Predicting and Optimizing Restorativeness in Campus Pedestrian Spaces based on Vision Using Machine Learning and Deep Learning

Abstract

Restoring campus pedestrian spaces is vital for enhancing college students’ mental well-being. This study objectively and thoroughly proposed a reference for the optimization of restorative campus pedestrian spaces that are conducive to the mental health of students. Eye-tracking technology was employed to examine gaze behaviors in these landscapes, while a Semantic Difference questionnaire identified key environmental factors influencing the restorative state. Additionally, this study validated the use of virtual reality (VR) technology for this research domain. Building height difference (HDB), tree height (HT), shrub area (AS), ground hue (HG), and ground texture (TG) correlated significantly with the restorative state (ΔS). VR simulations with various environmental parameters were utilized to elucidate the impact of these five factors on ΔS. Subsequently, machine learning models were developed and assessed using a genetic algorithm to refine the optimal restorative design range of campus pedestrian spaces. The results of this study are intended to help improve students’ attentional recovery and to provide methods and references for students to create more restorative campus environments designed to improve their mental health and academic performance.

1. Introduction

Given their high cognitive demands, educational environments are challenging, especially for university students requiring prolonged focus for learning. This makes them more vulnerable to psychological issues, including stress, anxiety, mental fatigue, and depression [1,2]. Research indicates that restorative environments aid individuals by regaining their depleted attention, alleviating psychological stress, and triggering positive physiological and psychological responses [3,4]. As pedestrian spaces are indispensable parts of campuses, their vitality and attractiveness significantly impact pedestrian students’ health. Optimizing environmental features and visual conditions can significantly reduce stress levels, whereas exposure to tense visual environments may decrease efficiency and focus [5]. Numerous studies have revealed the crucial role of visual environmental perception in restoring students’ attention, with visual perception quality correlating positively with restorative benefits [6,7,8], emphasizing the importance of restorative design in enhancing students’ attention restoration.

Studies have employed various methods to explore the impact of visual environmental perception on spatial factors, including field visits [9], scene photos [10], questionnaire surveys [11], street view images [12], and virtual reality (VR) technology [13]. Advancing computer vision technology has made the exploration of visual environmental perception as a quantifiable regression problem possible by analyzing feedback from millions of users on places. Virtual environments, with rich perceptual stimuli and vivid experiences, provide a high level of “presence”, effectively eliciting participants’ emotional responses [14]. Omar integrated physiological indicators to compare the relationship between physical and virtual walking [15]. Eye trackers enabled data collection on the number and duration of focal points and areas of interest (AOI), which was then used to filter architectural elements [16]. Combining VR and eye trackers enables efficient and accurate visual data collection from built environments [17]. Pei revealed visual attention patterns by analyzing physiological signals and eye-tracking data collected while roaming in a virtual environment, which were used to evaluate architectural design proposals [18]. Immersive Virtual Environments (IVEs) provide scenarios with high degrees of presence and immersion, and Ma’s study on lighting design proved the similarity in participants’ responses between IVE simulations and real environments [19]. Subjective data, obtained primarily through interviews and questionnaire surveys, often inaccurately reflect the effective connection between visual perception and spatial factors. Limitations in studying how spatial feature parameters impact the visual environment in real physical settings are significant. However, VR provides accurate data for visual environmental research.

In 1989, Stephen Kaplan and Rachel Kaplan introduced the Attention Restoration Theory (ART), focusing on the restorative and enhancing effects of environments on human attention [20]. Subsequently, Hartig developed the Perceived Restorativeness Scale (PRS) in 1997 to assess environmental impacts on attention restoration [21]. Perceived space quality is crucial in predicting space restorativeness [22]. Ulrich highlighted the decisive role of the quantity difference between natural- and human-made attributes in environmental perception and attention restoration [23]. Silvia revealed that natural factors (e.g., vegetation and water bodies) and constructed factors (e.g., community buildings, squares, and sculptures) significantly promote restoration [24]. They found, especially in urban settings, that landscapes were more conducive to restoration than some natural landscapes [24]. Jenny demonstrated how urban walking facilitates recovery in adults with poor mental health [25]. Catherine identified urban pedestrian environment features impacting pedestrian psychology; environmental characteristics had different health impacts on different groups [26]. Some researchers have compared the differences in restorative experiences during outdoor walking activities across seasons, thoroughly discussing restorative effects [27]. Lu identified key spatial perception factors within campus environments crucial for bolstering students’ attention restoration [28]. Yasser devised a comprehensive campus ART evaluation tool that synergizes restorative landscape design concepts, visual landscape preferences, and strategic campus planning [29]. Although existing research indicates a strong positive correlation between visual perception and psychological restoration and mature studies have explored urban pedestrian spaces, research on campus spaces generalizes spatial feature classifications too broadly, ignoring the contributions of specific physical dimensions (e.g., fences and paths) and the impact of walking activities on environmental perception. Moreover, studies have generally considered small sample sizes and have not gone in depth to fully understand how different spatial features in campus environments affect students’ stress recovery.

