The Non-Linear Impact of Green Space Recreational Service Performance on Residents’ Emotional States in High-Density Cities

Abstract

Amid accelerating global high-density urbanization, two pressing challenges have emerged: shrinking green space supplies in built-up areas and growing demand for residents’ emotional well-being. Notably, green spaces’ recreational function plays a pivotal role in alleviating emotional distress. This study aims to systematically assess Green Space Recreation Service Performance (GRSP) and unravel its non-linear impact on residents’ emotional states. Using Shijiazhuang—a representative high-density city in China—as a case study, we developed a GRSP evaluation framework integrating supply–demand balance and utilization efficiency. Natural Language Processing (NLP) techniques parsed social media texts, with continuous emotional scores quantifying residents’ emotional states. Finally, a Gradient Boosting Decision Tree (GBDT) model empirically explored the links between GRSP indicators and emotional states. Results show significant spatial differentiation and supply–demand mismatch in Shijiazhuang’s central urban GRSP: peripheral new districts have abundant green space supply but low utilization efficiency, while central built-up areas face insufficient supply paired with high usage intensity. Residents’ self-reported emotional health correlates with green space accessibility and crowding levels, with park distribution equity as the dominant driver. GRSP’s impact on emotional states exhibits non-linearities, threshold effects, and distinct interactions among core indicators. This study identifies key GRSP indicators influencing emotional states, clarifies their non-linear interaction mechanisms and critical thresholds, and provides empirical evidence for advancing emotional health theories in high-density urban contexts.

1. Introduction

The United Nations predicts that the global urbanization rate will surpass 60% by 2025. In high-density cities, intensive land use has caused a shortage of green spaces, reducing the frequency and quality of residents’ interactions with nature and thereby exacerbating mental health issues like depression, anxiety, and loneliness [1,2]. Data from the World Health Organization shows that anxiety disorders affect 4.4% of the global population, while depression impacts 4% [3]. According to China’s National Mental Health Development Report (2021–2022), the positive screening rate for depression risk among Chinese adults reaches 10.6% [4]. As a result, high-density urban areas have become high-risk environments for mental health problems.

Urban green spaces are scarce natural public resources that play a key role in alleviating emotional stress [5]. The concept of green space recreational services evaluates how well these spaces deliver leisure benefits to nearby communities. This assessment focuses on residents’ service benefits, considering both supply-side factors (e.g., spatial quantity and quality) and demand-side factors (e.g., actual access extent) [6], providing a comprehensive measure of recreational service supply and use. Meanwhile, developing an emotion-responsive green space system has become an increasingly important goal in park planning and renewal [7]. Studies consistently link green space exposure to more positive emotions among urban residents [8,9,10]. However, a major research challenge is that the relationship between green space recreational services and emotional benefits is not simply linear. Instead, it involves a dynamic coupling process between two complex systems: one representing green spaces’ spatial, functional, and perceptual features, and the other representing residents’ emotional responses [11]. A central question thus arises: how to effectively coordinate these two systems to maximize green space resource benefits and support residents’ emotional health [12].

Social media data serves as a dynamic analytical tool, offering a novel method to investigate public responses to green spaces [13]. User-generated content—such as comments—enables in-depth analysis of public behaviors and attitudes toward green spaces. Sentiment analysis of social media data can therefore reveal the potential impact of green spaces on residents’ emotions. Against this background, three critical tasks have emerged in the literature: establishing a multidimensional GRSP evaluation framework; using social media comments and NLP to explore residents’ emotional characteristics; and analyzing the non-linear relationships and threshold effects between GRSP and emotional states. These tasks require urgent methodological innovation and practical breakthroughs [14,15].

To address these gaps, this study focuses on the high-density urban areas of Shijiazhuang—a typical high-density city in northern China. First, we analyze GRSP dimensions from service supply and utilization perspectives, then construct a comprehensive evaluation system covering four dimensions: ecological coverage, spatial configuration, accessibility, and activity intensity. This system addresses the oversimplified performance evaluations in existing studies. Furthermore, using geographically referenced emotional big data, we apply NLP techniques to social media comments to uncover spatiotemporal patterns in residents’ emotional states. To overcome the limitation of traditional linear assumptions in capturing complex variable relationships, we employ the GBDT machine learning algorithm to quantify the non-linear relationships and threshold effects between GRSP and emotions. Theoretically, this research clarifies the non-linear mechanisms linking green spaces to emotional health. Practically, it provides a scientific basis for green space planning in high-density urban areas (Figure 1). Methodologically, it integrates multiple approaches to move beyond single-dimensional or linear analyses, offering new analytical perspectives and empirical tools.

Based on this understanding, this study aims to explore the following three core research questions:

(1)

Which GRSP dimensions exert the most critical impacts on emotional health?

(2)

How to quantify residents’ emotional states using social media data?

(3)

Do these impacts exhibit non-linear relationships and threshold effects?

2. Review

2.1. The Benefits of Urban Green Spaces for Emotional Health

The positive role of green spaces in emotional health is empirically supported, with well-being improvements occurring through three primary pathways: physiological regulation, psychological restoration, and social interaction.

At the physiological level, volatile organic compounds released by plants help lower stress hormone levels. Visual exposure to greenery can slow heart rate and blood pressure, reduce overactivation of the autonomic nervous system, and consequently alleviate anxiety and tension [16]. Psychologically, based on Attention Restoration Theory, natural landscapes mitigate directed attention fatigue and promote positive emotions. Studies show that office workers with long-term access to community green spaces score 15–20% lower on the Generalized Anxiety Disorder-7 scale compared to those without such access, with the strongest effect during early morning exposure on workdays [17]. Furthermore, green space type and structure influence benefits: multi-type green spaces with over 60% tree coverage (including shrubs, lawns, and water features) have a stronger positive effect on improving mood in individuals with low mood than single-type lawn spaces. This is likely due to better shade, richer landscape layers, and a greater sense of immersion in nature [18], partly explaining the observed correlation between vegetation diversity and happiness indices [19]. Socially, as public spaces, green areas facilitate neighborhood interaction and reduce social loneliness [20,21], serving as a key pathway through which community greening enhances residents’ mental health [22].

Research has also identified a complex relationship between green space use intensity and emotional benefits. In terms of frequency and duration, emotional health benefits peak when park visits reach 4–5 times per week, after which additional effects diminish [23]. Similarly, exposure to natural environments for about 120 min per week effectively increases subjective well-being, with marginal returns declining thereafter [24]. Additionally, emotional improvements are more pronounced during visits shorter than 10 min compared to 30 min stays, suggesting acute sensitivity to short-term exposure [25]. Regarding recreational experience quality, a significant negative correlation exists between park crowding and emotional health. In high-density urban areas with limited park supply, excessive crowding undermines recreational comfort and the stress-reducing function of green spaces, negatively impacting emotional states [26].

However, most existing studies focus on confirming the overall positive effects of green spaces, paying less attention to the boundary conditions under which these benefits emerge. Findings concerning the relationships between usage frequency, duration, and outcomes remain inconsistent. Moreover, the majority of related research relies on cross-sectional data or short-term experiments, lacking longitudinal follow-up to clarify the dynamics and long-term persistence of these effects.

2.2. Existing Limitations of GRSP Evaluation

The emotional health benefits of urban green spaces stem primarily from their recreational functions. Methodologically, existing GRSP studies predominantly use GIS spatial analysis techniques—such as network analysis and kernel density analysis—for evaluation. These studies typically assess two major dimensions: supply and demand. The supply dimension focuses on objective green space attributes, including spatial accessibility [27,28], landscape esthetic quality [29,30], and ecosystem service value [31,32]. The demand dimension primarily captures residents’ subjective recreational satisfaction via questionnaires [33] or analyzes their actual usage patterns [34]. Achieving a supply–demand balance in urban green space spatial allocation relies on a comprehensive assessment of dimensions like service accessibility and spatial equity [35,36].

