Huangshan City is located at the southern end of Anhui Province in eastern China, between 117°02′~118°55′ E and 29°24′~30°24′ N. It covers a total area of 9,678 km². The region experiences a humid subtropical monsoon climate characterized by short springs and autumns, long summers and winters, abundant heat, and plentiful rainfall. The mean annual precipitation ranges from 1,400 to 2,200 mm, with rainfall concentrated from May to August. The mean annual temperature is between 15.5–16.4 °C, with 1,647.6 hours of sunshine and a frost-free period of 210–230 days. The terrain is predominantly mountainous, with the greatest elevation difference exceeding 1,700 meters (Figure 1). The primary soil types are yellow soil and mountainous yellow-brown soil, and the city boasts a forest coverage rate of 82.9%. As the core area of Huizhou, the birthplace of one of the three major local cultures in ancient China, Huangshan City is world-famous for its numerous natural and cultural heritage sites (Ma et al., 2024; Wang et al., 2024b).

The foundational data on ancient trees were provided by the Huangshan City Forestry Bureau, encompassing species, age, spatial coordinates and morphological factors of DBH (diameter at breast height, measured at 1.3m above ground), tree height, crown diameter (CD). These data were collected between 2017 and 2022 in accordance with the technical standards outlined in the Technical Regulation for Surveying of Old and Notable Trees. The complete inventory encompasses 10,457 individual ancient trees across 178 species, comprising 20 gymnosperm species and 158 angiosperm species. From this pool, we selected the six most abundant species as our study subjects: the gymnosperms Torreya grandis and Taxus wallichiana var. mairei, the evergreen angiosperms Castanopsis sclerophylla and Camphora officinarum, and the deciduous angiosperms Liquidambar formosana and Aphananthe aspera. Among these, Torreya grandis and Taxus wallichiana var. mairei (also known as Taxus mairei) are classified as Level II and Level I protected species under China’s National Key Protected Wild Plant List, respectively (Laghari et al., 2020). Furthermore, Taxus mairei is listed as Vulnerable (VU) on the International Union for Conservation of Nature (IUCN) Red List (https://www.iucnredlist.org/species/191659/1991533).

Based on the predominantly mountainous terrain of Huangshan City, crucial environmental factors known to influence plant growth (Upadhyay et al., 2022), and existing literature on ancient trees (Xie et al., 2022; Yang et al., 2024), we selected the following environmental factors for the expansion of predictors: altitude and slope (topography), mean annual precipitation (MAP) and annual solar radiation (ASR) (climate), and soil bulk density (SBD) and soil organic matter content (SOMC) (soil properties).

The boundary data of the study area were sourced from the website of the Ministry of Natural Resources of China (https://mulu.tianditu.gov.cn/). DEM (Digital Elevation Model) data were obtained from the Geospatial Data Cloud (https://www.gscloud.cn) using the ASTER GDEM (Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model) 30 m resolution digital elevation dataset. MAP, ASR, SBD, and SOMC data were obtained from the Earth Resources Data Cloud (GRDC) (www.gis5g.com). These datasets and the location information of the selected ancient trees were imported into ArcMap 10.8 and unified to the WGS-84 coordinate system (Figure 2).

The results of the Global Moran’s I analysis (Table 1) indicate that all six species exhibit statistically significant spatial clustering within the study area. Kernel density analysis further reveals the specific clustering patterns for each species (Figure 3), with each species containing multiple density cores. However, the spatial distribution of these core areas varies substantially, demonstrating the unique species-specific spatial domains.

The Chinese official regulations mentioned in the introduction not only define 100 years as the threshold for ancient trees but also classify ancient trees into three protection grades based on age: Grade III: 100–299 years, Grade II: 300–499 years, and Grade I: ≥500 years. As shown in Figure 4, C. officinarum has the highest proportion of Grade I and II individuals, while L. formosana possesses the highest proportion of Grade III individuals.

All morphological factors showed significant positive correlations with age across species. In contrast, the strength and significance of correlations varied considerably for environmental factors (Table 2). All factors retained for subsequent modeling had VIF values below 10, with the majority (94%) below the more stringent threshold of 5, indicating that multicollinearity was not a critical issue for model stability (Table 3).

The performance of the four modeling methods applied with the full predictor set is compared in Table 4 for each species. From MLR to GWR, all three metrics (R², RMSE, MAE) showed significant improvement for every species. From GWR to RF, performance changes were species-specific. Clear gains were observed only for C. sclerophylla, L. formosana, and A. aspera. Finally, the GWRF model outperformed both GWR and RF across all species, achieving the best overall results.

Given the varying contributions of predictors, we simplified the models by removing low-contribution factors and compared the performance metrics of the reduced models (Table 5). For the two OLS-based models (MLR and GWR), predictor screening was performed using backward stepwise selection based on the Akaike Information Criterion (AIC). For the two decision-tree-based models (RF and GWRF), factor importance derived from the MDG value was used as the contribution measure. We tested screening thresholds at cumulative contribution rates of >90%, >80%, and >70%. Comparisons of the evaluation metrics indicated that, except for T. mairei, which performed best with the full predictors, the optimal model performance for all other species was achieved using the >80% threshold. Therefore, the results presented in Tables 5 and 6 are based on the >80% screening threshold.