In recent years, advancements in technology have opened new avenues for exploring space perception and its effects on student well-being. Among these advancements, machine learning (ML) and deep learning (DL) have emerged as powerful tools for analyzing complex data and uncovering patterns that are not immediately apparent through traditional statistical methods [30,31,32]. Machine learning techniques are widely applied in built environment studies, including building energy consumption [33,34], thermal comfort [35,36], wind environment [37], and environmental pollution [38]. DL, particularly suited for handling high-dimensional and large-scale datasets, is used in image and video analysis for object recognition, image segmentation, and behavior detection, aiding in monitoring environmental changes and user interactions [39,40]. For instance, analyzing video data can assess the impact of changes in natural environments on users’ psychological states [41]. Convolutional Neural Networks (CNN) are adept at automatically extracting and learning features from images, making them ideal for vision-oriented spatial studies [42,43]. The application of ML and DL in user behavior analysis helps understand activity patterns and preferences in different environments [44]. This information can optimize environmental design and enhance the effectiveness of restorative environments [45]. Despite the widespread application of ML and DL in architectural and environmental studies, research on spatial quality prediction based on visual perception remains limited. Most studies have focused on urban or indoor environments, often neglecting campus-specific scenarios. Additionally, horizontal comparisons between the predictive accuracies of different ML and DL models are scarce, limiting the understanding of model performance and the further application and development of these technologies.

The main purpose of this study was to comprehensively understand the factors in the campus pedestrian environment that influence students’ attention restoration and to propose optimization schemes. This study aimed to achieve the four scientific objectives. 1. Key Spatial Factors for Attention Restoration: To systematically identify and extract the restorative spatial factors that significantly impact students’ attention restoration during campus walks, using eye-tracker experiments and questionnaires to establish a foundational understanding of these factors.2. Mechanistic Analysis of Spatial Factors: To investigate the underlying mechanisms by which the identified spatial factors influence the attention-restoration process in college students, employing restorative orthogonal experiments to uncover causal relationships and interaction effects. 3. Predictive Modeling of Spatial Restorativeness: To develop the spatial restorativeness prediction model, we compared the performance of four machine learning models: Decision Tree (DT), k-Nearest Neighbors (KNN), Artificial Neural Networks (ANN), and CNN. We selected the model with the best training performance as the final predictive model. 4. Optimization of Spatial Design Parameters: To explore the optimization mechanisms of spatial design parameters for campus pedestrian spaces using Genetic Algorithms (GA), we aimed to uncover the theoretical underpinnings and empirical validation of design features that enhance attention restoration.

2. Mterials and Methods

2.1. Study Design and Restoration States Assessment

2.1.1. Study Design

This study investigates the influence of campus spatial factors on college students’ attention restoration during walking and determines the optimal parameter range of campus pedestrian space restorativeness. The methodology mainly consists of three parts, as shown in Figure 1. First, an SD questionnaire was developed to gather subjective data from college students in pedestrian spaces, followed by a correlation analysis to identify key restorative perception factors. We then conducted eye-tracking experiments in both real-world and VR scenarios to validate the findings of the SD questionnaire. In the second phase, orthogonal experiments systematically varied these factors to determine their impacts and mechanisms on attention restoration by analysis. Finally, we used data from these experiments to build predictive restorativeness models and employed genetic algorithms to optimize the design parameters for restorative campus pedestrian space.

We recruited subject students through a combination of online and offline methods; we released information on experimental recruitment through social media and campus bulletin boards. The questionnaire was disseminated to a total of 537 students (267 females) across 10 universities. Additionally, 200 university students aged 18–35 (mean age, 22.4 years), including 106 females, participated in the offline experiments, which encompassed both the eye-tracking and orthogonal experiments. These 200 participants could engage in one or two experiments, with at least 36 h between sessions to mitigate fatigue and practice effects. The specific number of participants for each part of the experiment will be introduced in the corresponding sections. Participation was voluntary, and all participants provided informed consent prior to participation. The Medical Ethics Committee of Harbin Institute of Technology (ethics number: HIT-2024003) approved the experimental protocol.

2.1.2. Restoration State Assessment Method

The restorative state was evaluated using the Restorative State Scale (RSS), composed of nine statements as

Table 1. Items in the restorative state scale.

Restorative State Scale
1. My mind is not invaded by stressful thoughts.
2. I can take time out from a busy life.
3. I can lose all sense of time.
4. I am thinking about everything and nothing at the same time.
5. I can make space to think about my problems.
6. I can leave all my problems behind me.
7. My mind just wanders in infinity.
8. I can imagine myself as part of the larger cyclical process of living.
9. I feel connected to the natural world.

[46], based on the Perceived Restorativeness Scale (PRS) [47]. PRS is widely used in studies of restorative environments, but because PRS can only be used to assess environmental settings (e.g., scene photographs), it is not suitable for studies that focus on changes in restorative states over time [46]. Instead, RSS settings focus on capturing the overall experience of nature and measuring different levels/functions of the restorative experience of nature [21,48]. In the present study, restorative state measurements were recorded during the SD questionnaire, at the orthogonal experiment baseline (S1), after the pressure source (S2), and after the simulated walk (S3). The restorative states were measured in different orders during the three varying measurement phases for each participant; thus, the 9-item RSS is more applicable to studies with multiple assessments of restorative states than other, more voluminous scales. Hence, the nature and volume of RSS are much more suitable for our research aim of examining the interrelation and temporal dynamics of restoration levels with the environment. Response options of RSS ranged from 1 to 5, as shown in

Table 2. Adjective pairs in SD questionnaire.

