Quantifying the Link Between 3D Vegetation Structure and Plant Diversity in Urban Parks Using Fused Multi-Platform LiDAR Data

Highlights

What are the main findings?

• Fusing Unmanned Aerial Vehicle (UAV) and handheld LiDAR precisely captures the complete 3D vegetation structure from understory to canopy in urban parks.
• Vertical structural complexity (Hstd) and 3D vegetation density (VDI) synergistically drive plant diversity, exhibiting significant non-linear relationships and ecological thresholds.

What are the implications of the main findings?

• The stronger predictive power of 3D structure for species richness (R) over the Shannon–Wiener index (H) indicates that physical structural metrics alone struggle to fully capture the spatial variation of community species relative abundance.
• The multi-platform LiDAR fusion framework provides methodological references for urban green space management, transitioning from traditional ‘2D green quantity’ to precise ‘3D green quality’ for biodiversity conservation.

Abstract

Traditional field surveys of urban park biodiversity lack efficiency and scale, whereas LiDAR offers precise 3D vegetation quantification. This study investigates how 3D vegetation structural complexity impacts urban park plant diversity. We integrated Unmanned Aerial Vehicle (UAV) and handheld LiDAR data with ground-based quadrat surveys to capture comprehensive vegetation structures. Using six key 3D structural metrics, we modeled their relationship with plant diversity via Random Forest. Results indicate canopy height standard deviation (Hstd) primarily influences cultivated plant diversity, while the vegetation density index (VDI) drives spontaneous diversity. The model predicted species richness better (30.63% variance explained) than the Shannon index (7.63%). These drivers exhibited significant non-linear effects and potential ecological thresholds. A strong synergy emerged: when the vertical structure is complex and the 3D spatial density is high, the predicted plant diversity initially exhibits a trend of saturation and stabilization. Ultimately, multi-dimensional 3D vegetation structure proves to be a robust indicator of plant diversity. Our proposed multi-platform LiDAR fusion framework enables rapid, precise ecological assessments, providing methodological references and support for the transition from 2D to 3D green quality evaluation in the fine-scale management of similar cities.

1. Introduction

Cities, as the most concentrated areas of human activity, are profoundly reshaping terrestrial landscapes and ecological processes through their rapid expansion [1]. In the course of urban development, natural habitats are often replaced by impervious surfaces, leading to severe habitat loss, fragmentation, and degradation. This poses a significant threat to biodiversity at regional and even global scales [2,3]. Consequently, within these highly anthropogenically disturbed ecosystems, the scientific assessment, conservation, and management of biodiversity have become a central issue and a key challenge for achieving urban sustainability [4,5,6].

In this context, urban parks, as core components of urban green infrastructure, have increasingly prominent ecological and social values [7]. Hosting a rich variety of plant species and complex community structures, urban parks not only act as “urban green lungs” that regulate local climate and purify the air but also serve as vital “refuges” for urban biodiversity [8,9,10]. Plant diversity, in particular, is the foundation of urban park ecosystem functions and directly determines their capacity for environmental regulation and ecosystem stability [11]. Urban plant diversity is intrinsically linked to the urban living environment and sustainable development [12]. By creating varied natural landscapes, abundant plant diversity offers invaluable opportunities for recreation, environmental education, and nature experiences for urban residents. These interactions with nature are crucial for enhancing public physical and mental health and well-being, and for building livable cities where humans and nature coexist harmoniously [13,14,15]. Given the vital role of plant diversity in urban parks, developing efficient and precise monitoring and assessment methods is imperative for guiding the scientific planning and fine-scale management of urban green spaces [16,17,18,19].

Traditional plant diversity surveys primarily rely on methods such as quadrats, transects, or comprehensive botanical inventories [20]. These methods provide high-precision species information, they are inefficient and labor-intensive [21,22]. Particularly in environments with high spatial heterogeneity of vegetation, the limitations of quadrat placement make it difficult to capture the complete information of plant communities across an entire park [23,24]. Moreover, for larger parks or entire urban green space systems, achieving large-scale, continuous, and repeated monitoring is challenging [21]. The rapid advancement of remote sensing technology has made large-scale ecological monitoring feasible. Early studies often utilized passive two-dimensional (2D) remote sensing data, such as the Normalized Difference Vegetation Index (NDVI), to indirectly assess vegetation health, biomass, or horizontal coverage [25,26]. In urban ecology, some studies have used satellite imagery combined with land-use indicators like NDVI to explore the relationship between the spatial distribution of urban green spaces and plant diversity, revealing the impact of urbanization gradients on biodiversity [27]. Other research has found a positive correlation between urban vegetation cover and species richness by analyzing the NDVI of different land-use types [28]. Additionally, some studies have investigated the relationship between diversity and landscape metrics such as patch type, shape index, density, and connectivity [9,29,30], which reflect the degree of landscape fragmentation and the complexity of spatial patterns from various dimensions [31,32]. However, these 2D projection-based metrics fail to effectively capture the vertical structure and complexity of vegetation communities, which is a key driver of biodiversity [33,34].

In ecology, the “habitat heterogeneity hypothesis” is widely accepted. This hypothesis posits that an increase in environmental gradients—as well as in the types of habitats, resources, and structural complexity—expands the available niche space, thereby allowing more species to coexist and supporting higher levels of species diversity [35,36]. In urban environments, this structural heterogeneity can be summarized as the horizontal and vertical structure of vegetation [37,38], which affects the provision of ecosystem services in different ways. The horizontal structure of vegetation primarily refers to the spatial distribution patterns and arrangement of vegetation patches on a two-dimensional plane (e.g., clustering, connectivity, and patch size). This planar layout forms the basis of the park landscape, thereby influencing functional zoning and residents’ activity trajectories [39,40]. The vertical structure of vegetation is manifested not only in the vertical stratification of trees, shrubs, and herbs but also in canopy density, gaps, and the vertical distribution of foliage [41]. Research has shown that structurally more complex parks tend to support a greater abundance and diversity of plant and animal species, thereby maintaining healthier ecosystems [42,43]. However, this positive correlation may not be indefinitely linear. According to niche theory, as habitat complexity continues to increase, the carrying capacity for species may reach saturation due to constraints such as limited spatial resources [44], light competition, or specific anthropogenic management disturbances [45]. Consequently, the growth rate of diversity may slow down or even stagnate. This critical turning point, marking the shift from linear growth to a saturation state, is referred to as an “ecological threshold” [46,47]. In highly engineered urban park ecosystems, identifying this threshold is crucial for determining the “optimal” structural complexity required to maintain high biodiversity; however, few studies have systematically investigated this issue. Although existing research tends to support the ‘habitat heterogeneity hypothesis’ in urban settings, a significant global research gap remains: very few studies have systematically validated it using complete, three-dimensional (3D) structural data within the unique ecosystem of designed and managed urban parks.

Light Detection and Ranging (LiDAR), an active remote sensing technology, offers an unprecedented tool for precisely quantifying 3D vegetation structure and estimating above-ground biomass (AGB) due to its unique ability to penetrate canopies and directly acquire high-density 3D point cloud data [48,49,50]. Its working principle involves emitting short laser pulses and precisely measuring their return times. Combined with data from an Inertial Measurement Unit (IMU) and a Global Positioning System (GPS), each ranging point is converted into 3D coordinates, forming a high-density point cloud on the surface of complex vegetation and thereby accurately reconstructing the vegetation’s 3D structure [51]. Airborne LiDAR has been proven effective in providing detailed vertical vegetation profiles to characterize the structural complexity of natural landscapes [52,53]. This data can be collected over large areas at high spatial resolution [54], and compared to traditional field surveys and passive remote sensing, LiDAR data acquisition is less constrained by accessibility, weather conditions, and illumination requirements [51].