However, such research has two fundamental limitations. First, evaluation frameworks often suffer from disconnected dimensions and mechanical indicator superposition. Existing studies generally treat supply and demand as independent categories, lacking an integrated system that systematically combines objective supply quality, subjective user experience, and actual behavioral data. Consequently, it is difficult to accurately identify genuine emotional health benefits in mismatched scenarios, such as “high supply but low usage” areas. Furthermore, while studies frequently measure multiple indicators, they often remain at the level of listing functions or simply combining dimensions, without in-depth analysis of internal mechanisms or interactive effects among these dimensions.

Second, methods for analyzing impact mechanisms are often oversimplified and inadequate for addressing the complexity of high-density urban areas. Most studies use linear regression models when examining the relationship between green space attributes and emotional health. These methods over-rely on linear assumptions [37,38,39,40] and neglect potential non-linear relationships and threshold effects. For instance, they cannot determine whether emotional benefits saturate or diminish once GRSP exceeds a critical threshold. Although some recent studies have begun using structural equation models to explore specific impact pathways [41] or adopted spatial strategy perspectives such as network connectivity [42,43], their scope is mostly limited to a single type or function of green space. As a result, they fail to fully reveal the combined impacts generated by the synergistic effects of multidimensional attributes in high-density urban settings.

In summary, the current mainstream methodological framework—dominated by GIS analysis and questionnaire surveys—struggles to fully uncover the complex, non-linear interaction mechanisms between green spaces and emotional health. This limitation stems from its fragmented structure and constrained analytical methods, restricting the practical guidance that existing findings can provide for refined planning and management. There is therefore an urgent need for an analytical approach capable of integrating multi-source heterogeneous data, clarifying the mutual influences among various factors, and deciphering the complex mechanisms underlying multidimensional interactions and non-linear relationships.

2.3. Emotion Analysis Method Based on Social Media Data

Using social media data to analyze public emotional states and behavioral interactions in green spaces has become a prominent research trend [44,45]. This includes analyzing text from platforms such as Twitter, Weibo, Douyin, and Instagram, alongside posts, images, and videos from Facebook. These studies typically employ large-scale spatiotemporal analysis, with a core methodology that integrates machine learning and NLP to extract semantic and visual features from extensive datasets. For instance, sentiment analysis tools like SnowNLP determine emotional tendencies in Chinese-language texts [46], while image recognition algorithms identify activity types within green spaces. By correlating users’ geographic locations, activity content, and emotional labels, researchers can quantify associations between green space usage patterns and mental health indicators, exploring the underlying mechanisms governing this relationship.

Compared to traditional questionnaire surveys, social media data offers distinct advantages by circumventing inherent sampling and recall biases. Leveraging user-generated content with spatiotemporal tags—such as geotagged tweets [47,48], photos [49], and check-in data [50]—enables real-time, large-scale capture of the public’s spontaneous behavioral and emotional expressions in natural environments. This approach reveals subtler interaction patterns [51] and facilitates urban spatial preference analysis, providing new tools for urban planning in the digital era [52].

Despite the unprecedented scale and granularity of research enabled by social media data, its data-generating mechanism contains inherent biases that often compromise sample representativeness. Most datasets exhibit significant user demographic bias, as active users are predominantly young, highly educated, and technologically proficient. Consequently, their emotional expressions and behavioral patterns cannot represent the entire population, particularly digitally marginalized groups (e.g., some elderly and low-income individuals). Therefore, there is an urgent need to adopt mixed-methods approaches that combine social media data with field surveys. This helps prevent research conclusions from being distorted by platform-specific user characteristics, prevailing cultural trends, and commercial algorithmic recommendations. More importantly, it is essential to recognize that associative mechanisms revealed by social media data do not equate to genuine causal mechanisms. Only through cross-validation using multiple methodologies can the underlying logic of core research questions be uncovered in a scientifically robust and comprehensive manner.

2.4. Application Status of Explainable Machine Learning Frameworks

Machine learning methods are currently recognized for significant advantages in fields such as urban geography, environmental epidemiology, and public health, owing to their capacity to handle high-dimensional and non-linear data [53,54]. Among them, ensemble learning techniques (e.g., Random Forest, XGBoost) are widely used to explore non-linear relationships between urban environmental factors (e.g., green spaces, noise, air quality) and residents’ physical and mental health. Their strong resistance to overfitting and ability to capture complex interactions make them ideal for this purpose [55]. Relevant empirical studies have been conducted across several major global cities, including Beijing, New York, and London [56].

However, most of these mainstream models operate as “black-box” systems. Although they can effectively identify key environmental variables and predict overall impacts, they do not clearly reveal the specific functional forms through which individual factors influence health outcomes—such as non-linear curves, inflection points, or thresholds. This limitation reduces their practical usefulness in guiding precise environmental interventions.

To address this gap, this study introduces an explainable machine learning framework. Designed to intuitively visualize non-linear relationships between individual environmental factors and residents’ emotions, as well as accurately identify impact thresholds, it compensates for the lack of mechanistic interpretability in traditional “black-box” models and clarifies the non-linear mechanisms through which individual factors operate.

3. Materials and Methods

3.1. Study Area

Shijiazhuang, the capital of Hebei Province and a key city in the Beijing–Tianjin–Hebei Urban Agglomeration, is a representative high-density city in northern China. This study focuses on its high-density urban core (Figure 2), with a population of approximately 4.043 million. The research scope encompasses four municipal districts: Chang’an (around 1.11 million), Xinhua (around 0.8 million), Qiaoxi (around 0.98 million), and Yuhua (around 1.153 million). Covering about 467.2 km², this area is characterized by high population concentration (density of 10,986 persons/km²) and an intensive built environment, forming a typical high-density urban zone.

According to the Classification Standard for Urban Green Spaces (CJJ/T85-2017) [57], the study area’s green space system includes 228 parks: 30 comprehensive parks, 59 community parks, and 139 neighborhood pocket parks. The green space rate in the built-up area is 41.15%. These spaces collectively constitute a pivotal natural resource for urban leisure and recreation.

As a medium-sized provincial capital in North China, Shijiazhuang is undergoing an “optimization and upgrading stage” in urbanization. This process reflects a distinctive approach in China’s urban planning: initial rapid residential construction in newly planned provincial capitals is followed by a “rational scale adjustment” phase to align community facilities with the eventual population size. Consequently, while park green space construction in Shijiazhuang has established a foundation, it confronts common challenges of high-density cities, including uneven spatial distribution and significant disparities in service efficiency. Furthermore, the city faces practical constraints due to tight land availability in its built-up areas. This context makes Shijiazhuang a typical and representative setting for investigating the impact mechanism of GRSP on residents’ emotions.