Spatial Factors	Factor of Neutrality	Adjective Pairs
Buildings and trees	Height	Short–Tall
Buildings	Height difference	Small–Large
Buildings and ground	Distance	Near–Far
Buildings	Angle	Small–Large
Buildings and ground	Color–Hue	Cold–Warm
Buildings and ground	Color-Contrast	Low–High
Shrubs	Area in vision	Little–Much
Ground	Texture	Sparse–Dense
Sky	Openness	Small–Large
Buildings and ground	Material	Artificial–Natural

, aligning with 5 adjectives from “Do not feel at all” to “Feel very strongly”. At the end of each experimental stage, participants assessed the environment based on their stress state using a 5-point scale. The average score of the nine items was considered the current evaluated restorative state.

2.2. Selection of Restorative Factors in Campus Pedestrian Spaces

2.2.1. SD Questionnaire

An SD questionnaire was employed to screen for spatial factors (space components and their attributes and characteristics) in the campus pedestrian space that significantly impact students’ restorative states. This questionnaire was designed based on the basic components of campus pedestrian spaces. It consisted of 23 items, including 9 questions from the Restorative State Scale (RSS) used to assess the respondents’ restorative states. Participants answered the questionnaire based on the pedestrian space they were in or the pedestrian space they most frequently visited.

To characterize the attributes of the spatial factors, 10 sets of adjectives were developed based on general visual perception and cognition, as detailed in Table 2. The remaining 16 questions combined these spatial factors and adjective groups to describe the restorative attributes of the campus pedestrian spaces. The questionnaire utilized a 5-point scale, with 1 representing very poor and 5 representing very good. The responses to the 9 RSS questions were averaged to quantify the participants’ restorative states, providing a clear measure of how different spatial factors influence restoration. The SD questionnaire is available in the Supplementary Materials. The questionnaire was distributed to 537 students (267 females) across 10 universities through field and online surveys, with 29% collected in the field. A total of 446 valid responses were received, yielding an 83.05% validity rate. The data collected includes students’ ratings of spatial perception attributes and their corresponding restorative states. Correlation analysis will be conducted to identify spatial elements strongly associated with restorative states.

2.2.2. Eye-Tracking Experiment

To validate the subjective data from the SD questionnaire, we conducted eye-tracking experiments in both real-world and VR scenes to obtain objective gaze characteristics for verification. Additionally, the feasibility of VR technology was confirmed by comparing eye-movement data between real-world and VR scenes, supporting subsequent experiments.

Eye-Tracking Experiment in Real Scene

To investigate the visual environmental factors of campus pedestrian spaces that concern college students, we selected five pedestrian spaces from Harbin Institute of Technology, Harbin Engineering University, and Harbin Normal University as research sites, as shown in

Table 3. Pedestrian spaces in investigation.

Case	1	2	3	4	5
Location	HIT 1	HIT	HEU 2	HNU 3	HNU
Type	Square	Street	Square	Path	Sidewalk
Feature	More artificial	More natural	More artificial	More natural	More natural
Picture of Vegetation period
Picture of Frost period

¹ Harbin Institute of Technology. ² Harbin Engineering University. ³ Harbin Engineering University.

. These sites were chosen based on the type of institution and the construction period of the buildings. Images of these spaces were captured during both the vegetation and frost periods to account for seasonal variations. However, the analysis focuses on the overall visual environmental factors affecting pedestrian spaces rather than seasonal differences.

In this study, eye-tracking data were recorded and analyzed using Tobii Pro Glasses 2 and Tobii Pro Lab (Tobii, Stockholm, Sweden), as shown in Figure 2b [49]. Eye movements were recorded in pixels using the top-left screen corner as the coordinate origin [50]. Raw data were manually mapped to scene snapshots in Tobii Pro Lab [51], with snapshots categorized into AOIs: ground (20%), buildings (30%), trees (15%), shrubs (10%), sky (15%), facilities (5%), and pedestrians (5%). Visual attention was assessed using total fixation (TFD) and glance (TGD) durations, percentages of fixation (PFD) and glance (PGD) durations, number of fixation points (NF), and number of visits (NV) [52]. And, for this study, 32 students participated, with 6–7 students recruited from each site. Participants were briefed on the tasks before the eye-tracking experiments, which were conducted outdoors between 9 a.m. and 5 p.m. The procedure and participants are shown in Figure 2a,c. After being equipped with calibrated eye-tracking devices, participants walked freely in designated campus pedestrian spaces for 2 min while their eye movements were recorded. To ensure data quality, the accuracy of the eye-tracker was evaluated, confirming its sufficient precision, as detailed in Appendix A.

VR Eye-Tracking Experiment in VR

To validate the VR scenarios, virtual environments were developed based on five pedestrian spaces, where eye-tracking experiments were conducted. Ensuring that participants’ gaze behavior in VR matched real-world contexts was crucial. Participants navigated these virtual landscapes by physically walking in a secure indoor space. The VR and real-world eye-tracking experiments were aligned, as shown in Figure 3a. For consistency, VR scenarios were based on both the vegetation and frost periods of the real pedestrian spaces, as shown in Figure 3c, maintaining visual proportions of elements like buildings, ground, sky, trees, shrubs, and amenities. For this part, 35 students participated, with 7 students recruited from each site. This study was the first to use HTC VIVE Pro Eye, a VR headset with integrated eye-tracking [53]. The Software Development Kit (SDK, version 1.7.1.1081) was integrated into Unity [54], with unfiltered data accessed via Tobii Pro SDK and filtered data processed using the Vive SRanipal SDK. The eye-tracking method aligns with that of Tobii Pro Glasses 2, mapping gaze data to panoramic images, as shown in Figure 3b.

