Exploring Optimisation Pathways for Underground Space Quality Under the Synergy of Multidimensional Perception and Environmental Parameters

Abstract

With the acceleration of urbanisation and the increased utilisation of underground space, providing a comfortable and healthy environment in public underground areas has emerged as a significant research topic. This study constructs a comprehensive decision-making framework for underground space environments by integrating human perception evaluations with physical environmental parameters. Using Shanghai Wujiaochang as a case study, field data collection and questionnaire surveys were conducted to evaluate key factors such as temperature (22.63 °C–26.39 °C), wind speed (0.26 m/s–0.67 m/s), and sound levels (59.68 dB–61.21 dB) for commercial-oriented spaces, and 63.15 dB–75.45 dB for transport-oriented spaces) to users’ perceived experiences. The appropriate ranges for key parameters were identified through single-indicator fitted regression analysis and the XGBoost machine-learning model, revealing the relationship between environmental parameters and human perception. The results indicated significant differences in user needs across various functional spaces, with commercial-oriented areas emphasising environmental attractiveness and comfort, while transport-oriented spaces prioritised access efficiency and safety. This study provided quantitative design benchmarks for underground spaces’ dynamic regulation and sustainable management, proposing a precise and adaptive environmental decision-making framework that combines physical parameters with user-perception feedback.

1. Introduction

1.1. Underground Space in the Perspective of Globalisation

With the acceleration of urbanisation, cities worldwide are confronting numerous challenges, including population growth, land resource shortages, and pressures on environmental sustainability. Traditional aboveground spaces fail to meet urbanisation demands, highlighting the potential of underground spaces to alleviate urban strain. Research indicates that underground space development can effectively increase the liveable area by 25% to 40%, providing enhanced functional support for high-density urban environments [1,2]. Broere et al. have highlighted the unique potential of underground spaces in traffic diversion, commercial development, and emergency evacuation, emphasising that their development is vital for enhancing the overall carrying capacity of cities [3]. Nonetheless, the distinct environmental conditions of underground spaces—such as limited ventilation and inadequate natural light—place significant demands on design and management, which can directly impact user experience and usability [4,5].

In urban spatial-quality research, numerous studies have concentrated on above-ground public spaces, such as parks, streets, and transport hubs. Xu et al. have developed an urban vitality evaluation model utilising Point of Interest (POI) data to investigate the impact of spatial characteristics on urban functions from multiple dimensions [6]. Biljecki and Ito systematically analysed the environmental characteristics of urban streets through streetscape image data, establishing a methodological foundation for urban design and spatial optimisation [7]. Despite the significant advancements in methodology and technology, research on underground spaces remains insufficient, and a unified evaluation framework has yet to be established [8]. Additionally, Yap et al. identified, through international research, that public concern regarding the environmental quality of underground spaces primarily revolves around air quality and health and safety issues, underscoring the importance of environmental regulation in developing underground spaces [9]. The environmental quality of underground spaces impacts user comfort and is closely linked to psychological perception and behavioural patterns. Gao et al. demonstrated the significant influence of building density and functional zoning on psychological perception through quantitative analysis of public perceptions in metro station areas [10]. Sun et al. explored the visual saliency of the guidance system in an underground shopping street using eye movement experiments, emphasising the crucial role of visual elements in environmental design [11,12]. These studies indicate that traditional methods for evaluating physical environments have proven inadequate for comprehensively assessing the environmental quality of underground spaces. Consequently, integrating human perception with environmental regulation has emerged as a vital approach to addressing the complex environmental challenges associated with underground spaces.

1.2. Environmental Decision-Making with a Human Perspective

User-centred evaluation of human perception is a key focus in environmental regulation research. The classical Predicted Mean Vote (PMV) model, introduced by Fanger in 1970, is an assessment method based on physical parameters, including air temperature, relative humidity, radiant temperature, and clothing insulation coefficient, to measure human thermal comfort [13]. The PMV model has been extensively utilised to optimise thermal environments within buildings. For instance, Hong et al. assessed thermal comfort and energy balance in residences using PMV models, and their proposed personalised thermal environment adjustment strategies significantly enhanced the living experience [14]. Additionally, Zhang et al. have improved the PMV model by incorporating metabolic rate adjustments, thereby increasing its adaptability and predictive capability in complex scenarios [15]. The PMV model plays a prominent role in thermal comfort research. However, its focus on single physical dimensions limits its ability to capture users’ multidimensional perceptual experiences in complex environments.

To address the limitations of a single-dimensional approach, the Indoor Environment Quality (IEQ) model offers a more comprehensive framework by integrating multidimensional parameters, including air quality, lighting, thermal comfort, and noise, into the assessment [16]. Wong et al. explored the application of the IEQ model within the air-conditioning systems of office spaces, demonstrating its ability to balance user comfort with energy-saving efficiency, thereby providing a valuable reference for environmental regulation [17]. Dunleavy et al. further employed the IEQ model to investigate the psychological impacts of aboveground and underground working environments, revealing the significant influence of underground environmental conditions on users’ psychological stress and emphasising the importance of psychological perception in environmental evaluation [18]. However, despite its broad applicability in integrating multidimensional parameters, the IEQ model exhibits limited responsiveness to individual user perceptions in dynamic environments, particularly in confined and complex scenarios such as underground spaces. This limitation indicates a need for further enhancement of the model’s systematicity and relevance, as it may impact the accuracy of evaluation results and constrain its practical application in dynamic environment regulation.

