Introduction

University students globally experience higher levels of psychological pressure in an unequal manner than their non-university counterparts based on academic workload, financial constraints, hyper-competitive conditions, and limited time resources1,2,3,4. Additional challenges, such as illness, social discrimination, and emotional fatigue, intensify this burden5, turning student mental health into a global concern6. WHO reports and large-scale studies across Europe, Asia, and North America estimate that 25–60% of university students experience psychological distress5,7,8,9,10,11, which could reduce academic performance and overall well-being8.

This growing mental tension has led researchers to investigate the restorative potential of physical environments, with particular attention to outdoor campus spaces. Informal open areas—including green pockets, shaded niches, and courtyards—provide students with critical opportunities for relaxation and spontaneous social interaction, which have been empirically shown to foster psychological recovery12,13. Furthermore, such social engagements demonstrably enhance mental well-being, academic success, and socio-emotional development14,15. Realizing these benefits necessitates a campus design approach that extends beyond conventional built structures and emphasizes outdoor environments enriched with shaded seating, spatial diversity, and opportunities for informal social engagement16. A theoretical extension of these findings is provided by Thwaites et al. (2007), who introduced the Socially Restorative Urbanism (SRU) framework to emphasize the role of spatial components—such as “Center,” “Direction,” “Transition,” and “Area”—in shaping public environments that foster social interaction and psychological restoration17. The SRU framework directs our investigation toward interactive rest spots: semi-structured outdoor campus spaces engineered to concurrently enable relaxation and informal social encounters.

SRU theory intersects with two foundational theories in environmental psychology: Attention Restoration Theory (ART) and Stress Reduction Theory (SRT). ART claims that natural and visually engaging environments restore depleted attentional capacity through mechanisms such as soft fascination and reflective thought18. SRT highlights preconscious emotional reactions to nature, activating parasympathetic recovery mechanisms19. Both theories state that restorative environments engage the senses and emotions, promoting psychological well-being. Consequently, environments that provide activities, resources, and spaces for stress and fatigue relief are perceived as beautiful20. Aesthetic preferences—individuals’ subjective evaluations of features such as vegetation, materials, color, texture, and spatial layout—are pivotal in shaping restorative experiences. Berghman et al. (2017) conceptualize aesthetic experience as a two-stage process: it begins with perceptual and emotional reactions and proceeds to a cognitive appraisal of pleasure or satisfaction21. These reactions are shaped by unconscious processes rooted in evolutionary adaptation, rather than being entirely subjective or random20.

University planning persistently prioritizes built structures over outdoor environments, neglecting experiential qualities of open campus spaces despite growing recognition of their psychological and aesthetic value22. Consequently, many outdoor areas remain underutilized due to misalignment with student needs and visual preferences23. Comprehensive empirical evidence linking specific environmental features to aesthetic preferences in interactive rest spots remains limited13, despite partial identification of individual drivers like plant diversity or spatial openness. A growing body of research in architecture and environmental design highlights the relevance of students’ visual preferences in shaping campus spaces24,25,26, yet this field has largely relied on narrow variables27, qualitative methods28,29, or traditional statistical models25,30.

Recent advances in machine learning (ML) provide powerful tools for modeling complex, subjective phenomena such as aesthetic preference. Algorithms like Artificial Neural Networks (ANNs), Random Forest, Convolutional Neural Networks (CNNs), and Bayesian networks have been applied to predict landscape aesthetics in parks31,32, forests33,34, and urban streetscapes35. Jahani (2019) applied ANNs for landscape beauty assessment in the forest36, and Havinga et al. (2021) combined ResNet-50 with Random Forest to study public scenic perceptions via crowdsourced imagery37. Similarly, Kerebel et al. (2019) adopted a Bayesian approach to identify key visual features influencing aesthetic judgments38, and Jahani et al. (2022) implemented a Multilayer Perceptron (MLP) to quantify the impact of variables such as land slope, vegetation, and built structures on perceived beauty in urban park environments39.

A significant methodological disparity persists between the proliferation of machine learning in general landscape aesthetics research and its limited adoption for campus aesthetic preferences40, despite intensifying scholarly attention to campus spatial qualities41,42. Soydaner and Wagemans45, for example, utilized Explainable Artificial Intelligence (XAI) to investigate the visual drivers of aesthetic perception in campus environments, particularly through the SHAP (Shapley Additive Explanations) method. Similarly, Wen et al.42 applied Fully Convolutional Networks to identify restorative visual variables in outdoor campus spaces. These pioneering studies underscore the potential of ML-based approaches in academic environments, while also highlighting the need for further research in this emerging field.

Ensemble methods remain underutilized in machine learning studies of landscape aesthetics, particularly for modeling campus environment preferences—a gap this study addresses through a novel modeling framework. This research offers three key contributions. First, it centers on students’ aesthetic preferences for interactive rest spots, a spatial typology largely overlooked in environmental design literature; second, it adopts an average-based ensemble of SVR, MLP, and RF models—optimized for small, image-based datasets and more interpretable than black-box methods like XGBoost, thereby enabling theory-informed sensitivity analysis; third, it links predictive modeling with environmental psychology by interpreting feature importance through the lenses of ART, SRT, and SRU. This study contributes novel interdisciplinary insights spanning environmental psychology, campus design, and machine learning, while explicitly accounting for Tehran’s distinct academic-urban context—marked by elevated stress levels and spatial density11,43,44—and documented cultural influences on aesthetic perception45. Figure 1 illustrates the end-to-end framework, operationalizing our approach from dataset curation to modeling.

