The concept of “Healthy Cities” was formally introduced by the World Health Organization (WHO) in 1984, emerging from research on the impact of public health systems on residents’ well-being. One of its definitions is that a healthy city should create and improve physical and social environments, and enhance social support through expanded community resources (17). With the advancement of the “Healthy Cities” movement, theoretical discussions between environment and health have increased. Among these, socioecological theory, which provides a framework for sustaining the “individual-environment” system, emphasizes the interactive relationship between individuals and environment (18). It posits that health is influenced by multiple layers, including personal and environmental factors, where environmental factors include the built, natural, and social environment (18, 19). Based on the theoretical explanation of “environment-health,” numerous scholars have explored the mechanisms through which multidimensional environments impact health. Among these, the neighborhood environment, where urban residents (particularly adolescents) live for extended periods, has been widely recognized as having a significant impact on health. Recent studies on neighborhood environments and adolescent health commonly measure neighborhood conditions in two dimensions, objective and subjective environments, due to differences in research content and measurement approaches (such as urban spatial big data or questionnaire survey data).

On one hand, the objective environment reflects the physical spatial structure of the community and the opportunities for accessing service facilities, and it can be categorized into natural and built environment based on specific indicators and research focus. The natural environment includes the density, quality, and exposure levels of blue and green spaces (13, 20–22). The built environment includes the road design (e.g., street connectivity, road and intersection density) (7, 8, 13, 23), density and accessibility of service facilities (e.g., parks, restaurants, and sports and recreational facilities) (6, 23, 24), land use mix (7, 25, 26), density and accessibility of transportation (e.g., bus stop and subway station) (7, 8, 13, 25). Regarding the natural environment, appropriate exposure to nature and well-designed landscapes can positively influence adolescents’ physical and mental health. For instance, the coverage and proximity to parks, forests, or green spaces have been shown to improve sleep efficiency, reduce body mass index (BMI) (6, 20, 27), and improve mental well-being (including reductions in depression, stress, anxiety, and behavioral health issues) (28, 29). Regarding the built environment, it influences adolescent health by shaping the overall quality of the community (e.g., residential quality, land use mix, and facility distribution). For instance, mixed land-use patterns (7, 25), appropriate intersection density (25), and the proximity of parks and sports facilities (7) promote physical activity, support social interactions, and reduce sedentary behavior (24), thus positively influencing adolescents’ physical health and emotional well-being. Conversely, significant environmental pollution, including air and noise pollution, can adversely affect physical health and lead to negative emotions (9, 20, 30). For instance, excessive concentrations of pollutants like NO₂ and PM_2.5, negatively affect adolescent sleep health (20), and are associated with adolescent psychiatric experiences (31). Although the objective environment effectively represents the physical spaces that adolescents inhabit, it does not fully capture the comprehensive subjective assessments of environmental factors that influence health status or behaviors (32, 33).

On the other hand, the subjective environment reflects adolescents’ perceptions and evaluations of the physical and social contexts, which indicates that individual differences may result in varied interpretations of similar neighborhood environments. This includes assessments and perceptions of the natural environment (10, 34), built environment (e.g., walkability, cognitive aspects of buildings, and service facilities) (10, 35), neighborhood order (e.g., crime and safety) (11, 36, 37), community support and interaction (e.g., cohesion, trust, and social capital) (6, 9, 38, 39). Among them, perceptions of the physical environment can indirectly influence adolescent health by promoting or hindering physical activity levels. For instance, perceived favorable landscape conditions (10) and proximity to recreational facilities (35) provide opportunities for positive interactions with the environment, thereby fostering healthy emotional responses and habits. In contrast, perceptions of the social environment significantly impact adolescents’ social behaviors and interactions. For example, perceived community safety is positively associated with reductions in adolescents’ body mass index (BMI) and well-being (6, 36). Conversely, negative perceptions of neighborhood trust, social cohesion, and social capital are associated with a higher likelihood of adolescents reporting internalizing and externalizing problem behaviors (40, 41).

Previous studies highlight that both objective and subjective environments are critical factors influencing adolescent health. Recognizing the limitations of assessing environmental influences on adolescent health through a single-dimensional approach, a few studies have started to examine the relationship between the environment and health by considering both subjective and objective environmental factors (6, 9, 42). However, variations in indicator selection and research focus have led these studies to primarily explore the effects of significant environmental variables. Recently, studies on neighborhood environment and residents’ health have increasingly focused on the joint analysis of objective and subjective characteristics (10, 33, 43, 44). Due to differences in measurement approaches and representational meanings, the relative influence of objective and subjective environments on health may vary. For instance, certain study indicates that the objective environment may be more significant (45), as it reflects the actual physical space of the community, thereby confirming the real meaning of the perceived environment and helping to guide planning practices (33). Other studies, however, indicate that the subjective environment is more significant (15, 43, 46), as perceptions of the environment reflect individual cognitive processes and have a particularly significant impact on health (47). Although the relative importance may vary by research aim and context, both objective and subjective environments contribute to reflecting the significant influence of the built environment on health through environmental perception and the promotion of health behaviors.