In recent years, research methods in multi-sensory integration have opened new avenues for human-centred perception studies. Mittal et al. highlighted the significant role of visual perception in enhancing quality of life and optimising urban spatial environments, underscoring the core importance of the diversity and prominence of visual features in spatial design [19]. Gibson’s Sensory Systems Theory further broadens the scope of perception research by advocating for the integrated analysis of multi-sensory data, including vision, hearing, and touch, to capture users’ experiences of spatial environments more comprehensively [20,21]. The multi-sensory approach is particularly relevant in the study of underground spaces, where the enclosed nature of these environments often significantly influences users’ perception patterns. Yao et al. experimentally categorised space into distinct features and identified substantial differences in perceptual preferences across various spatial orientations. This has provided a quantitative foundation for employing multidimensional perceptual measures in underground space [22]. Despite advancements in human-centred perception research at both theoretical and applied levels, several limitations persist. The interface between theoretical frameworks and practical perception evaluation methods remains imperfect, and the absence of a standardised framework results in inconsistencies between data collection and analysis outcomes. Furthermore, the capacity for real-time capture of subjective perceptual feedback in dynamic environments requires enhancement, particularly in the context of the closed and multifunctional nature of underground spaces, where evaluation results struggle to inform design optimisation directly. In this study, we constructed a framework combining multidimensional perception evaluation and environmental parameter regulation to cope with the above problems and further explore the interaction between user perception and environmental quality.

1.3. Machine Learning Applied to Environmental Evaluation Systems

With the advancement of big data technology, machine learning has emerged as a pivotal tool in environmental evaluation and optimisation research. Unlike traditional statistical methods, machine learning supports perceptual evaluation in dynamic environments through nonlinear modelling and efficient data processing. For instance, Liu et al. have automated the evaluation of large-scale urban environmental quality by analysing street view images using convolutional neural networks (CNNs), significantly enhancing research efficiency [23]. Similarly, Ye et al. have employed the SegNet algorithm to assess the visual quality of streets, further elucidating the potential mechanisms through which environmental design influenced user perceptions [24]. These studies illustrate that machine learning can uncover latent patterns within complex data, providing a scientific foundation for environmental optimisation.

XGBoost, an integrated learning algorithm based on the Gradient-Boosting Decision Tree framework, has gained widespread application in environmental data modelling due to its high efficiency and predictive capabilities [25]. Tian et al. have utilised XGBoost to analyse psychological assessment data, successfully predicting users’ moods in response to specific environmental changes, thereby offering dynamic modelling support for perceptual evaluation [26]. Furthermore, He et al. have integrated XGBoost with SHAP algorithms to examine the exact effects of environmental parameters on user perception, yielding highly accurate and interpretable results for multidimensional data correlation analyses in complex environments [27]. Additionally, Chen et al. have developed a prediction model for air quality in underground spaces using deep multitask learning, which has enhanced pollution monitoring accuracy and facilitated dynamic regulation of the spatial environment with quantitative support [28]. Collectively, these studies have demonstrated the significant technical value of machine learning in optimising underground space environments, providing crucial theoretical support for dynamic perceptual evaluation and design optimisation.

In summary, underground space has emerged as a focal point for research and application due to its potential to alleviate the pressures of urbanisation and enhance the functionality of cities. However, the unique environmental conditions of underground spaces have presented significant challenges for environmental management and user experience. While existing studies have made strides in evaluating single-dimensional physical parameters and multi-sensory fusion, there have remained notable deficiencies in dynamic environmental adaptation, integrating subjective perception with physical parameters, and developing a comprehensive optimisation framework. Furthermore, the techniques for acquiring and analysing perceptual evaluation data are not yet fully developed, complicating efforts to provide adequate guidance for practical environmental optimisation. To address these shortcomings, this paper has integrated human-centred perception with machine-learning techniques to propose a unified framework that explores the intricate relationship between environmental parameters and user perception, aiming to enhance underground spaces’ environmental quality. The subsequent method section provides a detailed description of the research design, including data acquisition and processing, the extraction of multidimensional perception indicators, and the application of machine-learning models.

2. Materials and Methods

2.1. Case Study

Shanghai, the economic centre of China, is committed to leveraging urban renewal to enhance urban functions and focuses on urban renewal to improve urban functions, the living environment, and sustainable development. The spatial utilisation of urban underground spaces can effectively address urban challenges such as land shortages and traffic pollution [29]. With the rapid expansion of the metro system, the development of underground spaces in Shanghai has accelerated. Researchers have argued that the underground public spaces in Shanghai serve essential functions and necessitate careful planning and design to optimise their performance [30,31]. For this investigation, the study focused on four specific locations in Shanghai as the case study: (a) the underground space of Shanghai People’s Square; (b) the underground space of Hongqiao Hub; (c) the public underground transport space at Jing An Temple; and (d) the underground space at Wujiaochang. The analysis revealed that case (a) was constructed earlier and exhibits limited commercial activity. Case (b) is still in the early stages of construction, with many facilities yet to become operational. Case (c), primarily serving as a transportation hub, has a relatively singular function. In contrast, case (d) has undergone a reorganisation of spatial resources, possesses diverse functional attributes, and has remained vibrant over time. Consequently, case (d) was selected as the focus of this study, as shown in Figure 1.

Wujiaochang comprises five distinct roads that collectively delineate five representative commercial areas, as outlined in

Table 1. Basic information about the studied site.

No.	Name	Commercial Area × 10,000 m2	Building Area × 10,000 m2
A	Wanda Plaza	21	33
B	Bailian Youyicheng Shopping Mall	9	12
C	Heshenghui West Building	16	36
D	Youmai Life Plaza	3	4.5
E	Suning Plaza	1.5	3.5

. From a human-centred perspective, a walking distance of approximately 500 m is generally acceptable [32]. Consequently, this study concentrates on the underground space of Wujiaochang within a 250 m radius as its primary research area. The autumn season was selected for data collection due to the mild climate conditions in Shanghai during this period, characterised by an average temperature close to the optimal 25 °C for human comfort. This temperature minimised the disparity between indoor and outdoor environments, thereby reducing the temperature-adaptation burden on participants and facilitating the collection of natural and stable environmental perception data.

