et al., 2025). SSP2–4.5 (intermediate development with moderate emissions) and SSP5–8.5 (fossil-fuel-driven growth with very high emissions) were used to capture a range of plausible mid- and late-century climate futures, from intermediate to more extreme warming trajectories. Projections were limited to the 2041–2060 and 2061–2080 periods to emphasise short- and medium-term changes in habitat suitability.

All 19 bioclimatic variables were clipped to the study area (Fig. 1) using the translate tool of QGIS v3.40.8. Since using too many predictors can reduce model performance and transferability, and increase the risk of overfitting, particularly when highly correlated variables are included, we reduced the variable set to those most relevant, as recommended in Elith et al. (2011). Multicollinearity among climatic variables was assessed using Pearson correlations in R v4.5.1. Analyses were run in RStudio v2025.5.1.513 (Posit Team, 2025). Following recommendations on handling collinearity in distribution models (Dormann et al., 2013), variables with a Pearson correlation coefficient |r|>0.75 were considered highly correlated, and one variable from each correlated pair was excluded from further analysis. The final set of six predictors was selected on the basis of the Pearson correlation analysis and the ecological relevance of each variable, and these variables were then used directly in all three algorithms (GLM, GAM, and MaxEnt).

Species distribution models (SDMs) were developed using the advanced ensemble modelling framework biomod2 (Guéguen et al., 2026) combined with the terra package for optimised spatial data processing within the R environment. To characterise the environmental profile of the species, a multi-algorithm approach was implemented, combining regression techniques and machine learning algorithms: generalised linear models (GLMs; Nelder and Wedderburn, 1972), generalised additive models (GAMs; Wood, 2017), and maximum entropy (MaxEnt v. 3.4.4; Phillips et al., 2025). MaxEnt modelling was executed natively by integrating the maxent.jar application via the dismo Java interface on R. Feature classes were left at the automatic MaxEnt setting, which for the training sample size enabled linear, quadratic, product, hinge, and threshold features. To account for the absence of true absence data, a background selection strategy was adopted. A single pool of 10 000 pseudo-absences (PAs) was randomly generated across the study area. For GLM and GAM algorithms, automated formula and hyperparameter selection native to biomod2 were used to prevent artificial overfitting.

Occurrence data comprised presence records from GBIF and the additional herbarium sources described earlier. Historical bioclimatic variables (1970–2000) were used as the baseline conditions, whereas future climate variables for both SSPs and time periods served as projection layers. Model performance and predictive robustness were assessed using a repeated split-sample cross-validation scheme. For each algorithm, the dataset was randomly split into 10 independent evaluation replicates, generated using 25 % of occurrence points for testing and 75 % for training, with a random seed set for reproducibility. Model accuracy was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve (Pontius and Parmentier, 2014) and the true skill statistic (TSS) metrics (Allouche et al., 2006), with a default prevalence of 0.5. AUC values range from 0 to 1, where 0.5 indicates a model performing no better than random and 1.0 indicates perfect discrimination (Fielding and Bell, 1997). To minimise individual algorithmic biases and reduce structural uncertainty, an ensemble model (EM) consensus was built by combining all calibrated individual models and cross-validation runs into a single global consensus framework. To ensure predictive quality, only individual models achieving a TSS score greater than or equal to 0.4 were retained for the ensemble assembly. Final potential suitability maps for both baseline (current) and future environmental scenarios were projected across the spatial grid and converted onto a standardised continuous scale ranging from 0 (completely unsuitable) to 1000 (maximum suitability).

Full documentation of the modelling workflow following the ODMAP protocol (Zurell et al., 2020) is provided in Appendix D.

In addition, predictor influence was evaluated via a permutation-based variable importance procedure executed with three repetitions per model run. In biomod2, this is calculated as 1-r between the reference prediction and the prediction obtained after randomly permuting each predictor in turn, computed independently for each variable. The resulting values therefore express the influence of each predictor separately and are not shares of a common total, so they do not sum to 100. Therefore, model outputs were used to identify the most influential climatic variables to generate current and future habitat suitability maps and to visualise potential shifts in the geographic distribution of O. europaea var. sylvestris. Final suitability maps were based on the EM output (0–1000) across the 30 replicates (10 runs with GLM, GAM, and MaxEnt) to ensure robust prediction and reduced sensitivity to sampling variability (Phillips et al., 2006; Elith et al., 2011).

