Exploring Suitable Urban Plant Structures for Carbon-Sink Capacities

Abstract

Urban parks, a type of urban green space, help mitigate environmental pollution and climate change by absorbing and storing atmospheric carbon. Optimizing their carbon-sink capacity requires thoughtful plant community design considering multiple factors. This study analyzed South Korean urban parks using QGIS and i-Tree Eco, integrating satellite imagery with field surveys at both spatial and tree scales. Park spaces were classified into six types based on the biotope criteria established in this study. Random forest regression was applied to each type to identify key variables influencing annual carbon sequestration and storage. The relationship between maturity and sequestration was examined for ten dominant tree species, offering insights for plant selection. Higher tree coverage and more deciduous species were linked to efficiency in carbon sequestration and storage. While variable importance varied slightly across biotope types, tree density was most influential for sequestration, and diameter at breast height and age were key for storage. These findings provide integrated insights into short-term sequestration and long-term storage, as well as strategic directions for structuring plant communities in urban ecosystems. The study offers empirical evidence for designing carbon-efficient urban parks, contributing to sustainable landscape strategies.

1. Introduction

Quantifying the carbon-sink capacity of urban green spaces (UGS) is crucial for understanding their actual and potential contributions to CO₂ management [1]. In the era of carbon neutrality, this has become a key research area. Studies quantifying UGS carbon at national and regional levels can inform future carbon reduction strategies [2]. Since the 1990s, countries—starting with the United States—have increasingly focused on UGS carbon quantification, expanding this research field [3].

Despite these advances, most carbon-sink studies have focused on forests, wetlands, grasslands, and soils [4,5,6,7], with relatively limited research on UGS [8,9] and the factors influencing their carbon-sink capacity [2,9,10]. In South Korea, spatial studies on UGS carbon sinks remain scarce. Although some regions have implemented UGS tree digitization projects, data inconsistencies persist in coverage, classification, and detail. Given the urban scale, digitization approaches are still in the early stages. With 62.7% of South Korea’s land covered by forests, research has primarily examined forest carbon [11,12] or developed models focusing on individual urban trees [13,14,15]. However, estimating carbon sequestration in UGS is challenging due to high spatial heterogeneity and complex urban interactions, which surpass those in forested areas [9,16]. This complexity highlights the need for further investigation. While UGS covers a smaller area and stores less carbon than forests, it plays a vital role in carbon uptake and ecosystem services in densely populated urban areas.

This study aimed to estimate the carbon sequestration and storage (CSS) of UGS and analyze how plant community variables contribute to carbon uptake. CSS can be estimated using sample plot measurements, model estimations, and remote sensing, with method selection depending on study scale, data characteristics, and methodological strengths and limitations [17]. Among model-based approaches, i-Tree Eco is widely used due to its high accuracy in small-scale urban settings, despite limited validation outside the United States [18]. The model provides CSS estimates at both the spatial and individual tree levels, making it suitable for this study. Given the complex interactions between human activities and urban nature, identifying the relative importance of influencing factors is essential [19]. Random forest regression (RFR) is particularly effective for high-dimensional and nonlinear environmental analysis [20]. Among various UGS types, such as gardens and street trees, urban parks provide extensive green space and play a pivotal role in urban carbon offsetting [21]. This study estimated carbon sinks in urban parks and identified key influencing factors using RFR. The findings provide quantitative evidence to support landscape architects, urban planners, and policymakers in integrating carbon reduction strategies into urban park planning.

2. Literature Review

As the role of urban parks as carbon sinks gains increasing attention, there remains limited research exploring strategies and key factors that incorporate ecological and quantitative approaches into urban park planning. In this context, UGS biotopes have been recognized as essential spaces for flora, fauna, and urban populations. They also serve as fundamental units in planting design, particularly within landscape ecological frameworks [22,23,24]. For instance, Wang et al. [25] studied urban parks in Beijing, China, and developed a hierarchical classification system considering individual vegetation and biotope units to optimize planting strategies for carbon sequestration. Species and size were selected as key indicators for individual trees, whereas land cover (gray, green, and blue), tree coverage (open, partly open, and closed), and dominant plant category (mainly grass, shrubs, evergreen trees, and deciduous trees) were designated as biotope level indicators. Using this framework, 16 biotope types were defined, and their relationships with carbon-sink efficiency were analyzed. Similarly, Zhao et al. [16] classified urban parks in Jinan, China, by biotope type and examined their association with carbon-sink efficiency. Several studies have reorganized urban parks into biotope units within a hierarchical framework to evaluate their carbon-sink efficiency. However, research on the relative influence of variables in specific biotopes remains limited. While previous studies have primarily examined individual variable effects on CSS, a comprehensive assessment of multiple variables is necessary for more effective UGS planning.