Initially, spatial data were collected on-site, and the analysis was conducted using Depthmap X for quantification, employing the DBSCAN clustering function to delineate the functional spaces (specific operational processes are detailed in Appendix A) [33,34,35]. Ultimately, collection points exhibiting distinct functional characteristics were identified, as shown in

Table 2. Number of collection points.

Space Type	Point Number	Questionnaire Number
Commercial point	8	144
Transport point	10	164

. These areas include underground public walkways, commercial walkways, sunken plazas, and other public spaces. The collection points are categorised into commercial-oriented and transport-oriented points, with composite points indicating areas that serve both traffic and commercial functions, as depicted in Figure 2 and Table 2.

2.2. Research Framework

This study develops a decision-making model for the environmental management of underground public spaces, focusing on human-centred perception. The research is organised into three phases: (i) site selection and research—suitable research subjects are identified and examined through site investigations; (ii) indicator selection and definition—appropriate evaluation and measurement indicators are determined using literature analysis and the Delphi method; (iii) environmental data collection and measurement—on-site measurements emphasise the human-centred collection of spatial perception and physical environment data; and (iv) environmental data processing and analysis—the collected data are processed to identify suitable environmental parameters.

The study investigates the influence of physical environmental factors on users’ perceptions of underground spaces and ultimately proposes strategies for regulating the spatial environment. The specific research steps are shown in Figure 3.

2.3. Research Methodology

2.3.1. Step 1: Site Selection and Research

This research focuses on human interactions and employs the PSPL (Public Space and Public Life Survey) method. This research focuses on human interactions and employs the PSPL method. The PSPL method is frequently used to assess the quality of urban public spaces and remains a vital tool in contemporary studies of urban public spaces (Gehl et al.) [36]. Matan et al. [37] evaluated urban spaces in Australia using the PSPL method and proposed optimisations for these areas. Similarly, Fan and Shi [38] derived recommendations for enhancing urban spaces based on field studies of street-scale façades. A comparative analysis of various underground urban research cases in Shanghai revealed that the underground space in Wujiaochang exhibits a diverse spatial layout and significant pedestrian activity. Moreover, it is characterised by rich spatial functions and extensive traffic coverage. These attributes offer diverse environmental conditions and sufficient experimental samples, making it suitable for underground-space research.

2.3.2. Step 2: Indicator Selection and Definition

The existing research findings were systematically organised using a literature analysis method, identifying indicators of human-centred perception and the physical environment closely related to evaluating spatial quality. This literature analysis method was used to assess the quality and reliability of existing studies while identifying their biases and limitations, and it provided a theoretical basis and reference framework for this study. The Delphi method was then employed to refine the identified indicators.

This study involved a panel comprising five postgraduate students and eleven in-service faculty members, all from related fields, such as architecture and town and country planning. All panel members were required to be researchers in urban and environmental disciplines. Postgraduate members included MSc and PhD students in architecture and environmental studies with strong expertise and research backgrounds. Meanwhile, the in-service faculty members are senior professors and experts in architecture, characterised by their substantial practical experience and teaching history. In the initial consultation round, the expert group identified pertinent perceptual and environmental indicators based on the existing literature. These indicators included dimensions such as vision [39], auditory [40], olfactory, somatosensory, taste, and other established international research categories. After a thorough screening process, 35 perceptual and 10 environmental indicators were initially recognised. The expert group members reevaluated and classified these preliminary indicators in the second consultation round. After extensive discussions, a refined selection emerged, comprising 22 perceptual and 7 environmental indicators, focusing on security, open, and wind strength. The third consultation round involved gathering anonymous feedback and facilitating targeted discussions based on the input received from the second round, which enabled the expert panellists to reach a consensus. Ultimately, three dimensions were established, consisting of 13 core perceptual indicators and six environmental indicators, which served as the evaluation criteria for assessing sensory perception of spatial quality in this study. Furthermore, a five-level Likert scale was utilised in the questionnaire design, and an additional dimension of ‘perception satisfaction’ was included as a criterion for evaluating spatial perception suitability, as shown in

Table 3. Spatial-quality sensory-perception evaluation indicators.

	Human-Centred Perception Indicator Evaluations
	Visual-perception evaluation	Strongly disagree	Disagree	Neutral	Agree	Strongly agree
1.	Security	□1	□2	□3	□4	□5
2.	Gorgeous	□1	□2	□3	□4	□5
3.	Non-repression	□1	□2	□3	□4	□5
4.	Beauty	□1	□2	□3	□4	□5
5.	Interesting	□1	□2	□3	□4	□5
6.	Open	□1	□2	□3	□4	□5
7.	Comfortable Lighting	□1	□2	□3	□4	□5
8.	Visual perception satisfaction	□1	□2	□3	□4	□5
	Auditory perception evaluation	Strongly disagree	Disagree	Neutral	Agree	Strongly agree
9.	Quiet	□1	□2	□3	□4	□5
10.	Varied	□1	□2	□3	□4	□5
11.	Pleasant	□1	□2	□3	□4	□5
12.	Auditory perception satisfaction	□1	□2	□3	□4	□5
	Somatosensory perception evaluation	Strongly disagree	Disagree	Neutral	Agree	Strongly agree
13.	Good ventilation	□1	□2	□3	□4	□5
14.	Wind strength	□1	□2	□3	□4	□5
15.	Warm	□1	□2	□3	□4	□5
16.	Somatosensory perception satisfaction	□1	□2	□3	□4	□5
	Physical environment indicator measurements
	Measurement indicators	Average value			Unit
1.	Aspect ratio	-			-
2.	Humidity	-			Percentage (%)
3.	Illumination	-			Lux (lx)
4.	Sound	-			Decibels (dB)
5.	Temperature	-			Degrees Celsius (°C)
6.	Wind speed	-			Meters per second (m/s)

. The specialised definitions of each indicator are provided in Appendix B.