3.2. Data Sources

This study integrates multi-source datasets characterized by high precision and timeliness. Road infrastructure data were obtained from the OpenStreetMap open platform (2024 release; https://openstreetmap.org/), providing a comprehensive vector road network for Shijiazhuang’s administrative area. The dataset includes expressways, arterial roads, collector roads, and local roads, stored in SHP format with a spatial precision of 10 m. Green space data were derived from Landsat 8 OLI remote sensing images acquired in July 2024. Images from this period typically exhibit low cloud cover, ensuring high data quality. Extracted green space information was cross-verified against Baidu Maps and OSM data. Key attributes—including boundaries, location, name, area, and administrative district—were validated, forming an empirical basis for analyzing green space distribution patterns and service coverage. Socioeconomic data were sourced from the 2023 Shijiazhuang Statistical Yearbook, supplemented with updates from the Seventh National Population Census. This includes population density data refined to the sub-district level, offering a structural foundation for examining the relationship between population distribution and public service allocation. Social media data were collected via the Sina Weibo Open Platform API (2024; https://open.weibo.com/), comprising 173,668 user comment texts geotagged within Shijiazhuang. Each record contains over 15 key fields, such as posting time, geographic coordinates, and interaction metrics, stored in JSON format. Under compliant collection protocols and rigorous sample screening, these texts can reasonably reflect the emotional state of local residents. Additionally, spatiotemporal big data were incorporated from the 2024 Baidu Huiyan Platform (https://huiyan.baidu.com/), which offers a spatial resolution of 100 m and a temporal resolution of one hour. This dataset captures dynamic population aggregation during typical periods—including weekday morning peaks, weekday midday off-peak hours, and weekend afternoons—thus delineating fine-grained spatiotemporal patterns of population activity.

3.3. Evaluation of GRSP in High-Density Urban Areas

The core explanatory variable in this study is GRSP. Integrating multidimensional characteristics—including ecological coverage, spatial configuration, accessibility, and activity intensity—this variable systematically measures the actual performance of green spaces in delivering recreational services [6]. Grounded in public goods supply–demand theory and public service theory, and aligned with the conceptual connotation and functional features of GRSP [58,59], the performance measurement framework is divided into two dimensions: a supply achievement dimension (evaluating service provision levels) and a utilization efficiency dimension (assessing service use status) (

Table 1. Evaluation system for GRSP.

Evaluation Dimension	Category	Factor (Abbreviation)	Unit	Description
Recreational Service Achievement (RSA)	Ecological Coverage Achievement	Park Area Ratio (PAR)	%	The proportion of green space area to the total area, measuring the macro-level of green space resources.
		Vegetation Coverage Rate (VCR)	%	The proportion of vegetation-covered area to the total area, reflecting the ecological quality and naturalness of the region.
		Water Coverage Rate (WCR)	%	The proportion of water-covered area to the total area, reflecting the urban blue-green space structure.
	Spatial Configuration Achievement	Park Distribution Equity (PDE)	-	The sum of the proportions of service coverage areas of various park types to the total area, embodying the equity of residents in different locations in accessing park services.
		Per Capita Park Guarantee (PCPG)	%	The ratio of service coverage area of various park types to the served population, representing the level of per capita green space resource access.
Recreational Service Efficiency (RSE)	Accessibility Efficiency	Comprehensive Park Accessibility (CpPA)	m2/person	The per capita available effective service area of comprehensive parks for residents, calculated using the 2SFCA method with a service search radius of 2000 m.
		Community Park Accessibility (CmPA)	m2/person	The per capita available effective service area of community parks for residents, calculated using the 2SFCA method with a service search radius of 1000 m.
		Pocket Park Accessibility (PPA)	m2/person	The per capita available effective service area of pocket parks for residents, calculated using the 2SFCA method with a service search radius of 300 m.
	Activity Intensity Efficiency	Heat Index (HI)	-	The average heat value collected by Baidu Heat Map every two hours, identifying the usage vitality and functional hotspots of urban space.
		Comment Density (CD)	comments/person	The ratio of the number of Weibo comments to the per unit population, representing the level of public attention and discussion on the spatial quality and activity experience of the region.

).

Within the Recreational Service Achievement (RSA) dimension, the focus is on evaluating the ecological base and spatial configuration of park green spaces. The Recreational Service Efficiency (RSE) dimension, in contrast, emphasizes measuring residents’ actual access to and use of park green space services. This indicator system therefore addresses both the inherent ecological attributes of parks as green spaces and their operational efficiency and user experience as public service facilities.

Among the indicators under the accessibility efficiency sub-dimension, the Gaussian two-step floating catchment area (2SFCA) method is employed to evaluate the supply–demand matching between park green spaces and residential populations, incorporating a spatial distance-decay effect. This method uses the park’s effective service area—the net area available for recreational activities after excluding internal inaccessible zones—as the supply measure, and community population as the demand measure. First, the weighted population that each park can serve within its search radius is calculated to derive a supply–demand ratio. Then, the weighted supply of all reachable parks within each residential point’s search radius is aggregated, yielding the per capita effective park area available to residents at that location. A higher value of this indicator reflects a more abundant and conveniently accessible supply of effective park services for the corresponding residential point.

For the activity intensity efficiency sub-dimension, the Baidu Heat Index is used to characterize the actual frequency of green space use and the intensity of recreational service utilization.

Building upon the aforementioned GRSP evaluation system, this study employs the entropy weight method to objectively determine weights for each indicator. The Natural Breaks method classifies the quantitative data of each grid cell into five gradient levels. Finally, the weighted sum model is adopted to calculate the comprehensive GRSP score by integrating the two core dimensions—recreational service achievement and utilization efficiency. The specific formula is presented in Equation (1):

$${\mathrm{RSP}}_{\mathrm{i}}={\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\mathrm{X}}_{\mathrm{i}}\times {\mathsf{\omega }}_{\mathrm{i}}$$

(1)

where ${\mathrm{X}}_{\mathrm{i}}$ represents the dimensionless quantified value of the GRSP indicator m corresponding to grid cell i, and ${\mathsf{\omega }}_{\mathrm{i}}$ represents the weight of the individual evaluation indicator m for park recreational service efficiency.

3.4. Assessment of Residents’ Emotional States Based on Weibo Comment Data

This study selected Sina Weibo—China’s predominant social media platform—as the primary source for collecting online commentary. The platform features high user engagement in Shijiazhuang’s local “livelihood discussion” forums, where residents freely express sentiments about urban life and interact with one another. This provides an authentic, real-time, and rich textual dataset for assessing resident emotional states. Building upon established frameworks for using social media data in urban perception evaluation [60,61,62], the collected posts display clear emotional tendencies and strong geographical relevance, making them highly suitable for investigating the emotional status of residents in Shijiazhuang’s central urban area.

API access was obtained through the Weibo Open Platform, and a total of 173,668 geotagged original posts published from January to December 2024 within Shijiazhuang’s central urban area were collected. Following systematic data cleaning, 145,240 valid comments within the study area were retained.

Subsequently, leveraging the Tencent Cloud NLP Platform, this study employed its sentiment analysis interface (request domain: nlp.tencentcloudapi.com) to process and analyze the subjective emotional content. This sentiment analysis function, built upon web-scale corpora and deep neural network models, effectively identifies user emotional tendencies and outputs probabilities for positive, neutral, and negative sentiment. Each probability ranges from 0 to 1, with the sum of the three equaling 1.

The positive sentiment probability for a single Weibo post is defined as the “individual text sentiment score,” a continuous variable ranging from 0 to 1. Lower values indicate more negative sentiment, while higher values indicate more positive sentiment. Following common research practice, this score is classified into three sentiment polarity levels: 0–0.4 (negative), 0.4–0.6 (neutral), and 0.6–1 (positive).

To enable spatial representation of sentiment, the arithmetic mean of sentiment scores for all Weibo posts within a specific grid cell is defined as the “grid-level resident sentiment score,” characterizing the overall emotional state of residents in that cell. Through this method, all sentiment data are aggregated into grid-based spatial units, ultimately generating a resident sentiment spatial distribution map that visualizes the geographical pattern of emotions. The specific calculation formula is provided in Equation (2):

$${\overline{\mathrm{RS}}}_{\mathrm{j}}={\displaystyle \frac{1}{\mathrm{i}}}{\sum }_{\mathrm{j}=1}^{\mathrm{i}}\mathrm{R}{\mathrm{S}}_{\mathrm{ij}}$$

(2)

where $\mathrm{R}{\mathrm{S}}_{\mathrm{ij}}$ represents the sentiment score of the i-th comment data in the j-th grid. The closer the average sentiment score of a grid is to 0, the stronger the negative sentiment; the closer it is to 1, the stronger the positive sentiment.