The results demonstrate that reduced models, utilizing fewer predictors, can achieve performance comparable to or even slightly better than that of full models (e.g., for L. formosana), highlighting their practical advantage. Among all methods, MLR consistently underperformed relative to the others. Meanwhile, RF was consistently outperformed by GWRF across all metrics and species. Consequently, the final comparison focused on the GWR and GWRF models. As illustrated in Figure 5, for T. grandis, C. sclerophylla, C. officinarum, L. formosana, and A. aspera, a comprehensive comparison of evaluation metrics indicates that the GWRF model is the more suitable choice. For T. mairei, the optimal model depends on the management objective. If the goal is accuracy in individual-tree age estimation, requiring the most reliable single-tree value, the GWRF model is preferable due to its lower MAE. Conversely, if the goal is a robust assessment of the overall age structure at the population level, prioritizing control over extreme errors, the GWR model is the more robust choice.

The predictor importance ranking from the GWRF model (Figure 6) reveals considerable variation across species. These differences not only shape the optimal predictor composition for each species-specific model (as reflected in Table 6) but also provide insights into the distinct morphological traits and environmental preferences of different species. For each species, the two highest-ranked morphological and environmental predictors are defined as its key morphological and key environmental factors, respectively. These factors are the variables that contribute the most statistically to explaining variation in long-term growth accumulation (tree age) and are highly associated with the growth process of ancient trees within the study area. Key morphological factors serve as the optimal proxy parameters for externalized age progression. Meanwhile, key environmental factors can act as robust indicators of growth suitability or potential environmental stress for different species.

For T. grandis, T. mairei, C. sclerophylla, and C. officinarum, the key morphological factors are DBH and CD. For A. aspera, they are DBH and Height. For L. formosana, the pattern is distinct: both CD and Height are the lowest-ranked among all predictors and were eliminated during the model reduction process (Table 6). Consequently, only DBH is retained as its key morphological factor.

The key environmental factors for each species are listed in Table 7. Based on univariate regression analyses between these factors and age, the optimal growth range for each key environmental factor can be determined. By comparing these optimal ranges with the observed variation of the corresponding factors across the study area (Figure 2), these ranges can be qualitatively categorized as low, medium-low, medium-high, etc.

The advantage of decision-tree-based models lies in their ability to capture nonlinear relationships between predictors and the dependent variable. A comparison between Table 2 and Figure 6 reveals that for several species, environmental factors which showed no significant linear correlation with age in preliminary analysis were ranked highly in the predictor importance composition (e.g., MAP for T. grandis). To further investigate how the tree-based models captured nonlinear relationships in this study, we performed curve estimation for all key morphological and environmental factors across the six species to identify their optimal functional relationships with the dependent variable. The results show that nonlinear models were consistently preferred over linear ones, yet the associated improvement in R² was very small (Supplementary Table 1). This suggests that the superiority of tree-based models in capturing nonlinearity may not primarily stem from fitting a few strong, isolated nonlinear relationships. Instead, it is more likely attributable to their capacity for integrated learning, which appears to effectively capture and combine multiple weak nonlinear signals along with complex interactions among predictors (Cutler et al., 2007). When using the GWRF model, these subtle signals—difficult to detect in isolation—are effectively extracted and amplified through the ensemble mechanism of decision trees combined with the local weighting scheme of GWR, leading to more accurate predictions. This finding underscores the importance of employing machine learning methods capable of modeling high-order interactions and nonlinearity when addressing problems driven by multiple complex factors, such as predicting the age of ancient trees.

Additionally, it is noteworthy that the GWRF model consistently achieved the lowest MAE for every species, regardless of whether the full or reduced predictor set was used. This indicates that GWRF predictions possess superior accuracy or unbiasedness. Despite fluctuations in other performance metrics, GWRF most stably minimizes the total sum of absolute errors, providing the most reliable point estimates for individual tree age.

The comparative data analysis identifies a theoretical “optimal model,” yet in practice, the choice should be tailored to specific constraints, resulting in a “most suitable model.” When computational capacity is limited, MLR remains a viable option for species like T. mairei and C. officinarum, as its performance disadvantage is not marked. Similarly, if data acquisition is a constraint, models for species where morphological factors contribute high cumulative importance (e.g., T. grandis and C. officinarum) can omit environmental factors, simplifying the survey process. Conversely, when prioritizing higher accuracy and robustness, the model selection should follow the evaluation metrics highlighted in our analysis. For example, the reduced GWRF is indicated for L. formosana, whereas the full-factor GWRF is recommended for T. mairei. Alternatively, following the rationale of this study, further adjustments involving the addition or removal of predictors can be attempted.