Fig. 1
figure 1

Research framework including data preparation as the input to the machine learning models, model develpment, and model outputs.

Full size image

This paper aims to address the following research questions: (1) What is the predictive accuracy of Machine Learning (ML) models in assessing students’ visual aesthetic preferences for interactive restorative spaces within university campuses? (2) What is the extent to which Ensemble Learning Models can enhance the precision of predictive outcomes relative to standalone models? (3) What are the key environmental components that significantly influence students’ aesthetic preferences for interactive restorative areas?

Research steps and method

Sampling, photographic spots, and questionnaire administration

Tehran, the academic hub of Iran and a rapidly urbanizing metropolis, hosts 462 educational and research institutions46, including 44 universities operating across 72 independent campuses. A geospatial dataset was developed by mapping all 72 campuses using Google Earth Pro (version 7.3), covering a total area of 6.6 km2, with an average campus size of nine hectares. These measurements were verified against municipal land-use plans. The mapped dataset produced an initial shortlist of 16 universities whose campus area exceeded the average size. Four universities—Amirkabir University of Technology, University of Tehran, Shahid Rajaee Teacher Training University, and Tarbiat Modares University—were selected based on spatial diversity and sustainability criteria, particularly a green space-to-built area ratio ≥ 20%, which aligns with UN-Habitat’s 2018 guidelines for sustainable public spaces47 as detailed in Table 1.

Table 1 Summary of selected university campuses and inclusion criteria.
Full size table

All values represent approximate measurements. Gray spaces include impervious paved areas such as parking lots and non-vegetated courtyards. Green & blue spaces refer to areas with vegetation or water features.

Student-preferred rest spots were determined through a sequential mixed-methods approach combining spatial analysis and on-site surveys. Originally, campus maps were created using AutoCAD 2024 and purified in Adobe Photoshop 2023, with key milestones and adjacent photo stimuli embedded to improve orientation. These illustrated maps served as the primary tool for a map-based, in-person survey performed across the four selected universities through random sampling, following the protocol established by Asim et al.48. A total of 443 students from the four universities revealed their preferred group resting spots until data saturation was achieved (i.e., no new locations were identified). Field observations were then done to verify all identified locations.

Photographic documentation was conducted during early December (autumn), between 9:00 and 11:00 a.m., under constant natural lighting. Photos were captured using a smartphone camera, adhering to a standardized protocol: each photo was taken from a fixed eye-level perspective (1.60 m) with a 50 mm focal length to replicate students’ natural viewing angles and minimize distortion. A single high-resolution photograph (4608 × 3456 pixels) was taken for each of the 125 identified spots (Fig. 2), with compositions including approximately 50% ground surface to ensure visual consistency. Human figures were deliberately removed from the images with the purpose of eliminating socio-cultural distractions while drawing attention to environmental features—a process used in prior investigations (e.g., Jahani and Saffariha31). All images underwent post-processing in Photoshop 2023 to enhance contrast, decrease visual noise (e.g., litter, human figures), and highlight key landscape elements using the Patch Tool and Content-Aware Fill. The selection included 100 images that met the criteria of visual transparency, accessibility, and consistency with SRU spatial principles. Appendix A (Figures A1-a to A1-i) presents representative examples.

Fig. 2
figure 2

Student-identified interactive rest spots on campus maps of four selected universities in Tehran city. Maps were redrawn by F.S.K. based on official campus plans and publicly available reference imagery from Google Earth Pro (v7.3; https://www.google.com/earth/versions/#earth-pro), with annotations created using Adobe Photoshop 2023.

Full size image

These 100 finalized images served two main purposes: (1) as visual stimuli in an online photo-based survey adapted from49 and structured according to Kaplan’s visual preference framework18,50; and (2) as the basis for environmental feature extraction and quantitative analysis. Participants were recruited from a diverse range of Iranian universities, extending beyond the four selected campuses, to reduce spatial familiarity bias. Students rated each image on a 7-point semantic differential scale, ranging from 1 (“This landscape is not beautiful at all”) to 7 (“This landscape is very beautiful”), following the methodology described by Sarmad et al.51. The survey was disseminated in spring 2024 through student online communities. Fourteen responses with uniform ratings were excluded to ensure data quality.

The final dataset comprised 394 valid responses, exceeding the minimum required sample size of 385 calculated using Cochran’s formula (95% confidence level, p = 0.5, margin of error = 0.05) (Appendix B). This perceptual evaluation approach aligns with environmental psychology theories such as ART and SRT, which emphasize the relationship between aesthetic experience and psychological restoration. A sample questionnaire item is provided in Appendix C.

Extracting features from images

Aesthetic preferences were assessed alongside the extraction of eighteen environmental variables from 100 selected images (Fig. 3), each previously associated with psychological restoration. These variables spanned architectural features, landscape elements, and facilities, including tree cover31,52,53,54, soft landscapes55, waterscapes31,54, color diversity56, buildings31,48, pathways57, skyscapes31,48, flowers31,56, natural stone58, plant cover diversity56, environmental amenities59, lawns60, overhead shading systems61, sculptures62, tree density63, vertical greenery and shrubs64, landscape complexity65, and seating facilities54.