To validate the reliability and validity of the social media (Weibo) data, this study conducted on-site surveys using a life satisfaction questionnaire to cross-verify the quantitative results of residents’ emotional states derived from Weibo posts. Through simple random sampling, two residential communities were selected from each of the 60 sub-districts within Shijiazhuang’s high-density urban area (120 communities in total). A total of 1200 copies of the Satisfaction with Life Scale were distributed.

Questionnaire results were aggregated to the sub-district level, and Pearson correlation analysis was performed to examine the relationship between the questionnaire-based life satisfaction scores and the Weibo-derived measures of residents’ emotional states. The analysis revealed a significant positive correlation between the two datasets (r = 0.75, p < 0.01), thereby confirming the reliability and validity of the Weibo-derived emotional state data.

3.5. Exploring the Non-Linear Driving Mechanism of GRSP on Residents’ Emotional States

3.5.1. Data Matching and Model Input

To explore the non-linear impact of GRSP on residents’ emotions, this study began with data matching and model input construction. In terms of spatial units, and in accordance with the living circle planning concept outlined in China’s Standard for Planning and Design of Urban Residential Areas (GB 50180-2018) [63], a 500 m × 500 m grid was adopted as the basic evaluation unit. This resulted in a total of 1017 grid cells. This scale was chosen as it adequately covers a typical 10 min community living circle, while avoiding the masking of local characteristics by overly large units or data fragmentation caused by excessively small units. Consequently, it ensures that the performance measurement aligns closely with the actual recreational accessibility range of residents. For each grid, GRSP indicators—including individual metric values and the comprehensive score—were aggregated. Simultaneously, the annual average emotional value for each grid was calculated using Weibo text data to characterize resident emotions. Grids lacking emotional data were excluded, with the remaining valid grids retained for modeling. Subsequently, a GBDT model was constructed. Using the park performance indicators within each grid as feature variables and the average grid emotion value as the target variable, this study analyzed the associative mechanism between GRSP and resident emotional states.

3.5.2. GBDT Model

To investigate the non-linear relationship and potential threshold effects between GRSP and residents’ emotional states in Shijiazhuang’s central urban area, this study employed a GBDT model, implemented in Python 3.10. As a classic ensemble learning algorithm, the GBDT model operates on the core principle of iteratively training decision trees to fit the prediction residuals of the preceding models, thereby gradually reducing prediction error and enhancing overall model accuracy. A systematic interpretability analysis was subsequently conducted on its predictions to decipher the internal decision-making mechanisms of this high-performance machine learning model.

First, the model is initialized, and the corresponding formula is presented as follows:

$${\mathrm{f}}_0\left(\mathrm{x}\right)={\mathrm{argmin}}_{\mathrm{c}}\sum _{\mathrm{i}=1}^{\mathrm{N}}\mathrm{L}({\mathrm{y}}_{\mathrm{i}},\mathrm{c})$$

(3)

where ${\mathrm{f}}_0(\mathrm{x})$ denotes the initial model for ensemble learning, ${\mathrm{argmin}}_{\mathrm{c}}$ represents finding the parameter c that minimizes the subsequent expression, that is, identifying a constant c to minimize the total sum of the subsequent losses, $\mathrm{L}({\mathrm{y}}_{\mathrm{i}},\mathrm{c})$ is the loss function used to calculate the difference between the actual value and the predicted value, and c is the constant that minimizes the loss function.

After entering the iterative optimization phase, for each round m = 1, 2, …, M and each sample i = 1, 2, …, N, the prediction residual of the preceding model ${\mathrm{F}}_{\mathrm{m}-1}(\mathrm{x})$ is calculated. Since the core of gradient boosting lies in using the negative gradient of the loss function as an approximation of the residual, the expression for the approximate residual of the i-th sample in the m-th round is as follows:

$${\mathrm{r}}_{\mathrm{mi}}=-{\left[{\displaystyle \frac{\partial \mathrm{L}({\mathrm{y}}_{\mathrm{i}},\mathrm{f}({\mathrm{x}}_{\mathrm{i}}\left)\right)}{\partial \mathrm{f}({\mathrm{x}}_{\mathrm{i}})}}\right]}_{\mathrm{f}\left(\mathrm{x}\right)={\mathrm{F}}_{\mathrm{m}-1}\left(\mathrm{x}\right)}$$

(4)

For the squared error loss function, substituting it into the aforementioned formula yields ${\mathrm{r}}_{\mathrm{mi}}={\mathrm{y}}_{\mathrm{i}}-{\mathrm{F}}_{\mathrm{m}-1}\left({\mathrm{x}}_{\mathrm{i}}\right)$, that is, the residual equals the difference between the actual sentiment score of the sample and the predicted score of the preceding model. Subsequently, using this approximate residual ${\mathrm{r}}_{\mathrm{mi}}$ as the target value, the GRSP is segmented. The split point of the decision tree node is determined based on the minimum sum of squared errors, and then the leaf node regions of the regression tree are estimated. The approximate residual values within each leaf node region are fitted to complete the training of the m-th decision tree. Next, for each leaf node region j = 1, 2, …, J, the gradient descent step size ${\mathrm{c}}_{\mathrm{mj}}$ is calculated, with the formula as follows:

$${\mathrm{c}}_{\mathrm{mj}}=\mathrm{argmin}\sum _{{\mathrm{x}}_{\mathrm{i}}\in {\mathrm{R}}_{\mathrm{mj}}}\mathrm{L}({\mathrm{y}}_{\mathrm{i}},{\mathrm{f}}_{\mathrm{m}-1}({\mathrm{x}}_{\mathrm{i}})+\mathrm{c})$$

(5)

Then, the learning rate $\mathsf{\rho }$ is introduced to update the model, with the formula as follows:

$${\mathrm{f}}_{\mathrm{m}}\left(\mathrm{x}\right)={\mathrm{f}}_{\mathrm{m}-1}\left(\mathrm{x}\right)+\mathsf{\rho }{\mathrm{c}}_{\mathrm{mj}}\mathrm{T}\left\{\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right\}(\mathrm{x}\mathsf{ϵ}{\mathrm{R}}_{\mathrm{mj}})$$

(6)

Finally, multiple weak decision trees are combined into a strong predictor, and based on this, the overall expression of the GBDT model can be expressed as follows:

$$\mathrm{f}\left(\mathrm{x}\right)=\sum _{\mathrm{m}=1}^{\mathrm{M}}{\mathrm{f}}_{\mathrm{m}}\left(\mathrm{x}\right)=\sum _{\mathrm{m}=1}^{\mathrm{M}}{\mathrm{C}}_{\mathrm{m}}\mathrm{T}\left\{\mathrm{x};{\mathsf{\theta }}_{\mathrm{m}}\right\}$$

(7)

where ${\mathsf{\theta }}_{\mathrm{m}}$ denotes the residual coefficient of the m-th decision tree T, M represents the number of decision trees, and ${\mathrm{C}}_{\mathrm{m}}$ is the parameter that minimizes the loss function L. The loss function L can be expressed as follows:

$$\mathrm{L}=\sum _{\mathrm{i}=1}^{\mathrm{n}}\left[{\mathrm{y}}_{\mathrm{i}},{\mathrm{f}}_{\mathrm{m}}\left({\mathrm{x}}_{\mathrm{i}}\right)\right]$$

(8)

where ${\mathrm{y}}_{\mathrm{i}}$ represents the actual value of sample ${\mathrm{x}}_{\mathrm{i}}$, and ${\mathrm{f}}_{\mathrm{m}}\left({\mathrm{x}}_{\mathrm{i}}\right)$ denotes the predicted value of the m-th decision tree model, which correspond to the true value and predicted value of residents’ sentiment, respectively.