In summary, using indicators from both subjective and objective environments provides a comprehensive understanding of the neighborhood environments that adolescents experience. However, existing research often examines the relationship between neighborhood environments and adolescent health from a singular perspective (subjective or objective) and single dimension (such as two-dimensional spatial environment). Few studies have investigated the interplay between subjective and objective environments.

Existing research has mainly investigated the impact of neighborhood environment on adolescent health using linear assumptions, often focusing on single effects. However, variations in indicators and methodologies frequently lead to estimation bias (48, 87). For instance, natural environmental contact has been shown to have a positive effect on health in most studies, and air pollution has a negative effect. But these two environmental conditions have not been able to predict mental health outcomes in some studies (42). Likewise, while proximity to transit stations and road density are generally positively associated with adolescent health, some studies have found no significant correlations with physical health outcomes (7). Additionally, the effects of built environment elements may differ by spatial scale, and focusing only on linear relationships could overlook potential nonlinear effects, leading to inconsistencies in findings (12). For instance, land use mix is generally positively correlated with adolescent physical and mental health (7, 25), but another study indicates that its impact on adolescent BMI is not significant, and may exhibit a possible nonlinear relationship (26). In the only two studies we found that explore the non-linear relationship between neighborhood environment and adolescent health, the non-linear relationships and threshold effects of certain environmental variables have been confirmed. For instance, the distance to transit stations shows a locally significant positive effect on adolescent physical health within specific intervals (7, 13), while green spaces (e.g., park coverage and NDVI) are associated with reduced adolescent obesity rates within certain thresholds (13).

Most existing studies on the built environment and residents’ health are based on linear assumptions. Recently, there has been growing interest in the non-linear associations between neighborhood environments and various health-related factors, including health status (15, 49, 50), obesity (13, 51, 52), and positive health behaviors (16, 53, 54). Some studies have identified potential threshold effects for specific built environment variables, indicating that significant effects only emerge within certain ranges. For instance, optimal levels of built environment features, such as proximity to sports facilities and bus stops (15), road density (54), and land use mix (53), can effectively enhance residents’ physical and mental health while promoting active travel. Moreover, evidence suggests that the relationships between certain built environment factors and health outcomes can vary across different ranges of influence. For example, green exposure has been widely shown to have complex mechanisms affecting health. Adequate levels of green exposure offer more opportunities for interaction with nature and improved landscape conditions, fostering positive health outcomes and related behaviors. However, negative impacts may arise beyond certain thresholds (12, 13, 15). Similar effects have been observed with road connectivity (15, 55), bus stop density (14, 15), and population density (13, 14).

Kunming, the capital of Yunnan Province, serves as a key gateway to South and Southeast Asia and is one of the core cities in Southwest China. As of 2023, Kunming has a resident population of 8.68 million (56), and covers an area of 21,012.54 km² (57). The study area is defined by the spatial scope outlined in the “Urban Master Plan of Kunming (2011–2020).” It includes the area to the east of the Third Ring Road and enclosed by the Ring Expressway. This region encompasses the core built-up areas of Kunming’s four main districts (Wuhua, Panlong, Xishan, and Guandu) and serves as a significant hub for the city’s public service facilities (Figure 2).

The data for this study were obtained from the “Online Questionnaire Survey on Primary and Secondary School Students’ Commuting and Healthy Growth,” conducted in Kunming, China in January 2024. The survey covered five major districts in Kunming: Wuhua, Panlong, Xishan, Guandu, and Chenggong. First, we randomly selected 31 sample schools from a total of 402 primary and middle schools in five districts using a stratified sampling strategy (these schools are within the compulsory education stage). Subsequently, we contacted the Kunming Education and Sports Bureau to obtain permission for the online survey distribution and coordinated with the teachers in charge of each school. We randomly selected approximately 5% of students from grades 1–6 in primary schools or grades 1–3 in middle schools in the sample schools, and distributed a total of 1831 questionnaire QR codes. The online questionnaires were completed jointly by students and their parents. Given that younger students may not fully understand the questions, we specifically noted that sections to be filled out by students should be completed with parental assistance. The legal guardians of all participants signed informed consent forms after being briefed on the purpose of the survey, and their information is strictly protected in accordance with relevant regulations. The survey collected information on various aspects, including students’ personal and family backgrounds, health status, and subjective environmental perceptions. After excluding questionnaires with incomplete information (such as home and school addresses), anomalies, and those with low reliability, we obtained 1,583 valid questionnaires, yielding an effective rate of 86.45%. These samples were distributed across 606 residential communities in the five districts. Given the limitations and representativeness of the urban built environment data, we further refined the sample to 1,388 valid responses within the study area. Research indicates that children aged 7 to 12 develop more logical and organized ways of thinking compared to children under the age of 7 (58). Therefore, to more accurately capture adolescents’ self-perception of the environment and health conditions, we excluded students under 7 years of age, yielding a final sample of 1,355 for subsequent analysis.