In recent years, the rapid development of Unmanned Aerial Vehicle LiDAR (UAV-LiDAR) has brought new opportunities for ecological research [55]. Compared to traditional airborne LiDAR, UAV-LiDAR offers advantages such as a compact platform, flexible flight paths, high repeatability, and relatively lower costs, enabling fine-scale scanning of specific areas with extremely high point cloud density [56,57]. This low-altitude, high-density data acquisition method significantly enhances LiDAR’s ability to penetrate vegetation canopies and capture structural details near the ground [58,59]. It not only effectively assesses the distribution and spatio-vertical structural changes of urban green space vegetation [60,61] but also enables the automatic detection of individual trees and their physical parameters, achieving remarkable success in fine-scale vegetation monitoring and above-ground biomass (AGB) estimation [62]. For example, innovative voxelization methods can convert discrete point cloud data into 3D grids, thereby constructing novel 3D landscape indices to assess urban forest structure [63]. At the individual tree level, automated Individual Tree Detection (ITD) algorithms can accurately extract key physical parameters such as tree height and crown width from high-density point clouds. These structural parameters can be directly used in allometric models to achieve precise estimation of single-tree AGB [64]. Furthermore, machine learning and deep learning algorithms have been applied to integrate various data sources and improve the accuracy of tree structural parameter and biomass estimations. Studies have shown that the R² for AGB estimation can reach 0.82 ± 0.07 [23,48,65,66], and individual tree detection accuracy can reach 94% [23].

However, despite the immense potential of LiDAR technology in quantifying vegetation structure and AGB, its application to specific ecological questions—particularly in revealing the deep relationship between 3D vegetation structure and biodiversity in complex ecosystems—requires further exploration. To date, studies have successfully used airborne or terrestrial LiDAR point cloud data to investigate the quantitative relationship between vegetation structure and biodiversity in natural forest ecosystems [67,68]. Nevertheless, existing research in this area has significant limitations.

First, study sites are predominantly focused on natural or near-natural forests, with insufficient research on urban parks—unique ecosystems strongly influenced by human activities [69,70]. Unlike natural forests, the vegetation communities in urban parks are artificially designed and managed, exhibiting unique patterns in species composition, community structure, and functional traits. They are also significantly influenced by multiple anthropogenic factors such as management intensity, species configuration, and visitor activities [71]. Some research suggests that while LiDAR performs well in dense forests, its performance in monitoring vegetation in structurally complex yet sparse urban environments can be limited by signal interference [72]. Second, when using airborne or UAV-LiDAR data, existing studies often struggle to capture information about low-stature understory vegetation. This leads to a focus on the tall tree canopy while neglecting the contribution of shrubs and herbs to diversity [43,73]. In urban parks, the shrub and herb layers are vital components for maintaining species diversity and providing rich ecological niches, and their structural characteristics have a non-negligible impact on the entire plant community [30,74]. In fact, handheld or backpack Mobile Laser Scanning (MLS) has been proven effective in capturing high-precision understory structural parameters [75]. To overcome the limitations of a single viewpoint, studies in forestry have attempted to fuse UAV and terrestrial LiDAR data to construct complete 3D forest structures, yielding promising results [76,77]. However, the application of this multi-platform fusion technique remains scarce in urban parks—fragmented artificial ecosystems with complex vertical structures—especially in the context of biodiversity assessment. Therefore, developing a research paradigm that can integrate multi-platform LiDAR data to achieve a holistic perception of 3D structure from the ground to the canopy is crucial for a comprehensive understanding of urban green space ecosystems.

Based on the aforementioned research background and gaps, this study combines high-precision UAV LiDAR data and handheld LiDAR data with detailed ground quadrat surveys to quantify the complete vegetation community structure, including trees, shrubs, and herbs. It systematically investigates the quantitative relationship between 3D vegetation structure and plant diversity in urban parks. The research framework is illustrated in Figure 1. This study aims to answer the following key scientific questions:

(1)

What are the most effective 3D structural metrics for predicting plant diversity in an urban park environment?

(2)

What is the quantitative relationship between key 3D structural metrics and plant diversity, and do ecological thresholds exist?

(3)

Do the driving mechanisms of 3D structure differ across various diversity dimensions and different plant strata?

The findings of this study will not only test the applicability of the ‘habitat heterogeneity hypothesis’ in urban environments but also provide new technological methods and a scientific basis for the ecological management of urban green spaces.

2. Materials and Methods

2.1. Study Area

This study was conducted in East Lake Park, located in Hangzhou, Zhejiang Province, China (Figure 2). Established in 2002, the park is situated between 30°15′21″N to 30°15′36″N and 119°43′15″E to 119°43′34″E, within a typical subtropical monsoon climate zone, covering a total area of approximately 19.8 hectares.

East Lake Park features rich plant communities and configurations, including woodlands, shrublands, lawns, and aquatic vegetation, providing diverse habitats and ecological niches for various species [35]. Due to its long history, the park’s vegetation has been artificially designed and under long-term maintenance, making it significantly different from natural forests [78]. This provides a unique setting for this study to investigate the impact of managed vegetation structure on plant diversity. Additionally, East Lake Park’s prime location and high accessibility make it an important recreational space that accommodates a large number of daily visitors. This strong anthropogenic disturbance continuously influences the park’s vegetation communities, further increasing its ecosystem’s complexity [79,80]. Therefore, the selection of East Lake Park is both typical and representative for this research.

2.2. 3D Point Cloud Model Construction and Pre-Processing

In this study, we utilized a professional-grade DJI Matrice 350 RTK quadcopter (DJI, Shenzhen, China), equipped with a Zenmuse L2 LiDAR (DJI, Shenzhen, China) and a 4/3 CMOS visible-light mapping camera (DJI, Shenzhen, China), to perform oblique photogrammetry and LiDAR scanning for acquiring centimeter-level accuracy 3D scene point clouds. To ensure data quality, a crosshatch flight pattern was adopted at an altitude of 60 m Above Ground Level (AGL) and a flight speed of 6 m/s, with a lateral overlap rate set to 50%. For areas with significant occlusion, such as understory spaces and areas blocked by buildings, a LiGrip H300 handheld LiDAR scanner (GreenValley International, Beijing, China) was used for supplementary scanning. The handheld LiDAR provided high flexibility for capturing detailed point cloud data in complex environments [81]. During scanning, the operator maintained a walking speed of approximately 1.5 m/s and followed a “closed-loop” path. After processing with the SLAM algorithm, the closed-loop trajectory drift error was effectively controlled within 0.02 m, ensuring the spatial reliability of the understory 3D data. The imagery and point cloud data were acquired from 6 to 10 June 2025.

We then used ContextCapture Center (Version 10.21, Bentley Systems, Exton, PA, USA) software to integrate the point cloud data from both the UAV and handheld LiDAR. During the multi-source data registration phase, the UAV point cloud with absolute RTK georeferencing was used as the baseline, and the Iterative Closest Point (ICP) algorithm was applied for fine spatial alignment. The alignment accuracy between the two point clouds achieved a Root Mean Square Error (RMSE) of 0.04 m. To eliminate the local density bias in the merged point clouds, a voxel-based subsampling strategy was employed during data fusion. To match the true physical scale of branches and leaves, the voxel size was set to 0.03 m, retaining only the centroid point within each voxel. This ultimately resulted in the successful construction of a high-precision 3D point cloud model for the study area (Figure 3).

The raw point cloud data of the site was pre-processed in LiDAR360 software (Version 5.2, GreenValley International, Berkeley, CA, USA). This process included several steps: point cloud denoising, smoothing, and ground point classification. Subsequently, the non-ground points were normalized to generate a normalized height point cloud, thereby eliminating the influence of terrain fluctuations on vegetation height [82,83].