To enhance model accuracy, a grid search was employed for hyperparameter tuning, utilizing the GridSearchCV tool from the scikit-learn library to optimize these hyperparameters. In the hyperparameter optimization process, 5-fold cross-validation was employed for model validation. The coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE) were employed to assess the overall performance of the regression model.

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}})}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{y}}_{\mathrm{i}}-\widehat{\mathrm{y}})}^2}}$$

(9)

$$\mathrm{MAE}={\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\left|{\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}}\right|$$

(10)

$$\mathrm{MSE}={\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}})}^2$$

(11)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}}{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{({\mathrm{y}}_{\mathrm{i}}-{\widehat{\mathrm{y}}}_{\mathrm{i}})}^2}$$

(12)

where N is the number of samples, ${\mathrm{y}}_{\mathrm{i}}$ is the manually annotated true sentiment score of the i-th sample, and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ is the sentiment score predicted by the model.

3.5.3. Model Interpretation Methods

To examine the non-linear impact of GRSP on residents’ emotional states, this study utilized a GBDT model. This model facilitates the ranking of variable importance through SHAP (SHapley Additive exPlanations) and reveals how variables non-linearly influence park perception via Partial Dependence Plots (PDP) [64]. The associated model interpretation and visualization were efficiently conducted using the shap and pdpbox toolkits.

SHAP [65] offers a method for interpreting predictions from complex machine learning models by computing SHAP values. These values quantify the contribution of each feature to the model’s output by comparing predictions made with and without that specific feature across various subsets of features, thereby assessing feature importance. The calculation formula for SHAP values is presented in Equation (13):

$${\mathrm{\varnothing }}_{\mathrm{j}}=\sum _{\mathrm{S}\subseteq \mathrm{F}\backslash \left\{\mathrm{j}\right\}}{\displaystyle \frac{\left|\mathrm{S}\right|!(\left|\mathrm{F}\right|-\left|\mathrm{S}\right|-1)!}{\left|\mathrm{F}\right|!}}\left[{\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}\cup \left\{\mathrm{j}\right\}}\right)-{\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}}\right)\right]$$

(13)

where ${\mathrm{\varnothing }}_{\mathrm{j}}$ denotes the SHAP value of feature j; F represents the set of all features; S refers to the feature subset that does not include feature j; $\left|\mathrm{S}\right|$ is the number of features in subset S; $\left|\mathrm{F}\right|$ is the total number of features; ${\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}\cup \left\{\mathrm{j}\right\}}\right)$ is the predicted value obtained by the model when using the features in subset S together with feature j; and ${\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}}\right)$ is the predicted value generated by the model when only the features in subset S are employed.

PDP provides a visualization method for exploring how various factors influence predictions in machine learning models [66]. Unlike SHAP values, which convey only variable importance, PDP visualizes the marginal effects of features via linear or non-linear plots and evaluates the impact of independent variables on dependent variables. Its calculation method is shown in Equation (14):

$${\widehat{\mathrm{f}}}_{\mathrm{x}\mathrm{s}}({\mathrm{x}}_{\mathrm{s}})={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\widehat{\mathrm{f}}\left[\left({\mathrm{x}}_{\mathrm{s}},{\mathrm{x}}_{\mathrm{c}}^{\left(\mathrm{j}\right)}\right)\right]$$

(14)

where ${\mathrm{x}}_{\mathrm{s}}$ denotes a feature variable; ${\mathrm{x}}_{\mathrm{c}}$ represents the other feature variables not included in ${\mathrm{x}}_{\mathrm{s}}$; ${\mathrm{x}}_{\mathrm{c}}^{\left(\mathrm{j}\right)}$ refers to the i-th observation of ${\mathrm{x}}_{\mathrm{c}}$; and n stands for the total number of observations.

In addition to basic PDP, this study incorporates 2D Partial Dependence Plots (2D PDPs) to clearly characterize the interaction mechanisms between variables, further enriching the visual analysis. This method focuses on the joint interaction mechanism between two feature variables and reveals the comprehensive impact of their interactions on the model’s prediction results. Compared with basic PDPs that only reflect the marginal effects of a single variable, 2D PDPs effectively address their limitation of “ignoring variable interactions” and are particularly suitable for analyzing the comprehensive interaction mechanism of feature pairs with coupling relationships on the target variable.

4. Results

4.1. Spatial Heterogeneity Characteristics of GRSP

In the assessment of GRSP, distinct spatial differentiation patterns are evident across both the supply achievement and utilization efficiency dimensions (Figure 3).

Regarding supply achievement, areas with high ecological coverage are mainly concentrated in northwestern Chang’an District and central Yuhua District, both of which have distinct advantages in vegetation coverage and water body distribution. High spatial configuration scores are largely concentrated in suburban sub-districts, reflecting more even park distribution and greater per capita green space area.

For utilization efficiency, accessibility generally follows a polycentric agglomeration pattern. High-efficiency clusters appear in eastern Xinhua District, south-central Chang’an District, and the border zone between Qiaoxi and Yuhua Districts. Some older urban areas achieve high service levels due to their dense road networks and extensive service radius coverage. High activity intensity areas form multiple hotspots centered on large comprehensive parks, central business districts, and community parks within densely populated zones.

Both dimensions show relatively high standard deviations, indicating significant spatial imbalance in GRSP across the study area. In terms of spatial pattern, the supply achievement dimension generally exhibits a structure of “high in the center and north, low in the periphery and south,” with high-value areas mainly found in mature urban cores and emerging northern ecological corridors. In contrast, the utilization efficiency dimension is characterized by polycentric clustering and axial extension, with service efficiency and activity intensity generally higher within the influence zones of urban comprehensive parks and along major road corridors (Figure 4).

Overall, high-GRSP areas display a patchy, clustered spatial distribution. Among them, high-performance zones in Chang’an and Yuhua Districts are more concentrated and spatially extensive, indicating notably strong comprehensive GRSP in these districts. In comparison, high-performance areas in Xinhua and Qiaoxi Districts are relatively scattered, resulting in a more pronounced overall spatial imbalance. Low-performance areas are widely dispersed, particularly in the marginal zones of all four administrative districts and in certain non-core areas within districts. These locations demonstrate relatively weak recreational service capacity from green spaces and have not formed large-scale, concentrated, or contiguous clusters.

4.2. Construction and Spatial Distribution Characteristics of the Residents’ Emotional State Map

Based on large-scale public post data collected from the Weibo platform, this study quantified and visualized the spatial distribution of residents’ emotional states in Shijiazhuang’s central urban area (Figure 5). After calculating sentiment scores for the valid texts—categorized as positive, neutral, or negative—these scores were mapped to corresponding grid cells, generating an emotional spatial distribution map that highlights regional variations.

The results reveal that the spatial distribution of residents’ emotions exhibits clear agglomeration characteristics and regional heterogeneity. Positive and negative emotions are relatively balanced, with no single emotional extreme dominating the overall pattern. Spatially, emotional positivity is significantly higher in the northern part of the city compared to other regions. Notably, a continuous cluster of high-sentiment areas has formed along the emerging northern ecological corridors, likely linked to the area’s high-quality natural landscapes and recreational amenities. In contrast, older urban areas in the central and southern parts show widespread negative emotions, with consistently low sentiment scores—particularly in zones characterized by aging infrastructure and high population density. The southern parts of the city display a more mixed emotional polarity, with an overall moderate sentiment level. However, scattered pockets of extremely negative emotions are still present. Meanwhile, highly positive emotional areas are distributed in a dotted pattern across multiple sub-districts, reflecting the positive influence of localized environmental conditions.

