Geographical distribution data for C. speciosa were sourced from the China Virtual Herbarium (CVH, http://www.cvh.ac.cn/), the National Specimen Information Infrastructure (NSII, http://www.nsii.org.cn), and the Global Biodiversity Information Facility (GBIF, https://www.gbif.org). After screening the raw data, duplicate records, specimens lacking precise geographic locations, and specimens collected before 2000 were excluded. The Baidu Map Coordinate Extractor (https://lbs.baidu.com/maptool/getpoint) was used to supplement the latitude and longitude coordinates of specimens with missing coordinates, resulting in 163 initial distribution records. A significant level of spatial autocorrelation is present in environmental variables over short ranges (Dormann et al., 2007). To mitigate potential bias from spatial autocorrelation in species distribution modeling, the SDMTools package was employed within ArcGIS 10.8 to apply spatial thinning with a 1 km filter distance. This ensured that only the most representative single distribution record was retained per 1 km × 1 km grid cell (Wu et al., 2025). Following this process, the final dataset comprised 157 valid distribution points, which were used for all subsequent modeling and analysis (Figure 2).

Climate data were obtained for the contemporary period (1970–2000) and two future periods (the 2050s and 2070s). Owing to the pronounced time lag in the response of ecosystems to climate change, the 2050s and 2070s were selected as representative periods for the future scenarios. This approach enables the separate assessment of species distribution patterns, indicating both the near-term response to climate change and the long-term potential for habitat loss or shifts in adaptability thresholds owing to cumulative effects (Fu et al., 2024). Nineteen bioclimatic variables together with 12-monthly data on solar radiation, vapor pressure, and wind speed were obtained from the WorldClim database (https://worldclim.org/). Simultaneously, the extraction of topographic variables from the platform’s elevation data produced three key factors: elevation, slope, and aspect. Soil environmental data were sourced from the World Soil Database (HWSD, http://www.fao.org/soils-portal/soil-survey/soil-maps-and-databases/harmonized-world-soil-database-v12/en/), with selected variables including soil effective water content (AWC_CLASS), topsoil textural class (T_TEXTURE), soil drainage class (DRAINAGE), topsoil gravel content (T_GRAVEL), topsoil reference bulk density (T_REF_BULK), topsoil organic carbon content (T_OC), and pH (T_PH_H2O). All environmental layers underwent uniform spatial clipping, resolution standardization (2.5 arc-min), and format conversion using SDMTools. Bioclimatic data for future periods were sourced from the WorldClim platform, which integrates the subscale data from the BCC-CSM2-MR climate model within the Sixth Coupled Model Intercomparison Project (CMIP6). The BCC-CSM2-MR model employed in this study demonstrates good applicability to climate simulations in China (O’Neill et al., 2017). An analysis of the potential changes in suitable regions for C. speciosa under two representative Shared Socioeconomic Pathways (SSP245 and SSP585) was conducted (Daï et al., 2023). All environmental layers were uniformly processed to a 2.5-arc-min resolution, with format standardization and spatial cropping performed using SDMTools.

Based on 65 environmental factors (Supplementary Material 1), a MaxEnt model was constructed for trial runs. To mitigate potential overfitting resulting from high correlations among environmental variables (Huang et al., 2024), we initially assessed each factor’s contribution to model predictions using the jackknife test in MaxEnt, and subsequently excluded factors exhibiting low contribution rates. For the remaining 41 factors, correlation analysis was performed using ENMTools (Figure 3), with the Pearson correlation coefficient threshold set at 0.85. Environmental factors (|r| > 0.85) were considered to exhibit strong collinearity (Zhao et al., 2024). These factors were merged and filtered based on the Jackknife test results, with selection priority given to those with higher contribution rates. After completing this screening procedure, a collection of 10 environmental factors was obtained, which were utilized in the construction of the species distribution model. (Table 1). This procedure effectively mitigates the risk of model overfitting while ensuring the ecological significance of key environmental drivers (Meyerhenke et al., 2018).