Using the software’s built-in machine learning tools, we performed semantic segmentation on the pre-processed 3D scene point cloud data. The objective was to isolate the vegetation class point cloud and reduce interference from other classes, thereby improving the accuracy and efficiency of subsequent individual tree segmentation. The tool is based on a supervised learning method, with a core workflow that includes training sample generation, feature calculation, classifier training and prediction, and accuracy validation [84,85].

For this study, we defined five fine-grained semantic classes: vegetation, ground, buildings, roads, and water. The algorithm employs a supervised learning approach, utilizing geometric and radiometric features to train a Random Forest classifier [86]. The final classification achieved an Overall Accuracy (OA) of 94.4% and a Kappa coefficient of 0.91, indicating reliable classification performance [87]. For detailed information regarding feature calculation, the model training process, and the semantically segmented point cloud model, please refer to Supplementary Material S1.

2.3. Vegetation Information Acquisition

2.3.1. Quadrat Setup

In this study, vegetation information was acquired on a quadrat basis using a systematic sampling design. First, a 50 m × 50 m grid was used to partition the entire core green area of the park. Subsequently, at the grid intersections where vegetation was present, a 20 m × 20 m plant quadrat was systematically established, centered on the intersection. The quadrats were set up in a fixed pattern distribution [88,89].

A nested sampling design was implemented. A total of 32 tree quadrats (20 m × 20 m) were established. Within each tree quadrat, four 5 m × 5 m shrub quadrats were laid out (one at each corner), for a total of 128 shrub quadrats. Additionally, five 1 m × 1 m herb quadrats were established (one at the center and one at each corner), for a total of 160 herb quadrats. This design ensured that the quadrat distribution provided comprehensive and representative coverage of the different vegetation types within the park (Figure 4).

2.3.2. LiDAR-Based Vegetation Information Extraction

To ensure the comprehensiveness and accuracy of the vegetation data, this study employed a hybrid method combining automated extraction from LiDAR with detailed field surveys. Relying solely on field surveys to measure the structural parameters (e.g., Diameter at Breast Height (DBH), crown width, tree height) of every tree is extremely time-consuming and labor-intensive, especially in large park areas [90]. LiDAR and individual tree segmentation algorithms can rapidly and objectively acquire vegetation structural information, significantly improving data collection efficiency [91]. After pre-processing, the entire point cloud dataset, we first used the “Polygon Clip” tool in LiDAR360 software to precisely extract the corresponding point cloud subset for each of the 32 quadrats, based on their field-collected GPS coordinates. All subsequent automated parameter extraction was then performed independently within these point cloud subsets.

This study adopted a “bottom-up” individual tree segmentation method to partition the semantically segmented vegetation point cloud into individual plants, facilitating the precise extraction of structural parameters for each. The workflow is primarily divided into two steps: trunk detection and crown segmentation.

First, we employed the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm for the precise identification and localization of tree trunks [92]. We selected the point cloud subset within the 1.2–1.4 m height range as input, as this height corresponds to the Diameter at Breast Height (DBH) position, where the trunk point cloud is most concentrated and distinct. By setting the minimum number of points (MinPts) to 100 and adjusting the neighborhood radius (Eps) between 0.1 and 0.3 m according to the stand conditions, the DBSCAN algorithm effectively clustered the points of each trunk [93]. Subsequently, the Comparative Shortest-Path (CSP) algorithm was used for fine-grained crown segmentation. This algorithm identifies the highest point above each detected trunk as the treetop and then delineates the precise boundary of each crown by calculating and comparing the shortest paths from the treetop to the crown edges in 3D space. This method effectively addresses segmentation challenges caused by overlapping adjacent crowns or intersecting branches, thereby ensuring the integrity of individual trees [94] (Figure 5).

Upon completion of the individual tree segmentation, all points belonging to the same individual tree were assigned a unique ID. Using the software’s built-in measurement tools, we then quantified the parameters for each individual plant within the quadrats [95]. The Diameter at Breast Height (DBH) of trees was calculated by fitting the trunk point cloud clusters. Tree height was defined as the vertical distance from the highest point in each tree’s point cloud set to the ground, and the crown radius was calculated as the average horizontal distance from all crown points to the treetop. For the shrub and herb layers, their average height and crown width (coverage) were directly quantified from the point cloud [31,96] (Figure 6).

2.3.3. Species Identification and Accuracy Validation Data Collection

Although multi-platform LiDAR fusion technology can accurately capture the 3D physical structure of vegetation, the calculation of diversity indices still relies on accurate species information. Species identification using point cloud data alone remains a significant technological challenge [60]. Furthermore, species-specific individual tree segmentation requires training on extensive sample data, and the structural parameters for these samples are themselves often derived from field censuses.

Therefore, to ensure data integrity and accuracy, we conducted a field survey in the study area from 17 to 23 June 2025. The selection of this timeframe was based on two key considerations: First, we ensured a reasonable quasi-synchronization between the ground survey and LiDAR data acquisition. Constrained by the meteorological conditions required for UAV flights and the extensive workload of ground surveys, there was a temporal gap of approximately two weeks in data collection. However, the core structural metrics in this study are extremely stable over the short term, and no major pruning activities occurred within the plots during this period. This effectively minimized structural discrepancies and guaranteed the accuracy of model training. Second, June corresponds to the peak growing season in the subtropical monsoon climate zone, characterized by complex community structures and high biomass, making it a critical period for capturing the relationship between spontaneous vegetation diversity and 3D structure [97].

During the survey, we performed species identification and counted the population of all individual plants within the quadrats. The plant classification in this study was primarily based on the phylogenetic characteristics (e.g., degree of lignification, branching patterns) from the Flora of China [98] to categorize species into life forms. For vegetation whose spatial attribution could not be clearly determined based solely on qualitative criteria, this study strictly adhered to mainstream quantitative survey protocols in urban forestry and community ecology, which prioritize the “spatial niche” (e.g., the Urban Forest Effects (UFORE) model) [99]. Specifically, the few tree species in a juvenile stage (height < 3 m or DBH < 2.5 cm, i.e., not yet meeting the standard for occupying the ecological canopy space) were temporarily classified and recorded under the shrub layer. Woody plants with a height ≤ 30.5 cm (having not yet formed a stable lignified structure) were included in the herbaceous layer. These adjustments were made to reflect the actual spatial structure of the community [100]. This classification method, based on actual physical morphology, is a widely accepted standard paradigm in structural ecology assessments. It not only accurately reflects the true hierarchical structure and ecological functions of vegetation in 3D space but also ensures absolute spatial alignment between the ground-measured biodiversity data and the physical canopy features captured by LiDAR [33,101]. Additionally, we differentiated between spontaneous and cultivated species. Spontaneous species were defined as plants that were not intentionally planted but germinated naturally or were dispersed by vectors such as wind and birds. To be included in this category, these species were required to appear in at least two independent plots, with no artificial maintenance activities observed during the survey period. Cultivated species referred to plants intentionally planted by municipal greening departments or landscape architects, encompassing both native cultivated species and those introduced from other provinces or foreign countries. The identification and classification of cultivated plants were conducted based on the Catalogue of Cultivated Plants in China and Flora Reipublicae Popularis Sinicae, and were double-verified against the actual planting records of the study site.

We conducted a detailed survey of the 32 sampling plots. This survey involved using an RTK-GNSS device to precisely locate the base of each tree and individual shrub, a diameter tape to measure the Diameter at Breast Height (DBH), a laser rangefinder to measure tree height (H), and a measuring tape for crown width (CW) in both the east–west and north–south directions. This ground truth data was subsequently used for the accuracy validation of the LiDAR-derived measurements to ensure the reliability of the data.