Recent studies have applied random forest (RF), a machine learning method, to analyze carbon-related variables in urban environments. Breiman [26] introduced the RF algorithm, developed from the Classification and Regression Tree (CART) methodology. In RF, N bootstrap samples are generated from the original dataset, and multiple decision trees are built using randomly selected subsets of variables. The final prediction is derived by averaging outputs from all trees. Variable importance is determined by the increase in mean squared error (InMSE), which measures the change in prediction accuracy when variable values are randomly permuted [27]. RF is advantageous due to its high speed, accuracy, and ability to prevent overfitting and multicollinearity. It also effectively handles nonlinear and noisy data, making it widely applicable in landscape architecture and ecology research [20,28,29]. Given these strengths, RF is a robust tool for analyzing carbon-sink variables in urban environments [30].

Zhang et al. [31] applied RFR with SHapely Additive exPlanations (SHAP) to evaluate the influence of seven landscape pattern indices on carbon-saving capacity in Shangqiu, China. Wang et al. [30] conducted RF analysis of 10 variables in the Yangtze River Economic Belt, identifying the impact of city size and geographic features on carbon emissions. Ye et al. [32] used RF to determine key variables influencing carbon emissions and applied these findings to simulate 2025 land use changes using a cellular automata model. Shi et al. [27] analyzed Moso bamboo forests in Zhejiang, China, integrating RF and structural equation modeling (SEM) to identify key factors affecting carbon stocks. These studies demonstrate RF’s effectiveness in environmental research, particularly in carbon studies.

Research on planting community characteristics in UGS, including urban parks, is increasingly vital for landscape planning that integrates aesthetic, ecological, and scientific principles. This study suggests strategies for enhancing the carbon-sink efficiency of urban parks by analyzing individual trees and biotope units while applying RFR analysis to assess variable importance. The findings establish clear priorities for future urban park planning and contribute to sustainable landscape architecture.

3. Methodology

3.1. Study Areas

The study focused on nine parks in Dongdaemun-gu and Yongsan-gu, two of Seoul’s 25 districts (126°53′–127°02′ E, 37°31′–37°37′ N). Situated in a mid-latitude temperate region, Seoul experiences four distinct seasons, with an average annual temperature of 12.8 °C and precipitation of 1417.9 mm.

Dongdaemun-gu has undergone rapid urbanization since the 1970s, during which residential and commercial development was prioritized over green space to accommodate the growing population [33]. Similarly, Yongsan-gu, despite its central location in Seoul, has failed to ensure sufficient green areas throughout its urban growth, except for pre-existing natural forested areas. The ratios of green space to built-up areas in Dongdaemun-gu and Yongsan-gu are 21.8% and 61.5%, respectively, while the per capita park area remains as low as 3.4 m² and 7.3 m² [34]. These figures fall below the World Health Organization’s recommended minimum of 9 m² per capita and rank among the lowest across Seoul’s 25 districts. In this context, Dongdaemun-gu and Yongsan-gu can be regarded as representative districts of highly urbanized environments with a lack of park-based green space. Therefore, this study focused on urban parks within these two districts to quantitatively assess the CSS functions of urban parks in high-density urban areas.

The study aimed to reorganize urban parks of varying sizes into standardized biotope units to facilitate ecologically comparable analysis. To achieve this, parks with sufficient spatial capacity were required to allow meaningful subdivision into ecologically valid biotope units. Accordingly, nine parks were selected as the subject of analysis. Dongdaemun-gu, in eastern Seoul, contains medium-sized parks (10,000–20,000 m²) and smaller parks (<10,000 m²). In contrast, Yongsan-gu features large parks (>100,000 m²), alongside a medium-sized (10,000–20,000 m²) park and a smaller park (<10,000 m²). To ensure a representative sample of small, medium, and large parks, nine sites were selected based on available field survey data (Figure 1 and

Table 1. Characteristics of study sites.

