Uncovering the structural influence of urban park landscapes on psychological restoration via graph learning

Abstract

Over half of global population now resides in urban areas, where mental health challenges such as anxiety and depression are increasingly prevalent. Urban park landscapes are recognized as critical settings for promoting psychological restoration, yet a systematic understanding of what makes these spaces restorative remains limited. In this paper, we propose a novel graph learning-based framework that models and interprets the perceptual structure of urban park landscapes in relation to their restorative potential. The framework comprises two key components: a predictive module that quantitatively estimates perceived restorativeness based on scene-level visual and spatial features, and an interpretive module that identifies the importance of individual landscape elements and their spatial configurations. To support this approach, we introduce a novel dataset of urban park landscape images annotated with semantic, structural, and perceptual labels, facilitated by a large language model–assisted annotation pipeline. Experimental results demonstrate our framework’s predictive accuracy and its ability to provide interpretable visual explanations grounded in environmental psychology. Our findings offer actionable guidance for urban planners and landscape designers aiming to create psychologically supportive public environments.

Introduction

In recent years, mental health issues—including anxiety, depression, and cognitive fatigue—have risen worldwide¹, underscoring an urgent need for accessible and effective strategies to support psychological restoration^2–4. Urban parks are important to maintain livable and sustainable cities. Urban park landscape, as essential elements of city fabric, is widely valued for their capacity to enhance mental well-being by providing restorative environments that alleviate stress and replenish cognition^5–8. We all have had such an experience: when encountering a landscape with a tranquil lake, dappled sunlight through tree canopies, and open green space, we may experience a sense of calm and attentional relief, helping to ease cognitive fatigue and emotional tension. Building on the attention restoration theory (ART)⁹ and the stress reduction theory (SRT)¹⁰, psychological and landscape-architecture research has demonstrated the restorative effects of natural environments and urban park landscape on human health, particularly in enhancing attention and reducing stress. Consistent with these theoretical perspectives, cities are investing heavily on urban parks. According to the Trust for Public Land, the 2025 ParkScore Index shows record-high investment of major American cities in their urban parks, as well as growing usage¹¹. Urban parks are increasingly recognized by city dwellers as vital spaces for daily respite, recreation, and psychological well-being^12,13.

Psychological studies highlight attributes such as naturalness, enclosure, and vegetation as key drivers of positive emotional responses^2,14–20. Instruments such as the Restorative Outcome Scale (ROS)²¹ have been widely used to capture self-reported restorative experiences across diverse settings and populations, demonstrating robust psychometric properties^22–24. Recent work has further explored perceived sensory dimensions, the presence of other people, exposure duration and supportive built elements, showing that vegetation types and density, visual enclosure, spatial arrangement, serenity and species richness are all associated with stress recovery and improved well-being^25–33. Eye-tracking studies indicate the role of urban landscape images in alleviating emotional stress^34,35, and several recent investigations suggest that the visual proportion of landscape elements such as sky, trees, waters are positively related to restorative potential^35–38. In landscape designs, research has largely focused on broader factors such as accessibility, greenery presence, and openness^5,39–41. Despite this rich body of work, most studies rely on self-reported surveys and controlled experiments using pre-selected visual stimuli within urban parks. Although these approaches are predominantly qualitative or semi-quantitative, they linked survey-based reports to perceived environmental quality or health outcomes at aggregated community or city scales. Therefore, they are inefficient in scalable data acquisition and provide limited insight into how concrete combinations and spatial configurations of landscape elements jointly shape restorative outcomes at the human-eye level.

In recent years, the restorative potential of landscapes is increasingly understood to be governed not merely by its constituent elements, but by their spatial structure—the specific arrangement and relational geometry that shapes visual perception and cognitive processing. Qualitative work at the human-eye level confirms that structural features (such as the contiguity of green space, vertical stratification, and the connective integration of natural and built elements) are significant predictors of restorative outcomes^42–44. These observed features align with the mechanisms proposed by ART and SRT: spatial structures that offer coherent spatial organization (e.g., contiguous green patches with clear depth cues) foster soft fascination and alleviate psychological stress, while configurations that suggest prospect and refuge (e.g., a sheltered bench with an open view) can reduce cognitive load. Consequently, a systematic understanding of landscape spatial structure is not merely supplementary but fundamental to advancing the science of psychological restoration. It represents the critical pathway to move beyond descriptive, element-centric accounts toward predictive models and generalizable design principles. However, this understanding remains qualitatively constrained. The lack of computational models capable of representing and quantifying the multivariate relational syntax of landscapes impedes progress. As discussed later in Sect. "Basic analytics of graph annotations", considering landscape element types and even their pairwise relationships alone is insufficient to explain the observed variation in restorative quality, highlighting the need for a framework that can capture higher-order structural configurations.