2.3. Restorative Orthogonal Experiment

To thoroughly investigate the effects of the restorative spatial factors identified in the previous section on college students’ restorative states, as well as the underlying mechanisms, we employed an orthogonal experimental design. Orthogonal experiments are an efficient statistical technique that systematically analyzes the impact of multiple factors through a reduced number of trials by systematically arranging the experimental conditions. According to the generated orthogonal experimental scheme, the standard model was modified to generate dozens of pedestrian spaces. Subsequently, virtual walking and attentional recovery were performed in these VR pedestrian spaces.

2.3.1. VR Scenes Setting

To conduct subsequent orthogonal experiments, we first needed to design the necessary VR scenes and scene parameters. Using Rhino and the Unity 3D platform, we created a standard model of campus pedestrian spaces situated between driveways and campus buildings, reflecting typical campus environments with wide applicability.

Based on the restorative spatial factors identified in the previous section and the current situation of Chinese campuses, we determined the scene parameters for each restorative spatial factor obtained in the previous section, as detailed in

Table 4. Orthogonal experimental parameters.

Environmental Factors	Parameters in Orthogonal Experiment
Building height difference	0, 12, 24, 36, 48, and 60 m
Tree height	8, 12, 16, 20, 24, and 28 m
Shrub area	0%, 0.03%, 0.06%, 0.09%, 0.12%, and 0.15%
Ground texture	100 × 100, 200 × 100, 300 × 300, 600 × 300 mm2, 600 × 600, and 1000 × 1000 mm2
Ground hue	0°, 72°, 144°, 216°, 288°, and 360°

.

Shrub coverage in the visual field was calculated using Photoshop by measuring the pixel area occupied by shrubs, dividing it by the total pixel count, and multiplying by 100 to obtain the percentage of the view occupied by shrubs. Ground texture was manipulated by adjusting the size of paving tiles, incorporating six tile sizes based on common Chinese pedestrian walkway and plaza designs. This approach ensured a realistic and visually engaging simulation of pedestrian spaces, reflecting the complexity and diversity of urban landscapes. Ultimately, through the generated orthogonal experimental scheme, we established 49 experimental scenarios, as shown in Figure 4.

2.3.2. Experimental Procedure

In the experiment, as shown in Figure 5a, participants were briefed on the procedure and provided consent. A 2 min relaxation session followed to calm participants before commencing the experiment. Initial assessment of the participants’ restorative state was conducted using the RSS (S1). Participants were subjected to stress-inducing stimuli created using E-Studio, compiling visual and mathematical challenges to distract participants from orientation and reduce their restorative state. Visual challenge is a sequence of 15 images depicting violence, disasters, and extreme emotional states, chosen randomly from 50 images, as shown in Figure 5b. A total of 15 images displayed every 2 s, and 3 of these images were shown more than once; participants were to notify the experimenter immediately upon recognizing repeats. This method can make participants more focused on the content of the picture, so as to more effectively evoke stress by eliciting feelings of discomfort and anxiety. The mathematical challenge involved summing the number presented in a circle’s center with its predecessor and selecting the correct sum from options around the circle’s edge, all within a stringent 3 s timeframe (which was reduced to 2 s after 10 rounds). Correct answers scored points, whereas incorrect answers or timeouts were met with a noise deterrent. After stressor exposure, the participants’ restorative states were measured again (S2). Thereafter, participants embarked on a 2 min virtual walk through a simulated campus pedestrian space equipped with an HTC VIVE Pro Eye headset. After the virtual stroll, the RSS was administered again to evaluate the restorative state (S3). The duration of the experiment, designed to explore how environmental stimuli affect psychological restoration, was approximately 24 min.

A total of 180 university students participated in this experiment, with ages ranging from 18 to 35 years and an average age of 22.3 years. Among them, 106 were female. To prevent the subjects from becoming immune to the stress procedure, each subject could choose to participate in up to two orthogonal VR experimental scenarios. Additionally, to avoid VR-induced dizziness, fatigue, and practice effects, each session was limited to a maximum of 30 min, with a minimum interval of 24 h between sessions. In this part of the study, 227 experimental sessions were conducted, of which 200 yielded valid data, resulting in an effective rate of 88.1%.

2.3.3. Analysis of Influence Mechanism

To investigate the impact mechanisms of restorative spatial factors in campus pedestrian spaces on the restorative state (ΔS) of college students, we employed multivariate non-linear fitting. Using Origin, the dataset fitting extended beyond linear modeling by incorporating polynomial terms to capture potential non-linear interactions between variables. Additionally, a comparative analysis was conducted between linear and non-linear regression. The goodness-of-fit of the model to the data was evaluated using R². Its value ranges from 0 to 1, with an R² closer to 1 indicating a stronger explanatory power.

Multiple statistical analyses were employed to comprehensively analyze the data and address various research questions. Linear regression was used to identify overall trends and relationships between environmental factors and attention restoration. Additionally, non-linear models, such as polynomial regression, were applied to capture complex interactions and non-linear effects that could not be detected through linear models alone.

2.4. Training of Predictive Models

To develop the spatial restorativeness prediction model, this study evaluated and compared the capabilities of Decision Tree, KNN, ANN, and CNN in learning and extracting data features to select the surrogate model with higher accuracy of predication. The training phase utilized classification ML for predictions, with the RSS as the evaluation criterion for the restorative state, using a 5-point rating scale. In the ΔS range (S3 − S2), the training data were divided into five for training. The training dataset comprised 200 sets of selected environmental factors and their corresponding restorative capabilities, ΔS. The values of the environmental impact factors were used as the input dataset when training the surrogate model, with ΔS scores as the output.