4.3. Non-Linear Relationships and Threshold Effects of Variables

4.3.1. Relative Importance of Model Features

Upon validation and optimization, the final model achieved an R² of 0.92 on the training set and 0.88 on the test set, with a MAE of approximately 0.019. A comparison of error trends between the training and test sets confirmed the absence of overfitting. Based on the model fitting results, the relative importance of all explanatory variables is presented in Figure 6, where the sum of importance across all variables equals 100%. Global feature importance indicates that Park Distribution Equity (PDE) contributed the most (54.34%), emerging as the most critical predictor in the model. This was followed by Park Area Ratio (13.12%) and Comment Density (CD) (6.89%), while Pocket Park Accessibility (PPA) showed the lowest importance (1.34%), exerting the weakest influence on predictions. This study further examines how each GRSP indicator influences residents’ emotional states, sequentially analyzing the non-linear effects of the supply achievement dimension, the utilization efficiency dimension, and the comprehensive performance on emotional states.

4.3.2. The Impact of GRSP Factors on Residents’ Emotional States

Through systematic analysis of the relationship between GRSP-related variables and residents’ emotional states, the PDP (Figure 7) and predicted value scatter plots (Figure 8) reveal that all variables exhibit significant non-linear effects on residents’ emotional states. These effect patterns are closely linked to the degree of recreational supply–demand matching and the marginal effects of park supply.

Ecological coverage factors consistently influence residents’ emotional states: impacts are negative at low coverage levels and shift to a stable positive effect once specific thresholds are exceeded. Specifically, when the Park Area Ratio (PAR) is below 0.49%, it negatively affects emotions; between 0.49% and 11.28%, the effect becomes positive and strengthens considerably; beyond 11.28%, the positive impact persists but with diminishing marginal returns, and predicted values show a pattern of sharp initial increase followed by high-level fluctuation. The Vegetation Coverage Rate (VCR) exerts a strong positive effect in the range of 15.27% to 53%; above 53%, the effect weakens, partial dependence values stabilize, and predicted values display an increasing then stabilizing trend with minor fluctuation. For the Water Coverage Rate (WCR), impacts are weak below 9.27%; once exceeding 12.64%, partial dependence rises and predicted values increase rapidly before stabilizing.

All spatial configuration factors demonstrate the characteristic of “negative or weak impact at low values and diminishing marginal effects after surpassing a maximum threshold.” When the PDE is below 1.0896, it exerts a clear negative emotional impact; after equity improves, the effect turns positive rapidly and stabilizes at a high level beyond 2.1964, with predicted values showing a sustained upward trend. Below 40.0013, Per Capita Park Guarantee (PCPG) shows a negative impact; as guarantee levels increase, the effect gradually rises and stabilizes after exceeding 79.7203, while predicted values correlate in a pattern of rapid initial rise followed by high-level fluctuations.

Accessibility efficiency factors generally correlate positively with residents’ emotional states, though variations exist across park types. When Comprehensive Park Accessibility (CpPA) is below 6.1217 m² per capita, the effect is negative; beyond 10.6583 m² per capita, partial dependence stabilizes, and predicted values show marked initial improvement followed by slight further enhancement at a high level. Community Park Accessibility (CmPA) follows an “inverted U-shaped” relationship: negative below 0.1632 m² per capita, then positive, peaking at 1.7748 m² per capita, and declining toward zero when increased to 2.4839 m² per capita. The range of 0.3215 m² to 1.7748 m² per capita represents the optimal interval for emotional benefit, with predicted values rising rapidly before fluctuating at a high level. PPA has a weak effect below 0.4222 m² per capita; beyond 1.6241 m² per capita, partial dependence stabilizes, and predicted values increase initially before leveling off.

Activity intensity efficiency factors display distinct impact characteristics. When the Heat Index (HI) is below 6.0142, its effect is weak; above 15.8373, marginal benefits gradually diminish and eventually flatten. The corresponding predicted values suggest that continuous improvement beyond this threshold weakens the positive correlation, indicating an optimal interval. For CD, the impact is weak below 0.0115; above this value, partial dependence rises sharply, peaks at 0.0332, and then gradually weakens, with predicted values showing rapid initial increase followed by high-level fluctuation.

4.3.3. The Impact of Comprehensive GRSP on Residents’ Emotional States

To systematically assess the multidimensional effectiveness of ecological coverage governance, this study further investigated the correlation patterns between three integrated variables (RSA, RSE, and GRSP) and residents’ emotional states (Figure 9). Each demonstrates similar non-linear relationships with residents’ emotions and their corresponding predicted values. At low levels, all three variables exhibit a slight negative influence on emotional states. The positive emotional benefits are maximized when RSA lies within the range of 6.5 to 42.95. For RSE, once it exceeds 6.3682, it enters an effective interval characterized by a positive impact on residents’ emotional states.

The results indicate that GRSP also displays a non-linear association with residents’ emotional states, accompanied by diminishing marginal returns. The interval from 3.8484 to 21.7493 represents the optimal range for GRSP, within which its positive emotional effect strengthens progressively. Beyond the maximum threshold of 21.7493, however, the positive influence plateaus, and further increases do not enhance residents’ positive emotions. Meanwhile, the relationship between these factors and the actual predicted values follows a pattern of “rapid initial rise followed by stabilization”: initial increments significantly elevate predicted values, while at medium-to-high levels, values remain elevated with a decelerated growth rate. Across all analyses, the width of the 95% confidence interval remains relatively stable, suggesting that these factors exert consistent effects in most ranges and that individual variations are limited.

4.3.4. The Interaction of Dominant Factors on the Effect of Emotional Improvement

To clarify the interaction mechanisms between GRSP and residents’ emotional states, this study used 2D PDPs and SHAP interaction values. Results reveal significant synergistic effects among key performance indicators, with their interactive impacts on emotional improvement following distinct stage-specific variation patterns.

Specifically, VCR exhibited significant interactions with PDE, CpPA, and CmPA. At high VCR levels, increasing PDE, CpPA, or CmPA substantially enhanced positive emotional effects. In contrast, improving these indicators in isolation produced only limited emotional benefits when VCR was low (Figure 10).

Furthermore, PDE showed clear interactions with CmPA and HI. At high PDE levels, increasing CmPA strongly amplified positive emotional effects—an enhancing effect that weakened under low PDE conditions. For the PDE-HI interaction, PDE contributed most to emotional improvement when HI was moderate; excessively high or low HI values reduced this effect. Additionally, the CmPA-HI interaction followed a non-linear pattern: moderate HI levels facilitated emotional improvement via higher CmPA, but beyond a specific HI threshold, further increases in CmPA diminished emotional benefits (Figure 11).

In summary, GRSP influences residents’ emotional states not through isolated indicators, but through the coupling of multiple dimensions. The interactive effect of any single indicator depends on the levels of other relevant factors.

5. Discussion

5.1. Multidimensional Mechanisms of the Non-Linear Impact of GRSP on Emotions

This study advances prior research, which has mostly focused on the independent impacts of individual green space attributes on residents’ emotional states—an approach that overlooks the synergistic mechanisms of multiple factors within GRSP. Additionally, the common practice of dichotomizing emotions into positive and negative categories limits the accurate quantification of how various attributes gradationally influence emotional states [67,68], constraining theoretical understanding of the green space-emotion relationship.

Through a multidimensional evaluation of GRSP, this study accurately identifies key performance determinants. Simultaneously, using continuous emotional scores as the dependent variable overcomes the limitations of traditional binary classification, systematically uncovers the differentiated pathways through which various performance factors shape residents’ emotional states, and clarifies their dynamic impact mechanisms—providing empirical evidence to refine the theoretical framework in this field.