Fig. 3
figure 3

Feature extraction at interactive rest spot 62 (Fig. 2c), Shahid Rajaee University. Photo by first author (F.S.K.).

Full size image

Fourteen of these variables were quantitatively measured using AutoCAD and Excel by calculating their area coverage within each image. “Pathways” included all paved surfaces, staircases, and walkways. “Environmental amenities” referred to features such as waste bins, CCTV cameras, lighting, recreational equipment, and signage. “Seating facilities” comprised benches, gazebos, and informal seating elements such as ledges or planter edges. Color diversity was rated using Itten’s 12-color wheel. Tree density and plant cover diversity were visually evaluated by a panel of five environmental design experts. Tree density was categorized into four levels—none, low, medium, and high—based on the number of visible trees and the estimated spatial extent of planting beds in each image. Plant cover diversity was classified into five vegetation types: lawn, trees, shrubs, flowers, and herbaceous plants.

Landscape complexity was assessed using a structured 6-point bipolar scale developed for this study, drawing conceptually from Twedt et al.65, who emphasized perceptual complexity as a key component of environmental preference and restoration. A separate panel of 20 landscape perception experts rated each image across six anchored categories representing varying levels of visual order and complexity: “most simple,” “moderately simple,” “least simple,” “least complex,” “moderately complex,” and “most complex.” This approach enabled a multidimensional evaluation of spatial legibility, visual richness, and environmental coherence.

All assessed variables were theoretically grounded in Attention Restoration Theory (ART), Stress Reduction Theory (SRT), and Socially Restorative Urbanism (SRU), and adapted to the contextual characteristics of small-scale interactive campus rest spots. The extracted variables were subsequently used as predictive features in machine learning models to estimate students’ aesthetic preferences.

Environmental modeling

The systematic assessment of landscape aesthetic quality has increasingly integrated machine learning (ML) and computer vision techniques, offering strong options to traditional statistical models for capturing complex and subjective environmental evaluations32,57,66,67. A various set of models—including Random Forest (RF), XGBoost41, Artificial Neural Networks (ANNs)36, Multilayer Perceptrons (MLP)31,39,41,52,68, Support Vector Regression (SVR)31,41, Radial Basis Function Neural Networks (RBFNNs)31,52,53, and explainable AI (XAI) techniques41—have been used in aesthetic prediction tasks.

Research on designing interactive rest spots on university campuses remains limited, despite the growing application of machine learning and computer vision techniques in landscape aesthetic assessment. This study addresses this gap through a multi-model approach based on Support Vector Regression (SVR), Random Forest (RF), Multilayer Perceptron (MLP), and two Ensemble Learning Models. These models analyze 18 environmental variables extracted from 100 images of interactive rest spots located within four Iranian public universities. The aim is to identify the visual variables that significantly contribute to aesthetic-based mental restoration and assess their relative importance. The following sections will analyze the data preprocessing process, details of each model, and its underlying structure.

a. Data preprocessing

The dataset was randomly split into training (80%) and testing (20%) subsets, following the protocol used by Jahani et al.67. Feature standardization was conducted using Python’s StandardScaler(), with the scaler fitted exclusively on the training set and then applied to the test set to prevent data leakage (Eq. 1). After prediction, all model outputs were inverse-transformed to the original aesthetic preference scale (1 to 7) to ensure meaningful interpretation of the results.

$${\text{Standard value}} = \frac{x - \mu }{\sigma }$$
(1)

where \(x\) is the original feature value, \(\mu\) is the mean, and \(\sigma\) is the standard deviation calculated from the training data.

b. Support vector regression (SVR) model

Support Vector Regression (SVR) was implemented as one of the core predictive models in this study due to its strong ability to handle nonlinear relationships and its robustness in small to medium-sized datasets. SVR aims to identify a regression function that approximates the underlying data structure within a specified margin of tolerance (ε), while minimizing prediction errors and model complexity69,70. This is achieved by optimizing a hyperplane that fits the data, defined by a subset of critical data points known as support vectors. The model’s objective is to ensure that most data points lie within the ε-insensitive tube, while penalizing deviations beyond this margin71. The detailed mathematical formulation of SVR, including the optimization problem and constraints, is provided in Appendix D to maintain clarity and focus in the main text.

The performance of the SVR model is highly sensitive to hyperparameter tuning. Key parameters include the Regularization parameter (C), which balances model complexity and training error; the Gamma parameter (γ), which defines the influence of individual data points on the decision function; and Epsilon (ε), which sets the threshold within which prediction errors are ignored. The Radial Basis Function (RBF) kernel was ultimately selected due to its superior empirical performance in modeling complex visual patterns. Preliminary testing also involved other kernel types, such as linear, polynomial, and sigmoid. The mathematical formulations of these kernels appear in Appendix E (Equations E1–E4).

c. Multilayer perceptron (MLP) model

Multilayer Perceptron (MLP) is a widely used kind of feedforward artificial neural network inspired by the parallel processing mechanisms of the human brain. A subset of Artificial Neural Networks (ANNs), MLPs are especially influential in capturing intricate nonlinear relationships between environmental features and aesthetic preferences66,67. Structurally, an MLP comprises an input layer, one or more hidden layers, and an output layer. Each neuron computes a weighted sum of its inputs, adds a bias term, and passes the result through an activation function. The model was trained using the backpropagation algorithm, which updates weights and biases iteratively by minimizing a loss function based on prediction errors72,73.