2.3.4. Individual Tree Segmentation Accuracy Validation

To quantitatively evaluate the accuracy of the individual tree segmentation and parameter extraction algorithm described in Section 2.3.2, we systematically compared its automated extraction results with the field-measured ground truth data. First, we calculated the accuracy of individual tree detection (ITD) by spatially matching the tree positions detected by the algorithm with the ground truth data. The results demonstrated that the segmentation algorithm employed in this study performed excellently, achieving a Recall of 91.23%, a Precision of 93.69%, and an overall F1-score of 0.92. This proves that the algorithm can accurately identify the vast majority of individual trees [100,102].

For the successfully detected trees, we further evaluated the estimation accuracy of their key structural parameters [103,104]. Regression analysis was performed to compare the LiDAR-estimated values of tree height, diameter at breast height (DBH), and crown width with field-measured values. The results showed high estimation accuracy for tree height (R² = 0.96, RMSE = 0.85 m) and DBH (R² = 0.94, RMSE = 2.4 cm), while crown width also exhibited a good fit (R² = 0.90, RMSE = 1.02 m). Overall, the individual tree segmentation and parameter extraction methods adopted in this study demonstrated high reliability, providing precise data support for subsequent analyses. Detailed accuracy evaluation metrics, validation methods, and linear regression analysis plots are provided in Supplementary Material S2. To further exclude potential spatial bias, we mapped the spatial distribution of tree height estimation errors based on the validation plots (Supplementary Material S3, Figure S3A). The results indicated that height estimation errors increased slightly in areas with high vegetation density. However, the overall errors were uniformly distributed across the space and were not significantly affected by specific geographic environments.

2.4. Calculation of Diversity Indices

Based on their common application and representativeness in ecological studies, four species diversity indices were selected: species richness (R), the Shannon–Wiener diversity index (H), Simpson’s dominance index (D), and Pielou’s evenness index (J) [105]. The calculation formulas are as follows:

$$R=S$$

(1)

$$H=-{\displaystyle \sum _{i=1}^S{P}_i\ln P_i}$$

(2)

$$D=1-{\displaystyle \sum _{i=1}^S{P_i}^2}$$

(3)

$$J=H/\ln S$$

(4)

where S is the number of species in the quadrat. N_i is the number of individuals of the i-th species, where i = 1, 2, 3…S. P_i is the importance value of a species.

2.5. Acquisition of 3D Structural Metrics

In this study, a series of metrics were extracted from the point cloud data corresponding to each quadrat to quantify the 3D structure of the urban park vegetation.

Mean Canopy Height (Hm) refers to the average height of all vegetation points within a quadrat. It reflects the overall height level of the vegetation and is often used to measure the overall growth status and developmental stage of a community [38,106]. Research has shown that mature plant communities tend to have a higher mean canopy height [107]. As points with a height > 2 m are typically considered vegetation points from the tree/shrub layer, we set a 2 m threshold for this calculation, extracting points above this height from the point cloud model to exclude the influence of the herbaceous layer and other factors [108].

$$\mathrm{Hm}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{n}hi}}{n}}$$

(5)

where h is the height value of a vegetation point, and n is the total number of vegetation points.

The Standard Deviation of Canopy Height (Hstd) is the standard deviation of the height values of all vegetation points and serves as a key indicator of vertical vegetation structural complexity. A larger standard deviation signifies a more complex vertical structure, which may provide more ecological niches [109].

$$\mathrm{Hstd}=\sqrt{{\displaystyle \frac{{\displaystyle {\sum }_{i-1}^n(hi-\mathrm{Hm}{)}^2}}{n-1}}}$$

(6)

The Foliage Height Diversity (FHD) index is used to quantify the evenness of the vertical distribution of vegetation. A higher FHD value indicates that vegetation is more evenly distributed across various layers from low to high, signifying a more complex structure [110]. To calculate this index, the vertical space within a quadrat, from the ground to the highest point, is divided into several vertical layers (j) of equal 2 m height intervals. The number of vegetation points (n_j) falling within each layer is counted, and its proportion (P_j) relative to the total number of vegetation points is then calculated. The formula is as follows:

$$\mathrm{FHD}=-{\displaystyle {\sum }_{j=1}^{S}Pj} \ln (Pj)$$

(7)

where S is the total number of layers, and P_j is the proportion of points in the j-th layer.

The Canopy Gap Fraction (CGF) assesses the ability of light to penetrate the canopy and reach the understory [111]. Studies have shown that areas with higher porosity receive more sufficient understory light, which directly affects the growth and species composition of understory plants [112]. We calculated this metric by counting the number of ground returns and the total number of returns within the non-normalized, raw point cloud and then computing their ratio.

The Vertical Distribution Ratio (VDR) reflects the vertical stratification of vegetation by quantifying the proportion of vegetation points at different height levels [113]. It ranges from 0 to 1. A value approaching 1 indicates that the mean height is much lower than the maximum height, meaning the vegetation is primarily concentrated in the lower layers. Conversely, a smaller value suggests that the mean height is close to the maximum height, indicating that vegetation is concentrated in the upper layers. The calculation formula is as follows:

$$\mathrm{VDR}=\left(\mathrm{Hmax}-\mathrm{Hm}\right)/\mathrm{Hm}$$

(8)

The Layering Index (LI) is used to quantify the number of distinct layers formed by vegetation in the vertical direction. A vertical histogram of height distribution is generated from the 3D point cloud model of a quadrat, where peaks in the point cloud count typically represent different vegetation strata. The Layering Index can therefore be simply defined as the number of these peaks [114]. A higher LI value indicates more pronounced vertical stratification in the vegetation, providing more diverse ecological niches [38].

The 3D Green Volume (GV) refers to the three-dimensional spatial volume occupied by vegetation and is used to measure the overall abundance of vegetation [16]. To calculate it, the normalized vegetation point cloud data within a study quadrat is first divided into a series of uniformly distributed 3D grid cells, where each cell is a voxel (0.1 m × 0.1 m × 0.1 m). The total number of voxels occupied by vegetation points is then counted and multiplied by the volume of a single voxel to obtain the 3D Green Volume within the study area.

Building on this voxelization, the number of vegetation points contained within each voxel is counted. The Vegetation Density Index (VDI) is then calculated as the ratio of the total number of vegetation points to the total number of occupied voxels. It represents the average number of vegetation points contained per unit volume within the quadrat. This metric quantifies vegetation density in 3D space rather than on a 2D plane, thus more comprehensively reflecting the structural characteristics of the vegetation community [115].

$$\mathrm{GV}=Nv\cdot V$$

(9)

$$\mathrm{VDI}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^{N}ni}}{N}}$$

(10)

where N_v is the total number of voxels occupied by the vegetation point cloud. V is the volume of a single voxel. n_i is the number of points in the i-th voxel. N is the total number of voxels.

2.6. Statistical Analysis

First, a collinearity analysis was performed on all 3D structural metrics and diversity indices to prevent model instability resulting from multicollinearity [116]. We calculated the Variance Inflation Factor (VIF) for all predictor variables. Initial tests revealed severe multicollinearity among FHD, Hm, and GV (VIF > 5). Adhering to the standard statistical procedure of stepwise exclusion of variables with the highest VIF, we prioritized the removal of Hm and FHD while retaining GV. This trade-off balances ecological representativeness with practical management applications: the retained Hstd is fully capable of characterizing the vertical heterogeneity of the habitat; simultaneously, compared to FHD—an abstract metric based on information entropy—GV serves as an absolute physical quantity measuring green space volume, thereby offering more intuitive and practical guidance for the fine-scale management of urban parks. Following this iterative exclusion process, the remaining structural metrics exhibited no further multicollinearity (VIF < 5), resulting in the final selection of six metrics as predictor variables. Among the diversity indices, the Shannon–Wiener index (H) and the Simpson index (D) exhibited extremely high collinearity. To avoid information redundancy, we selected the more widely applied Shannon–Wiener index (H) and the fundamental species richness (R) as the core response variables for the subsequent modeling and analysis.