Parallel advances in computer vision and machine learning have enabled more scalable, human-centric assessments of urban environments from imagery, offering opportunities for quantitative analysis and automated reasoning regarding the restorative benefits of urban park landscapes^45–47. Deep convolutional neural networks (CNNs), Transformers and related models have been used to analyze street-level or aerial images, quantify visual exposure to greenery through metrics such as the Green View Index (GVI) and relate these measures to health or affective outcomes^48–51. Vision Transformer-based models have demonstrated correlations between visual elements, human activities, and positive psychological perceptions^52,53, while panoramic street-view analysis connect neighborhood-level visual cues with urban experiences such as wealth, boredom, and depression⁵⁴. Machine learning models have also been applied to identify specific landscape characteristics that moderate emotional restoration^32,55, and to integrate spatial parameters (for example, enclosure and walking distance) into design-support frameworks for assessing green spaces⁵⁶. More recently, graph neural networks (GNNs) have been adopted to estimate restorative outcomes from street-view imagery at large scales and to model urban networks in planning contexts^41,57,58. Collectively, these studies demonstrate the power of machine learning to capture complex visual patterns in urban landscapes and to link them with human perceptions. However, existing research methodologies are primarily qualitative and based on on-site survey, and focus on broad geographic or community-scale evaluations, lacking the specificity, granularity, and scalability required for practical and future-oriented applications.

The above limitations lead to the central research problem of this study: how to computationally model and explain the restorative potential of urban park landscapes in a way that is both accurate and interpretable for psychological and design practice? Specifically, we investigate three questions: (1) Can a learning framework reliably predict scene-level restorative outcomes from human-eye-level images of urban parks? (2) How can such a framework represent park landscapes as structured compositions of elements and spatial relations, rather than as unstructured feature vectors? (3) Can the learned model provide interpretable explanations that identify which elements and configurations contribute most to psychological restoration, thereby offering actionable guidance for landscape design?

Here we address these questions by introducing a graph learning-based framework that links the perceptual structure of urban park landscapes to psychological restoration. At its core, the framework comprises two complementary modules: Perceiver, a predictive module to predict restorative potential from visual landscape input, and Miner, an interpretive module to reveal the spatial and semantic factors driving these predictions. Perceiver is an effective graph neural network-based model, which accurately predicts psychological restoration scores by processing landscape images represented as structured graphs. Each graph captures landscape elements as nodes and their spatial relations as edges, with perceptual depth information integrated to reflect the viewers’ visual experience. Miner is a second deep learning model that learns to assign soft importance weights to graph components, identifying significantly influential elements and spatial configurations through a graph-masking strategy. Additionally, we develop a new dataset of urban park landscape images with restorative emphasis. Particularly, we leverage large language models (LLMs) to generate scene graph annotations from pre-segmented landscape images via a targeted prompting strategy, significantly reducing manual annotation effort while maintaining semantic and structural integrity. Together, these components yield a first-learn-then-interpret architecture, offering a novel, computational lens for quantitatively exploring how urban park landscapes influence psychological restoration. While this study does not aim to establish definitive theoretical conclusions, it introduces a practical methodology that enables automated prediction, interpretable analysis, and scalable data generation. This framework provide psychologists, landscape designers, and urban planners with a quantitative, real-time tool for refining theoretical models, informing evidence-based spatial design, and advancing urban development strategies that support mental well-being.

Dataset

To date, no existing datasets of urban park images are designed to support graph learning. In this section, we introduce ParkScape-PR (urban park landscape dataset emphasizing psychological restoration), comprising 1,346 high-resolution images of urban park landscapes. To ensure data diversity, the images were sourced from five types of urban park scenes: cultural heritage parks, amusement parks, comprehensive parks, community parks, and ecological parks. ParkScape-PR comprises 236 on-site photographs and 1,100 web-sourced images. On-site images were captured at eye level using a smartphone to approximate a pedestrian’s horizontal field of view; web-sourced images were selected to match this perspective. We removed blurry, distorted, or poorly composed images and considered variations in weather and season. We detected human faces in images by YOLOv5⁵⁹ and obscured them with the Gaussian blur⁶⁰. Different from previous urban imagery datasets, ParkScape-PR features multi-modal annotations, including instance segmentation, scene graphs, depth information, and survey-based ratings.