In addition to the survey data, we collected the following multi-source urban spatial data: (1) Road network data was obtained from OpenStreetMap.¹ The raw road network data were processed using ArcMap 10.8.1, including clipping, cleaning, and topological adjustments, with interruptions at intersections. (2) Points of Interest (POI) data was acquired from the Gaode Map.² (3) Street view image was collected from the Baidu Map.³ A Python program was used to sample street view panoramic images at 100-meter intervals from the road network, resulting in 54,431 images. These images were analyzed using a PSP Net model pre-trained on the MIT ADE20K dataset to identify and calculate the area proportions of various visible spatial elements, such as sky, roads, and vegetation. (4) Normalized difference vegetation index (NDVI) was obtained from the National Ecological Science Data Center.⁴

The Random Forest (RF) model, introduced by Leo Breiman in 2001, is a decision tree-based machine learning algorithm, particularly effective at handling large datasets while minimizing the risk of overfitting (64). In classification tasks, as illustrated in Figure 4, the core principle involves repeatedly drawing M sample sets from the original dataset using the bootstrap resampling method, generating M decision trees that collectively form a random forest. At each tree node, a random subset of n features is selected from the total of N features, and the feature that minimizes the Gini index is chosen to split the node. Each tree is fully grown, and the ensemble of M trees provides the final classification result, which is determined by majority voting across all trees.

In this study, we utilized the “Random Forest “package in Python 3.6, Jupyter Notebook 5.7.10. The model construction and validation process involves two main steps. First, using the ‘train_test_split’ module randomly allocates 70% of the analysis samples to the training set for model building, while the remaining 30% are reserved as the test set to assess the model’s performance. Second, to optimize the model and avoid issues of underfitting and overfitting, based on previous study (65), three key hyperparameters are fine-tuned. The search ranges for these hyperparameters are as follows: max_depth ranging from 3 to 7 (step size of 1); n_estimators ranging from 100 to 500 (step size of 100); and max_features ranging from 1 to 5 (step size of 1). Grid SearchCV from Scikit-learn is used for grid search and five-fold cross-validation to evaluate 125 possible hyperparameter combinations and determine the optimal set that delivers the best model performance: max_depth is set to 3, n_estimators is set to 100, max_features is set to 1.

To assess the suitability of the Random Forest model for this study, we trained and tested three commonly used machine learning models: RF, XGBoost, and LightGBM. The performance of these models is evaluated using commonly used metrics in supervised machine learning: Accuracy, Precision, Recall, and F1 scores. The value closer to 1 indicates better model performance (13, 15). As shown in Table 2, RF outperformed the other models in accuracy, recall, and F1 score. Although RF did not achieve the highest precision value among the models, it was very close to the maximum value. Overall, the Random Forest model demonstrated high accuracy, adaptability, and excellent predictive performance, making it effective for uncovering the complex nonlinear relationships between neighborhood environment and adolescent health. To further assess the robustness of the selected research units and study results, we constructed an 800-meter network buffer to measure the objective environment. The results indicate that the Random Forest model also exhibits strong robustness. Furthermore, the relative importance rankings of both subjective and objective environmental factors remained largely consistent. Specifically, with the exception of a few variables, the rankings of the top 12 factors remained stable across each control group.

Table 3 displays a comparison of the relative importance and rankings of all variables. Among the three types of independent variables, the objective environment has the greatest significance for both physical and mental health, with relative importance percentages of 51.03 and 53.30%, respectively. The subjective environment ranks next in importance, with its significance notably exceeding that of socioeconomic attributes. When comparing the relative importance rankings and percentages of various indicators, the following observations are made: (1) The greening view index, relative walking width, and ratio of traffic safety facilities in representing the three-dimensional environment of streets are comparable. (2) In contrast, the NDVI is more strongly associated with physical health, whereas intersection density, distance to bus stop, and community safety are more strongly associated with mental health.