Prior to formally constructing the predictive model, we observed that the data distribution of certain variables did not satisfy the assumption of normality (Shapiro–Wilk test, p < 0.05). Therefore, we employed Spearman correlation analysis to examine the bivariate correlations between the six selected 3D structural metrics and the core diversity indices (R and H). The objective of this analysis was to quantify the strength and direction of the linear relationship between each individual predictor variable and the response variables, providing a baseline reference for the subsequent interpretation of the more complex multivariate model results.

To further quantify the combined influence of multiple 3D structural metrics on plant diversity, this study employed a Random Forest (RF) regression model. RF is an ensemble learning algorithm based on decision trees that can effectively handle complex non-linear relationships and interactions among predictor variables. It does not require specific data distributions, is robust against overfitting, and has been widely applied in ecological research [117]. Using the RF package in the R (Version 4.3.3, R Core Team, Vienna, Austria), we constructed two independent regression models with the six selected 3D structural metrics as predictor variables. The response variables for these models were R and H, respectively. In the model construction, the number of trees was set to 500, and the number of candidate variables sampled at each node split was set to the default value for regression tasks within the package. The overall predictive performance of the models was evaluated by the percentage of variance explained on the out-of-bag (OOB) data (% Var explained). Given the relatively limited sample size in this study, we further performed a permutation test (1000 iterations) on the model to rule out the risk of overfitting to small-sample noise. To identify the core driving factors, we quantified the relative importance of each 3D structural metric by calculating the percentage increase in the Mean Squared Error of the OOB data (%IncMSE) after each predictor variable was randomly permuted [118]. It should be emphasized that the Random Forest model in this study is primarily positioned as an exploratory heuristic tool, aiming to identify the relative importance of variables and potential non-linear ecological trends, rather than to establish an absolute high-precision prediction benchmark.

To visualize how the key metrics influence plant diversity, we generated Partial Dependence Plots (PDPs) to analyze the potential non-linear relationships and threshold effects between the variables [119]. To evaluate the stability of the model and the uncertainty of its predictions given the current sample size, we further employed a bootstrapping approach to analyze these partial dependence relationships. This process involved resampling the original dataset 100 times with replacement, independently training a model, and calculating the PDP for each bootstrap sample. Finally, the results from the 100 iterations were aggregated to compute the mean trend of the partial dependence curve and its 95% confidence interval, thereby providing a more robust basis for the interpretation of the results.

3. Results

3.1. Overview of Community Characteristics

In terms of community structural features, the 3D vegetation structure of the study area exhibited significant heterogeneity. The mean Hstd was 2.80 ± 0.83, ranging from 1.68 to 4.65, which indicates that the sample quadrats covered a full gradient from structurally simple to highly complex. The CGF was generally low with a mean value of 0.166, suggesting that the forest canopy was relatively closed in most areas. Regarding vegetation volume and density, both GV and VDI varied greatly among quadrats (ranging from 888.52 to 6247.29 m³ and 110.7 to 349.25, respectively), revealing substantial spatial differences in vegetation biomass and compactness. From the perspective of vertical stratification, LI ranged from 1 to 5, while VDR ranged from 0.39 to 1.80. This indicates that the community structures covered a spectrum from upper-canopy dominant to vertically uniform, further revealing the community’s complex vertical morphology.

In this survey, a total of 149 vascular plant species were recorded, belonging to 81 families and 134 genera. Of these, 104 were cultivated species, belonging to 53 families and 89 genera. This indicates that to create diverse landscapes, landscape designers introduced plants from a wide range of origins and distant phylogenetic relationships. The most dominant families among the cultivated plants were Rosaceae, Oleaceae, and Sapindaceae. In contrast, the families and genera of spontaneous plants were relatively more concentrated. The 45 spontaneous species belonged to 22 families and 41 genera. Poaceae and Asteraceae were the two most dominant families, comprising a large number of “weeds” or native herbaceous plants. This reflects the patterns of natural succession, whereby only a few plant families and genera are particularly well-adapted to the anthropogenically disturbed environment of a park.

Broken down by strata, the tree layer comprised 54 species, belonging to 28 families and 49 genera. An analysis of the Importance Values (IV) showed that Cinnamomum camphora was the absolute dominant species, with an IV of 43.15. Other species with high IVs were Acer buergerianum (21.46), Choerospondias axillaris (15.30), and Liquidambar formosana (15.17). This indicates that the community’s foundational species are a mix of evergreen and deciduous broad-leaved trees. It is worth noting that a very small number of spontaneous woody seedlings (e.g., Celtis sinensis and Ligustrum lucidum) were observed during the survey. However, in highly artificial urban park environments, constrained by limited growing space and initial design, these spontaneous species struggle to reach mature tree size. According to the vegetation classification criteria of this study (see Section 2.3.3), these juvenile spontaneous woody plants were primarily categorized into the shrub or herb layers for analysis.

The shrub layer consisted of 46 species from 32 families and 41 genera. Shibataea chinensis (IV = 32.51) was the absolute dominant species, being extremely numerous though its distribution was relatively concentrated. Loropetalum chinense (28.02), Ruellia simplex (21.31), and Rhododendron indicum (16.4) also had high IVs, all of which are artificially configured ornamental plants. The herbaceous layer contained 49 species from 31 families and 47 genera. Among them, Ophiopogon japonicus (21.39) and Zoysia japonica (10.35) were the main ground cover plants. Oplismenus undulatifolius (10.33) was a common native herb in the understory, while the invasive alien species Alternanthera philoxeroides (7.4) was also widely distributed within the plant community.

Regarding the level of species diversity, R of the quadrats ranged from 5 to 16, with a mean of 10.1 ± 2.7. H ranged from 0.33 to 1.46, with a mean of 0.94 ± 0.28. This demonstrates that there is a certain degree of spatial heterogeneity in plant diversity among the quadrats, providing a solid data foundation for the subsequent investigation of its relationship with 3D structure.

3.2. Relationship Between 3D Vegetation Structure and Diversity

The results of the Spearman correlation analysis revealed a significant positive correlation between overall plant diversity and the complexity of the vertical vegetation structure (Figure 7). Specifically, Hstd showed a significant positive correlation with both total species richness (Total R) and the total Shannon–Wiener index (Total H), with correlation coefficients (r) of 0.59 and 0.62, respectively (p < 0.01). The VDI also exhibited significant positive correlations with both Total R (r = 0.56, p < 0.001) and Total H (r = 0.36, p < 0.05). This suggests that areas characterized by denser vegetation and more complex vertical structures tend to support higher overall plant diversity. GV showed a significant positive correlation with Total R (r = 0.36, p < 0.05).

When analyzed by different vegetation strata, the results varied. VDI exhibited significant positive correlations with R and H in both the tree layer (r = 0.44 and 0.47, p < 0.05) and the herb layer (r = 0.58 and 0.49, p < 0.01). Both Hstd and CGF were significantly and positively correlated with R and H in the shrub layer (r = 0.41 and 0.44, p < 0.05 for Hstd; r = 0.55 and 0.59, p < 0.01 for CGF). Additionally, CGF showed a significant negative correlation with H in the tree layer (r = −0.37, p < 0.05). In contrast, the correlations between VDR or LI and plant diversity were not statistically significant (p ≥ 0.05).