2.3.3. STEP 3: Environmental Data Collection and Measurement

(1)

Person-centred perceptual data acquisition

This study used a questionnaire to assess human-centred environmental perception on a defined scale. This approach demonstrated strong efficacy in research concerning urban spatial quality [41]. The study identified 13 locations within the site for evaluating human-centred environmental perception. Questionnaires were administered during three designated periods: 11:00–13:00, 14:00–16:00, and 17:30–19:30. Five architecture and planning students conducted the on-site questionnaires and gathered feedback from respondents. To enhance comprehension, the basic definitions of the indicators were explained to the respondents before the questionnaire. Additionally, respondents were reminded to perform on-site evaluations of their environment to ensure the relevance of their responses. The test subjects primarily consisted of approximately 200 passers-by at the venue, and the random selection of respondents ensured a diverse research sample.

(2)

Physical environment data measurements

To assess the physical environmental data within the site, this study conducted several real-time measurements across 13 designated points, focusing on key parameters such as Sound, temperature, illumination, wind speed, and humidity. The measurement equipment included a Kanomax hot-wire anemometer for precise wind speed measurements; a CDT-8820 environmental tester for recording temperature, illuminance, and humidity; and a laser range finder to measure the cross-sectional scale of the site, ensuring accurate recording of physical environmental scales at each point. Data measurements were synchronised with questionnaires to analyse the impact of physical environmental factors on human perceptions. Data were collected from 11 October to 13 October 2023, ensuring continuity and comprehensive coverage.

2.3.4. STEP 4: Environmental Data Processing and Analysis

(1)

Determination of indicator weights

This study employed a comprehensive weighting method that integrates the CRITIC and Entropy weight techniques to ascertain the weights of human-centred perception indicators across three dimensions. The Entropy weight method is advantageous, as it quantifies the information entropy of each indicator, enabling the determination of its importance based on the objective variability of the data, free from subjective influences [42]. Meanwhile, the CRITIC weight method ensures that the weight assigned to each indicator accurately reflects its overall impact by analysing the comparison coefficients and the intrinsic correlations among the indicators [43] (The specialised definitions of each weighting method are provided in Appendix C.). Consequently, combining these two methods facilitates a more precise determination of weights within each perceptual dimension, thereby establishing a robust scientific foundation for subsequent analysis.

(2)

Correlation analysis

This study employed Kendall’s tau-b (τ_b) correlation analysis to identify highly correlated human-centred perception indicators related to physical environment data. Kendall’s tau-b (τ_b) is a non-parametric statistical method that effectively measures the correlation between two variables. It is particularly well-suited for analysing hierarchical and non-normally distributed data [44].

(3)

Regression analysis

After identifying indicators that exhibit a high correlation with the physical environment data, this study conducted a regression analysis to investigate the relationship between these perceptual indicators and physical environment variables, such as illuminance. The regression analysis aims to fit curves that clarify this relationship. This method proves effective in the regression analysis of spatial perception [45]. For the scores of the perceptual indicators, this study established a suitability range between 3.5 and 5. Fitting curves inferred the optimal environmental parameters for each perceptual indicator within this range, and the suitability parameters were derived by integrating data from actual measurement points.

3. Results

3.1. Human-Centred Perception Questionnaire Collection

In total, 236 questionnaires were collected during this field study (as detailed in

Table 4. Collection of Basic Information on the Questionnaire.

	Collection of Basic Information on the Questionnaire
	Gender	Count	Percentage
1	Male	120	56.34%
2	Female	93	43.66%
	Age group	Count	Percentage
1	Below 18 years	20	9.39%
2	18–29 years	106	49.77%
3	30–39 years	42	19.72%
4	40–60 years	28	13.15%
5	Above 60 years	17	7.98%
	Functional use	Count	Percentage
1	Shopping	94	44.13%
2	Dining	89	41.78%
3	Leisure	87	40.85%
4	Transportation	83	38.97%
5	Walking	64	30.05%
6	Entertainment	62	29.11%
7	Work	35	16.43%
8	Rest	32	15.02%
9	Chatting	22	10.33%
10	Fitness	2	0.94%
11	Others	2	0.94%

), with 213 valid responses retained after screening. Among the respondents, 120 were male (56.34%) and 93 were female (43.66%). Regarding age distribution, the 18–29 age group represented the most significant proportion at 49.77%, followed by the 30–39 age group at 19.72%. Respondents under 18 (9.39%) and those over 60 (7.98%) constituted smaller portions of the sample. These results show that young and middle-aged adults, particularly those aged 18–39, form the primary user group of the underground walkway.

The underground space was primarily utilised for commercial and transportation purposes. Specifically, shopping (44.13%), dining (41.78%), and leisure activities (40.85%) represented the main commercial functions, highlighting the space’s emphasis on commercial usage. Transportation accounted for 38.97%, further demonstrating the space’s dual role in supporting commerce and pedestrian mobility. In contrast, less frequent activities included resting (15.02%), chatting (10.33%), and fitness-related pursuits (0.94%), indicating that the space is more conducive to active, purpose-driven activities rather than passive or social interactions. Consequently, this study seeks to examine the relationship between environmental influences and human-centred perception by focusing on two functionally oriented spaces—commercial and transportation—as the core research contexts.