For clarity in interpretation and spatial analysis, the continuous suitability output was rescaled from 0–1000 to 0–1 and reclassified into five equal-interval habitat suitability classes of 0.2 width: unsuitable (p<0.2), low (0.2≤p<0.4), moderate (0.4≤p<0.6), high (0.6≤p<0.8), and very high (p≥0.8). The class boundaries are equal intervals adopted for presentation and area-based comparison, not ecological thresholds, following classification schemes commonly used in ecological modelling studies (Li et al., 2020; Khan et al., 2022; Alegria et al., 2023). This classification enhanced the visualisation of habitat patterns, enabled comparison across scenarios, and facilitated the area-based analysis of habitat shifts under projected climate conditions. To quantify habitat suitability under current and future climate scenarios, the classified suitability rasters were processed in R using the terra package (Hijmans, 2025). For each raster, the number of cells per suitability class was extracted using the freq() function, and the area of each class was calculated by multiplying the cell count by the pixel area, derived from the raster resolution using the res() function. For area calculation, the classified rasters were reprojected to a cylindrical equal-area projection, giving a cell size of approximately 739.59m×739.59m, so that each pixel corresponded to approximately 0.547 km2. The resulting areas (km2) were computed for all scenarios to allow quantitative comparison of habitat suitability, expressed as the estimated area within each suitability class across different time periods and climate conditions.

Six bioclimatic variables were retained on the basis of low intercorrelation (|r|<0.75) (Fig. A1), and ecological relevance to the distribution of O. europaea var. sylvestris in Mediterranean climates. The relative permutation importance of these variables is summarised in Table 1. Minimum temperature of the coldest month (bio6) showed the highest permutation importance. Mean temperature of the warmest quarter (bio10) had the second-highest permutation importance, whereas precipitation of the warmest quarter (bio18) indicated the third-highest permutation importance (complete information of permutation importance by algorithm can be found in Table S1 in the Supplement).

Using these six variables, the ensemble model revealed strong predictive performance, with a mean calibration AUC of 0.9 and a mean calibration TSS of 0.7 (Fig. B1), along with a sensitivity of 87.4 % and a specificity of 79.3 %. Performance on the held-out validation data was comparable for AUC and slightly lower for TSS, indicating good discrimination with little overfitting. The ROC curve indicated good discrimination (mean AUC=0.9), well above random expectation (AUC=0.5).

The response curves (Fig. 2; complete information in Table S2) show the relationship between environmental variables and the predicted suitability of O. europaea var. sylvestris, outlining the climatic ranges associated with higher suitability. For bio2 (mean diurnal range), suitability was lowest at approximately 7.5 °C and then rose with increasing values, reaching its highest values at the upper end of the observed range (approximately 16 °C). Bio6 (minimum temperature of coldest month) revealed unsuitable (p<0.2) conditions where temperatures were below approximately 0.5 °C, with a sharp increase above 1.5 °C and the highest suitability at approximately 9 °C. For bio10 (mean temperature of warmest quarter), suitability began to increase at approximately 16 °C and reached its peak at 20 °C, followed by a steep decline beyond 24 °C, becoming unsuitable at temperatures over 25 °C.

Bio12 (annual precipitation) was associated with maximum suitability of around 500 mm, with a decline on both sides of this peak, with unsuitable conditions below 300 mm and above 1250 mm of annual precipitation (Table S2). Bio15 (precipitation seasonality) had a pronounced peak at a coefficient of variation of about 50, followed by a decline and a moderate secondary rise near 110. Bio18 (precipitation of the warmest quarter; Fig. 2) rose steeply from 0 mm to a peak at approximately 30 mm, declined to a plateau by approximately 75 mm, and remained within the low-suitability class across the remainder of its range, with a secondary rise near 175 mm. Suitability did not fall below 0.2 at any point across the observed range of this variable.

Under current climatic conditions, large parts of the Iberian Peninsula, particularly southern and central Spain and Portugal, exhibited extensive areas of high to very high suitability for O. europaea var. sylvestris (Fig. 3). Southern Italy; the western coastal zones of Greece, northern and central Morocco; and parts of Algeria and Tunisia also revealed considerable suitable habitat. Smaller but distinct patches of suitability were found on Mediterranean islands such as Sicily, Corsica, and Crete, as well as along coastal fringes of the Mediterranean Levant, including parts of Lebanon and Palestine.