Analysis based on plant origin (Figure 8) indicated that spontaneous species richness was significantly and positively correlated with VDI (r = 0.43, p < 0.05). The overall plant diversity of cultivated plants was significantly and positively correlated with Hstd (p < 0.05). Breaking down the cultivated plants by strata: the diversity index of the tree layer was significantly negatively correlated with CGF (p < 0.05) and significantly positively correlated with VDI (p < 0.05). The diversity index of the shrub layer showed a high positive correlation with both Hstd and CGF (p < 0.05), while the herbaceous layer had no significant correlation with any of the vegetation structure metrics (p ≥ 0.05).

3.3. Importance Ranking of Driving Factors

The RF model revealed a moderate predictive power for Total R. The results show that the percentage of variance explained by the out-of-bag (OOB) data (% Var explained) was 30.63%, meaning that the six 3D structural metrics collectively explained 30.63% of the variation in Total R. The permutation test further confirmed that the model’s explained variance was significantly higher than the null model distribution under completely random conditions (p = 0.002), effectively ruling out statistical artifacts arising from small-sample fitting (Figure S4a). In contrast, the model yielded a lower explained variance for the overall Shannon–Wiener index (Total H) (% Var explained = 7.63%). Although its permutation test still achieved statistical significance (p = 0.037), confirming a genuine ecological association between 3D structure and this index (Figure S4b), the low explained variance indicates that relying solely on 3D physical structural metrics is insufficient to fully capture the spatial variation in relative species abundance within the community. This suggests that in our study area, while 3D structural features are robust predictors of species richness, their independent predictive power for the Shannon–Wiener index—which measures comprehensive community diversity—is relatively limited.

To further evaluate the spatial applicability of the model predictions, we mapped the spatial distribution of the prediction residuals for Total R (Figure S3b). The results indicated that the prediction residuals for the 32 sampling plots were randomly distributed across the entire park, exhibiting no significant clustering patterns (e.g., bias near road edges or in specific functional zones). This suggests that the RF model was not significantly subject to bias from specific geographic environments and maintained robust performance throughout the study area.

By calculating the importance of each predictor variable, we identified the key 3D structural metrics influencing plant diversity (Figure 9). For Total R, VDI (%IncMSE = 13.01) and Hstd (%IncMSE = 11.00) were the most critical driving factors, with their contributions far exceeding those of other metrics (p < 0.01). This indicates that vegetation density and vertical structural complexity jointly determine the Total R in the area. For Total H, Hstd was the only significant and important driving factor, with an importance value of 11.67 (p < 0.01). This reconfirms the central role of vertical structural complexity in shaping community structure.

3.4. Relationship Patterns of Key Factors and Diversity

We generated Partial Dependence Plots (PDPs) for the most important driving factors in each model to explore their specific patterns of influence (Figure 10). The results reveal that the effect of VDI on Total R follows a complex, non-linear, step-wise pattern of increase. In the VDI range of 110 to 200, the curve exhibits its first phase of rapid ascent. Between VDI values of approximately 200 and 270, the slope of the curve decreases. Subsequently, as VDI increases from 270 to 300, a second, steeper increase occurs. The curve finally plateaus, reaching its maximum value when VDI exceeds 300.

The relationship between Total R and Hstd also exhibited a pattern of initial increase followed by stabilization. As Hstd increased from 1.9 to 3.1, the predicted value of Total R showed a steady rise. At an Hstd of approximately 3.1, this upward trend briefly stagnated or slightly declined. Subsequently, when Hstd exceeded 3.6, the predicted value of Total R entered another phase of rapid ascent until Hstd reached about 4.2, at which point the curve peaked and then showed a slight downward trend.

The relationship between Total H and Hstd also displayed a similar non-linear pattern. When the Hstd value exceeded 2.1, Total H began to increase significantly. After a brief decline at 2.3, it resumed its increase at 2.5 and entered a core interval of stable growth. Within the range of 2.5 to 4.0, Total H steadily rose with the increase in Hstd. When Hstd surpassed 4.0, the curve exhibited a final rapid increase, ultimately peaking at 4.2, after which it entered a saturation phase.

Since VDI and Hstd jointly influence Total R, we plotted a two-dimensional Partial Dependence Plot to further investigate the interaction effect between these two factors (Figure 11). The results revealed a significant positive interaction effect between the two variables on Total R. According to the numerical analysis, when VDI >~ 295.3, Total R remains at a consistently high level. This suggests that when VDI is sufficiently high, variations in Hstd over a wide range do not significantly lower the high predicted values of Total R. The model predicted higher Total R when Hstd was in the 3.2 to 3.4 range, and when Hstd >~ 4.3, the predicted Total R reached its maximum and stabilized. Concurrently, Total H also peaked when Hstd >~ 4.3. Therefore, the current model analysis indicates that plant diversity tends to be maximized only when both VDI and Hstd are simultaneously within high-level ranges.

4. Discussion

4.1. Applicability of LiDAR Technology

By fusing UAV LiDAR and handheld LiDAR, this study successfully achieved high-precision, quadrat-based individual tree segmentation and 3D structural quantification within a complex urban park environment. The accuracy achieved meets or exceeds that of similar research (F1-score = 0.92, R² for parameter estimation = 0.87–0.96) [23,48]. This outcome not only validates the significant potential of multi-platform LiDAR fusion technology as an efficient and precise tool for urban green space inventories but, more importantly, it overcomes the limitation of single UAV platforms, which often struggle to accurately capture understory vegetation structure.

Among the individual tree parameters, tree height was the most accurately estimated, with an R² of 0.96 [120]. The fine-scale scanning from the fused UAV and handheld LiDAR provided an extremely high point cloud density, ensuring that the algorithm could precisely capture the highest point and the ground level for each tree. The tree height parameter is less susceptible to edge effects in small-scale sample plots [121]. Spatial error analysis indicated a slight increase in tree height estimation errors in areas characterized by high vegetation density. This is primarily attributed to high canopy closure limiting the penetration capability of laser pulses, which results in reduced ground point density, thereby affecting, to a certain extent, the construction accuracy of the Digital Terrain Model (DTM) and the subsequent tree height extraction.

Diameter at Breast Height (DBH) also benefited from this approach (R² = 0.94), showing improved measurement accuracy compared to results from studies in larger forest areas [122]. In large-scale UAV or airborne scans, it is often difficult for lasers to penetrate the canopy and densely sample the tree trunk. In contrast, the fine-scale scanning from low-altitude flights fused with ground-based handheld LiDAR allows more point cloud data to reach and cover the trunks, making the method of calculating DBH by fitting a cylinder to the point cloud far more reliable than conventional methods.

The accuracy for crown width also reached a high level (R² = 0.90) but exhibited a clear systematic underestimation bias. For trees with crowns located entirely within the quadrat, the high-density point cloud could accurately delineate the crown edge. However, the crowns of some trees were rendered incomplete by the clipping effect of the quadrat boundaries. This caused the algorithm to calculate a smaller crown width than the tree’s actual width, which in turn led to a systematic underestimation in the overall results [121].

In addition to its excellent performance in estimating individual tree parameters, the applicability of LiDAR technology is also reflected in its ability to capture the spatial heterogeneity of the urban park environment. Traditional methods often rely on field measurements within quadrats, resulting in qualitative or semi-quantitative descriptions of spatial structure [89], or they are based on 2D remote sensing imagery to classify different habitat types and calculate landscape pattern indices for patches. These methods possess a degree of subjectivity and neglect the vertical dimension, which is critical for defining ecological niches. In contrast, the 3D point cloud data acquired through LiDAR technology enables objective, repeatable, and precise measurements of vegetation structure [41,123]. This allows us to test the “habitat heterogeneity hypothesis” with greater precision and to uncover more fine-grained non-linear relationships and threshold effects between structure and diversity.