3.2. Data Standardisation and Empowerment

Table 5. Results of the combined weights method.

Entropy Weight Method				CRITIC Weight Method				Final Weight (%)
Item	Information Entropy Value e	Information Utility Valued	Weight (%)	Indicator Variability	Indicator Conflict	Information Quantity	Weight (%)
Visual dimension	0.977	0.023	32.62	1.081	0.646	0.699	30.51	31.57
Auditory dimension	0.974	0.026	37.91	1.094	0.728	0.796	34.77	36.34
Somatosensory dimension	0.980	0.020	29.46	1.049	0.758	0.795	34.72	32.09

presents the weight relationships among the three dimensions derived using the CRITIC and Entropy weighting methods. The auditory dimension has the highest weight, reaching 36.34%, signifying its dominant role in multidimensional perception. The somatosensory dimension is closely followed by a weight of 32.09%, highlighting the importance of somatosensory factors. In contrast, the visual dimension weighs 31.57%, slightly lower than the first two dimensions.

3.3. Correlation Analysis of Human-Centred Perception Evaluation and Physical Environment Data

Figure 4 shows that the various space types exhibit distinct correlation patterns. Aspect ratio significantly influenced commercial-oriented spaces, showing a strong positive correlation with wind strength (correlation coefficient = 0.171). Additionally, humidity showed a moderate positive correlation with good ventilation and wind strength, with correlation coefficients of 0.157 and 0.118, respectively. Furthermore, illumination was significantly correlated with comfortable lighting and security, with correlation coefficients of 0.178 and 0.138. Although some correlation trends resemble those in transport-oriented spaces, significant differences exist in commercial-oriented spaces. The correlation between aspect ratio and wind strength remains notable, with a correlation coefficient of 0.140. Illumination plays a more significant role in the transport-oriented space, exhibiting positive correlations with security and openness, with correlation coefficients of 0.185 and 0.160, respectively. Additionally, wind speed emerges as the dominant factor, displaying the highest correlations with good ventilation and wind strength, with correlation coefficients of 0.291 and 0.288.

Nevertheless, certain factors, such as temperature and sound, exhibit weaker correlations in transport-oriented spaces than commercial-oriented spaces. For instance, the correlation between temperature and wind strength is only −0.080, indicating a limited impact on user perception under typical conditions. However, the subtle differences in these correlations across spatial contexts underscore the importance of prioritising key parameters for various functional types of environments.

Through the screening and organisation of the correlation results between human perceptual indicators and physical environment parameters,

Table 6. Relevance attributions for the human-centred perception indicators.

Physical Environment Data	Perceptual Dimension	Commercial-Oriented Space’s Perceptual Indicators	Transport-Oriented Space’s Perceptual Indicators
Aspect ratio	Visual, somatosensory	Security, non-repression, interesting, open, comfortable lighting, good ventilation, wind strength, warm	Wind strength
Humidity	Somatosensory	Good ventilation, wind strength	Good ventilation, wind strength
Illumination	Visual, somatosensory	Security, gorgeous, non-repression, open, comfortable lighting, warm	Security, gorgeous, non-repression, beauty, interesting, open, comfortable lighting, warm
Sound	Auditory	Quiet	Varied
Temperature	Visual, somatosensory	Comfortable lighting, good ventilation, wind strength	Good ventilation
Wind speed	Somatosensory	Good ventilation, wind strength, warm	Good ventilation, wind strength, warm

illustrates the relationships between various physical environment data (aspect ratio, humidity, illumination, sound, temperature, and wind speed) and the perceptual dimensions (visual, auditory, and somatosensory). Additionally, it categorises the specific perceptual indicators (e.g., security, comfortable lighting, good ventilation, and wind strength) under the commercial and transport dimensions.

3.4. Cross-Validation Analysis Based on Traditional Regression and Machine Learning Regression

Through correlation and regression analyses, it is found that different physical environment parameters have significant nonlinear effects on users’ perceptual experiences. The physical environment parameters such as aspect ratio, humidity, and illumination significantly increase the satisfaction of perceptual indicators within the appropriate range. Different parameters show differences in the commercial-oriented and transport-oriented spaces. Aspect ratio, humidity, illumination, and other physical environment parameters significantly increase satisfaction in the appropriate range.

3.4.1. Single-Indicator Fitted Regression Analysis

This section integrates single-indicator fitting regression analyses of commercial and transport-oriented spaces. It examines the performance trends of each parameter across different spatial function orientations and refines the recommended ranges.

In the commercial-oriented space (as shown in Figure 5), the aspect ratio is significantly positively correlated with indicators such as wind strength. When the aspect ratio falls between 1.59 and 2.81 or exceeds 3.14, the satisfaction levels of several perception metrics surpass 3.5, particularly wind strength at an aspect ratio of 1.83, which reaches 3.75. This suggests that this ratio optimises users’ perception of ventilation within the space. Similarly, humidity significantly enhances the satisfaction of good ventilation and wind strength within the range from 52.31% to 59.22%, with the highest recorded satisfaction at 4.40 occurring at 59.20%. This indicates a strong sensitivity of humidity levels to perceptions of airflow. Illumination levels are notably higher, within the range from 808.11 lx to 1338.24 lx, with a peak satisfaction score of 3.5 for wind strength at an aspect ratio of 1.83, indicating that this ratio also optimises user perceptions of ventilation. Illumination performs best within the 808.11 lx to 1338.24 lx range, as users report significantly greater satisfaction regarding comfort and spatial perception of the lighting, especially for the gorgeous and open metrics, which scored over 4.00. Analysis of sound levels indicates that when the noise level is below 61.21 dB, the satisfaction level for quiet generally exceeds 3.5, particularly at 59.60 dB, where satisfaction reaches 3.00, and at 59.20 dB, where satisfaction increases to 4.40. Regression analysis of temperature indicates that when temperatures range from 22.63 °C to 26.39 °C, the satisfaction levels for good ventilation and wind strength reach 4.50 and 4.00, respectively, underscoring the importance of temperature regulation in enhancing comfort perceptions. Wind speed has a particularly significant impact on user experience, with good ventilation and wind strength scores exceeding 3.5 for wind speeds ranging from 0.26 m/s to 0.67 m/s, further validating the critical role of wind speed in optimising perceptions of ventilation in business-oriented spaces.