2.4.1. Decision Tree (DT)

Decision Tree, consisting of nodes and directed edges, are conditional probability distributions defined on feature and decision spaces and are used for both classification and regression [55]. Formula $p\left(i|t\right)$ is the ratio of $i$ in node $t$, $i$ represents the proportion of $i$, and $c$ indicates the category.

$$E(t)=-{\displaystyle {\sum }_{i-1}^{c}p(i|t)}{\log }_{2}p(i|t)$$

(1)

2.4.2. k-Nearest Neighbor (KNN)

KNN is a non-parametric, instance-based learning algorithm used for classification and regression [56]. Inputs consist of the k-closest training examples in feature spaces. This study used KNN classification, where the object is assigned to the class most common among its k-nearest neighbors. The region adjacent to x covering k point is referred to as $N_k\left(x\right)$. Class y of x is determined in $N_k\left(x\right)$ according to the classification decision rule. $x_i$ is the eigenvector of the instance; $y_i$ is the class of the instance; $i=\mathrm{1,2},\cdots ,N$.

$$y=\arg \max {\displaystyle {\sum }_{x_i\in N_{k(x)}}I(y_i,c_j)}$$

(2)

$$i=1,2,\dots ,N;j=1,2,\dots ,K$$

(3)

where I is the indicator function; I = 1 when $y_i=c_j$. Otherwise, I = 0.

2.4.3. Artificial Neural Network (ANN)

ANN models contain an input layer, one or more hidden layers, and an output layer [57], where each Multilayer Perceptron (MLP) neuron is connected to all neurons in the next layer, as shown in Figure 6.

The MLP network of Input $u\left(k\right)$ and Target $h\left(k\right)$ in hidden layer 1 was used for restorative prediction of the campus pedestrian space with different parameters, as follows:

$$h(k)={\displaystyle {\int }_2(w_2}\times x(k)+b_2)$$

(4)

$$x(k)={\displaystyle {\int }_1(w_1}·u(k)+b_1)$$

(5)

where $x\left(k\right)$ is the output vector of the hidden layer; $w_1$ is the connection weight matrix from the input layer to the hidden layer; $w_2$ is the connection weight matrix from the hidden layer to the output layer; and $b_1$ and $b_2$ are the numbers of deviations in the hidden and output layers, respectively [58]. The following transfer algorithm was used between the hidden and output layers:

$${\displaystyle \int (P)}={\displaystyle \frac{1-e^{-2P}}{1+e^{-2P}}}$$

(6)

$$P={\displaystyle \int ({\displaystyle \sum w_i·x_i})}$$

(7)

During dataset building, the dataset containing five spatial factors (HDB, HT, AS, TG, and HG) as the input, ΔS (S3 − S2) was the output, and 200 sets of experimental datasets were gradually fed into the ANN model. During ANN training, the data were usually divided into training and validation datasets. Regarding the results, 70% and 15% of the datasets were used for training and validation, respectively; the rest were used for testing.

2.4.4. Convolutional Neural Network (CNN)

The image dataset consisted of 200 scene screenshots, with the location, angle, and size of each screenshot maintained consistently. After preprocessing, the input image size for the CNN was set to a 224 × 224 pixel RGB three-channel image (3, 224, and 224), with the corresponding 200 pedestrian space restorative capabilities ΔS as the CNN’s output. The dataset was divided into 70% training, 15% validation, and 15% testing. The CNN architecture used in this study included an input layer, convolutional, pooling, and fully connected layers, and an output layer, as shown in Figure 7.

Once processed by the convolutional layer, the input images, as detailed in Equation (8), underwent activation through the ReLU function in Equation (9), followed by max pooling, as shown in Equation (10). Throughout this phase, the configuration maintained a 3 × 3 filter size and strides of 1 and 3 for the convolutional and pooling layers, respectively. Following multiple iterations of convolution, activation, and pooling operations, the resultant feature maps (32, 7, and 7) were flattened. Subsequently, the architecture integrated a fully connected layer (fc1) with a dimensionality of 512, followed by a dropout layer with a 0.09 dropout probability, aimed at reducing overfitting, while the ReLU activation function was utilized to introduce non-linearity into the model. The data then advanced through another fully connected layer (fc2) with a dimensionality of 256, where the ReLU activation function was applied again, leading to the final predictive output through a fully connected layer (fc3) with a dimensionality of 1. In the concluding stages, the fully connected layers further refined the data processing, employing the ReLU activation function to ensure non-linearity, culminating in the output of predictive results for the classification task through the dimensionally singular fully connected layer (fc3).

$$(I\times K)(i,j)={\displaystyle {\sum }_m{\displaystyle {\sum }_{n}I(i+m,j+n)}}·K(m,n)$$

(8)

$$f(x)=\max (0,x)$$

(9)

$$f(R)={\max }_{x\in R}x$$

(10)

where $(i,j)$ denote the position within the output feature map, and m and n represent the dimensions of the convolutional kernel.