Regarding GRSP specifically, this study confirms that ecological coverage and spatial configuration factors systematically affect residents’ emotional states. Results show the marginal effect of the green space area ratio on emotional improvement is most pronounced at low values; as the ratio rises to a medium level, emotional gains gradually decelerate and stabilize at a relatively high scale. This finding not only verifies the threshold effect in green spaces’ emotional regulatory function [69] but also theoretically challenges the linear assumption that “area expansion equals benefit enhancement” [70], indicating that simply expanding green space area is ineffective for continuously optimizing residents’ emotional experiences [67].

The impact of vegetation coverage displays distinct stage-specific characteristics: when coverage is insufficient, increasing it yields the strongest emotional improvement; however, beyond a specific threshold, emotional gains weaken significantly. This result clarifies the saturation range between vegetation coverage and emotional benefits [38,71], providing a theoretical basis for the quantitative planning of vegetation configuration and cautioning against the blind pursuit of high coverage in practice. Analysis of water coverage shows that its emotional value improves most during the initial configuration stage (from absence to presence), followed by a period of stable growth, and ultimately plateaus at relatively high coverage levels. This finding systematically uncovers the pattern of residents’ positive emotional responses to blue spaces [72], addressing a gap in prior quantitative analyses of water landscapes’ emotional regulatory mechanisms.

The positive impacts of ecological elements like vegetation and water bodies on emotions follow the law of diminishing marginal returns. This may stem from the human perceptual system’s adaptation to continuous environmental stimuli. Once coverage rates or water access opportunities reach a certain level, residents’ basic recreational needs become saturated, shifting focus to other dimensions such as internal landscape quality and facility completeness [73]. Excessively high PAR may even crowd out functional spaces like residential or commercial areas, indirectly weakening positive emotional effects [74]. At the same time, the novelty and restorative benefits of green or water exposure gradually saturate; additional increments fail to elicit equivalent positive responses, and higher water coverage may even correlate with declining water quality, leading to negative impacts [75]. From a cognitive psychology perspective, this explains why a “reasonable range” exists rather than a “more is better” principle. Notably, this study uses Shijiazhuang’s central urban area as a case, where water resources are relatively limited and coverage is generally low. This context may restrict the full manifestation of emotional effects at high coverage levels and influence the completeness of the observed marginal change pattern. Therefore, extending these conclusions to cities with abundant water resources or significantly different spatial structures requires considering local contexts.

Tests on park spatial configuration indicate that when spatial balance is inadequate, its positive impact on emotions is limited; as balance enters a reasonable range, the positive emotional effect rises rapidly; further improvement to a higher level leads to a decelerated growth rate, though emotional value can still be steadily enhanced under high-balance conditions. This dynamic characteristic highlights the role of spatial configuration optimization in emotional regulation [76,77]. The impact of per capita park area follows the law of diminishing marginal returns: significant improvement occurs initially, followed by decelerating growth, with emotional benefits weakening once supply becomes sufficient. This finding provides key empirical evidence for accurately formulating park supply standards and can mitigate the risks of excessive or insufficient supply in resource allocation [78]. The accelerated positive impact of spatial balance after reaching a basic equity threshold reflects the psychological mechanisms of social comparison and fairness perception. Perceived equity in park services enhances community identity and well-being, while perceived inequity can trigger negative emotions such as relative deprivation.

Regarding green space recreational service efficiency, this study found that accessibility across different park types has significant differences in emotional impact. Comprehensive parks substantially enhance emotional benefits within their effective service radius, beyond which gains gradually plateau—confirming a matching relationship between service radius and emotional benefits [40]. The impact of community park accessibility follows a complex pattern: emotional benefits increase with improved accessibility at low levels, fluctuate in the middle range, and gradually recover beyond a threshold. This characteristic reveals the non-linear mechanism between community park accessibility and emotional benefits, emphasizing the need to optimize service radii [27,75]. The emotional benefits of neighborhood pocket parks display a typical distance decay pattern: optimal regulatory effects are achieved at short distances, weakening as distance increases. This confirms the location sensitivity of pocket parks as daily-use facilities [24,25], providing empirical support for the “proximate service” principle in their layout planning [28].

Underlying these differences is a potential cost–benefit evaluation mechanism in residents’ recreation decisions, which are often loss-averse [79]. When park resources are scarce, residents may plan special trips despite poor accessibility, valuing the novelty and sense of escape. However, when supply is abundant and homogenized, the marginal emotional benefits of improved accessibility diminish, shifting value dependence to a park’s uniqueness, ecological quality, and activity offerings. Excessively high accessibility—bringing green spaces too close to residences—may also trigger negative externalities such as noise and privacy loss, thereby suppressing emotional benefits [80].

In terms of usage intensity, an inverted U-shaped relationship exists between crowd density and emotional benefits. Moderate agglomeration stimulates positive social interactions and place vitality, but excessive crowding triggers resource competition, privacy disturbances, and stress, undermining emotional restoration [81]—confirming that parks have social carrying capacity limits. The impact of residents’ online activity frequency within parks also follows a curve of initial rise followed by decline [82], indicating an optimal range for recreational activity intensity. Beyond this range, overuse may reduce experience quality due to information overload. These findings collectively reveal the balancing mechanism between social support and stress perception in park usage: appropriate density maximizes GRSP by balancing social interaction and comfort demands [81], while excessively high intensity leads to crowding and reduced service efficiency [83].

Interaction effects across recreational green space indicators show that VCR—a measure of landscape quality—acts as a fundamental regulator. Only at relatively high VCR levels do improvements in PDE or spatial accessibility (CpPA, CmPA) produce significant synergistic effects, enabling the emotional and health benefits of high-quality green spaces to be widely realized. In contrast, insufficient VCR constrains improvements in other dimensions due to basic environmental quality limitations. Meanwhile, PDE, as a major contributing dimension, is highly context-dependent. It works synergistically with CmPA to increase recreational opportunities and maximize emotional benefits, while its interaction with HI follows an inverted U-shaped relationship. This indicates that PDE’s positive role is fully realized only under comfortable, uncrowded conditions; extreme crowding weakens recreational willingness, undermining the value of equitable distribution.

Results from non-linear analysis and threshold effect testing indicate that various GRSP factors collectively explain the majority of variation in residents’ emotional states, confirming that GRSP is a significant driver of this variation. This study further reveals that non-linear relationships and interaction effects generally exist between each GRSP factor and emotional states. Consequently, green space planning does not need to pursue scale expansion but should focus on maintaining optimal ranges for various GRSP factors. This approach avoids weak emotional regulation due to insufficient GRSP while preventing resource waste from excessive allocation.

5.2. Spatial Optimization Pathways Based on the Threshold Effects of Driving Factors

Based on the identified non-linear relationships and clear thresholds between GRSP and residents’ emotions, a community emotional diagnosis and zoning system was constructed using the threshold effects of integrated variables. This system avoids reliance on a single indicator; instead, it classifies the study area into four distinct types (Figure 12) based on integrated performance indicators such as RSA and RSE, enabling rapid identification and categorization of regions. This framework provides a decision-making basis for differentiated planning interventions and proposes corresponding spatial optimization pathways by integrating the characteristics of each zone with the specific factor thresholds influencing GRSP.

The Priority Intervention Zone (RSA < 6.5, RSE < 6.3682, GRSP < 3.8484) faces a severe shortage of comprehensive service supply, indicating extremely low supply scale and system efficiency. Most of these zones are located in densely populated central city communities, which are key factors restricting improvements in residents’ emotions. Therefore, optimizing this zone is a top priority, requiring mandatory, rapid intervention measures focused on addressing basic deficiencies and meeting essential standards. Specifically, planning should prioritize rapidly raising ecological coverage indicators such as PAR and VCR to above-threshold effective ranges through measures like creating green spaces in vacant areas and converting illegal constructions into green spaces.