Selecting an appropriate activation function plays a key role in improving the performance of neural networks, as these functions affect the convergence of the model, the learning speed, and the prevention of problems such as vanishing gradients72. The logarithmic sigmoid function was selected for both hidden layers of the MLP architecture in this study, as it facilitated the learning of nonlinear relationships in the normalized input data74. Other activation functions such as ReLU and Tanh, although commonly employed in neural network models, were excluded in the final configuration due to the favorable convergence behavior and predictive accuracy achieved with the sigmoid–linear combination. The mathematical formulations of network operations, learning rules, and activation functions are presented in Appendix F (Equations F1–F7).

d. Random forest model (RF)

Random Forest (RF) is a supervised ensemble learning algorithm commonly used in classification and regression tasks. The method builds multiple decision trees using randomly sampled subsets of data and features through bootstrap aggregation, which improves model stability and predictive accuracy compared to individual decision trees68,75,76. The algorithm selects a random subset of features at each decision node and calculates the final output by averaging the predictions across all trees in regression problems. This mechanism reduces overfitting and enables the model to capture complex, nonlinear relationships within high-dimensional data. Two hyperparameters—the number of trees and the number of features considered at each split—significantly affect the model’s performance. RF regression demonstrated robustness and strong predictive capabilities in handling structured environmental variables and was therefore selected to estimate students’ visual aesthetic preferences for interactive campus rest spots in this study77,78. Additional technical details and the mathematical formulation of the RF model are provided in Appendix G.

e. Ensemble Learning Models: MLP, SVR, and RF

This study developed two ensemble learning models aimed at enhancing prediction accuracy and reducing model variance by combining SVR, MLP, and RF—either utilizing all three models or selecting the two with the highest individual performance. Averaging served as the aggregation method, appropriate for regression tasks, where the final output corresponds to the mean of predictions from individual models (see Appendix H). The ensemble learning models exhibited superior predictive performance compared to individual models by effectively leveraging complementary strengths and mitigating overfitting, which aligns with findings from previous studies79.

Model accuracy evaluation

Evaluating model performance is a critical step in the machine learning pipeline. In this study, both individual models and ensemble learning models were assessed using three standard evaluation metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), and the coefficient of determination (R2). These metrics provide complementary insights into the prediction accuracy and error distribution of each model. The mathematical formulations of these metrics are presented in Appendix I (Equations I1–I3).

Sensitivity analysis

Sensitivity analysis is an essential step in evaluating machine learning models, as it identifies which input variables most influence the output predictions80. Quantifying uncertainty related to input parameters and model response enhances understanding of model behavior and assists in selecting critical features for robust prediction81. This study performed sensitivity analysis on the two most accurate models—Multilayer Perceptron (MLP) and Support Vector Regression (SVR)—whose predictions were combined to improve result robustness, following established methods82,83. The combined prediction was computed as the average of MLP and SVR outputs. A bootstrap-based perturbation approach with 100 iterations was employed. Each iteration involved: (1) generating a bootstrap dataset through sampling with replacement; (2) perturbing each input feature by adding random noise within ± 2 standard deviations; (3) measuring the change in model predictions before and after perturbation; and (4) calculating sensitivity scores as the mean absolute difference in predictions.

Statistical significance was evaluated using t-tests (p < 0.05), and 95% confidence intervals were computed to assess the reliability of sensitivity estimates84. Variables were subsequently ranked based on aggregated sensitivity scores, reflecting both the magnitude and statistical significance of their influence on model output. This ranking enabled the prioritization of environmental features most influential on students’ aesthetic preferences. The formula for computing sensitivity scores is provided in Appendix J.

Results

Analyses based on demographic characteristics of participants

A total of 394 students from 20 universities and 34 academic disciplines participated in the study. Table 2 identifies key demographic characteristics—age, gender, education level, and residency status—that may shape aesthetic preferences and help interpret survey results on rest spots on campus.

Table 2 Demographic characteristics of participants.
Full size table

Both the Kolmogorov–Smirnov and Shapiro–Wilk tests (Table 3) indicated that none of the demographic variables followed a normal distribution (p < 0.001). Consequently, Spearman’s rank-order correlation was employed to examine associations between demographic factors and aesthetic preferences.

Table 3 The normality tests of demographic variables.
Full size table

Spearman’s correlation analysis (Table 4) revealed statistically significant negative correlations between aesthetic preference scores and two demographic variables: age (ρ = –0.533, p < 0.001) and education level (ρ = –0.547, p < 0.001). These results imply that aesthetic preference scores decrease as participants’ age and education level increase. In contrast, no statistically significant correlations were observed for gender (ρ = 0.096, p = 0.056) or residence status (ρ = 0.039, p = 0.442), suggesting these variables had minimal influence on participants’ visual aesthetic evaluations.