2.5. Settings of Restorative Optimization

The Genetic Algorithm (GA) is a sophisticated optimization technique that mimics evolution observed in nature [59]. Inspired by natural selection and genetic inheritance in biological evolution, it forms part of the broader family of Evolutionary Algorithms [60]. Within the GA framework, the fitness level of each individual in a population is determined using a specific fitness (objective) function. This selection favors higher-quality solutions by making the selection probability directly proportional to the individual’s fitness level. To prevent the algorithm from becoming stuck in local optima, a mechanism occasionally selects fewer fit solutions, thereby broadening the algorithm’s capacity for global exploration.

GA optimization was performed using an integrated MATLAB environment. The predicted restorative state (ΔS) was the GA fitness index, pinpointing the most conducive environmental parameter range for optimizing campus pedestrian spaces. Following existing guidelines, GA parameters were defined as having a population size, crossover rate, and migration rate of 200, 0.8, and 0.2, respectively. With inputs of 200 distinct HDB, HT, AS, TG, and HG sets, 200 respective ΔS values were evaluated to ascertain the peak value in each generation, until a satisfactory convergence level or the maximum generation limit was reached.

3. Results

3.1. Selection of Influencing Factors in Campus Pedestrian Spaces

3.1.1. Restorative Factors Correlation Analysis

The mean of the nine questions from each questionnaire was determined to assess restorative states ($\overline{\mathrm{S}}$); Spearman’s rank correlation was utilized to examine relationships between the 12 factors within campus pedestrian spaces and the restorative state, as shown in Figure 8. The R value was used to determine correlation strength. Among the studied environmental factors, significant correlations were observed for |R_HDB| = 0.31 > 0.20, |R_HT| = 0.33 > 0.20, |R_AS| = 0.27 > 0.20, |R_TG| = 0.32 > 0.20, and |R_HG| = 0.28 > 0.20. Consequently, the building height difference (HDB), tree height (HT), shrub area (AS), ground hue (HG), and ground texture (TG) exhibited strong correlations with the restorative state, suggesting that optimizing the design parameters associated with these factors in campus pedestrian spaces may enhance environmental restoration.

3.1.2. Eye-Tracking Experiment Analysis

We analyzed data from 32 eye-tracking experiments of real scenes and hotspot distribution maps, as shown in Figure 9. Metrics such as total fixation duration (TFD), total glance duration (TGD), percentages of fixation (PFD) and glance (PGD) durations, number of fixation points (NF), and number of visits (NV) were evaluated. Ground surfaces attracted the greatest attention, followed by buildings, trees, shrubs, the sky, pedestrians, and facilities, with significantly greater attention given to ground, buildings, and vegetation compared to the sky, pedestrians, and facilities. These findings align with the spatial component results from the questionnaire, confirming the validity of the screening process.

Hotspot distribution maps of the VR eye-tracking experiment from the 35 sets of valid experimental data are shown in Figure 10, which also illustrates the NF and NV results. A visual attention hierarchy existed across environmental factors, with buildings garnering the highest attention, followed by ground surfaces, trees, shrubs, pedestrians, the sky, and facilities. The TFD, TGD, PFD, and PGD showed similar trends, with buildings being the most salient, followed by ground surfaces, trees, shrubs, pedestrians, the sky, and facilities, albeit with varying degrees of significance.

These results validated that the five restorative spatial factors identified through the SD questionnaire—HDB, HT, AS, TG, and HG—are the primary elements in pedestrian spaces that attract students’ attention and influence their restorative state. The eye movement characteristics in the VR experiment are similar to those in the real world, as seen by examining the gaze characteristics in Figure 9 and Figure 10, indicating that the VR experiment is sufficiently credible. However, there was a significant difference in the level of attention to the ground between the two eye-tracking experiments. This may be because, in the real-world scenario, participants needed to focus on the ground to identify potential safety hazards. In contrast, the VR experiment provided a spacious and secure environment, reducing the need for participants to pay as much attention to the ground.

3.2. Restorative Factors Influence Mechanism

The correlations between environmental factors such as HDB, HT, AS, TG, and HG and the restorative state (ΔS) were examined, as shown in Figure 11. The coefficient of determination (R²) for the quadratic fitting function between HDB and ΔS (R² = 0.283) surpassed that of the linear fitting function (R² = 0.147). Notably, HDB and ΔS correlated negatively (i.e., as HDB increased, ΔS decreased). However, beyond a critical point (ΔS = 3), ΔS increased with increasing HDB. The HDB range conducive to restoration (ΔS = 5) was 3.42–15 m. Similarly, the R² value for the quadratic fitting function between HT and ΔS (R² = 0.225) exceeded that of the linear fitting function (R² = 0.114). HT and ΔS correlated positively, with ΔS increasing and then decreasing with increasing HT. The optimal HT range for achieving a maximum ΔS (ΔS = 3) was 16–20.9 m. Regarding AS and ΔS, the quadratic fitting function (R² = 0.213) exhibited a higher R² than the linear fitting function (R² = 0.117). The relationship between AS and ΔS first increased and then decreased with increasing AS. The peak ΔS value (ΔS = 5) was within the AS range of 0.07–0.12%. For TG and ΔS, the quadratic fitting function (R² = 0.225) demonstrated superior performance compared to the linear fitting function (R² = 0.146). TG and ΔS correlated positively, with ΔS decreasing as TG increased. The maximum ΔS value (ΔS = 5) occurred within the TG range of 0.31–0.62 m². The quadratic fitting function between HG and ΔS (R² = 0.160) exhibited a higher R² than the linear fitting function (R² = 0.026). The relationship between HG and ΔS first increased and then decreased with increasing HG. The optimal HG range for achieving a maximum ΔS (ΔS = 3) was 136–224°.