In this section a classified categorical map is shown (Fig. 4). The map shows five suitability classes: unsuitable (<0.2), low (0.2–0.4), moderate (0.4–0.6), high (0.6–0.8), and very high (≥0.8), under the current (Fig. 4a) and future (Fig. 4b–e) conditions. To check the continuous suitability maps, see Fig. C1 in Appendix C. Using the six selected bioclimatic variables, projections for 2041–2060 and 2061–2080 under SSP2–4.5 and SSP5–8.5 revealed consistent shifts in habitat suitability (Fig. 4b–e). All future scenarios indicated a reduction in total suitable area and a general northward shift in optimal habitat, particularly across northern Spain, southern France, northern and central Italy, and parts of Greece and Türkiye. Under SSP2–4.5, both 2041–2060 (Fig. 4b) and 2061–2080 (Fig. 4c) displayed broadly similar patterns in suitability and range shift, with moderate contraction of highly suitable areas. SSP5–8.5 for 2041–2060 (Fig. 4d) showed a comparable pattern, but SSP5–8.5 for 2061–2080 (Fig. 4e) resulted in a much stronger loss of highly suitable habitat.

Many areas classified as highly or very highly suitable under current conditions were downgraded to lower-suitability categories or became unsuitable, especially in the southern Mediterranean regions. However, some coastal areas, including parts of Portugal, northern Morocco, and parts of Spain, persisted as isolated hotspots of high or very high suitability. Portugal retained relatively high suitability across all future projections.

Table 2 summarises the estimated area (×104km2) under each suitability class for current and future scenarios. All future projections show a decline in the area classified as moderately, highly, or very highly suitable compared with current conditions. The SSP5–8.5 scenario for 2061–2080 showed the most pronounced decline, with a strong increase in the area classified as unsuitable and a marked reduction in high and very high suitability classes, suggesting a substantial loss of climatically suitable habitat for O. europaea var. sylvestris by the end of the century. Highly suitable area (≥0.6) declined from 29.7×104km2 under current conditions to 5.3×104km2 under SSP5–8.5 for 2061–2080, a reduction of more than 80 %.

Only ecologically meaningful and non-redundant bioclimatic predictors were retained, which improves model interpretability, reduces overfitting, and enhances transferability (Phillips et al., 2006; Elith et al., 2011). The resulting model achieved a mean calibration AUC of 0.9 and a TSS of 0.7, indicating good discrimination between suitable and unsuitable conditions. This value is unlikely to be much higher for a broadly distributed taxon, as previous work has shown that species with wide geographic ranges often exhibit lower AUC scores because it is more difficult for the model to distinguish suitable from unsuitable habitat across diverse environments (McPherson and Jetz, 2007).

In this study, bio6 (minimum temperature of the coldest month) and bio10 (mean temperature of the warmest quarter) emerged as key contributors to habitat suitability for O. europaea var. sylvestris, underlining the importance of temperature extremes in shaping its distribution. This is consistent with the physiology of O. europaea: minimum winter temperatures directly affect survival and seedling recruitment, as freezing conditions can cause cold stress and mortality (Petruccelli et al., 2022; Ruiz et al., 2006), whereas excessively high summer temperatures can impair photosynthesis and fruit development, limiting reproductive success in overly hot environments (Haworth et al., 2018).

Together, these two thermal variables exert opposing pressures on the species' projected distribution under a future climate. Warming winter minima (bio6) make previously too-cold areas to the north and at higher elevations newly suitable, driving the northward and upslope shifts seen across the projections. At the same time, rising summer temperatures (bio10) push the warmer, low-elevation, and inland parts of the current range beyond the species' heat tolerance threshold, rendering them unsuitable. Because bio10 warming affects a substantially larger portion of the current range than the area newly opened by bio6 warming, the net outcome is a pronounced loss of suitable habitat. In other words, more of the currently suitable range becomes too hot in summer than is gained from milder winters in newly suitable areas farther north and upslope, which explains both the northward shift and the overall reduction in suitable area found in this study. A similar pattern has been reported for olive elsewhere in the Mediterranean. Özdel et al. (2025) projected a northward and upslope redistribution for cultivated olive in Türkiye, with the high-suitability class contracting sharply (from 4.4 % to 0.2 % of the country under SSP5–8.5 by 2081–2100) even as lower-suitability classes expanded, and Arenas-Castro et al. (2020) reported net reductions in suitability for olive varieties in southern Spain. Our basin-wide results for the wild variety show a comparable direction of poleward displacement, although the underlying species, extent, and predictors differ between studies. Precipitation of the warmest quarter (bio18) was also an important variable influencing modelled habitat suitability. This aligns with other studies (Arenas-Castro et al., 2020). In their assessment of climate change impacts on olive varieties, including O. europaea var. sylvestris, Arenas-Castro et al. reported that changes in water availability driven by decreasing precipitation and increasing evapotranspiration were the dominant factors reducing future suitability. They found that much of the projected loss of suitable area was associated with drier conditions, especially in autumn.