Table 4 The Spearman’s correlation test between landscape score and demographic variables.
Full size table

Performance of models for the research dataset

The first model employed was Support Vector Regression (SVR), optimized using the Radial Basis Function (RBF) kernel with ε = 0.0475, γ = 0.0260, and C = 1.3. Table 5 indicates that the SVR model maintained constant performance across the training, testing, and total datasets. The model acquired a total Mean Squared Error (MSE) of 0.1753 and explained about 82.49% of the variance (R2 = 0.8249). The relatively low Mean Absolute Error (MAE) of 0.2698 further reveals its robustness in predicting students’ visual aesthetic preferences.

Table 5 The performance of the models of the study.
Full size table

The second model, a Multilayer Perceptron (MLP), was implemented using scikit-learn’s MLPRegressor. It comprised a single hidden layer with 20 neurons and employed the logistic activation function. Training was carried out using the L-BFGS optimization algorithm with early stopping enabled to prevent overfitting. All input variables were standardized using z-score normalization. The MLP demonstrated strong performance, achieving R2 = 0.8142, MSE = 0.1860, and MAE = 0.3226 on the total dataset (Table 5), confirming its suitability for capturing complex visual-perceptual patterns.

The third model, Random Forest (RF), was optimized through systematic hyperparameter tuning. The model was built without bootstrap sampling, meaning each tree used the entire training dataset. The splitting criterion was set to squared error, and the maximum tree depth was limited to 3. The number of features considered at each split was defined as the square root of the total features. Additional parameters included a minimum impurity decrease of 0.001 and a minimum sample size of 7 for both leaf nodes and internal splits. Node splitting was restricted to branches with a minimum sample weight of 0.025. The final RF model consisted of 14 trees whose outputs were averaged to generate predictions. The RF model showed lower predictive power compared to SVR and MLP, with the lowest R2 and highest error values among the three models (see Table 5), indicating its limited ability to capture complex nonlinear patterns in aesthetic judgments.

Based on the results of the individual models, two ensemble learning models were developed to combine predictive performance. The SVR–MLP achieved the highest accuracy among all tested configurations, yielding R2 values of 0.8505 (train), 0.7672 (test), and 0.8286 (total). This model also delivered lower error rates, with MSE values between 0.1378 and 0.3070 and MAE ranging from 0.2603 to 0.4026. The three-model ensemble (SVR–MLP–RF) showed comparably strong performance, with R2 scores of 0.8412 (train), 0.7623 (test), and 0.8205 (total), but did not exceed the predictive accuracy of the SVR–MLP combination. The metrics presented in Table 5 confirm that the SVR–MLP ensemble offers the most effective trade-off between accuracy and generalizability.

The scatter plots in Fig. 4 illustrate the performance of all study models, including SVR, MLP, RF, the two-model ensemble (MLP–SVR), and the three-model ensemble (MLP–SVR–RF). The plots compare model outputs against target values for the test, train, and total datasets. These visualizations highlight subtle but important differences in performance among the modeling approaches.

Fig. 4
figure 4

Scatter plots of the outputs of five models utilized in the study versus actual values.

Full size image

Comparison of all models

The performance of five predictive models—three individual models (SVR, MLP, and RF) and two ensemble models (SVR–MLP and SVR–MLP–RF)—was evaluated using R2, MSE, and MAE across training, test, and total datasets (Table 5) to assess predictive accuracy and generalization.

Among the individual models, SVR demonstrated strong overall performance, with a relatively small difference of approximately 0.08 in R2 between training and test sets. Although the MAE difference was somewhat larger (about 0.17), the model maintained acceptable generalization performance. MLP showed slightly more stable error behavior, with a difference of about 0.11 in MAE, suggesting smoother adaptation to complex inputs through its capacity for modeling non-linear patterns. RF exhibited the largest discrepancies between training and test results, with a difference of around 0.13 in MAE, reflecting limited generalization likely due to restricted model depth and structural simplicity. Although RF provides interpretability, its accuracy and stability were comparatively inferior.

The ensemble models outperformed individual models in most metrics. The SVR–MLP ensemble achieved lower overall error and preserved a small train-test gap, with differences of about 0.083 in R2 and 0.142 in MAE, indicating both accuracy and robustness. This combination yielded a slight increase of approximately 0.004 in total R2 compared to the best-performing individual model (SVR), which, while numerically small, can be considered practically meaningful due to simultaneous reductions in error.

Including RF in the ensemble did not improve performance; instead, it led to a relative decline, with the overall MAE increasing from 0.289 to 0.303 and the total R2 decreasing by approximately 0.008 compared to the SVR–MLP ensemble. This addition did not enhance model stability, as evidenced by similar or slightly increased train-test gaps. The findings indicate that ensemble learning is most effective when based on careful selection and complementarity of models, rather than merely increasing the number of components. Visual comparisons of predicted and actual values are shown in Fig. 5.

Fig. 5
figure 5

The comparison between the predicted and actual states for the five models across the test, training, and total datasets.

Full size image

Subsequently, the accuracy of these models was assessed by comparing the residual errors between predicted and actual values (Fig. 6), and the coefficient of determination, R2 (Fig. 7).

Fig. 6
figure 6

Comparison of the residual value in test, train, and total values ​​among five research models.

Full size image
Fig. 7
figure 7

Comparison of the R2 value in test, train, and total values ​​among five research models.