3.3. Prediction of Restorativeness

3.3.1. Machine Learning

This section presents a comparative analysis of various classification algorithm models. Utilizing the MATLAB Classification Learner toolbox, Decision Tree (DT), k-Nearest Neighbor (KNN), and Artificial Neural Network (ANN) were deployed to train the prediction model. This model takes environmental factors as inputs to predict the restorative state, denoted as ΔS. A crucial tool for assessing the performance of these classification models, the confusion matrix, is shown in Figure 12, which presents the accuracy confusion matrices for each algorithm in predicting ΔS, thereby comprehensively summarizing their average prediction accuracies. The ANN model utilized a two-layer neural network with 10 neurons per layer, employing the ReLU activation function. Remarkably, the accuracy rate achieved by the ANN was 94.5%, exceeding the other classification results. Conversely, the Decision Tree exhibited a relatively lower accuracy rate (78.0%), whereas the KNN performed similarly, albeit inferior, to the ANN, with a 90.5% accuracy rate.

3.3.2. Deep Learning

Environmental factors were utilized as labels to train a predictive model using a Convolutional Neural Network (CNN) for the restorativeness of campus pedestrian spaces. The model’s efficacy was validated using a dedicated test set derived from the study’s dataset. The model convergence during training is depicted in accuracy curves for both the training and validation sets. The model’s accuracy improved consistently over 200 epochs, achieving a peak of 97.2%, as shown in Figure 13a. Concurrently, the error rate stabilized at a low level toward the end of training iterations.

To assess the predictive performance of the model, we examined confusion matrices, as shown in Figure 13b. The confusion matrices delineated the predictive outcomes across different categories along with their respective proportions within the dataset. The high precision and recall rates across classes indicated the model’s robust classification capability. Specifically, the model demonstrated formidable proficiency in forecasting the restorativeness of campus pedestrian spaces, with an overall accuracy of 95.2%.

3.4. Optimization of Restorativeness

To optimize the restorativeness of campus pedestrian spaces, denoted as ΔS, this study constructed 49 VR scenarios based on different environmental parameters, including HDB, HT, AS, TG, and HG. For each scenario, we focused on its impact on the restoration state, ΔS, which served as the output metric for the Genetic Algorithm (GA). The best fitness value rapidly decreased from an initially high value and then stabilized, as shown in Figure 14. This indicated that, after the algorithm found a preliminarily suitable solution, it gradually refined the search to enhance the solution quality. The change in mean fitness was relatively smooth, demonstrating that the algorithm maintained diversity within the population.

Through GA optimization of the datasets, we successfully obtained the values of environmental factors that most significantly impacted the restoration state. We summarized and analyzed the values of these key environmental factors, and the results are presented in

Table 5. Optimization thresholds of environmental factors.

	Solution Set 1	Solution Set 2	…	Solution Set 20	Thresholds
HDB (m)	27.543	20.819	…	26.382	20–28
HT (m)	8.813	15. 368	…	12.792	8.5~15.5
AS (%)	0.147	0.129	…	0.135	0.12~0.15
TG (m2)	0.347	0.496	…	0.265	0.20~0.50
HG (°)	236.597	245.013	…	146.823	140~245
ΔS (S3 − S2)	4 (4)	4 (3.9)	…	4 (4)	4 (predict)

. According to the outcomes of the GA, in the restorative design of campus pedestrian spaces, the recommended ranges for HDB, HT, AS, TG, and HG were 20–28 m, 8.5–15.5 m, 12–15%, 0.20–0.50 m², and 140–245°, respectively.

4. Discussion

4.1. Restorative Environmental Factors in Campus Pedestrian Spaces

The spatial factors that significantly affect the recovery of students’ attention during walks on campus are the building height difference (HDB), tree height (HT), shrub area (AS), ground hue (HG), and ground texture (TG), identified using an eye-tracking experiment and a questionnaire. Specifically, the results of our experiment showed that natural elements such as trees and shrubs attracted significant attention from students on campus, consistent with the strong restorative perception associated with natural environments [46]. At the same time, buildings and grounds also had a very significant effect on the state of restoration, which is further evidence of the restorative potential of urban elements as well as their positive impact on health [24]. This is also consistent with the findings of similar studies on the critical role of these elements in urban and campus environments [61]. Moreover, the strong link between spatial factors and students’ state of recovery evidenced by the study supports the Attention Recovery Theory (ART) [62,63].

However, some differences emerged between the eye-tracking results from real and virtual environments. In the virtual setting, participants showed slightly longer fixation durations and fewer saccades than in the real world. This may be due to the VR scenes’ realism and fidelity, which influence how participants perceive and interact with the environment [54]. The controlled and possibly less stimulating VR environment could result in more focused and prolonged fixations, with fewer distractions than the real-world setting [64]. Future studies could use VR to simulate diverse auditory and thermal conditions, offering deeper insights into how various environmental factors impact students’ restorative states [65].

4.2. Restorative Influencing Mechanism

Our second scientific goal was to investigate the underlying mechanisms by which these spatial factors influence the restorative state. The fitting results showed that HDB, HT, AS, TG, and HG all have unique impacts on attention restoration, which can be described by quadratic relationships rather than simple linear ones. These results suggest that moderate variations in appropriate ground hues and textures can significantly enhance visual appeal and comfort, contributing to better attention restoration [6,25]. And the building height difference indicated that moderate height variations create a visually stimulating yet not overwhelming environment. Similarly, the obtained range of tree heights was most beneficial for attention restoration, likely due to their ability to provide a sense of enclosure and connection with nature without obstructing views [66,67]. Additionally, a moderate increase in the presence of shrubs within the visual field benefits attention restoration. It is clear that an excessive number of shrubs may convey a sense of wildness and neglect of care, leading to unease among students and, consequently, diminishing attention restoration [46,68].