4.2. Driving Mechanisms of 3D Vegetation Structure on Plant Diversity

This study identifies Hstd as the primary structural factor influencing plant species richness in urban parks. In both the correlation analysis and the RF model, Hstd consistently demonstrated the strongest predictive power. This provides robust empirical support for the “habitat heterogeneity hypothesis” within a managed urban ecosystem [35]. The significant positive correlation found in this study between Hstd and the diversity index of cultivated plants (p < 0.05) precisely reveals the core role of Hstd within the ‘designed ecosystem’ of a park: it serves as a physical manifestation of human design intent. In urban park planning, designers often configure plants of varying heights and life forms to create rich seasonal aesthetics and layered landscapes [31,33]. Consequently, an area designed to have a greater diversity of cultivated plants will naturally exhibit higher vertical structural complexity, resulting in a higher Hstd. Our data explicitly confirm this, showing that plant communities with high Hstd are a product of landscape design and that Hstd can serve as an effective proxy indicator for evaluating the diversity of a landscape’s design.

Our study found that the positive effect of Hstd on diversity operates within an “optimal range.” In the unique context of an urban park, this may reflect the dual driving mechanism of ‘design’ and ‘ecological processes’. For example, the core positive effect on Total R was observed in the Hstd range of 1.9 to 4.2. The rise in Total R within this range primarily stems from rich landscape plant design: more layered configurations of cultivated plants directly lead to a simultaneous increase in both Hstd and cultivated plant richness. When the curve flattens or slightly declines after Hstd exceeds 4.2, it may be due to two factors. On one hand, the diversity of cultivated species within the design may be approaching saturation. On the other hand, it may be related to intensified competition in an overly complex understory [124]. Therefore, the observed stagnation and decline in Total R is the net effect of the saturation of cultivated plants and a decrease in spontaneous plants, clearly demonstrating how artificial design modulates natural ecological processes.

To visually demonstrate the actual vegetation community characteristics represented by different Hstd values, we selected two representative sample plots for comparison (Figure 12). As shown in Figure 12A, the community with an Hstd of 4.2 exhibited typical structural characteristics of a “multi-layered mixed forest.” Such communities are typically composed of a tall canopy layer (e.g., Cinnamomum camphora > 20 m), a dense sub-canopy/shrub layer (e.g., Osmanthus fragrans, Magnolia denudata), and a diverse herbaceous layer. Correspondingly, the LiDAR point cloud displayed high dispersion in the vertical direction. In contrast, communities with lower Hstd (e.g., 1.6) (Figure 12B) appeared as structurally simple open woodlands or single-layered forests, offering relatively limited vertical niche space.

As the other key driving factor for Total R, VDI’s importance was comparable to that of Hstd (%IncMSE of 13.01 and 11.00, respectively). Unlike Hstd, VDI reveals more about the driving mechanisms of natural ecological processes. As the key predictive driver of spontaneous plant diversity, VDI quantifies the density and biomass of vegetation in 3D space, reflecting the abundance and availability of resources within the niche [125]. The diverse microhabitats provided by high-VDI areas increase niche dimensions, creating favorable conditions for the establishment and growth of highly adaptable native herbs or invasive plants (such as Alternanthera philoxeroides, Causonis japonica, and Cyperus rotundus) [30,88]. This explains why the diversity of spontaneous plants (all of which were herbaceous) is significantly and positively correlated with VDI (p < 0.05). The study found that the positive effect of VDI on species richness exhibits a “step-wise” threshold pattern. Within two specific intervals—from a VDI of 110 to 200 and again from 270 to 300—species richness showed rapid jumps. This may reflect the phased processes of species immigration and establishment [126].

This study found a significant positive interaction effect between Hstd and VDI, which jointly drive the enhancement of species richness. This reveals a deeper mechanism: the total plant diversity in urban parks is the result of the superposition of ‘designed diversity’ and ‘spontaneous diversity’. A structurally complex framework (high Hstd) provides habitat for a large number of cultivated plants and when this framework is also resource-rich (high VDI), it creates opportunities for the immigration and establishment of numerous spontaneous plants. Therefore, this synergistic effect indicates that the maximization of species richness arises neither solely from structural complexity nor solely from resource abundance, but from the coupling of the two [124]. This also provides a novel interpretation for the application of the habitat heterogeneity hypothesis in artificial ecosystems—one that combines both design and ecology. In our study, areas where VDI and Hstd were synergistically high were typically ‘near-natural’ patches with less anthropogenic intervention, where communities were allowed to undergo self-succession. This high degree of physical structural heterogeneity and resource availability allows spontaneous vegetation to fully utilize diverse microhabitats, ultimately maximizing overall diversity. Therefore, the combination of VDI and Hstd can serve as an effective proxy of significant managerial importance for identifying biodiversity hotspots or core ecological functional zones in urban parks [127].

When analyzed by stratum, CGF was found to have a positive effect on the diversity of the shrub layer (r = 0.52, p < 0.01; r = 0.38, p < 0.05), but a negative correlation with the diversity of cultivated trees (r = −0.36, p < 0.05; r = −0.39, p < 0.05). This finding quantitatively reveals the ecological effects of canopy gaps as a horticultural management tool. Canopy gaps are considered a form of micro-scale ecological disturbance that can alter the understory light environment [128]. By selectively pruning upper-canopy trees to create these gaps (thereby increasing CGF), more incoming light reaches the understory. This increased availability of photosynthetically active radiation, in turn, promotes the establishment and growth of shrub species [129]. In our study, correlation analysis revealed that GV was significantly and positively correlated with Total R, suggesting that a larger spatial volume occupied by vegetation tends to support a greater number of species. In contrast, neither VDR nor LI demonstrated a significant influence on plant diversity in either the correlation analysis or the RF models. This reflects the limitations of these traditional forest structural metrics when applied to anthropogenically managed urban park ecosystems. Fundamentally, VDR quantifies the vertical “center of gravity” of vegetation biomass (i.e., whether it is concentrated towards the canopy or the understory). However, in urban parks dominated by human design and management, a significant “decoupling” phenomenon has emerged between the vertical centroid distribution of vegetation and species richness. For instance, a manicured hedge composed of a single species (high VDR) and a natural shrub-herb patch consisting of multiple mixed species (high VDR) may be highly similar in their physical vertical centers of gravity, yet they exhibit fundamental differences in species composition. This implies that VDR, as a ratio-based metric, lacks sufficient sensitivity to distinguish between artificial monodominant communities and natural complex communities. Similarly, LI, which quantifies the number of layers by identifying peaks in the vertical distribution of the point cloud, may also be prone to misinterpretation in artificial landscapes. This is because the system cannot distinguish between natural stratification and artificial physical layering; consequently, simply counting the number of layers does not effectively represent true habitat complexity.

This study found that the driving mechanisms of 3D vegetation structure differ for various dimensions of plant diversity (R and H). The predictive power of the RF model for Total R (R² = 30.63%) was substantially better than its power for Total H (R² = 7.63%). Research has shown that, unlike in natural habitats, vegetation dynamics in urban ecosystems are predominantly governed by social decisions and human activities, with landscape design being the main driver of species composition in urban parks [33]. The initial species pool in a park is therefore largely determined by the “planting list” used in its landscape plan [130]. The diverse artificial plantings in urban parks create complex vertical structures and resource bases; this complexity, in turn, provides ecological niches for the establishment and survival of more spontaneous and adapted species, further increasing the overall species richness [16]. Consequently, Total R exhibits a stronger coupled relationship with structural metrics, which results in the model having higher predictive power.