The response curve in bio15 seems to highlight the strong dependency of O. europaea var. sylvestris on Mediterranean precipitation seasonality regimes. The primary predictive optimum (at a coefficient of variation of ∼50) aligns with the balanced climate conditions of the Mediterranean Basin where the species is ecologically and climatically dominant. The species being a sclerophyll, the first peak fits the Mediterranean climatic adaptation of O. europaea var. sylvestris (Guerrero Maldonado et al., 2016). Conversely, the secondary moderate rise observed at high seasonality values (∼110) suggests the species' capacity to persist under scenarios of extreme intra- or inter-annual climate variability, indicating tolerance of highly seasonal eastern Mediterranean climates. However, this secondary peak may also indicate an artefact of sparse and spatially uneven occurrence data. Given the wider uncertainty around this part of the curve, the secondary peak should be interpreted cautiously.

Although factors such as soil properties, topography, and grazing pressure may also influence wild olive distribution (Tsiftsis et al., 2025), this study intentionally focuses on climatic suitability as a standalone driver.

Suitability tends to be lower in inland, mountainous, or more arid regions, where colder winter temperatures and reduced water availability are likely to constrain the species' distribution. Large portions of the eastern Mediterranean, such as Syria and inland regions of Türkiye, are predicted to be unsuitable, although these predictions should be interpreted cautiously given the absence of occurrence data from these regions, as discussed in the limitations.

It is not always possible to determine whether each occurrence record truly represents the wild form O. europaea var. sylvestris because many collectors and databases record specimens simply as O. europaea without specifying the variety. This uncertainty is reinforced by the fact that O. europaea var. sylvestris (Mill.) Lehr is currently treated as a synonym of O. europaea subsp. europaea in major taxonomic sources, such as POWO (2025) and Dimopoulos et al. (2016). Hybridisation between wild and cultivated olives is also widespread, especially in marginal or abandoned agroecosystems such as old olive groves (Ben-Dor et al., 2024), where feral individuals may be misclassified. Countries such as Italy and Greece contributed fewer records than Spain, but their points were relatively well distributed, which supports more reliable predictions in those regions. Overall, Spain, Portugal, Morocco, France, Italy, and Greece provided occurrence data that broadly cover a representative range of bioclimatic conditions.

Under the intermediate climate scenario SSP2–4.5, some regions such as parts of southern France and central Italy are projected to retain moderate suitability, but areas of high suitability retreat from several core regions, including parts of southern Spain and the Maghreb. These patterns are broadly consistent with studies suggesting northward or upslope shifts of Mediterranean taxa in response to warming (Gritti et al., 2006). Projections for 2041–2060 under SSP5–8.5 are similar to those under SSP2–4.5 for 2061–2080, reflecting the relatively modest divergence between the scenarios during mid-century, when projected changes in temperature and precipitation partly overlap. After 2060, however, SSP5–8.5 diverges more strongly, with substantially greater warming and sharper declines in summer precipitation, as shown for the SSP framework by Fuss et al. (2014). This likely explains the pronounced loss of suitable habitat under SSP5–8.5 for 2061–2080.

Across all future projections, many coastal zones remain suitable. Coastal regions in the Mediterranean tend to warm less than inland or mountainous areas, even under high-emission scenarios (El-Samra et al., 2018). This pattern links closely to mean temperature of the warmest quarter (bio10), as sea breeze effects and maritime influence buffer extreme summer temperatures. Coastal areas also often receive higher summer precipitation than nearby inland regions (Curtis, 2019), which relates to the precipitation of the warmest quarter (bio18) and may provide a climatic buffer against the increasingly frequent droughts characteristic of Mediterranean summers. Portugal, along with several other Mediterranean coastal zones, retains areas of high suitability across all future scenarios. The persistence of these areas likely reflects a combination of moderated summer temperatures and comparatively higher summer rainfall. Although a similar geographical signal has been reported for Portugal, where the cultivated form (O. europaea var. europaea) may remain comparatively resilient under projected climate change (Fraga et al., 2022), this resilience partly reflects irrigation and active grove management. Wild populations lack these interventions and may therefore prove more vulnerable, therefore the comparison indicates a shared climatic signal rather than equivalent resilience. These coastal zones are identified as areas of projected climatic stability rather than validated refugia. The inference rests on the modelled persistence of suitable climate and on maritime buffering of summer temperature and precipitation (El-Samra et al., 2018; Curtis, 2019), not on direct microclimate, demographic, or stress-physiology data from coastal populations of O. europaea var. sylvestris. Confirming a refugial function will require field validation, including local microclimate data and demographic or genetic monitoring of coastal versus inland populations of the species.