Instance annotation

In each image, objects that are visually distinguishable to human eyes are considered individual instances. We identified major instances and annotated each with a textual label, forming the instance-level annotation for that image. In ParkScape-PR, instance classes follow a two-level hierarchical taxonomy, as defined in Table 1. For example, a tree is coarsely labeled as “plant” at Level-1 class and as “tree” at the fine-grained Level-2 class. When multiple instances of the same class appear in an image, each is assigned a unique number to distinguish them as individual instances.

Table 1.

Hierarchical classification (two levels) of the objects in UPL images.

Level-1 class		Level-2 class	Definition or examples
Water	C1	Lake	Lake, pond, reservoir
	C2	River	River, urban canal (width)
	C3	Fountain	Tiered fountain, interactive water jet
Architecture	C4	Civil architecture	Residential building, commercial building, bridge
	C5	Landscape architecture	Sculpture, park pavilion
Facility	C6	Civilian facility	Streetlight, bin, fencing, toilet, vending machine
	C7	Leisure facility	Bench, fitness facilities
Plant	C8	Tree	Usually well-defined, upright trunks, like pine and maple trees
	C9	Shrub	Usually multiple stems in bushy appearance, like boxwood, holly
	C10	Lawn	Grass, like meadow fescue
	C11	Flower	Rose, tulip
Sky	C12	Sunny sky	Clear sky (mostly free of clouds)
	C13	Overcast sky	Thick clouds
Person	C14	Exercising person	Persons in walking, jogging, cycling, or stretching
	C15	Sitting person	Persons resting on benches, or sitting to read
Road	C16	Road	Road, pebble path, sidewalk, asphalt walkway
Animal	C17	Pet	Dogs, cats
	C18	Wild animal	Gray squirrels, house sparrows

Using the pre-defined instance classes in Table 1, we applied semantic segmentation to identify instances within images. Specifically, we used the ISAT tool⁶¹, which integrates with Meta’s Segment Anything Model (SAM)⁶² for efficient, interactive segmentation. In our scenario, however, SAM faces challenges with visually irregular and diffuse instances (such as the sky, water, and tree canopies), often resulting in inaccurate boundaries between adjacent elements. Additionally, small objects, such as trash cans, often blended into the background, leading to mis-segmentation or omission. All instance annotations underwent a thorough manual review to maintain high annotation quality and consistency. Fig. 1 demonstrates four sample images and their instance annotations.

Fig. 1.

Examples of instance annotations for ParkScape-PR. Here, for instance, “plant-tree : 8-1” indicates that the image contains an instance classified as a Level-1 class “plant” and a Level-2 class “tree”, uniquely numbered as “8-1”.

Graph annotation

We constructed a directed graph that captures the spatial relations between instances. In this graph, nodes represent the instances, while edges indicate the spatial relations between them. The spatial relations used in ParkScape-PR are defined in Table 2 from viewer’s perspective. In some cases, the spatial relationships between two instances cannot be easily described using directional terms such as “left of” or “right of”. To address this, we introduce two additional relation types: “adjacent to” and “around”. These two relation types are exemplified in Fig. 2, where the bank lawn and tree cluster are both classified as “around” the lake in the left image, and the bank lawn and the lake are “adjacent to” each other in the right image.

Table 2.

Nine types of spatial relations between objects in an UPL image.

Relation		Definition	Examples
T1/T2	left of / right of	Horizontal planar relationship from viewer’s perspective	A bench is located left of the tree
T3/T4	below / above	Vertical alignment with an elevation difference of 0.5 m	A street lamp is positioned above the sidewalk, and a river flows below a bridge
T5/T6	in front of / behind	Depth relationship relative to observer’s position	Fence is behind a trash can, a statue in front of the lawn
T7	adjacent to	Objects sharing common boundary (1m apart)	A path is adjacent to the flower bed
T8	around	Encircling spatial configuration (270° coverage)	Flowers are around the statue, a lawn is around the pavilion
T9	on	Surface contact with supporting object	A person sits on bench, a dog is on the lawn

Fig. 2.

Examples of relation types “around” and “adjacent to” exhibited within two semantically-segmented images.

