Sky and trees shape campus stress recovery, but more is not better

A walk across a university campus can either steady a stressed mind or fray it further, and a new study argues that the difference may come down to how much sky and greenery a pedestrian actually sees — and how those elements are arranged. Researchers in Nanning, China, paired more than a thousand street-level photographs with students’ ratings of restorative quality and then trained machine-learning models to predict which campus walkways feel most restorative. Their results, published in Computational Urban Science, challenge the intuitive assumption that greener is always better: the perceived restorative benefit of vegetation did not rise indefinitely with the amount of green in view, while open sky showed its own limits.

The study rests on Stress Recovery Theory, a framework holding that visual exposure to natural settings can trigger restorative psychological responses through emotional regulation, physiological relaxation, and the activation of effortless attention. Earlier work in this tradition concentrated heavily on dramatic natural landscapes such as forests and lakes, and most previous campus studies relied on single-site analyses or questionnaire surveys that are difficult to scale across an entire university. The authors set out to build a faster, broader method: instead of asking students to walk every path and fill out forms, they used street-view imagery as a stand-in for what a pedestrian actually sees, then linked those images to perceived restorativeness scores.

The empirical work covered four university campuses in Nanning, in China’s Guangxi province: the main campus of Guangxi University, the main campus of Guangxi Medical University, and two campuses of Guangxi Minzu University, one at Xiangsi Lake and one at Siyuan Lake. The sites were deliberately chosen for contrast. Guangxi University features dense urban morphology, Guangxi Medical University offers a quieter, specialised setting, and the two Minzu University campuses provide extensive green and waterfront space with lower building density. In March 2025, the team collected 2,250 geo-tagged street-view images along campus pedestrian paths at roughly 20-metre intervals using the Baidu Street View API, filtering to summer captures to limit seasonal bias in vegetation density. Half of these — about 1,060 images — were randomly selected for human evaluation.

Students rated each image using the Restoration Outcome Scale, a validated questionnaire covering four restorative outcomes: emotional relaxation, attention restoration, mental clarity, and vitality. Participants were asked to imagine themselves walking within the depicted scene before scoring each item on a 7-point scale. The online survey yielded 312 valid responses from people aged 18 to 25, with a mean age of 21.3 and 51 percent female. To check that ratings made from a screen genuinely reflect on-site experience, the researchers ran a field validation at 80 randomly chosen sampling points, where 96 students completed the same questionnaire in place. Agreement between online and on-site ratings was strong — an intraclass correlation coefficient of 0.81, with a 95 percent confidence interval of 0.73 to 0.87 — so the online ratings alone served as training labels for the models.

Each image was then decomposed into quantifiable visual features across four dimensions. A panoptic-segmentation model, MaskFormer fine-tuned on the ADE20K dataset, measured the pixel share of semantic elements such as sky, vegetation, and roads, yielding 16 built-environment variables. An object-detection model, YOLOv8, counted cars, bicycles, and people. A ResNet-50 network fine-tuned on the Place Pulse 2.0 dataset produced perceptual scores for how beautiful, safe, wealthy, lively, depressing, or boring a scene appeared. Finally, pixel-level measures — hue, saturation, brightness, edge density, and a threshold value — were extracted with the OpenCV library and normalised to a 0-to-1 range.

The predictive engine was a Random Forest regression model, an ensemble of decision trees chosen after comparison with gradient boosting and k-nearest-neighbours approaches. Trained on 80 percent of the annotated images and tested on the remaining 20 percent, with hyperparameters tuned by grid search under five-fold cross-validation, the model reached an R² of 0.602, a root-mean-square error of 0.078, and a mean squared error of 0.049 on restorative-quality scores normalised to a 0-to-1 scale — an error corresponding to roughly 7.8 percent of the scale range. A spatial block cross-validation variant, which kept geographically proximate sampling points within the same fold, cost only marginal performance, suggesting that spatial autocorrelation in the data did not seriously inflate the reported accuracy.

The maps that emerged from the model revealed pronounced heterogeneity both within and between campuses. The Siyuan Lake campus of Guangxi Minzu University recorded the highest average restorative quality score, 0.452, while Guangxi University recorded the lowest, 0.391. High-quality zones clustered around lakes, large green spaces, and water features, whereas the lowest scores appeared in densely built areas with little vegetation and few open vistas — near administrative and laboratory buildings at Guangxi University, for instance, and around sports fields dominated by hard paving on the Siyuan Lake campus. A global Moran’s I test confirmed moderate positive spatial autocorrelation (0.342, z-score 8.721, p < 0.001), meaning nearby sampling points tended to score similarly.

The feature-level results carried the study’s most consequential message. A multiple linear regression confirmed positive associations for natural elements: the sky-view factor was the strongest predictor (β = 0.313, p < 0.01, 95 percent CI 0.248 to 0.376), followed by roads (β = 0.175) and vegetation (β = 0.128). But the random forest’s partial-dependence analysis uncovered nonlinear patterns the linear model could not. As vegetation coverage rose from 0.2 to 0.5, the predicted restorative score actually declined from 0.43 to 0.37 before stabilising — a pattern the authors suggest may reflect dense plantings reducing ambient light in street views. Sky visibility, by contrast, consistently enhanced restorative quality, yet scores fell sharply once sky exceeded roughly 0.6 of the frame, hinting that overly exposed, sparsely vegetated spaces can feel bleak rather than open. Perceived safety showed a U-shaped relationship, and scenes with warmer colour tones and stronger perceived vitality scored higher. The authors argue this undermines a simple ‘more trees is better’ prescription: spatial configuration and perceived quality of environmental features, rather than sheer quantity, appear more decisive for stress recovery.

The study’s limitations are acknowledged by its authors and worth keeping in view. Street-view imagery is static and purely visual, ignoring the sounds, smells, and movement that shape a real walk; the restorative scores are self-reported perceptions rather than direct physiological measurements such as heart-rate variability or cortisol; and the sample covers a single city and a single population — university students — which constrains generalisability. The design is also cross-sectional, so the findings describe associations between visual features and perceived restorativeness rather than experimentally demonstrated causal effects. Still, the framework offers campus planners something previous methods could not: a scalable, image-based way to map where restorative quality is high and low across an entire campus, and to identify which visual features — sky exposure, vegetation structure, perceived safety and vitality — are worth tuning. The authors suggest that targeted interventions, such as punctuating large plazas with tree groves or balancing openness with enclosure, could follow from such maps, moving healthy campus design beyond simple greening metrics toward a quality-focused approach grounded in how students actually perceive their everyday walking environments.

Subject of Research: Street-view imagery and machine learning analysis of how campus pedestrian visual features relate to perceived stress recovery

Article Title: More trees, more restorativeness? Exploring the influence of campus pedestrian space on stress recovery using street view imagery

Article References: Mai, S., & Ma, H. (2026). More trees, more restorativeness? Exploring the influence of campus pedestrian space on stress recovery using street view imagery. Computational Urban Science, 6(1), Article 48. https://doi.org/10.1007/s43762-026-00286-1

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00286-1

Keywords: stress recovery, campus green space, street-view imagery, machine learning, restorative environments, Stress Recovery Theory, vegetation cover, sky visibility, random forest, urban perception, student well-being, Nanning