All projected scenarios in this study indicate a net loss of suitable habitat for O. europaea var. sylvestris. This contrasts with Kassout et al. (2022), who reported increased suitability for the wild olive in Morocco, with the largest suitable area under RCP8.5 in 2070. The most likely reason for this difference is the role of temperature in each model. Both studies included minimum temperature of the coldest month (bio6), but its importance differed greatly: it was the strongest predictor in our ensemble (permutation importance 64.9), whereas it accounted for only 0.6 % of per cent contribution in Kassout et al. (2022). The opposite held for annual precipitation (bio12), which ranked low here (permutation importance 8.8) but dominated their model at 76.7 %. These metrics are not directly comparable in magnitude, as MaxEnt per cent contribution is normalised to sum to 100, whereas biomod2 permutation importance is computed independently per variable; the comparison is therefore one of rank and direction rather than of absolute values. Our basin-wide model is therefore driven mainly by temperature, whereas their model for Morocco is driven mainly by rainfall. This likely reflects the different study areas: across the whole Mediterranean Basin, cold winters limit the species at its northern and inland edges, making bio6 highly informative, while in the Kassout et al. (2022) study, Moroccan winters are rarely cold enough to limit O. europaea var. sylvestris, so water availability matters more. Because our model is temperature led, warming has two opposing effects: milder winters (bio6) let the species spread northward, while hotter summers (bio10) make southern areas unsuitable; this warm-edge loss outweighs the northern gain, producing a net decline. In a rainfall-driven model where temperature is not limiting, warming does not impose this penalty, so suitability can instead increase. The studies also differ in their climate data (CMIP6 under SSPs here versus CMIP5 under RCPs) and number of predictors (6 versus 12), which can further affect the results. Notably, Kassout et al. (2022) also found declining suitability for the endemic O. europaea subsp. maroccana. Arenas-Castro et al. (2020), assessing several olive varieties including O. europaea L. subsp. europaea var. sylvestris, reported reduced suitability for six of seven varieties, with marked declines for the wild variety in Spain under the 2040, 2070, and 2100 projections, which agrees with the trends found here.

Furthermore, Hidalgo-Triana et al. (2023) showed that western Mediterranean O. europaea var. sylvestris forests are highly exposed to climatic change. This corresponds with the strong reduction in suitable areas in the southern Iberian Peninsula found in our study. However, the authors argue that, despite this high exposure, these forests are less vulnerable to climate change because the species possesses functional traits that confer resilience to increasingly dry and hot conditions.

Projections in this study focused on the mid- and late-century horizons (2041–2060 and 2061–2080) to provide more actionable information for conservation planning and management of the species. Longer-term projections could also be informative for assessing potential range collapse or the need for assisted migration under more extreme warming.

Model output should be interpreted carefully, as it reflects the specific combination of bioclimatic variables used in this study, and a different variable set may cause variations in the distribution patterns. Future projections here rely on a single CMIP6 model; incorporating additional models (for example, IPSL-CM6A-LR and MRI-ESM2-0) would allow between-model climate uncertainty to be quantified and is a priority for further work. At the same time, many records are spatially clustered, for example, in specific parks or forest stands, which can introduce sampling bias and lead to overestimation of suitability near well-sampled locations (Anderson and Raza, 2010; Fourcade et al., 2014). Because the final cleaned dataset contained only 558 unique presence points, spatial thinning was not applied in order to retain sufficient environmental coverage, but this may amplify the effects of clustering. Several Mediterranean countries lacked occurrence data in this study, including Algeria, Tunisia, Palestine, Israel, Lebanon, Syria, and especially Türkiye. This absence does not imply that the species is ecologically absent in these regions. For Türkiye, 28 herbaria were contacted, but no responses were obtained and no herbarium-preserved records were found in GBIF. A similar lack of accessible herbarium data applied to several of the other countries mentioned above. However, O. europaea var. sylvestris is known to occur naturally throughout the eastern Mediterranean, and multiple studies document its presence in these areas (Fanelli et al., 2022). In fact, the Mediterranean Levantine region is widely considered the centre of origin of the wild olive, from which it spread westwards across the Mediterranean Basin (Besnard et al., 2013; Besnard et al., 2018). Given this, the “unsuitable” predictions there should be interpreted with caution and treated as an artefact of incomplete occurrence sampling rather than a genuine ecological signal. Future iterations of the model would benefit substantially from incorporating herbarium records and field survey data from these under-represented countries, which would likely expand the predicted suitable area eastward and improve model transferability across the full Mediterranean Basin.

To improve future projections and obtain more accurate SDM outputs for the whole Mediterranean Basin, it will be necessary to expand occurrence datasets, particularly by incorporating records from currently under-represented countries.