Graph annotations essentially encode the structural and spatial features of their corresponding images. To efficiently generate graph annotations for all images, we leveraged ChatGPT, an LLM, to interpret visual scenes based on predefined prompts. An overview of the annotation pipeline is presented in Fig. 3. ChatGPT was used to identify not only the instances within each image but also their spatial relationships. To mitigate the tendency of LLMs to lose contextual continuity, we prompted ChatGPT individually for each image, rather than relying on a single initial prompt. For each image, ChatGPT was provided with two inputs: (1) a list of instance class labels identified for the given image, and (2) all spatial relation types defined in Table 2. This approach guides the ChatGPT to semantically focus on relevant instances and infer spatial relationships within a constrained relational framework. ChatGPT then generated a textual description of the spatial relationships present in each image. These descriptions were manually reviewed for accuracy and then, saved in plain text format. Any descriptive errors or omissions were corrected to ensure fidelity to the actual spatial layout. Based on the textual descriptions, ChatGPT subsequently produced a corresponding graph structure, in which nodes represent instances and edges represent their spatial relationships.

Fig. 3.

LLM-assisted pipeline of generating a graph annotation from an image and its instance annotation.

Figure 4 shows a graph annotation example, where four segmented instances—named as , , x, and v—serve as the nodes of the graph. A directed edge is formed between two nodes if a spatial relationship exists between the corresponding instances in the image. Each node is associated with an instance class, and each edge is associated with a relation type. In Fig. 4, node (instance class C8/“tree”) and node v (instance class C7/“leisure facility”) are connected by a directed edge with relation type of T3/“above”, reflecting the visual arrangement in which the tree appears above the chair in the image. To facilitate subsequent structural analysis, we introduce two descriptors:

• Instance-relation triplet: For any edge e(u, v), its relation type, and the two instance classes, and , together define an instance-relation triplet . Obviously, the triplet encodes the basic semantic configuration between two individual elements by jointly specifying their classes and spatial relation. In this study, there are 18 instance classes and 9 relation types, resulting in a total of possible triplet types.

• Edge equivalency: Two edges, whether from the same graph or from different graphs, are consdiered equivalent if their corresponding triplets are identical in all three dimensions. In Fig. 4, for example, the edges and are equivalent, as both describe “a tree is below the sunny sky”; in contrast, the edges and are not equivalent.

Fig. 4.

Demonstration of generating a graph annotation from an example image.

In ParkScape-PR, the resulting graph annotations were stored in JSON format and structured according to the Visual Genome annotation standard⁶³, enabling seamless integration and compatibility with existing vision-language processing tools. By combining instance annotations with LLM-driven spatial reasoning, our semi-automated workflow substantially reduces the manual effort required and improves the overall efficiency of generating high-quality graph annotations for the dataset.

Depth annotation

In urban landscape design, depth not only contributes to visually pleasing compositions but also about fosters a connection between viewers and space. Well-designed landscapes with clear or dynamic depth cues offer a range of aesthetic benefits and promote psychological restoration by encouraging relaxation, engagement, and a sense of well-being^20,64. To quantify the scene depth within each image, we extracted the depth information using the Depth Anything V2 model⁶⁵, which measures the depth variation in the scene and adds an additional annotation layer for ParkScape-PR. In implementation, we simplify the depth maps using a statistical matrix which will be detailed later in Sect. "Depth extraction". Figure 5 illustrates two example images along with their corresponding depth annotations.

Fig. 5.

Demonstration of the depth annotations for two example images.

Survey annotation

We designed a comprehensive survey comprising three key sections: (1) scene preference, (2) perceived restorative benefits, and (3) personal attributes, as outlined in Table 3. The scene preference section includes a single-item assessing overall landscape preference for each image. The second section evaluates perceived restorative benefits using ROS, which captures three core dimensions of restorative experience—clearing one’s thoughts, inducing relaxation and calmness, and restoring attention—through six statements rated on a five-point Likert scale²¹. The third section gathers subjects’ demographic information, including age, gender, and so on.

Table 3.

Questionnaire accessing restorative benefits from exposure to urban park landscape images.

Category	Question/attribute
Scene preference (15 points)	Visual preference
	Q1. How much do you like this scene?
Restorative benefits (15 points)	Clearing one’s thought
	Q2. Viewing this image makes my thoughts feel clear.
	Q3. Viewing this image allows me to forget daily worries.
	Relaxation and calmness
	Q4. Viewing this image makes me feel restored and relaxed.
	Q5. Viewing this image makes me feel calm.
	Q6. Viewing this image piques my interest.
	Attention restoration
	Q7. Viewing this image makes me feel focused and alert.
Demographics (single choice)	Gender: □ Female □ Male
	Age: □ 18–24, □ 25–34, □ 35–44, □ 45–55, □ 56–64, □ 65+
	Education: □ K10-12, □ Higher education, □ Vocational education