A correlation exists between a species’ age structure and both model performance and the distribution of predictor importance. T. grandis, T. mairei and C. officinarum all possess more than 5% Grade I individuals and over 20% Grade II individuals. In contrast, these proportions are considerably smaller in the other three species, particularly in L. formosana and A. aspera, where Grade III individuals constitute over 85% of their populations (Figure 4). We distinguish these groups as having an older age structure (OAS) and a younger age structure (YAS), respectively. Compared to OAS species, the performance advantage of tree-based models is more pronounced for YAS species. During the model reduction process, regardless of the method employed, YAS species retained a greater number of environmental factors (Table 6). Furthermore, the cumulative importance of morphological factors is significantly higher for OAS species than for YAS species (Figure 6), and a significant positive correlation exists between mean age and the cumulative importance of morphological factors (P< 0.01). In summary, prediction models for YAS species rely more on environmental factors than those for OAS species. Differences in age structure reflect distinct ecological strategies. The long-lived, older individuals in species like T. grandis and T. mairei are “historical survivors”; their large sizes directly result from prolonged growth accumulated over extended lifespans after successful establishment in suitable habitats. Consequently, their age information is encoded more profoundly within their morphological traits. In contrast, populations of YAS species are more strongly influenced by contemporary environmental conditions, and their growth is constrained by complex, non-linear environmental factors.

Among the OAS species, T. grandis and T. mairei belong to the gymnosperms of the family Taxaceae. Gymnosperms predominantly live and often dominate in stressful and somewhat stochastic environments. A long life under chronic or episodic stress favors conservative growth and reproductive strategies, which skew strongly toward the “live long and prosper” (Fossdal et al., 2024). T. mairei and other trees in the genus Taxus are commonly known as Yew trees. The longevity of Yew trees also believed to be partly due to their ability to ward off fungal diseases by accumulation of Taxol in their bark (Talbot, 2015). Similarly, as a unique survival strategy, C. officinarum can enhance its resistance to pests and bacterial diseases by releasing camphor oil from its leaves and branches (Lee et al., 2006). Additionally, social and cultural attributes as external drivers also influence the abundance of older age individuals. For instance, T. grandis has a long history as a cultivated economic species in the study area, with its seeds constituting a vital local oil source (Chen et al., 2024b). Meanwhile, C. officinarum holds profound cultural and religious significance in many regions of southern China, where old-growth specimens are often protected and revered (Zhou and Yan, 2016).

This difference not only informs the selection and construction of age prediction models, but also provides a scientific basis for managing designated ancient trees. For OAS species, conservation efforts should prioritize maintaining their existing stable habitats and preventing physical damage. For YAS species, however, management requires proactive intervention to monitor and optimize key environmental factors. The goal is to facilitate the successful survival of more individuals into higher grades, thereby ensuring the long-term persistence of their populations.

As discussed in Section 4.1, providing suitable prediction models for different practical needs represents the primary application of this study for ancient tree management. Beyond this, the identified key morphological and environmental factors can also offer valuable guidance for ancient tree resource surveys and habitat management. Key morphological factors serve as crucial criteria for the official designation and grading of ancient trees. In large-scale field surveys, relying on these key factors effectively balances the demands of operational efficiency and assessment accuracy.

If key morphological factors are applied to the discovery of unknown or potential individuals, then key environmental factors are applied to the conservation and management of known ones. Specific applications include, but are not limited to, the following three aspects. First, diagnosing environmental stress. For example, C. sclerophylla is sensitive to SBD. In the routine management of this species, especially for declining individuals, close attention should be paid to changes in soil compaction within the root zone to prevent soil hardening. Second, anticipating climatic risks. If MAP or ASR is a key factor, integrating future climate change trends allows for preliminary predictions regarding which areas may expose ancient trees to greater water or thermal stress. Individuals in these areas should then be prioritized for intensified monitoring. Third, guiding ex situ conservation. When ex situ conservation of an ancient tree becomes necessary due to natural or societal reasons, priority should be given to relocating it to a “high suitability” site that matches its key environmental factors.

A core challenge in this study lies in the potential inaccuracies inherent in the “tree age” data used as the dependent variable. Some data were derived from literature records or interview estimates rather than direct dating measurements. This potential error may propagate during model training and validation, affecting the absolute assessment of prediction accuracy. However, it should be noted that the primary objective of this study is to compare the relative performance of different methods within a unified existing data framework, an aim that is less susceptible to such systematic data bias. Additionally, the temporal scale of environmental factors may also influence the findings. The growth of ancient trees spans centuries, yet our models relied on contemporary environmental data. The “temporal mismatch” between current conditions and the historical climate fluctuations and land-use changes these trees experienced throughout their lifespans means the models may not fully capture critical long-term growth drivers.

To address these limitations, future research will pursue several promising directions. Non-destructive techniques such as CT scanning and l LiDAR (Martin and Valeria, 2022) will be explored for precise dating of key individuals, thereby supplementing existing data to construct more reliable training datasets. Additionally, efforts will focus on employing multi-source fusion technologies, such as using remote sensing time series to reconstruct recent crown growth dynamics, thereby calibrating and supplementing existing age records. Regarding the temporal scope of environmental factors, a highly promising direction involves reconstructing historical climate sequences (e.g., annual precipitation and temperature data spanning the past century) and integrating them with tree-age data. This approach would enable models to reflect the long-term impact of climate fluctuations on growth.