In contrast, the transport-oriented space (as shown in Figure 6) exhibits a more refined trend in physical environment parameters. The aspect ratio, ranging from 3.12 to 3.20, significantly correlates with increased user satisfaction regarding wind strength, achieving a peak rating of 4.00. The recommended ranges for humidity and illumination are from 50.09% to 59.20% and from 142.19 lx to 480.03 lx, respectively. Notably, humidity is highly correlated with perceptions of ventilation, while lower illuminance levels (e.g., 158.80 lx and 474.00 lx) align more closely with user preferences for transport spaces. The regression analysis for sound indicates that user satisfaction with quiet is significantly enhanced when noise levels fall between 63.15 dB and 75.45 dB. Temperature is most effective in the range from 21.95 °C to 26.20 °C, optimising perceptions of ventilation and warmth, thereby improving overall satisfaction. Furthermore, wind speed ranging from 0.18 m/s to 0.78 m/s is the most effective in enhancing user satisfaction with good ventilation and warmth. Specifically, wind speed in this range demonstrates a marked increase in ratings for good ventilation and wind strength, with satisfaction peaking at 4.75 for a wind speed of 0.67 m/s.

These results further emphasise the synergies among the parameters and their significant differences across various functionally oriented spaces. Such differences are evident in specific physical environmental parameters and the diverse user needs for different scenarios.

3.4.2. Regression Model Analysis Based on XGBoost Algorithm

This study further investigates the nonlinear relationship between physical environmental parameters and human perception indicators, as shown in Figure 7. Although the predictions generated by XGBoost are generally effective, the construction of the model demonstrates a degree of randomness, with R² scores proving insufficient in certain instances. The inadequate R² scores can be attributed to several factors. (1) Data characteristics: The data distribution in specific dimensions, such as aspect ratio and sound, is more scattered and exhibits more significant variability, complicating the capture of complex relationships. (2) Sample size limitation: The limited number of samples in these dimensions further restricts the model’s generalizability and stability. (3) Model-specific limitations: While XGBoost excels in modelling structured data, it is less responsive to weakly correlated features or noisy data, which may account for the model’s diminished predictive capability in these dimensions.

This indicates that result stability needs improvement. Nonetheless, the model shows promise in providing auxiliary references for the future optimisation of environmental parameters. The study specifically selected model results with an R² greater than approximately 0.4 and compared these with those observed for single-indicator fits. A detailed analysis of the specific regression results is presented below (as shown in Figure 7).

(1)

Aspect Ratio

In the commercial-oriented space, the XGBoost regression model exhibits R² values of less than 0.3 for all relevant indicators. This indicates that it does not provide valid predictions and suggests instability within this space. This limitation implies that the XGBoost model struggles to accurately capture the relationship between aspect ratio and human-oriented perception indicators in this context. Consequently, the prediction results for this dimension should be interpreted with caution and regarded as exploratory rather than definitive findings. In the transport-oriented space, the fitted equation for wind strength has an R² value of 0.1886, indicating that the model’s predictions exhibit insufficient stability to offer informative ranges. As a result, the aspect ratio model prediction results in both spaces lack validity.

(2)

Humidity

In the commercial-oriented space, the XGBoost regression model demonstrates that good ventilation’s fitted equation has an R² value of 0.5212, indicating a good predictive capability, with a recommended humidity range from 39.2000% to 59.2000%. Compared to the recommended humidity range for the single-indicator fit (52.31% to 59.22%), the model extends the range under low-humidity conditions, showcasing greater adaptability. In the transport-oriented space, good ventilation’s fitted equation presents an R² value of 0.4043, and the model also predicts effectively, recommending a humidity range from 39.2000% to 59.2000%, demonstrating broad applicability under low-humidity conditions.

(3)

Illumination

In the commercial-oriented space, the XGBoost regression model demonstrated an R² value of 0.4077 for the fitted equation concerning gorgeous, indicating enhanced model prediction capabilities. The recommended illuminance range is between 23.5000 lx and 1184.8990 lx. In contrast, the single-indicator fit suggests a narrower range, from 808.11 lx to 1338.24 lx. The model predicts a lower bound, indicating a more robust user perception under low-illumination conditions. In the transport-oriented space, gorgeous’s fitted equation yields an R² value of 0.3818, with the model also demonstrating compelling predictions, recommending an illuminance range from 23.5000 lx to 8525.0000 lx. This range is significantly broader than that of the single-indicator fit, which is between 142.19 lx and 480.03 lx, particularly highlighting the model’s capability to predict under high-illuminance conditions.