To simulate potentially suitable distribution regions for C. speciosa, we imported its distribution records, along with filtered environmental variables, into the MaxEnt model (version 3.4.4). Using a Logistic output format to evaluate model performance, we implemented a random split of the occurrence records; 75% was the training set for model development. The remaining 25% comprised the validation test set (Dong et al., 2022). To enhance the robustness and generalizability of our predictions, a bootstrap procedure with 10 replicates was used (Ouyang et al., 2022). An iteration ceiling of 500 was established as a safeguard against model overfitting, and the environmental context of the study area was characterized by generating 10,000 background points. When species presence data are limited, using a larger set of background points can help mitigate possible sampling bias. The predictive accuracy of our model was quantified by analyzing the receiver operating characteristic (ROC) curve and calculating the area under the curve (AUC). The AUC values range from 0 to 1, with higher values indicating stronger correlations between the predicted species distributions and environmental factors, indicating superior model performance. The AUC accuracy grading criteria employed in this study were as follows: 0.50–0.60 (failing), 0.60–0.70 (poor), 0.70–0.80 (fair), 0.80–0.90 (good), and 0.90–1.00 (excellent) (Chen et al., 2025). Furthermore, we employed the Jackknife test to quantify the independent contribution of each environmental variable to model construction, thereby identifying the key ecological factors influencing the distribution of C. speciosa.

To address the potential limitations of the AUC metric in specific contexts, we introduced the true specific sensitivity (TSS) statistic. By integrating model sensitivity and specificity, the TSS effectively mitigates sample imbalance bias in species distribution data, resulting in a more balanced evaluation (Allouche et al., 2006). TSS values range from –1 to 1, with values approaching 1 indicating better model performance. The TSS value between 0.6 and 1 indicates good predictive capability. All TSS computations in this study were conducted using the R software (version 4.5.1, R Foundation for Statistical Computing, Vienna, Austria).

The regularization level of MaxEnt is primarily determined by two parameters: the regularization multiplier (RM) and feature combinations (FC) (Yan et al., 2021). The model may employ five fundamental feature functions: linear (L), quadratic (Q), hinge (H), product (P), and threshold (T). This study employed a systematic experimental design to identify the optimal parameter combinations. The regularization multiplier was set incrementally over the range 0.5–4.0, with increments of 0.5. Six distinct feature-class combinations (L, H, LQ, LQH, LQHP, and LQHPT) were evaluated using a cross-validation framework. For model fit assessment, the Akaike Information Criterion Correction (AICc) metric was employed as the evaluation criterion (Muscarella et al., 2014; Wu et al., 2025). The specific computational process was implemented using the ENMeval package in the R programming platform, resulting in AICc quantification values for the maximum entropy model under various parameter configurations. During model selection, the AICc metric was used to assess the relative fit of the candidate models to the optimal model. This value was calculated by subtracting the minimum AICc value of the candidate models from the AICc value of each model. When ΔAICc is <2, the model is considered to have significant statistical support; if ΔAICc exceeds 10, the model’s credibility is considered significantly insufficient (Olofsen and Dahan, 2013). In this study, the optimal modeling scheme with ΔAICc=0 was ultimately selected for simulating C. speciosa habitat.

To visually represent the potential habitat suitability regions for C. speciosa, the model-predicted average species occurrence probabilities were imported into ArcGIS for spatial mapping and classification. Natural breakpoint classification was employed to segment the suitability probabilities. This method determines classification intervals based on the inherent natural groupings of the data, thereby minimizing intra-class variation. Ultimately, four suitability classes were generated: high (0.5–1.0), medium (0.3–0.5), low (0.1–0.3), and unsuitable (0–0.1). This classification scheme, which partitions the continuous probability of occurrence (0–1) into distinct ecological suitability levels, is a well-established practice in species distribution modeling (Zhang et al., 2022). Widely adopted for area calculation and spatial planning, these thresholds are tied to interpretable gradients in habitat suitability. Compared to merely analyzing fluctuations in the area of each suitability class, the migration trajectory of their distribution centers more profoundly reveals the temporal restructuring process of the spatial pattern of suitable regions (Kong et al., 2019). Based on this, the present study quantified areas with different suitability grades across periods. Furthermore, using ArcGIS spatial statistics tools, the present study transformed the continuous suitable distribution surface of C. speciosa into representative distribution centers (Cong et al., 2020; Ahmadi et al., 2023) to characterize the spatial aggregation centers and dynamic shifts of its overall suitable habitat.