On the other hand, the Shannon–Wiener index (H) accounts not only for species richness but also incorporates the relative abundance (i.e., evenness) of species within the community [88]. The low explained variance of the model (7.63%) indicates that relying solely on 3D structural metrics is insufficient to effectively capture the spatial variation in community species evenness. In highly artificial urban parks, relative species abundance depends largely on the initial landscape planting configurations [33], rather than being filtered or determined by current 3D structural features [131]. Although the complexity of the physical structure provides potential living spaces for various plants, it cannot predict the specific quantitative distribution of species. This not only explains the model’s weak predictive performance for Total H, but also corroborates that in ecosystems subjected to high anthropogenic intervention, relying solely on physical structural metrics is inadequate for a comprehensive assessment of overall diversity.

4.3. Limitations and Future Outlook

This study has revealed potential non-linear relationships and threshold effects between 3D structural metrics and plant diversity, offering new insights into the influence of community structure on plant diversity. However, we also recognize the limitations of our research at its current stage. First, the sample size in this study is relatively limited, which poses a certain risk of overfitting for machine learning models. Although we confirmed the statistical validity of the core variables through Out-of-Bag (OOB) error validation and rigorous permutation tests, and employed a bootstrapping approach to quantify model uncertainty, the confidence intervals for some variable relationships remained wide. This suggests that under these ecological conditions, their relationships may be more complex or heavily influenced by other unmeasured factors. Therefore, the non-linear threshold effects and variable relationships revealed in this study should be regarded as exploratory findings and approximate estimates based on a specific dataset, rather than absolute ecological boundaries. Future research with larger sample sizes is necessary to further validate and refine these ecological relationships.

In terms of technical methods, to achieve physical spatial alignment with the LiDAR point clouds, this study adopted a non-standard, height-based vegetation stratification strategy during the ground survey. Future studies could conduct sensitivity analyses by setting different height classification thresholds to quantitatively assess the potential impact of this approach on model stability. Although the LiDAR technology employed in this study performed excellently overall in quantifying vegetation information and 3D structure, species identification remains a key challenge [132]. This study relied on field surveys to obtain species information, which, to some extent, limited the efficiency and scale of monitoring. Currently, progress has been made in using LiDAR-derived 3D structural information for tree species classification. For example, Yao et al. [60] successfully classified major urban tree families with an overall accuracy of 85% by extracting 33 structural and geometric features from individual tree point clouds and applying a RF algorithm. However, the classification ability that relies solely on structural information is still limited, particularly for cases where species are structurally similar but taxonomically different. Therefore, future research could explore the fusion of hyperspectral imagery with LiDAR point cloud data and utilize advanced algorithms such as deep learning to achieve a precise, automated identification of major tree species in urban parks [133]. This would enable more convenient and efficient large-scale remote sensing prediction and dynamic assessment of plant diversity.

As an exploratory single-site study, the specific ecological thresholds revealed herein are constrained by the unique design concepts and high-intensity management practices of the investigated parks. Future research is urgently needed across broader spatial scales (e.g., varying climatic zones and urbanization gradients) to verify the generalizability of these findings. This study was limited by its single-season survey design (June). Although this period corresponds to the peak growing season, the survey may inevitably miss certain spring ephemerals or autumn-specific species; thus, the findings primarily reflect vegetation characteristics during the peak growth period. Future research should consider conducting multi-season monitoring combined with multi-temporal LiDAR data to capture the dynamic patterns of vegetation communities associated with seasonal phenological changes.

Urban ecosystems are complex systems co-driven by multiple factors. Anthropogenic management is an indispensable variable in shaping community structure and function [134]. However, the lack of quantitative management covariates is a major limitation of this study. Furthermore, the distribution of spontaneous herbaceous plants in urban parks is often highly sensitive to specific habitat types and characteristics (e.g., proximity to water bodies or roads, edge effects, and specific topographical features) [88]. Although the 3D structural metrics used in this study (such as Hstd) reflect light availability and vertical niche differentiation to a certain extent, they cannot fully capture or replace these specific habitat attributes. The absence of these fine-scale habitat variables in the current dataset limits, to some degree, our comprehensive explanation of the distribution mechanisms of spontaneous vegetation. Additionally, VDI identified as a key driver in this study, underscores the importance of microhabitat conditions. It is closely linked to site microclimate and soil conditions. High-density plant communities can form unique and stable microclimatic environments [135], and their abundant litterfall can improve soil physicochemical properties, which in turn supports more diverse plant communities [136]. Subsequent research could incorporate multi-source data—such as microclimate (e.g., understory temperature and humidity), soil physicochemical properties, and specific maintenance measures (e.g., weeding, fertilization, irrigation frequency)—and integrate them with LiDAR-derived 3D structural metrics. This would allow for the construction of more comprehensive predictive models, thereby more fully elucidating the driving mechanisms of biodiversity in urban green spaces.

5. Conclusions

This study establishes an innovative 3D assessment and analytical method to evaluate and manage plant diversity within intensively managed urban ecosystems, offering exploratory insights for urban conservation efforts in similar climatic and management contexts. By fusing multi-platform LiDAR data, we successfully moved beyond traditional 2D metrics to reveal the complex structural drivers of biodiversity. The main conclusions are as follows:

(1)

Habitats with high vertical structural complexity and high 3D spatial density are fundamental for maintaining high biodiversity. The standard deviation of canopy height (Hstd) and the vegetation density index (VDI) are the most critical structural factors driving plant diversity in urban parks, but their driving pathways differ: Hstd primarily influences the diversity of cultivated plants by representing the complexity of anthropogenic design, whereas VDI affects the diversity of spontaneous plants by reflecting resource availability. This finding specifies the classic ‘habitat heterogeneity hypothesis’ in the context of urban parks as a superposition of ‘design’ and ‘ecological’ processes.

(2)

The influence of key structural factors on plant diversity is significantly non-linear, exhibiting potential ecological thresholds suggested by the current dataset alongside synergistic effects. The study quantitatively identified a synergistic effect between Hstd and VDI on plant diversity, demonstrating that the coupling of high ‘design diversity’ and high ‘resource abundance’ is a prerequisite for maximizing total diversity. Based on the current model analysis, the predicted plant diversity tends to be maximized and stabilized only when both VDI and Hstd reach high-level ranges.

(3)

The driving mechanisms of 3D structure differ across various diversity dimensions, plant origins, and vegetation strata. In the context of urban parks, the predictive power of 3D structure for species richness (R) is far superior to that for the Shannon–Wiener index (H). This indicates that 3D physical structural metrics struggle to fully capture the spatial variation in relative species abundance (evenness), and their independent predictive power for comprehensive diversity indices is relatively limited. Furthermore, different structural metrics exhibit trade-offs or even antagonistic effects on different plant strata; for example, the canopy gap fraction is beneficial for shrubs but potentially detrimental for trees.

Based on these conclusions, this study provides the following exploratory scientific references for the planning, design, and fine-scale management of urban parks: In the field of urban green space planning and management, we recommend a gradual shift from a focus on ‘2D green quantity’ to ‘3D green quality.’ We advocate incorporating 3D structural metrics (e.g., Hstd and VDI) into urban planning evaluation systems within similar environments. The methodology presented herein offers an efficient tool for dynamic monitoring, enabling managers to make informed decisions by referencing ecological threshold analyses and trade-offs among different strata. This approach provides a technical reference for fine-scale management interventions, helping urban parks under comparable high-disturbance conditions to gradually evolve into more effective biodiversity refuges.