(4)

Sound

In the commercial-oriented space, the XGBoost regression model reveals that the fitted equation for noisy has an R² value of 0.3654, which is close to the stability threshold but does not meet the desired criteria. The recommended noise range is from 59.68 dB to 71.06 dB, which is broader than the single-indicator fit range, from 59.60 dB to 61.21 dB. This wider prediction range, particularly in high-noise environments, suggests that the machine-learning model is better equipped to handle varying noise conditions. In the transport-oriented space, the fitted equation for varied has an R² value of 0.2263, indicating that the model predictions are not sufficiently stable to provide a reliable reference. Given the low R² score, the projections for sound should be considered preliminary and indicative of the need for further refinement of the model or additional data collection to improve stability.

(5)

Temperature

The results of the XGBoost regression model indicate that the fitted equation for good ventilation in the commercial-oriented space has an R² value of 0.5269, predicting a temperature range from 21.75 °C to 27.70 °C. The single-indicator fit predicts a narrower range, from 22.63 °C to 26.39 °C. This finding suggests that the machine-learning model offers a slightly broader range and encompasses the recommended temperature parameters of the single-indicator fit. This indicates that the model is more flexible and stable in predicting temperature parameters. In the transport dimension, good ventilation’s fitted equation has an R² value of 0.3602, close to the threshold value, with a recommended temperature range from 21.95 °C to 27.70 °C.

(6)

Wind Speed

In the commercial-oriented space, the XGBoost regression model shows that good ventilation’s fitted equation has an R² value of 0.4828, and the model predicts better, with recommended wind speeds ranging from 0.11 m/s to 0.67 m/s. Compared to the results of the single-indicator fit (0.26 m/s to 0.67 m/s), the model demonstrates higher predictive ability at low wind speeds, indicating that it can accommodate a broader range of environmental conditions. In transport-oriented space, good ventilation’s fitted equation has an R² value of 0.3681, with a recommended wind speed range from 0.11 m/s to 0.78 m/s, showing a more excellent range of acceptability.

However, the predictive capability is somewhat limited and lacks a high level of stability. This single-indicator fitting regression analysis effectively identified the suitability parameters strongly correlated with the human-oriented perception indicators for each physical environment parameter in commercial-oriented and transportation-oriented spaces. Based on these theoretical calculations and actual data points, the recommended ranges for physical environment parameters are adaptable and serve as a reference for subsequent spatial environment optimisation.

In summary, the XGBoost model demonstrates high stability in predicting humidity, illumination, and other key parameters, with its recommended ranges being more flexible than those derived from single-indicator fitting—particularly in the low and high parameter intervals. Ultimately, this study integrates the results of the single-indicator fitting curves and XGBoost model predictions to create a summary table of the recommended ranges (as shown in

Table 7. Summary table of recommended ranges predicted by single-indicator fit curve and XGBoost model.

Commercial-Oriented Space					Transport-Oriented Space
No.	Parameter	Single-Indicator Recommended Range	XGBoost Model Recommended Range	Final Recommended Range	No.	Parameter	Single-Indicator Recommended Range	XGBoost Model Recommended Range	Final Recommended Range
1	Aspect ratio	1.59–2.81; 3.14–3.22	Not Applicable (R2 < 0.3)	1.59–2.81; 3.14–3.22	1	Aspect Ratio	3.12–3.20	Not Applicable (R2 < 0.4)	3.12–3.20
2	Humidity	52.31%–59.22%	39.20%–59.20%	52.31%–59.20%	2	Humidity	50.09%–59.20%	39.20%–59.20%	50.09%–59.20%
3	Illumination	808.11 lx–1338.24 lx	23.50 lx–1184.90 lx	808.11 lx–1184.90 lx	3	Illumination	142.19 lx–480.03 lx	23.50 lx–8525.00 lx	142.19 lx–480.03 lx
4	Sound	59.60 dB–61.21 dB	59.68 dB–71.06 dB	59.68 dB–61.21 dB	4	Sound	63.15 dB–75.45 dB	Not Applicable (R2 < 0.4)	63.15 dB–75.45 dB
5	Temperature	22.63 °C–26.39 °C	21.75 °C–27.70 °C	22.63 °C–26.39 °C	5	Temperature	21.95 °C–26.20 °C	21.95 °C–27.70 °C	21.95 °C–26.20 °C
6	Wind speed	0.26 m/s–0.67 m/s	0.10 m/s–0.67 m/s	0.26 m/s–0.67 m/s	6	Wind Speed	0.18 m/s–0.78 m/s	0.11 m/s–0.78 m/s	0.18 m/s–0.78 m/s

).

4. Discussion

4.1. Research Innovation

This study contributes to the literature by addressing gaps in research on the multi-sensory environmental impacts of underground spaces. It systematically reveals the complex nonlinear relationship between physical parameters and perceptual experiences. Compared with previous studies that mainly focus on the association between behavioural patterns and environmental features, such as Sevtsuk A et al.’s assessment of urban facility construction through pedestrian flow distribution and so on [46,47], and Istrate A’s study on the activity of small-scale businesses and buildings along the street [48], the present study introduces a multi-sensory perspective to make up for the lack of research on the mechanism between physical environmental parameters and perceptual experience. By introducing visual, auditory, and somatosensory dimensions, the study constructed a correlation model between perception and physical environment parameters, providing theoretical support for underground space optimisation.

The results show that physical parameters not only affect a single perceptual index but may also affect multiple perceptual experiences at the same time. For example, humidity significantly increased the satisfaction of good ventilation and wind strength in the range from 52.31% to 59.22% and enhanced overall comfort by modulating the sense of airflow. This finding provides an important basis for fine-tuning the design of physical parameters. This study also found significant differences in lighting and temperature requirements between traffic and commercial spaces. The recommended illuminance range for traffic space is from 142.19 lx to 480.03 lx, while the recommended range for commercial space is from 808.11 lx to 1338.24 lx, which reflects the difference in the light environment demanded by the users in different scenarios in terms of temperature; the results of this study are basically the same as the range of the working environment temperature proposed by Wu and other scholars (from 22.0 °C to 27.3 °C) [49], but the range in this study differs slightly. It is basically consistent with the range proposed by Wu and other scholars (from 22.0 °C to 27.3 °C) [49], but the range of this study is more specific, and it is further found that the commercial space has a higher demand for temperature stability, while the traffic space is more affected by the synergistic factors such as humidity and wind speed.