Initial modeling using default parameters (RM = 1, FC = LQH) resulted in a ΔAICC of 287.87, indicating a risk of overfitting. To optimize model performance, parameters were systematically tuned using ENMeval in R. Setting RM = 3 and FC = LQHPT reduced ΔAICC to 0, indicating this configuration as the optimal model setting (Figure 4). With these optimized parameters, the predictive performance of the model significantly improved, achieving an average AUC of 0.908 (Figure 5) and a mean TSS of 0.674. Both evaluation metrics confirmed the excellent predictive capability and robustness of the proposed model.

The model results (Supplementary Material 2) indicated that Precipitation of Driest Month (Bio_14), Temperature Seasonality (Bio_4), elevation, and October solar radiation (Srad_10) collectively explained 86.3% of the variation in habitat suitability for this species. Specifically, the Precipitation of Driest Month was identified as the most explanatory environmental factor, accounting for 39.1% of the contribution. The jackknife test further confirmed that Temperature Seasonality (36.9%) and elevation (36.2%) had the highest replacement importance in the model. A comprehensive analysis indicated that moisture conditions, temperature range, and solar radiation energy collectively constituted the dominant environmental influences that determined the geographical distribution patterns of this species (Figure 6).

Based on univariate marginal effect analysis, response curves were generated for the four dominant environmental factors, as determined by their highest contribution rates. Suitable ranges and their extremes were extracted when the probability of occurrence exceeded 0.5 (Wei et al., 2023; Hou et al., 2023). Precipitation of Driest Month (Bio14) exhibited an optimal range of 7.64–214.50 mm (Figure 7A); Temperature Seasonality (Bio4) had an optimal range of 153.85–850.84 (Figure 7B); Elevation exhibited an optimal range of 0–271.65 m (Figure 7C); October solar radiation (Srad_10) exhibited a bimodal response pattern, with suitable ranges of 6835.50–7883.75 and 11370.31–14287.18 W/m² (Figure 7D).

According to the optimized MaxEnt model, the overall climatically suitable zone for C. speciosa under current climatic conditions was approximately 328.40 × 10⁴ km². Within this region, the location of a highly suitable region was 88.20 × 10⁴ km², accounting for 26.86% of the total suitable region. This was primarily concentrated in Shanghai, eastern Shandong, eastern Sichuan, central Yunnan, eastern Jiangsu, northeastern Zhejiang, eastern Hubei, and eastern Hunan, and formed the core region for the growth and development of the species. The moderately suitable region was the largest, reaching 148.89 × 10⁴ km² and accounting for 45.34% of the total. This formed a belt surrounding the highly suitable region and extended outward, covering Anhui, Henan, Jiangxi, western Shandong, southern Hebei, western Taiwan, southeastern Guizhou, and southern Liaoning. The low-suitability region covered 91.32 × 10⁴ km², accounting for 27.81% of the total region. This was primarily located at the periphery of the suitable regions and extensively distributed across Shaanxi, Shanxi, Liaoning, and southern Gansu. This distribution pattern indicates that C. speciosa exhibits broad ecological adaptability under the current climatic conditions, with its suitability range radiating outward from a core region in central-eastern China (Figure 8).

The following projections for the 2050s and 2070s are based on climate data derived from a single global climate model (BCC-CSM2-MR). While this model has demonstrated good applicability for climate simulations in China (O’Neill et al., 2017), it is important to note that projections from different GCMs can vary due to differences in their inherent physical assumptions and parameterizations. Therefore, the distribution patterns and trends presented below should be interpreted as one plausible scenario under each SSP pathway, rather than a definitive forecast. The core trends are robust, but the magnitude and fine-scale spatial details may be subject to model-related uncertainty.