We recruited 19 college students as subjects. Table 4 shows the demographic statistics. Each viewed all ParkScape-PR images displayed on a 72-inch screen, with 15 seconds allocated per image to complete the corresponding survey. All experimental protocols were reviewed and approved by the Human Study Ethics Committee of Beijing Forestry University. These experimental protocols, conducted with informed consent from all subjects, are in accordance with relevant guidelines and regulations. For each image, the ROS ratings from all subjects were aggregated to form the final survey annotation. Specifically, the score for each survey question was computed as the average of all individual responses. For instance, an image may receive a ROS score vector of (2.4, 3.0, 4.2, 4.6, 2.7, 3.3, 4.0), where the i-th value corresponds to the average rating for the i-th survey question. The average ROS score of an image is calculated as the average of its scores over all seven dimensions. Figure 6 presents two example images and the distribution of their scores in each dimension.

Table 4.

Demographic statistics of subjects in survey.

Characterics	Category	n (%)
Gender	Female	8 (42.1%)
	Male	11 (57.9%)
Age (years)	1822	5 (26.3%)
	2230	14 (73.7%)
Education	Undergraduate	6 (31.6%)
	Postgraduate	13 (68.4%)

Fig. 6.

Survey annotation statistics of two example images.

Basic analytics of graph annotations

We performed a graph analysis of the graphical structures extracted from all images of ParkScape-PR to explore whether specific structural features are associated with ROS score variations. As shown in Fig. 7, we examine the influence of five graphical attributes on each image’s total ROS score. These attributes are (1) the number of instances (“#instance”), (2) the number of relations (“#relations”), (3) the average degree (“avg. degree”, i.e., the average number of relations per instance), (4) the number of Level-2 instance classes (“#instance classes”), and (5) the number of relation types (“#relation types”). Across all five subfigures in Fig. 7, no consistent or interpretable patterns emerge. The ROS scores exhibit considerable variability across different attribute values, with overlapping error bars and no clear monotonic trends. These results suggest that simple graph analytics alone is insufficient to account for variations in perceived restorative quality, reinforcing the need for more advanced methodologies to uncover deeper and more informative structural patterns.

Fig. 7.

Effect of five basic graphical attributes on the averge ROS scores of the ParkScape-PR images.

Methods

We propose a graph learning-based framework to quantitatively assess images and uncover structural relationships between landscape composition and psychological restoration. Figure 8 illustrates this framework, which consists of two complementary modules: Perceiver and Miner. The Perceiver module leverages graph convolutional operations to accurately predict ROS scores from each image in a manner closely aligned with human perceptual judgments. Complementing the Perceiver, the Miner module provides critical interpretability by analyzing and uncovering key features and structural patterns that contribute to predicted ROS scores. Together, these two modules offer a computational framework, systematically combining predictive accuracy with insightful structural and semantic interpretability.

Fig. 8.

Overall architecture and workflow of the proposed framework.

Design of Perceiver

Graph embedding

As described in Sect. "Dataset", images were annotated with graph structures. To make these annotations computationally tractable, we first transform them into graph embeddings to support subsequent graph convolution operations. To ease expression, we denote the graph annotation for the i-th image by , where is the set of nodes corresponding to identified instances, is the set of directed edges representing pairwise spatial relationships, and is the set of edge attributes qualitatively describing those relationships. Each node of is encoded as an 18-dimensional one-hot vector based on its class label. For unrecognizable labels, an all-zero vector is assigned. This yields a node embedding , where . is encoded into an edge embedding , where and each row vector indicates a directed edge from u to v. Given nine predefined spatial relation types, is encoded into an attribute embedding , where each row is a one-hot vector indicating the relation type. During training, a batch of b images is processed, resulting in b corresponding sets of graph embeddings.

Graph convolution

We propose a graph learning layer that scores input images by learning from their graph embeddings. This layer includes three sequential processing units: two edge-aware GCN (EaGCN) units followed by a pooling unit. They form a pipeline integrated with ReLu nonlinear activation and dropouts. The basic design principle behind this layer is to hierarchically encode node-level features into global graph-level representations by incorporating attention mechanisms into graph convolution process.