Region	Park ID	Area (m2)	Land Cover Accuracy (%)
Dongdaemun-gu	Jangan	14,744.8	89.0
	Jangpyeong	15,698.4	90.0
	Saesam	8910.1	82.0
	Tteokgol	5282.7	75.0
	Yongdu	17,182.0	76.0
Yongsan-gu	Giwato	2225.0	88.0
	Hyochang	171,277.0	79.0
	Seobingo	11,879.0	96.0
	Ungbong	187,834.0	82.0
Total			84.1

).

3.2. Satellite Image Processing and Biotope Classification

Integrating high-resolution remote sensing data (≤10 m) with field survey data is necessary for refining urban planting strategies and land cover analysis [35,36,37]. Accordingly, 2 m resolution Compact Advanced Satellite (CAS) 500-1 orthorectified images were incorporated to verify data consistency and analyze land cover [38]. Satellite images from 15 April 2022 (Dongdaemun-gu) and 19 October 2022 (Yongsan-gu) were selected as the temporally relevant datasets.

Land cover classification for each park was performed using QGIS v3.36 to compute the normalized difference vegetation index (NDVI). The NDVI was calculated using near-infrared (NIR) and red (R) spectral bands with the formula NDVI = (NIR − R)/(NIR + R) [39]. Areas with NDVI ≥ 0.3 were classified as green spaces, while those with NDVI < 0.3 were considered non-green spaces. Unlike other urban settings, parks typically contain a higher proportion of green spaces than gray or blue spaces [16,25]. Therefore, the study focused on green land cover. To assess classification accuracy, 100 random points were selected within each of the nine parks (totaling 900 points) and cross-checked against Google Earth images. The classifications achieved an accuracy rate of 84.1% (Table 1).

Following the land cover classification, all parks were subdivided into 20 m × 20 m grids, with each grid considered a valid biotope unit when green land cover exceeded 50%. Based on these units, biotope classification was structured around key ecological characteristics. Openness is a key factor in plant structure design, influencing both visitors’ psychological perception of space and plant community density, which in turn affects carbon sinks [40]. Additionally, evergreen trees contribute to year-round carbon absorption, underscoring their importance in UGS planning [41]. Based on these considerations and previous studies [16,24,25,42], biotope classification in this study was structured around green land cover, tree coverage, and dominant plant category.

In the first tier (openness), biotopes were categorized by tree coverage: open (≤30% tree coverage), partly open (30–70% tree coverage), and closed (≥70% tree coverage). In the second tier (dominant plant category), each biotope was further classified based on tree dominance: deciduous-dominant and evergreen-dominant (

Table 2. Biotope type classification.

Tree Coverage	Dominant Plant Category	Biotope Types
Open	Deciduous trees	Open and Deciduous-dominant space (OD)
Open	Evergreen trees	Open and Evergreen-dominant space (OE)
Partly open	Deciduous trees	Partly open and Deciduous-dominant space (PD)
Partly open	Evergreen trees	Partly open and Evergreen-dominant space (PE)
Closed	Deciduous trees	Closed and Deciduous-dominant space (CD)
Closed	Evergreen trees	Closed and Evergreen-dominant space (CE)

).

3.3. Field Survey and Carbon-Sink Estimation

Tree inventory data collected through field surveys were spatially integrated into the areas identified as valid biotope units. A field survey conducted in Dongdaemun-gu from July to November 2022 collected tree data from urban parks [43]. During this period, a comparable tree inventory was compiled for Yongsan-gu parks. These datasets were used to estimate individual tree Annual Carbon Sequestration (ACS) and Carbon Storage (CS). Given that trees contribute more significantly to UGS carbon sinks than shrubs or ground cover [44,45] and have a theoretical advantage in estimation over soil [21,46,47], the analysis primarily focused on trees.