The distribution centroid of the high-suitability area for C. speciosa remained within Hubei Province across all evaluated periods (Figure 11). Currently centered in Yidu, Yichang (111°22’41” E, 30°30’51” N), its future trajectory shows divergence: under SSP245, a northwestward shift of 47.59 km will position it in Yiling, Yichang (111°15’19” E, 30°55’49” N) by the 2050s. However, under SSP585, the suitable habitat will redistribute northeastward, leading to an 87.82 km movement of the centroid to Dongbao, Jingmen (111°58’19’’ E, 31°7’6’’ N).

From the 2050s to the 2070s, under SSP245, the center of mass will migrate 76.43 km northeastward, reaching Dongbao, Jingmen (112°0’60’’ E, 31°8’37’’ N); Under SSP585, the center of mass will shift 20.00 km northwestward, reaching Yuan’an, Yichang (111°47’2’’ E, 31°11’54’’ N) (Supplementary Material 4).

The escalating pace of climate change makes precisely projecting the geographic distribution of medicinal plants essential for informing both their conservation and sustainable management (Pecl et al., 2017). Chaenomeles speciosa, an important traditional Chinese medicinal herb, is experiencing increasing market demand that threatens its limited wild resources, underscoring the need for a comprehensive scientific assessment of its future suitable habitats. This study focused on C. speciosa by integrating geographic distribution data with critical environmental factors. The optimized model exhibits outstanding predictive capability and robustness, validating the success of the parameter tuning process. This study provides a reliable theoretical basis for scientifically formulating resource conservation strategies for C. speciosa and achieving precise habitat management in the context of global climate change.

The dominant drivers were Precipitation of the Driest Month (Bio14), Temperature Seasonality (Bio4), elevation, and October Solar Radiation (Srad_10). Bio14 and Bio4 together explained approximately 66% of the distribution variance, indicating that water availability and temperature fluctuations are the primary factors shaping the distribution of C. speciosa. This aligns with the species’ strong adaptability to water stress, a trait suited to seasonally arid environments (Mattos et al., 2023).

Bio14 had the highest contribution rate. Its suitable range indicates that while C. speciosa tolerates seasonal drought, it still requires a minimum amount of water to complete its life cycle. The species possesses xeromorphic adaptations like deep roots, supporting the view that water availability is a key limiting factor for this genus (Ren et al., 2023). Bio4 was the second most influential variable. Although direct physiological evidence for C. speciosa is limited, studies on related temperate perennials offer insights. For example, experimental winter warming in Ribes nigrum delayed dormancy release, advanced budburst, and reduced fruit yield by 14–41% (Pagter et al., 2015). By analogy, pronounced temperature seasonality likely regulates dormancy, spring phenology, and reproductive success in C. speciosa. Furthermore, the model shows a clear preference for low-elevation habitats, consistent with its distribution in plains and hilly areas. This altitudinal restriction is common among temperate Rosaceae species to avoid climatic extremes. For instance, Pyrus pashia is also confined to low elevations (<1000 m), where warmer temperatures prevent frost damage during spring flowering and ensure sufficient heat for fruit maturation (Kouser et al., 2021). This shared preference highlights that avoiding the combined risks of frost and insufficient heat at high altitudes is key to the lowland distribution of several Rosaceae species, including C. speciosa. Srad_10 exhibits a distinct bimodal pattern. We hypothesize that these two peaks of high suitability may correspond to divergent ecological or physiological strategies. One plausible interpretation is that the lower radiation range supports carbohydrate accumulation prior to dormancy, while the higher range could be associated with enhanced synthesis of photoprotective or medicinal compounds, such as flavonoids, which are abundant in this species (Xu et al., 2023). However, this mechanistic interpretation remains speculative and stems from correlative model output combined with known phytochemical traits; it requires direct experimental validation to confirm any causal relationship between October radiation levels and specific metabolic pathways in C. speciosa.