Compared with traditional indoor space evaluation models that focus on a single physical environment indicator [50], this study systematically introduces a multi-sensory evaluation framework, which strengthens the depth of the ‘environment–human’ linkage study. Previous studies on underground spaces, such as Meng Q’s study, have shown that background music in traffic spaces can mask unwanted noise, and the comfort level is optimal between 65 dB and 70 dB [51], which is consistent with the results of this study, and further reveals that users of commercial spaces are more sensitive to noise than those of traffic spaces.

In summary, this study not only validates the established research results but also presents new contributions at the level of fine-grained understanding of perceptual mechanisms and guidance for optimisation of physical parameters. Through the systematic framework of multi-sensory perception, this study provides a comprehensive and scientific theoretical approach to the design of underground space environments with different functional orientations.

4.2. Limitations and Future Directions

This study explores key issues in the environmental regulation of underground spaces based on human perception evaluation. While the research methodology was effective in achieving the research objectives, the findings also revealed several critical problems in the field that need to be addressed, thus providing important directions for future research.

Firstly, during the data selection process, most of the existing studies have paid relatively more attention to visual perception indicators while covering less of the other perception dimensions, such as visual, auditory and somatosensory. This imbalance limits the comprehensiveness of perception data, as well as the representativeness of the findings. Future research could focus on the balanced collection of multi-sensory data. In addition, environmental indicators do not exist in isolation, and their interactions and their comprehensive influence mechanisms on user perception have not been fully investigated. Future research should focus on the dynamic interactions between these metrics and explore how they synergistically affect the overall user experience.

Second, this study used machine learning to model and analyse environmental data. These results demonstrate the potential of these techniques for modelling environmental perception in underground spaces. However, model stability and generalisation remain limited due to the small size and single source of training data. Therefore, future research should expand the size and diversity of the dataset to cover a wider range of environmental conditions and user perceptions and diversify the data collection channels. Meanwhile, introducing more advanced techniques, such as neural networks or multi-modal fusion approaches, can further improve the adaptability and prediction accuracy of the model in complex environments.

Once again, environmental data collection was focused on specific periods, and this limitation in timeframe may have resulted in differences in seasonal environmental conditions and their impact on user perceptions not being adequately captured. For example, user perceptions of temperature and humidity may be significantly different in winter versus summer, a limitation that may affect the applicability and extrapolation of the study results. For this reason, future studies should conduct long-term and multi-seasonal data collection covering different temporal scenarios to more comprehensively assess the impact of dynamic changes in environmental factors on user perceptions. This broader temporal coverage can provide a more solid foundation for understanding the dynamic relationship between environmental parameters and user perception and enhance the generalisability of the research results.

In addition, the findings suggest that emerging technologies, such as virtual reality (VR) and augmented reality (AR), can be combined in the future to break through the limitations of physical experiments. By simulating multiple physical environment parameters through virtual scenarios, researchers can conduct large-scale experiments more efficiently, collect users’ perceptual feedback in multiple scenarios, and enhance the breadth and depth of experimental data. Virtual experiments can not only complement traditional field studies but can also be used to quickly test hypothetical design scenarios and accelerate the design process for underground space optimisation. The rapid development of artificial intelligence (AI) and the Internet of Things (IoT) is driving underground space management towards greater intelligence and adaptability. For example, the combination of intelligent sensor networks and deep-learning algorithms can realise real-time monitoring and adaptive regulation of environmental conditions. Future designs should also achieve a deeper integration between intelligence and sustainability to make underground spaces efficient, adaptable, and environmentally friendly urban infrastructures.

5. Conclusions

Based on the human-centred perception evaluation framework, this study explored the relationship between the physical environment parameters of underground spaces and users’ perceptual experiences. It revealed the complex nonlinear effects of different physical environment factors through multidimensional data collection and analysis. Using a combination including a questionnaire survey, correlation analysis, traditional regression analysis, and machine-learning model, we systematically analysed the differentiated performance of visual, auditory and somatosensory dimensions of commercial and transportation functional spaces and their key influencing factors.

The results showed that the role of physical parameters is not a single dimension but affects the user’s experience through multidimensional perceptual mechanisms. For example, in the commercial functional space, the aspect ratio significantly enhanced the visual comfort of security and open in a specific range; the appropriate range of humidity and illumination can significantly optimise the user’s physical comfort of wind strength, good ventilation, and so on. In the transportation functional space, the physical parameters are not one-dimensional; instead, work together through multidimensional perceptual mechanisms. However, different results were presented in the transport functional space. The study further validated the suitability ranges of these key parameters through regression and machine-learning modelling and observed significant differences in user needs between commercial and transport spaces regarding temperature, humidity, and illumination.

Additionally, the XGBoost model demonstrated flexibility and adaptability in predicting the impact of humidity, illumination, and temperature, especially at low and high parameter intervals. This provides a more comprehensive theoretical reference for the optimal design of physical environments.

In summary, this study addressed gaps in the environmental perception mechanism of underground spaces and provided a theoretical basis for optimising spaces with different functional orientations. The results not only validate the established theoretical framework; they also further deepen the understanding of the complex interaction between physical environment parameters and perceptual experience, which is of great academic value and practical application significance.