The structure of the EaGCN block is shown in Fig. 9. EaGCN begins with a GATConv unit, a widely used graph attention mechanism that captures hierarchical relationships through attention-weighted neighborhood aggregation⁶⁶. This unit aggregates node information by assigning learnable attention weights to neighboring nodes. Subsequently, a fully connected neural network (FCN), consisting of a single layer of neurons, is applied to transform and refine intermediate features. These features are then passed to an Edge-aware Attention unit, which extends the conventional attention mechanism by incorporating both node and edge attributes. This allows EaGCN to differentiate between various types of node interactions. While edge attributes often encode key structural or relational cues, in Perceiver, they primarily represent spatial orientation between nodes. To fully leverage edge information, we explicitly integrate edge features into the attention computation. This design enables the network to adaptively modulate attention weights based on the semantic content of edges, thereby enhancing its ability to capture spatial structure and improving the representational capacity of the learned graph features. Finally, dropout and ReLU activation functions are applied to prevent overfitting.

Fig. 9.

Structure of the EaGCN block.

In Perceiver’s graph convolution component, we apply EaGCN twice, motivated by the need to effectively capture both local and global structural patterns within the graph. Specifically, the first EaGCN extracts initial features and models node-edge interactions using edge-aware attention, which dynamically adjusts attention weights based on node and edge attributes. The second EaGCN refines these weights by incorporating a broader structural context across the graph. The edge-aware attention mechanism in both stacked EaGCN further enhances the model’s ability to leverage spatial and relational information encoded in edge attributes.

Depth extraction

In the depth extraction component, we employ Depth Anything V2⁶⁵, a state-of-the-art monocular depth estimation model known for its strong generalization across diverse visual scenes. However, the raw depth feature maps output by Depth Anything V2 are not directly compatible with graph features generated by the graph convolutional component, due to mismatched dimensionalities and representation formats. To this end, we introduce an extractor block that transforms the depth feature map into a structured depth matrix of size . As detailed in Table 5, the four row vectors—each with five dimensions—correspond to distinct statistical descriptors computed from depth map.

Table 5.

Information within the depth matrix.

No.	Descriptor	Statistics (in row vector)
1	depth value	mean	std	median	min	max
2	quantile	q-0.1	q-0.25	q-0.75	q-0.9	q-0.95
3	histogram	mean	std	median	min	max
4	texture	contrast	energy	homogeneity	correlation	ASM

The texture descriptor includes five statistical variables extracted from the depth feature map: contrast, energy, homogeneity, correlation, and Angular Second Moment (ASM). “Contrast” captures local intensity variation, reflecting differences between light and dark regions. “Energy” quantifies the smoothness of depth map, with higher values indicating more uniform patterns. “Homogeneity” measures the consistency of depth distribution, while “correlation” evaluates the relation between depth values in different regions. ASM captures texture roughness, where higher values typically indicate fine-grained or repetitive patterns. these metrics form a compact and informative depth matrix that balances descriptive richness with computational efficiency. Before feature fusion, the depth matrix is flattened into a 20-dimensional vector representing the global depth feature.

Feature fusion and prediction

Effective feature fusion plays a critical role in our scene understanding with multiple modalities. In Perceiver, we need to integrate the structural semantics extracted by the graph convolution layer with the depth information derived from the depth extraction layer. To this end, we introduce a feature fusion layer that ultimately generates the ROS score for the input image. Specifically, we first concatenate the depth matrix and the pooled graph feature to form a unified representation. However, due to the difference in dimensionality of these two features, direct concatenation may lead to over-representation and feature imbalance. To address this, we employ an FCN with seven neurons to both fuse the modalities more effectively and align their dimensionalities. This layer also performs implicit dimensionality reduction, mitigating the overfitting risk while enhancing the quality of the fused representation.

Following the FCN, a sigmoid activation is used to map the fused feature into a predicted ROS score vector of size , denoted as , where each element is real number falling within [0,1] for . Recall that the ground-truth ROS scores range from one to five points. We transform the normalized prediction to in the original [1, 5] scale as the final output of the Perceiver. This transformation follows the rule of for .

Perceiver training

Our Perceiver operates under a regression framework, aiming to minimize the deviation between the predicted and ground-truth ROS scores. During training, the loss function is defined as the mean squared error (MSE): , where denotes the ground-truth ROS scores of the j-th sample and is the corresponding prediction result. The model is trained using the Adam optimizer with an initial learning rate of . The batch size is set to four to balance computational efficiency with resource demands of graph convolution and feature fusion.

Design of Miner

The Perceiver module presented above is capable of predicting ROS scores from landscape images by learning the structural feature and depth feature. From a statistical learning perspective, Perceiver, hereafter denoted as , learns a conditional probability distribution , where Y is a random variable representing ROS scores and G is the graph annotation derived from the input image. As demonstrated later in Sect. "Evaluation of Perceiver", Perceiver is able to predict ROS scores with high accuracy. However, it alone cannot identify which elements or spatial patterns play a critical role in determining the restorative capacity of a landscape. Therefore, we introduce the Miner module to understand the factors that predominantly influence Perceiver’s ROS score predictions.