i-Tree Eco (Urban Forest Effects Model), developed by the U.S. Forest Service Northern Research Station, quantifies urban forest ecosystem services, including CSS. The model integrates field data with meteorological data, including local hourly air pollution data [48]. Although originally designed for the U.S., i-Tree Eco supports databases from Canada, Europe, Australia, South Korea, Colombia, and other countries, making it widely applicable to multidisciplinary research. Consequently, it has been extensively used for carbon estimation at national and regional levels [10,49,50,51]. It provides high accuracy for small-scale urban analyses [18] and simultaneously generates both spatial and individual tree-level results [48]. This study required assessments at both structural and individual tree-level, based on the urban park inventory dataset. Therefore, i-Tree Eco was selected over alternative estimation methods. A complete inventory approach was employed, inputting tree species, diameter at breast height (DBH), latitude, longitude, and local climate data from the nearest observation stations. Field surveys in 2022 collected data on DBH, tree age, and geographic coordinates but did not include direct crown measurements. Therefore, in this study, tree crown width was estimated using the i-Tree Eco model, which derives values from detailed biological and environmental inputs. The study used i-Tree Eco v6.0.35.

The final dataset included 395 biotope units across six types, encompassing 4589 trees (

Table 3. Distribution of biotope units and trees by biotope type.

Biotope Types	Number of Biotopes	Number of Trees
Open and Deciduous-dominant space (OD)	56	278
Open and Evergreen-dominant space (OE)	26	90
Partly open and Deciduous-dominant space (PD)	160	1878
Partly open and Evergreen-dominant space (PE)	75	825
Closed and Deciduous-dominant space (CD)	52	1037
Closed and Evergreen-dominant space (CE)	26	481
Total	395	4589

). An initial classification identified three additional biotope types with balanced deciduous and evergreen distribution. However, these contained fewer than 10 datasets and were excluded to maintain statistical robustness.

3.4. Random Forest Regression

Six independent variables were selected for RFR analyses: tree density, tree coverage, species richness, Shannon–Wiener diversity index (Shannon index), mean DBH, and mean age. These variables represent key physical and structural components of urban park plant communities. Tree density and coverage were considered indicators of biotope structure. Previous research suggested that horizontal structures have a greater influence on plant design than vertical structures [52]. Therefore, this study focused on horizontal attributes. Richness and diversity are essential aspects of plant community composition. Among diversity indices, the Shannon index is widely used in biodiversity research for its versatility [53]. Accordingly, these values were calculated using the vegan package in R [54]. Mean DBH and mean age were included as representative characteristics of individual trees. The importance of these six variables for ACS and CS was evaluated across biotope types.

RFR analyses were performed in R 4.2.2 using the rfPermute package, which operates alongside RF to assess variable importance, p-values, MSE, and R² [55]. The primary RFR parameters—ntree (number of trees), mtry (number of variables tried at each split), and node size (minimum node size)—were set to 1000, 2, and 5, respectively, based on data characteristics and sample sizes per biotope type [27]. The variable importance permutation test was conducted with 500 iterations.

3.5. Research Workflow

The research workflow of this study can be seen in Figure 2.

4. Results

4.1. Carbon-Sink Efficiency by Species

A total of 4589 trees representing 70 species were analyzed. The average DBH was 16.83 cm for deciduous trees and 20.27 cm for evergreens (Figure 3). Most deciduous trees fell within the small-to-medium size range (10 cm < DBH < 25 cm; 44.33%), whereas the majority of evergreens were classified as medium-sized (15 cm < DBH < 35 cm; 60.67%).

Based on individual tree count, the most dominant among the 61 deciduous species were Acer palmatum (16.8%), Zelkova serrata (16.3%), Prunus serrulata (10.6%), Cornus officinalis (7.2%), and Ginkgo biloba (6.1%). Among the nine evergreen species, the most prevalent were Pinus densiflora (64.4%), Pinus koraiensis (18.8%), Pinus strobus (9.5%), Juniperus chinensis (3.5%), and Platycladus orientalis (2.5%). These five species from each group accounted for 57.0% of all deciduous trees and 98.7% of all evergreen trees in the dataset.