Our projections reveal a nuanced spatial dynamic: an expansion of the total suitable area coupled with a significant contraction of high-suitability core zones and a non-linear shift of the distribution centroid within Hubei Province. This complex pattern differs from a simple poleward or upward migration often reported (Dai et al., 2022).

Although direct evidence for C. speciosa is currently unavailable from the literature, a study on another important Rosaceae species, Prunus armeniaca, under future climate scenarios similarly projected a northward expansion of its range in China, but concurrently predicted a substantial loss and fragmentation of highly suitable habitat in its traditional core cultivation regions (Ma et al., 2025). This parallel suggests that for some temperate Rosaceae species with specific climatic optima, climate change may not merely shift their range, but also degrade the quality of their habitat within historically optimal areas. Future changes in precipitation patterns are spatially more heterogeneous than uniform warming, potentially driving the complex, non-directional centroid movement observed for C. speciosa. The initial northeast shift under SSP585 may reflect a temporary expansion into regions like the North China Plain, where projected changes in seasonal precipitation temporarily enhance suitability, while the longer-term westward trend may indicate a tracking of more favorable future moisture regimes.

It is crucial to contextualize these projected shifts within the limitations of our modeling framework. Our reliance on a single climate model (BCC-CSM2-MR) means that the precise trajectory, distance, and regional details of the centroid migration, as well as the exact magnitude of habitat gain and loss, carry inherent uncertainty. Future studies employing multi-model ensemble approaches from CMIP6 would help quantify this uncertainty and distinguish between robust, cross-model trends and model-specific artifacts.

This study employed a center-of-mass migration analysis to reveal the spatial restructuring dynamics of C. speciosa habitats under future climate change scenarios. The simulation results indicated that across different climate scenarios and time periods, the distribution center of high-suitability regions for C. speciosa remained within Hubei. However, the migration direction and distance varied significantly. These nonlinear migration trajectories indicate that C. speciosa’s response to climate change is not a simple unidirectional latitudinal migration but rather a complex process influenced by the synergistic effects of multiple environmental factors (Xie et al., 2025).

However, most studies suggest that global warming drives species migration toward higher latitudes or elevations to track suitable ecological niches (Steinbauer et al., 2018; Pecl et al., 2017). The northwestward migration observed in this study may be linked to future changes in key limiting factors in central China, such as seasonal temperature (Bio4) and precipitation during the driest month (Bio14). These changes may considerably enhance suitability in certain northwestern regions, such as Shaanxi and southern Gansu. Additionally, the notable northeastward shift observed in the 2050s under SSP585 may result from the short-term enhancement of the thermo-hygric conditions under extreme emissions in regions such as the North China Plain. This is consistent with the eastward expansion of temperate tree species observed under specific scenarios in several previous studies (Fei et al., 2017; Wang et al., 2023). The shift in the direction of the center-of-mass migration revealed the time-varying effects of different climate change intensities and pathways on species distributions. Consequently, future conservation strategies should consider this spatial dynamic complexity, providing comprehensive spatiotemporal recommendations for the ex situ conservation and habitat management of C. speciosa germplasm resources.

The observed nonlinear shifts in the distribution centroid carry significant implications for conservation planning. First, the centroid’s oscillation within Hubei Province, despite the overall expansion of suitable range, indicates that the core climatic niche of C. speciosa is undergoing spatial restructuring rather than a simple translation. This challenges static conservation approaches focused solely on current core areas. Second, the divergent migration paths under different SSP scenarios highlight the uncertainty in predicting long-term refugia. This variability cautions against deterministic, long-range assisted migration based on short-term projections. Instead, it underscores the need for a dynamic, adaptive management strategy that prioritizes the protection of currently suitable habitats while monitoring and facilitating natural colonization into newly suitable, climatically stable regions identified over longer timeframes.