In terms of physical health, the most significant variables in the neighborhood environment are sense of community belonging, distance to subway station and greening view index. The relative walking width, sky view index and land use mix are also closely associated with physical health. However, the relative importance of community safety and perceived sanitary condition is comparatively lower.

In terms of mental health, the most significant variables are distance to bus stop and subway station, sense of community belonging. Next in importance are intersection density, greening view index, and land use mix. Conversely, the relative importance of perceived landscape environment and distance to sports facility is comparatively lower.

Additionally, among socioeconomic attributes, housing area plays a significant role in influencing the physical health of adolescents, with its relative importance ranked 3rd for physical health and 2nd for mental health. Other personal socioeconomic attributes, such as gender and household registration, do not exhibit a significant impact on adolescent health.

We visualize the non-linear relationships and threshold effects between various variables and adolescent physical and mental health using partial dependence plots. Additionally, we apply a fitting curve to smooth the effects of objective environmental factors and some socioeconomic attributes on health outcomes (66). Based on the relative importance ranking of factors influencing adolescent health, we individually analyze the top 12 most significant neighborhood environment variables.

Compared to subjective environments, objective environments have a more significant impact on adolescent physical and mental health. This finding contrasts with existing research, which emphasizes the substantial influence of subjective environments on residents’ health and older adults health (15, 67, 68), considering them as crucial predictors of health status. Unlike adults, who possess fully developed judgment and cognitive abilities, adolescents are still undergoing developmental changes, and their mental and cognitive faculties are not yet fully mature. Therefore, their health status is more directly influenced by objective material conditions than by perceptions of social environment (e.g., neighborhood relationships, facility conditions and community image).

Additionally, we observed notable differences in the contribution of specific key variables between physical and mental health. For instance, NDVI plays a more significant role in physical health. One possible explanation is that vegetation coverage directly reflects the air quality, heat, humidity conditions (e.g., local heat island effects), and opportunities for natural contact, which are closely associated with adolescents’ respiratory and cardiovascular health (69), thereby having a more pronounced impact on physical health. In contrast, distance to bus stop and community safety are more influential for mental health. One possible explanation is that the distance to bus stop reflects the convenience of accessing social places, which plays a crucial role in the development of adolescents’ independent social capabilities and participation in activities. Similarly, during the adolescent mental development stage, fostering independent travel abilities also contributes to enhancing self-efficacy. Environmental psychology theory posits that a safe and comfortable environment can reduce anxiety and stress (70), thereby enhancing personal well-being and a sense of security. Similarly, for adolescents, the security level of the neighborhood environment (e.g., the frequency of criminal or violent incidents) indirectly influences patterns of social interaction and emotional stability, which may lead to increased anxiety and depression (63, 71, 72).

We found that the subjective environment typically has a beneficial impact on promoting adolescent health. For instance, among the key variables related to community management, such as community safety is positive associated with adolescent mental health. Poor maintenance of facilities, littering, and other signs of environmental disorder can lead to increased concerns about public safety and crime (71, 72). In contrast, a well-maintained and safe community environment enhances adolescents’ sense of security and comfort during outdoor activities, which can alleviate potential sleep difficulties and reduce life stress (6).

Among the key variables related to community image, a quieter noise environment is beneficial to physical and mental health. Previous studies have indicated that urban traffic noise leads to sleep disturbances (82). Similarly, exposure to high levels of noise (including traffic and activity noise) can induce stress responses in adolescents, potentially affecting academic performance and sleep quality. Air quality can directly affects respiratory and cardiovascular health (69). Exposure to low levels of air pollution (e.g., NO₂, PM_2.5), is more likely to lead to physiological issues such as insomnia and cardiovascular diseases during adolescence (6, 23).

Among the key variables related to community capital, sense of community belonging positively influences both adolescent physical and mental health. Existing research indicates that social cohesion among neighbors helps alleviate psychological stress and enhances social connections in adults (83, 84). Similarly, a supportive and emotionally connected neighborhood environment promotes positive health behaviors and psychological well-being in adolescents (40, 85).

This study has two limitations due to the study area and sample constraints. First, due to the sample size of the survey data, it is challenging to control for geographic environmental differences that may arise from residential self-selection (16). For instance, parents of students living in commercial properties or rental accommodations might choose residences with better scenic environments and amenities based on their economic status and social position. Second, although we explored the optimal impact range of the built environment on adolescent physical and mental health, the threshold effects of the built environment may differ under varying regional conditions. Therefore, future research should aim to enhance the generalizability of the findings by conducting comparative studies across different cities and increasing sample size. Nonetheless, considering the challenges of data acquisition, the non-linear and threshold effects of neighborhood environments highlighted in this study offer valuable insights for adolescent-friendly community planning, construction, and management, particularly for large cities along the southwestern border of China.