Carbon-sink capacities were estimated by analyzing the relationship between DBH and ACS for these ten dominant species (Figure 4). Juniperus chinensis, Pinus densiflora, Ginkgo biloba, and Zelkova serrata exhibited a linear increase in carbon sequestration with DBH. In contrast, Prunus serrulata, Acer palmatum, Platycladus orientalis, and Pinus koraiensis displayed a nonlinear pattern, where sequestration declined when the DBH exceeded 30 cm. Cornus officinalis and Pinus strobus followed roughly linear trends, though their narrow DBH distributions limited growth trajectory assessment.

4.2. Carbon-Sink Efficiency by Biotopes

A consistent pattern emerged when comparing ACS and CS across biotope types (Figure 5). The CD biotope demonstrated the highest efficiency. Closed biotopes had the strongest carbon-sink capacity, followed by partially open biotopes, while open biotopes exhibited the lowest efficiency. Regarding openness, biotopes with tree coverage above 70% showed a particularly strong carbon-sink effect. In terms of species composition, deciduous-dominant biotopes generally had higher average carbon-sink efficiency than evergreen-dominant ones.

To identify key factors influencing ACS and CS, 12 RFR models were developed, each with a biotope type. The ACS models explained an average of 49.4% of variance, with OD showing a lower explanatory power (34.7%) and CE the lowest (22.1%). In contrast, the remaining four biotopes had explanatory powers of approximately 60% (Figure 6).

Tree density was the most influential variable across all biotopes. Tree coverage, representing structural attributes, was highly significant in certain biotopes. Composition-related variables such as richness and the Shannon index had moderate importance but influenced most biotopes. Conversely, mean DBH and age, which reflect individual tree characteristics, had minimal impact on ACS.

For the PD biotope, a comprehensive planning approach incorporating all variables was recommended. The ranking of variable importance in PD was as follows: tree density (37.59%, p < 0.01) > Shannon index (26.33%, p < 0.01) > mean age (22.12%, p < 0.01) > DBH (22.06%, p < 0.01) > tree coverage (21.97%, p < 0.01) > richness (16.51%, p < 0.01). In PE, all variables except mean DBH were significant, with tree density (39.50%, p < 0.01) > tree coverage (29.62%, p < 0.01) > mean age (13.01%, p < 0.05) > richness (12.30%, p < 0.01) > Shannon index (10.91%, p < 0.05). In CD, the top three contributors were tree density (35.27%, p < 0.01) > Shannon index (14.51%, p < 0.05) > richness (14.49%, p < 0.01). For OE, the order was tree coverage (20.36%, p < 0.01) > tree density (17.28%, p < 0.05) > richness (11.19%, p < 0.05) > Shannon index (9.47%, p < 0.05). In OD, the order was tree density (24.35%, p < 0.01) > mean age (16.42%, p < 0.05) > richness (10.00%, p < 0.05), while in CE, tree density (14.50%, p < 0.05) > richness (9.98%, p < 0.05).

In the CS-RFR analyses, the CE biotope exhibited minimal explanatory power (<5%) with no significant predictors. Excluding CE, the average explanatory power across the remaining five biotopes was 45.4% (Figure 7).

OD, PD, PE, and CD followed similar variable importance patterns, with individual characteristics (mean DBH and age) being highly influential. OE and CE exhibited different significance levels, likely due to smaller sample sizes. Structural attributes played a larger role in partially open biotopes but had lower importance than in ACS findings. While community composition variables had lower importance, they remained statistically significant across biotopes.

All variables exerted a combined influence on PD, not only in ACS but also in CS. The ranking for PD was as follows: mean age (27.95%, p < 0.01) > tree coverage (24.15%, p < 0.01) > mean DBH (23.56%, p < 0.01) > tree density (20.63%, p < 0.01) > Shannon index (15.84%, p < 0.05) > richness (14.96%, p < 0.05). Similarly, in CD, all variables were significant, with mean age (24.15%, p < 0.01) > mean DBH (18.26%, p < 0.01) > richness (15.27%, p < 0.05) > Shannon index (14.09%, p < 0.05) > tree density (14.00%, p < 0.05) > tree coverage (10.53%, p < 0.05). For PE, it was tree coverage (27.02%, p < 0.01) > mean age (23.16%, p < 0.01) > tree density (20.22%, p < 0.05) > mean DBH (19.98%, p < 0.01) > richness (9.64%, p < 0.05), indicating a reduced impact of composition variables. For OE, the results were tree coverage (19.88%, p < 0.01) > mean DBH (12.58%, p < 0.01) > Shannon index (8.92%, p < 0.05) > richness (6.83%, p < 0.05), whereas OD had mean age (22.25%, p < 0.01) > mean DBH (17.93%, p < 0.01) > tree density (15.71%, p < 0.01).