However, reliance on self-reported measures introduces subjectivity and individual variance. To address this, future research could employ wearable devices that measure physiological indicators such as heart rate variability, electrodermal activity, electromyography, body temperature, and breathing rate, facilitating an objective evaluation of stress states [66,69].

4.3. Machine Learning and Deep Learning for Restorativeness Prediction

This study leverages advanced machine learning and deep learning techniques to predict the restorativeness of campus pedestrian spaces. The Convolutional Neural Network (CNN) provided the highest accuracy; this finding was aligned with previous research on the application of machine learning, which highlighted the capability of CNNs to handle high-dimensional data and capture intricate patterns in environmental variables [70]. It also further proves the effectiveness of CNN in predicting environmental influences on psychological outcomes [71].

Future research could focus on refining the deep learning models to improve their interpretability and applicability to environmental studies [61]. Also, we could explore other advanced techniques, like reinforcement learning and generative adversarial networks (GAN), to open new avenues for understanding and predicting restorative environments [72].

4.4. Implications for Restorative Design in University Campuses

By employing CNN and Genetic Algorithms (GA), this study identified restorative parameter thresholds for campus pedestrian spaces that favor attention recovery for college students. The recommended ranges for HDB, HT, AS, TG, and HG provide concrete guidelines for designing restorative campus environments. These findings are consistent with research that highlights the importance of specific natural and architectural features in promoting restorative experiences [25]. Through the design of the built and natural elements present in the construction of a restorative campus, pedestrian spaces can be effective in reducing mental fatigue and enhancing cognitive function [48,73,74]. Our study adds to this body of knowledge by providing specific, quantifiable guidelines that can be directly applied to campus design.

Future research should continue to refine these design strategies and explore their long-term impacts on student well-being. Incorporating additional factors such as seasonal changes, auditory stimuli, and thermal comfort could provide a more comprehensive understanding of restorative environments [19].

5. Conclusions

5.1. Restorative Environmental Factors Selection

We utilized SD questionnaires to identify key restorative spatial factors in campus walkways. Spearman’s rank correlation was employed to analyze the questionnaire data, successfully isolating five spatial factors that showed a strong correlation with university students’ restorative states: building height difference (HDB), tree height (HT), shrub area within the field of view (AS), ground hue (HG), and ground texture (TG).

To verify the reliability of the questionnaire results, we conducted eye-tracking experiments in both real and VR environments. By tracking eye movements, we assessed the extent to which different spatial elements captured attention. The findings revealed that students predominantly focused on the ground and surrounding buildings during their walking activities, with natural elements such as trees and shrubs identified as secondary points of interest. These results not only validated the five restorative spatial factors identified by the SD questionnaire but also confirmed the reliability of the VR experiment.

5.2. Interrelation of Restorativeness and Environmental Factors

The research is based on 49 virtual reality (VR) campus pedestrian space restorative orthogonal experiments of 5 restorative pedestrian space factors that impact the state of the recovery of college students’ study. The experimental data reveal the influencing mechanism of non-linear fitting. In the analysis of campus pedestrian spaces, it was identified that HDB was negatively correlated with the restorative state (ΔS), suggesting that, as the height differential among buildings increases, the restorative potential of these spaces diminishes. Interestingly, beyond the minimum ΔS value, an increase in HDB led to some improvement in the environment’s restorative capacity. Positive correlations were established between ΔS and HT, AS, HG, and TG. Specifically, as HT increased, ΔS first increased and then decreased. This pattern was mirrored in the trends observed with increases in AS and TG, where ΔS first increased and then decreased. Similarly, an increase in HG resulted in an initial upsurge in ΔS, which then tapered off.

5.3. Restorative Prediction of Campus Pedestrian Space

A comparative analysis of various machine learning models—namely, Decision Tree (DT), k-Nearest Neighbors (KNN), and Artificial Neural Networks (ANNs), alongside a deep learning model, a Convolutional Neural Network (CNN)—revealed that the CNN exhibited superior performance. Specifically, the CNN achieved a classification prediction accuracy of 95.2%, surpassing the ANN (94.5%), KNN (90.5%), and Decision Tree (78.0%). This outcome underscores the exceptional efficacy of CNNs in predicting the restorative capabilities of campus pedestrian spaces, highlighting their potential to accurately capture and model the complex interactions between multiple spatial factors and their impact on attention restoration.

5.4. Restorative Optimization of Campus Pedestrian Space

Furthermore, the results from the Genetic Algorithm (GA) optimization for the VR motion visual environment provide valuable insights into the threshold values for key environmental factors. The GA findings recommend specific ranges for the restorative design of campus pedestrian spaces: the HDB should be within 20–28 m, the HT should range from 8.5 to 15.5 m, the AS should be between 12 and 15%, the TG should fall within 0.20–0.50 m², and the HG should be 140–245°. These recommendations are instrumental in enhancing the restorative quality of campus pedestrian environments, providing a scientifically grounded basis for the design and optimization of spaces that promote student well-being and attention restoration.