Principles underlying Miner

As illustrated in Fig. 8, Miner employs the trained model as a frozen, callable predictor, and takes a graph G along with its corresponding predicted ROS scores as two inputs, where . Fundamentally, Miner is a machine learning model that learns a graph mask to assess which components of G (individual nodes, edges, and subgraphs) are more important in contributing . During inference, Miner assigns soft importance weights for each node and edge of G. Furthermore, based on these weights, a weighted subgraph (referred to as a perceptual subgraph) can be extracted, by a -thresholded strategy, to represent the original full graph G. The weighted nodes, edges, and perceptual subgraph collectively serve as an interpretable explanation of the Perceiver’s prediction, offering insights into which landscape elements and structural patterns contribute most to the predicted ROS score. Our Miner is built on the GNNExplainer model⁶⁷, a widely-used post-hoc explanation method for any graph neural networks. It aims to identify graphical features that are most influential for a given prediction.

Without practical constraints, one could theoretically enumerate all possible weight assignments over the input graph G, and evaluate each candidate assignment by the trained model to identify the optimal one. However, such exhaustive search is computationally intractable, as the number of possible assignments grows exponentially with the size of G. As a learning-based scalable alternative, Miner introduces a graph mask that acts as a soft, differentiable filter over G. This approach formulates the discrete combinatorial problem of weight assignment into a continuous optimization task over the graph mask parameters, enabling efficient end-to-end learning.

Workflow and learning process of Miner

The graph mask in Miner is denoted by , and it consists of two components: an edge mask and a node mask . is an m-dimensional vector, and is an n-dimensional vector, where m and n are the number of edges and nodes in the input graph G, respectively. The mask is initialized with random reals within [0, 1], and it is treated as a set of learnable continuous parameters that encode the importance of each edge and node, and are iteratively updated during Miner learning.

Figure 10 describes Miner’s major components and iterative learning scheme. At each iteration, Miner applies an element-wise multiplication between G and the edge and node masks, formulated as and , respectively, where sigmoid operation is used for restraining mask parameters within [0, 1]. This multiplication generates soft importance weights for all nodes and edges, which collectively represent a weighted graph . By feeding into the trained Perceiver, a prediction ROS score is obtained. Thus, by comparing with , the prediction of on G, we can evaluate how well the current importance weights reflect the Miner’s interpretation, i.e., assessing the quality of current graph masks. We next articulate Miner’s loss function for learning.

Fig. 10.

Structure and learning process of the Miner module.

The core task of Miner, is to learn a graph mask capable of identify an optimal that most effectively explains the Perceiver’s prediction. Miner evaluates the importance of using mutual information, defined as , where H(Y) represents the entropy of Y and is the conditional entropy of Y given . This metric quantifies the reduction in uncertainty about Y when a weighted graph is sampled or observed. A higher mutual information value indicates that contains more informative content with regarding to the prediction of model , and is thus more likely to represent a meaningful explanatory substructure. Actually, H(Y) remains constant across all iterations because the trained Perceiver is fixed. Thus, Miner only needs to minimize the conditional entropy to optimize its graph mask. However, since is a regression model whose output ROS score vector is continuous, falling within , directly minimizing the conditional entropy is nontrivial and impractical as a loss function for Miner training. Following the design of GNNExplainer, the loss function of Miner, denoted as , consists of three components: (1) the discrepancy between and , and (2) two regularization terms derived from the -norm and entropy that are applied to the learned graph masks. In detail, this loss function is defined as

1

where and are hyperparameters to control tradeoff between two types of regularizations, and both masks are constrained within by using the sigmoid operation. In Eq. (1), lower sparsity indicates that fewer nodes and edges are selected by the mask ; similarly, low entropy reflects the reduced uncertainty in predictions. We use the Adam optimizer to minimize with respect to the two masks and .

After learning completes on G, Miner generates an optimal graph mask for G, which effectively assigns an importance weight to each node and edge of G, resulting in a weighted graph . These importance weights explain how much nodes and edges contribute to ROS scores. Subsequently, Miner applies a -thresholding policy to extract a perceptual subgraph from . This subgraph is structurally simpler than G, yet retains the most salient components that contribute to restorative benefits, thus serving as an effective and interpretable representation of G.