5. Discussion

5.1. Hierarchy of Environmental Variables

The ACS analyses identified structural characteristics of urban parks as critical factors in carbon sequestration, with tree density emerging as the most influential variable across all biotope types. Higher tree density and canopy coverage per unit area enhanced photosynthetic activity, increasing sequestration. However, in closed environments, the contribution of tree coverage was negligible once canopy coverage exceeded 70%, indicating its diminishing independent effect. Composition-related variables were moderately significant across most biotope types. Notably, in closed environments, not only tree density but also species richness and the Shannon index had a significant effect. This suggests that in high-density settings with limited sunlight, a mixed-species composition enhances ecological stability more effectively than a homogeneous composition. While previous studies have linked plant community variables to carbon sequestration [52,56], this study observed variations across biotope types, underscoring the need for further research on biotope-specific environmental influences.

The CS analysis highlighted the significance of individual tree attributes, particularly age and DBH. Since CS represents cumulative carbon sequestration over a tree’s lifespan, the increase in aboveground biomass with maturation has a substantial impact. Conversely, when trees become unhealthy or die, the carbon they have stored may be released back into the atmosphere [45,57,58]. This emphasizes the need for continuous health assessments and long-term management strategies for both young and mature trees. Plant community composition variables significantly influenced PD, CD, OE and PE, though their overall impact remained relatively low. Nevertheless, plant community composition remained relevant to ACS. Numerous studies have highlighted species diversity as a key factor in ecosystem resilience [59,60]. Larger-scale landscape units beyond the scope of this study may offer stronger explanatory power for biodiversity dynamics and carbon sequestration efficiency.

5.2. Tree Species Selection Based on ACS

Deciduous species in the study area exhibited relatively high diversity, with Acer palmatum and Zelkova serrata comprising significant proportions. In contrast, evergreen species, predominantly Pinus densiflora, demonstrated lower diversity. Species dominance is influenced by ecological adaptability, cultural preferences, management efficiency, and policy frameworks. Analysis of the relationship between carbon sequestration and DBH for the five most dominant evergreen and deciduous species revealed a nonlinear trend in Prunus serrata, Acer palmatum, Platycladus orientalis, and Pinus koraiensis. Previous studies have similarly noted that carbon sequestration plateaus or declines beyond a maturity threshold due to physiological environmental factors [61,62].

From a user perspective, large, mature trees are often preferred [63]; however, their introduction during planning does not always maximize sequestration benefits. Instead, incorporating species diversity and a mix of tree sizes while considering long-term growth rates is recommended [25]. Species exhibiting a linear relationship between maturity and sequestration may offer strategic advantages as their carbon uptake does not decline with increasing DBH. Given that the variability in sequestration potential varies across climatic zones, regions, and environments, further research should explore a broader range of species and locations.

5.3. Planting Design Strategy for Carbon Efficiency

This study found that biotope types with higher tree coverage and deciduous species dominance were more effective in CSS. However, Wang et al. [25] identified CD and PE types as the most effective for ACS, while CE and PD ranked among the top-performing categories. Some studies also suggest that evergreen trees can be more effective carbon sinks than deciduous trees in the ecosystem [64,65]. Notably, sequestration efficiency varies due to regional contexts, tree distribution, and environmental interactions. This aligns with previous studies emphasizing the complexity of urban tree carbon sequestration [9,16], highlighting the need for further research to address the high heterogeneity of UGS.