Full size image

Sensitivity analysis

The sensitivity analysis was conducted using the SVR–MLP ensemble learning model, trained on 18 scaled input features (x) to predict aesthetic preference scores (y). This model was selected due to its balanced accuracy, interpretability, and stable performance across both training and testing sets (Total R2 = 0.8286, Total MAE = 0.2887), making it suitable for identifying robust feature effects. Predictions from SVR and MLP were aggregated using simple unweighted averaging to maintain transparency and avoid overfitting, particularly due to the limited sample size.

Bootstrap resampling (100 iterations) was applied, with perturbations within ± 2 standard deviations of each feature. Feature importance was defined as the average R2 change post-perturbation. Statistical significance was tested using two-tailed t-tests (α = 0.05). Results were visualized via a bar plot (Fig. 8a) and a heatmap of log₁₀ p-values (Fig. 8b).

Fig. 8
figure 8

Sensitivity analysis of environmental features influencing student aesthetic preferences using MLP and SVR models.

Full size image

The top six influential features were: tree (Importance Score = 0.148), soft landscape (Importance Score = 0.110), waterscape (Importance Score = 0.098), color diversity (Importance Score = 0. 066), building (Importance Score = 0. 053), and pathway (Importance Score = 0. 051). Although building and pathway ranked among the top features, their perturbation was associated with a negative effect on model output, suggesting an inverse relationship with aesthetic preference. Figure 8b compares importance scores across SVR, MLP, and the ensemble model. Warmer tones indicate high significance, particularly for trees, soft landscapes, and color diversity; cooler tones represent weaker features such as seating facility (0.0220). Feature magnitude reflects the extent of each feature’s impact on model predictions, while significance levels confirm that key effects are unlikely to be due to chance. Overall, top-ranked features showed strong agreement across models, with statistical significance aligning with importance scores.

These findings align with theoretical perspectives on environmental restoration. According to Attention Restoration Theory (ART), natural elements such as trees, soft landscapes, and color diversity boost unintended attention and cognitive recovery, mainly through the mechanism of “soft fascination.” Similarly, Stress Reduction Theory (SRT) highlights the evolutionary attraction of vegetation and water features, which elicit positive affective responses and lessen physiological stress. The prominence of these visual attributes in our results supports the notion that visually rich, nature-oriented environments can improve perceived aesthetic quality and restorative potential.

Figure 9 presents the results of a sensitivity analysis performed on data derived from the ensemble learning model combining SVR and MLP. The analysis focuses on the six most influential features identified in this study, highlighting their relative impact on the model predictions. These features were selected based on their significant contribution to the models’ performance, as determined through rigorous evaluation and ranking processes.

Fig. 9
figure 9

Sensitivity analysis of the six most influential features in the ensemble learning model combining MLP and SVR.

Full size image

Discussion

This research integrates two classical theoretical frameworks, Attention Restoration Theory (ART) and Stress Reduction Theory (SRT), along with the novel Social Restorative Urbanism (SRU) theory, aiming to extend the concept of environmental restoration from individual experience to interactive-social contexts within university settings18,19,85. Machine learning algorithms, particularly ensemble models, were used as quantitative tools to analyze users’ visual preferences in complex socio-spatial environments. The following discussion first assesses the performance of these models, then examines the role of key environmental features, and finally discusses the resulting design implications.

Machine learning model performance

Three individual models—Support Vector Regression (SVR), Multi-Layer Perceptron (MLP), and Random Forest (RF)—alongside two ensemble models (SVR–MLP and SVR–MLP–RF) were employed to predict students’ aesthetic preferences in interactive university campus spaces. Previous studies primarily relied on descriptive analysis or visual classification of environmental variables42; however, this investigation adopted a predictive and quantitative approach. Results indicated that SVR and MLP exhibited relatively similar and more accurate performance compared to RF, with R2 values of 0.824 and 0.814, respectively, versus 0.761 for RF. Ensemble models demonstrated comparable or slightly improved performance over individual models, achieving R2 values of 0.828 and 0.820. Although these differences were modest, the mitigation of the weaker RF model’s effect within the ensembles suggests potential improvements valuable for predicting complex human preferences and supporting design-oriented decision making.

Similar studies conducted in natural settings have demonstrated the strong performance of MLP models in evaluating the aesthetic quality of parks and forests, with reported R2 values ranging from 0.83 to 0.9431,32,52,53. Models such as MLP, SVR, RF, and XGBoost have also achieved high predictive accuracy in digital image aesthetics, with R2 values around 0.84 to 0.8741.

The application of SVM in forest landscape evaluation has similarly yielded promising results (R2 = 0.84)53. While some of these models outperformed those used in the present study in terms of accuracy, the achieved results remain comparable. One contributing factor to the difference may lie in the use of natural landscapes with more regular spatial structures and richer visual content, which enhances feature separability for machine learning algorithms. Earlier studies employed more complex algorithms such as XGBoost and CNN; however, this work prioritized models offering a balanced trade-off between predictive accuracy, interpretability, and design applicability. The proposed approach introduces a novel combination of ensemble learning models, a focus on student-centered interactive rest spaces instead of generic public environments, and a theory-driven feature selection approach grounded in ART, SRT, and SRU frameworks.