The Benefit-Low Zone (RSA ≥ 6.5, RSE < 6.3682, 3.8484 < GRSP < 21.7493) already has a certain supply scale, but exhibits insufficient resource utilization and service conversion efficiency, with emotional benefits entering a period of slow growth. For these zones, optimization should focus on structural improvement and breakthroughs, shifting the development model from scale expansion to quality enhancement. Simple scale expansion should be avoided; instead, efforts should target optimizing the internal structure and service quality of green spaces. For instance, focus can be placed on improving vegetation layers and biodiversity within green spaces, or increasing water landscape coverage to exceed 15.27%; in terms of spatial configuration, priority should be given to enhancing park layout balance, pushing PDE above 1.0, and raising the per capita park guarantee rate to over 40% to break through the current plateau of emotional benefits.

In the Benefit Optimization Zone (6.5 < RSA < 42.95, RSE ≥ 6.3682, 3.8484 < GRSP < 21.7493), supply scale and system efficiency are well-matched, and improvements in various variables within this range can effectively drive sustained growth in emotional benefits. The core of optimizing this zone is to consolidate existing advantages and prevent recreational overload, aiming to precisely maintain the current high-efficiency state. Planning should focus on refining landscape quality and continuously optimizing the service system, while maintaining indicators such as crowd HI and CD within optimal ranges through dynamic monitoring to preserve high-efficiency system operation and avoid service overlap from excessive investment.

In the Overload Risk Zone (RSA ≥ 42.95, RSE ≥ 6.3682, GRSP > 21.7493), emotional responses have entered a saturation range, with marginal returns from continued investment approaching zero—indicating potential risks of excessive resource input or system overload. The optimization direction for these zones is clearly defined as load reduction and balanced development, requiring regulatory and diversion measures. For example, for overcrowded parks with a crowd HI exceeding 15, measures such as activity diversion, spatial expansion, or reservation-based guidance can be implemented; for areas with overly dense community parks, resource allocation can be optimized through functional integration and differentiated positioning. The ultimate goal is to keep all indicators within the optimal range for emotional health benefits and ensure the long-term positive impact of recreational services.

In practice, spatial optimization paths must balance trade-offs under limited resources. For example, in resource-constrained old urban areas (classified as Priority Intervention Zones), spatial optimization requires prioritizing either increasing PAR or improving PDE. Current PAR in these areas is 0.0039, and PDE is 0.8017—both below the minimum thresholds identified in this study (0.0049 and 1.0896, respectively). Due to limited available space and funding, threshold effect analysis shows that raising PDE above 1.0896 is the fundamental threshold for generating emotional benefits, with greater urgency than marginally increasing PAR from 0.0039 to 0.0049. Therefore, resources should be prioritized to improve PDE. Specifically, the zone can deploy pocket parks at the centers of multiple residential clusters to first push PDE above 1.0896, while simultaneously using vertical greening and similar methods to raise VCR to the 15.27% baseline. This approach prioritizes the spatial equity of park service coverage, aligning with the primary goal of Priority Intervention Zones to “quickly address basic deficiencies.”

5.3. Global Applicability of the GRSP Framework

The GRSP framework demonstrates international applicability, yet requires adaptation to local urban contexts. Based on the dual dimensions of “supply level” and “utilization efficiency,” it is rooted in the core logic of green space recreation services, making it suitable for cities at various development stages and densities. Cities globally, from high-density metropolises to rapidly urbanizing areas, share common challenges in equitable green space supply, accessibility, and usage efficiency [36]. These shared issues establish a foundation for the framework’s worldwide application.

The specific findings of this study also have broad reference value. Phenomena such as the “law of diminishing marginal returns” and “threshold effects” are widely applicable. Cities with scarce green space resources, like those in arid Middle Eastern regions, can prioritize basic green space supply. This strategy is informed by the finding that ecological elements enhance emotional benefits even with minimal green coverage. For cities with sufficient but unevenly distributed green spaces, such as some in Europe, optimizing layout based on the importance of spatial balance (PDE) can improve equity.

GRSP indicator thresholds require local calibration based on natural conditions. For instance, in water-rich cities, the mood benefits of water coverage may be more sustained [72], necessitating threshold adjustments using local data. In low-density suburbs, parameters such as pocket park service radii should reflect local travel patterns.

Culturally, the framework must consider variations in recreational preferences. Western cities, which emphasize social and personalized green space use, could include metrics like social interaction types and facility usage frequency [68]. In collectivist cultures, greater weight should be given to accessibility and shareability [77].

5.4. Limitations and Future Explorations

This study has several limitations. First, relying solely on Weibo text data for emotional measurement has certain drawbacks, as it cannot fully represent the emotional states of the entire urban resident population. In particular, it inadequately covers groups such as the elderly, individuals with low educational attainment, and inactive social media users. Although survey data were used to supplement the dataset, the effectiveness of emotional measurement is still constrained by the sample’s uneven demographic distribution. Future research should integrate multi-source technologies (e.g., physiological sensor monitoring) to enhance the real-time performance and multidimensionality of emotional assessment [84].

Additionally, relying on data and contexts from a single city, this study cannot fully capture or disentangle the subtle impacts of local cultural norms, social habits, and residents’ collective preferences on emotional responses. Future research can conduct cross-city comparative analyses to further distinguish the universal mechanisms and regional moderating factors underlying the emotional benefits of urban park green spaces.

Furthermore, mediating mechanisms (e.g., social interaction and environmental perception) and moderating mechanisms (e.g., seasonal variations) between GRSP and residents’ emotional states were not included in the analysis. Future research could incorporate activity behaviors as mediating pathways, construct dynamic coupling models of the environment and emotions, advance cross-scale and multi-modal data fusion analyses [85], and verify the stability of performance-impact mechanisms across additional urban cases to enhance the universality and depth of conclusions.

6. Conclusions

Focusing on the relationship between GRSP and residents’ emotional states, this study clarifies the interaction mechanisms between these two factors in high-density urban settings. By developing a multidimensional evaluation framework and refining emotion quantification methods, we overcome the limitation of single-attribute assessment. The framework comprises two core dimensions: the supply achievement dimension (focused on ecological coverage and spatial configuration) and the utilization efficiency dimension (encompassing accessibility and activity intensity). Together, these form a quantifiable system that enables accurate identification of key influencing factors.

Going beyond the traditional emotional dichotomy, this study uses sentiment analysis techniques with continuous emotional scores as the dependent variable. This approach enables refined, gradational measurement of emotional states and significantly improves the accuracy of correlation analysis. Key findings confirm that GRSP has a significant non-linear impact on residents’ emotional states, consistent with the law of diminishing marginal benefits.

The supply achievement dimension plays a dominant role in explaining variations in emotional states, with PDE accounting for over 50% of the contribution—highlighting the central importance of spatial equity. Indicators such as PAR and VCR exhibit distinct thresholds beyond which emotional benefits plateau. While the utilization efficiency dimension has a relatively limited overall impact, it indirectly moderates emotional states through factors such as usage convenience and spatial vitality, also showing clear threshold effects. For example, comprehensive parks generate substantial emotional benefits within their effective service radius, moderate crowd aggregation enhances emotional well-being, and excessive crowding exerts an inhibitory effect.

This study provides clear empirical evidence and practical guidance for green space planning in high-density cities. Future healthy urban development should promote the transition of green spaces from mere “availability” to “high quality,” achieving scientific and cost-effective planning through precise resource allocation and human-centered design, while safeguarding social equity and residents’ emotional well-being.