Based on the ACS and CS findings, this study proposes integrating the insights into CSS strategies for carbon-efficient urban parks, particularly regarding biotope characteristics. Short-term sequestration can be enhanced by optimizing tree density and canopy coverage within each biotope type, while long-term storage can be improved through the strategic management of mature trees. Open biotopes, characterized by high user activity, require frequent pruning and maintenance. In such areas, introducing fast-growing species with high short-term sequestration efficiency and those with a linear ACS-DBH relationship can improve sink efficiency. Partly open biotopes benefit from maintaining moderate tree density through periodic thinning and supplementary planting while ensuring ecological stability with diverse species and tree sizes. Closed biotopes, which generally have high carbon-sink efficiency, require careful management of tree mortality and succession dynamics. Incorporating small and medium-sized trees supports future canopy development, while dense environments necessitate precise monitoring to mitigate risks from pests, diseases, and environmental disturbances. However, it is important not to interpret these strategies in isolation but rather to integrate them with the spatial and ecological characteristics of the site. These planting and management strategies can be further refined by incorporating threshold analysis to examine temporal variations in their impacts on biotopes and individual trees.

5.4. UGS Carbon-Sink Estimation Method

In South Korea, UGS carbon estimates typically use an allometric equation with a coefficient of 0.8 for forest vegetation. However, urban trees exhibit different growth characteristics based on local conditions, making a fixed coefficient unsuitable. In line with Yoon et al. [66], refined formulas tailored to regional conditions are necessary.

This study used the i-Tree Eco model to estimate carbon-sink capacities. Hauv et al. [67] noted that, while i-Tree has been widely applied across disciplines, it was originally developed in the United States, which may limit its accuracy in capturing annual environmental variations in international contexts. In this study, the most recent Korean meteorological data had not been incorporated into the model, and thus, the 2020 dataset option was applied. Upon consultation with the model development team, it was confirmed that climatic conditions of the two periods were highly similar and that this discrepancy would not have a significant impact on CSS estimations. However, potential inaccuracies may arise from coefficient substitutions in model calculations, spatial scale discrepancies in pollutant dispersion, physiological tree dynamics under varying climatic conditions, and environmental fluctuations beyond default observation periods [49,51,68,69]. Although statistical adjustments using sufficient input data can mitigate such errors [10,50,51,70], a domestic model incorporating local environmental conditions and species data would improve the precision of UGS carbon research at the national level. Some regions in South Korea have digitized tree data, yet inconsistencies in survey criteria and data quality persist. Establishing a standardized system under a central agency could enhance UGS carbon-sink analyses at regional and national scales.

5.5. Limitations of the Study

The study has some limitations, which provide avenues for future research. First, this study focused primarily on trees, the most significant contributors to carbon sequestration in urban parks. Future research could expand to other factors such as soil, shrubs, grass, and ground cover, considering their interactions with land use, vertical structures, and management systems. Second, this study specifically examined plant selection and community structure, as they are directly modifiable in urban parks. The focus was on physical aspects because estimating future socioecological changes remains challenging due to the inherent difficulty of controlling uncertainty and the complex interplay of unpredictable policy factors [71]. To address these limitations, future research could incorporate social and cultural dimensions [72] and adopt a scenario planning approach to systematically assess and compare diverse futures under uncertainty [73]. Expanding analyses to broader regional and national scales may also improve RF model performance, enabling more comprehensive interpretations and insights.

6. Conclusions

Urban parks are essential carbon sinks within urban ecosystems and play an increasingly critical role in achieving carbon neutrality. This study estimated ACS and CS using satellite imagery, field surveys, an ecosystem service analysis model, and spatial analysis tools. This study reclassified urban parks into biotope units and employed machine learning techniques to evaluate the relative influence of key variables across biotope types.

The findings revealed that tree density was the most significant factor influencing ACS, whereas DBH and tree age were the primary determinants of CS. Although plant community composition variables had lower significance, they still contributed to CSS. Among the tree species, deciduous trees exhibited greater diversity, with Acer palmatum and Zelkova serrata being dominant, while Pinus densiflora was the most prevalent evergreen, indicating lower diversity among evergreens. Some species exhibited nonlinear sequestration patterns as they matured, emphasizing the importance of species selection. When comparing ACS and CS efficiencies across the six biotope types, those with greater tree coverage and a higher proportion of deciduous species demonstrated superior CSS efficiency. This study also provides recommendations for improving UGS carbon assessments and synthesizes key findings on influential variables, biotope classifications, and tree species. These findings can inform carbon-efficient urban park planning and management strategies. Furthermore, incorporating the ecological and functional insights from this study into urban park planning will enhance sustainable urban ecosystems and advance data-driven, technology-supported UGS strategies.