This study differs from previous works that relied on computer vision algorithms to automatically extract limited visual indicators—such as sky percentage, depth estimation, or environmental openness—from public campus scenes42,57,86. The difference in input data granularity and theoretical approach, especially when compared to models like GBDT in Ma et al.57 (R2 = 0.726), likely contributed to improved prediction accuracy. Furthermore, this study focuses on directly predicting aesthetic preferences, whereas most previous studies addressed indicators such as happiness, sense of security, or psychological restoration57,86, or relied solely on qualitative classifications or neurological analyses42,48. A deeper understanding of model performance requires examining the key influential environmental components, which are discussed in the following section.

Environmental features and their restorative potential

This study’s correlation analysis showed that age and education level significantly affect students’ aesthetic preferences; as students grow older and advance in education, their tendency to appreciate university environments declines. This may stem from increased academic pressures and responsibilities that lessen focus on positive environmental aspects. These findings align with Lee et al.87 in a mental health and environment study. However, Jahani and Saffariha31 found no significant demographic effects, possibly due to differences in samples and contexts. These correlations, though not central to the study’s objectives, highlight the potential role of demographic variables in shaping aesthetic perception and may serve as a basis for further research in similar contexts.

Beyond demographic factors, the sensitivity analysis results of this study revealed that several key natural and built features play a critical role in shaping students’ visual preferences, including trees, soft landscapes, waterscapes, color diversity, and lower densities of buildings and pathways. These findings align well with the assumptions of Attention Restoration Theory (ART) and Stress Reduction Theory (SRT), reinforcing the importance of these factors in activating mental restoration processes18,19. Soft landscapes, comprising natural and semi-natural elements such as vegetation, water, and sky, have been widely recognized as significant indicators of environmental aesthetics 31,48,55. While this concept was initially defined as a general category of natural elements, its components—including vegetation, water, and sky—were analyzed separately in this study to clarify the individual contribution of each element to restorative processes31,48,55.

Although findings in both aesthetic studies31 and neuroscience research48 support our results by suggesting that the presence of buildings negatively affects mental restoration, and that lower building density can enhance visual openness and enable access to views of the sky and even natural landscapes beyond the campus (such as parks, lakes, or surrounding mountains), the quality of architectural design may still contribute to generating positive emotions and shaping favorable visual experiences59. Similar to buildings, the pathway feature—including walkways and staircases—showed a negative impact on students’ visual preferences despite its functional significance within campus environments. This outcome may be attributed to the low material quality observed across many areas in the sample, such as the use of gravel in certain pathways. The well-documented importance of path design57 and flooring material quality60 in enhancing sensory and visual experiences of outdoor settings provides a basis for interpreting this finding.

In contrast, the role of trees as a key component of green landscapes within soft environments has been frequently examined in environmental aesthetics literature, consistently aligned with the principles of Attention Restoration Theory (ART) and Stress Reduction Theory (SRT)31,32,53,88. Trees also represent one of the most influential environmental features contributing to students’ mental restoration29,57,89. Their orderly arrangement, balanced density, and visual harmony with other natural elements enhance the overall aesthetic appeal of the campus environment63. Moreover, by improving thermal and acoustic comfort, trees address users’ sensory and physiological needs90. Trees also contribute to spatial organization by acting as soft boundaries. This function aligns with the Social Restorative Urbanism (SRU) framework, which highlights their role in shaping interaction areas and shared pathways.

Water landscapes play a prominent role among the environmental features influencing students’ aesthetic preferences in university settings. Both static forms, such as pools and ponds, and dynamic forms, such as fountains and surface flows, contribute significantly to restorative experiences54,61,63,91. This element relates directly to Attention Restoration Theory (ART) and Stress Reduction Theory (SRT), and it also connects with the Social Restorative Urbanism (SRU) framework by creating focal points and edges, as exemplified by the ecological wetland at Universiti Sains Malaysia (USM)87. Static water features add depth to the environment and reflect the sky and surroundings, enhancing the sense of spatial extent and “being away” while stimulating soft fascination through their visual appeal, as described by ART30. Dynamic elements such as fountains can additionally contribute to reducing physiological stress by activating the parasympathetic nervous system, in accordance with Stress Reduction Theory (SRT).

Color diversity emerged as a significant environmental factor influencing visual preferences in this study. This finding aligns with previous research demonstrating that increased color variety in vegetation, especially through diverse species and colorful flowers, enhances visual attractiveness and improves the restorative quality of natural landscapes56,92. Additionally, in interactive university spaces, the presence of colorful plants has been reported to contribute positively to students’ psychological well-being59. According to Attention Restoration Theory (ART) and Stress Reduction Theory (SRT), color diversity can strengthen mental restoration processes by stimulating soft fascination and inducing positive emotions.

Finally, beyond the aforementioned natural and built factors and with lesser influence, the present study demonstrates that, consistent with previous research56,59,65,93, other variables such as landscape complexity, vegetation diversity, presence of stones, and environmental amenities (including trash bins, lighting, and sports equipment) also affect aesthetic preferences. Considering the importance of seasons in environmental aesthetics94 and the data collection conducted during autumn, the influence of certain natural variables, such as grass cover, may have been limited due to reduced vibrancy and seasonal changes. Furthermore, the use of two-dimensional images for variable control restricts the complete representation of the perceptual and emotional experience of the environment. Although some studies consider the impact of two-dimensional images comparable to real experience95, others regard methods such as virtual reality or 360-degree images as more effective95,96. Nevertheless, no platform can fully replicate a physical environment97, and controlled environmental simulations remain valuable for research purposes96. This study employed two-dimensional images to control confounding variables; however, this methodological limitation remains a key challenge.

Design Implications for University Campuses

This study provides a foundation for developing data-driven approaches to the design of interactive university spaces. Enhancing landscape aesthetics and mental restoration potential requires the integration of diverse natural elements across campus environments. For instance, a combination of deciduous and evergreen trees can diversify winter scenery and promote ecological resilience27,98. Alongside tree diversity, planting colorful flowers and a variety of vegetation56—combined with small fountains or larger water features—offers another effective way to enhance visual appeal and meet students’ aesthetic preferences.

In addition to visual richness, physical comfort also plays a significant role in shaping students’ spatial experience. Redesigning campus pathways plays a critical role in improving user comfort. The selection of appropriate paving materials60 such as brick, wood, stone, or tile—combined with grass strips instead of concrete and asphalt—along with shading elements like vegetation or semi-transparent canopies, contributes to thermal comfort. Seating and interaction nodes aligned with the Social Restorative Urbanism (SRU) framework85 may further enhance spatial experience, particularly when integrated along pathway edges without obstructing pedestrian flow. These design strategies not only offer improved aesthetic quality but also reduce mental fatigue and increase users’ sense of spatial presence.

Artistic design of functional elements—such as benches, trash bins, and lighting fixtures—represents another effective strategy for enhancing the visual quality of campus spaces. These components, beyond their practical utility, stimulate aesthetic perception and encourage prolonged user engagement. The visual character of campus environments can also benefit from careful design at larger spatial scales. A harmonious skyline, shaped by rhythmic architectural composition99 and controlled building heights, enhances both distant views and the internal aesthetics of university spaces.

Student dormitories are often located within university grounds, making nighttime environmental quality essential. Appropriate lighting improves both visual safety and perceived security100, while a pleasant night atmosphere contributes to anxiety reduction and overall spatial experience101. The use of soft, uniform lighting along pathways and seating areas, combined with focused spotlights, can meet both functional and aesthetic needs during evening hours.

This integrated approach not only enhances predictive accuracy but also grounds design decisions in well-established restorative theories, paving the way for evidence-based and user-centered campus planning.

Conclusion

University students frequently experience psychological stressors such as anxiety and mental fatigue, often exacerbated by academic pressures and social isolation10,43. Well-designed campus rest spots can alleviate these issues by offering visually appealing and socially supportive environments. However, identifying the specific visual attributes that shape students’ preferences remains a complex and underexplored task.

This study proposed a data-driven framework for predicting aesthetic preferences in university rest spots by integrating ensemble machine learning (ML) models with theories from environmental psychology. The analysis used images collected from four universities in Tehran and a large set of student ratings. Using this dataset, the SVR–MLP and SVR–MLP–RF ensemble models demonstrated high predictive performance (R2 = 0.828 and 0.820, respectively), comparable to the best-performing individual models (SVR = 0.824, MLP = 0.814). Although the ensembles did not significantly surpass the highest individual R2 scores, they provided improved stability and error compensation, particularly mitigating the weaker performance of the Random Forest model (R2 = 0.761, MAE = 0.38). These results support prior findings suggesting that combining models can enhance overall predictive robustness and reliability102. Future studies may further improve prediction performance by assigning optimized weights to individual models based on their error distributions and generalization ability.

Analysis of the 18 visual features revealed that six consistently emerged as the most influential: tree cover, soft landscapes, waterscapes, color diversity, fewer buildings, and pathway presence. These attributes are strongly aligned with restorative principles from Attention Restoration Theory (ART), Stress Reduction Theory (SRT), and Social Restorative Urbanism (SRU). Their prioritization in campus design can help create spaces that are both aesthetically appealing and psychologically supportive.

Additionally, the design of pathways, flooring, and stairs should be carefully considered to create spaces that are both visually stimulating and functionally effective. This study bridges computational prediction with theoretical insights from environmental psychology to offer a practical framework for integrating human-centered visual evaluation into campus planning. The proposed ensemble learning Models offer practical tools for architects and environmental designers, supporting both Pre-Design Phase analysis and evidence-based campus development. Furthermore, the accompanying aesthetic preference prediction framework and codes are readily deployable in applications.

While our dataset was conducted in autumn only, seasonal variations94 and cultural differences may influence preferences. Future studies should explore multi-seasonal data, indoor rest spaces, and cross-cultural validations to generalize findings. The integration of biometric indicators, such as eye-tracking or EEG, also holds potential for enhancing prediction accuracy and capturing more implicit aesthetic responses. This computational approach marks a paradigm shift from intuitive to evidence-based landscape design, with potential applications extending to the healthcare and academic environments.

Limitations

The substantial number of questionnaire items resulted in relatively low student participation rates. Additionally, limitations in online response storage occasionally caused technical errors, requiring researchers to periodically download and clear the data from the web server. Moreover, future studies are encouraged to automate expert-based features—such as tree density or vegetation diversity—through advanced image recognition, segmentation algorithms, or LiDAR-based analysis, thereby enhancing scalability and model generalizability.