Generation of perceptual subgraph

Figure 11 presents two example images from ParkScape-PR, along with their corresponding graph annotations and graphical explanations generated by Miner. We now turn to deriving a perceptual subgraph from the weighted graph. This process involves two main steps: (1) selecting a subset of important nodes base on their assigned importance weights, and (2) expanding this node set by incorporating relevant edges and nodes.

Fig. 11.

Examples of Miner’s input and output. For clarity, in the two graph annotations, the relation types for edges and the instance classes for nodes are all omitted, while in the weighted graphs, node and edge weights are partially omitted.

In our study, the selection of nodes for perceptual subgraph construction is motivated by the Theory of Visual Attention⁶⁸ and the Spotlight Theory of Attention⁶⁹, both of which suggest that individuals tend to focus on a limited set of salient elements or features within complex visual scenes. To simulate this constrained cognitive focus, we select the top- nodes—the top percentile of nodes ranked by node importance—as core nodes. At this step, the perceptual subgraph includes only these core nodes, without any connecting edges.

In the subsequent expansion step, we first identify core edges, defined as edges connecting two core nodes, and put them into the perceptual subgraph. Next, we examine all non-core edges incident to core nodes, and retain those with importance above a threshold . These retained non-core edges are referred to as extended edges, and the non-core nodes they connect, as extended nodes. The inclusion of extended nodes and edges completes the construction of a perceptual subgraph. We will demonstrate the representational capacity of perceptual subgraphs, as well as the setup of parameters and , in Sect. "Evaluation of Miner".

The above expansion strategy is inspired by both theoretic and empirical findings from visual attention research, showing that visual attention can extend from focal elements to spatially adjacent regions and, under certain conditions, be even distributed across discrete zones^70,71. The advantages of perceptual subgraph are as follows.

• Alignment with human attention. The -thresholding strategy draws on principles of visual attention theory by emphasizing focal components and their salient peripheral relations, while excluding redundant elements. As a result, perceptual subgraphs closely reflect human perceptual responses, thus ensuring psychological validity.

• Enhanced interpretability. These subgraphs isolate the most influential landscape elements and spatial relationships, facilitating the identification of structures that contribute to psychological restoration. By reducing visual and cognitive clutter, they better align with the mechanisms by which humans prioritize visual information in urban park settings.

• Structural simplification with semantic emphasis. Perceptual subgraphs simplify graph topology while preserving semantically rich, high-impact components. This abstraction improves the tractability of downstream tasks such as pattern mining and visual explanation, while minimizing computational overhead.

• Effective design guidance. Beyond analysis, perceptual subgraphs provide practical insights for landscape designs. They highlight not only which elements to include but also how to spatially configure them to maximize restorative potential, by revealing recurrent high-impact structural patterns across diverse scenes.

Results

In this section, we first present a series of experiments to evaluate the performance of our proposed framework. The Perceiver and Miner modules of the framework were trained on a computer system equipped with an Intel(R) Xeon CPU (24 cores, 60 GB) and an NVIDIA RTX 4080 GPU with 16 GB of memory. We then conduct an in-depth analysis of the framework’s outputs to analyze the underlying factors that influence the ROS scores of urban park landscape images.

Evaluation of Perceiver

We evaluate Perceiver alongside an ablated variant, termed Perceiver-wo-Depth, in which the depth extraction component is removed. The parameters used for training both models are outlined in Table 6. Both models are trained for 300 epochs, ensuring convergence under stable optimization dynamics.

Table 6.

Parameter setup for Perceiver training.

Parameter	Description	Perceiver	Perceiver wo depth
input_dim	Node feature dimension	18	18
hidden_dim	Hidden layer dimension	128	128
output_dim	Output dimension (questions)	7	7
edge_dim	Edge feature dimension	9	9
global_dim	Global feature dimension	20	N/A
heads	Number of attention heads	8	8
dropout	Dropout rate	0.3	0.3
num_layers	Number of EaGCN layers	2	2

Training convergency

Figure 12 illustrates the convergence performance of Perceiver and Perceiver-wo-Depth. The Perceiver, which incorporates depth information, consistently demonstrates faster convergence and achieves a lower final loss compared to the ablated model. Specifically, the Perceiver rapidly reduces its training loss in the initial epochs and continues to improve steadily, ultimately stabilizing around 0.30. In contrast, Perceiver-wo-Depth shows slower convergence and plateaus at a higher loss near 0.35. These results highlight the critical role of depth features in enhancing model optimization and learning stability.

Fig. 12.

Convergence of the Perceiver and Perceiver-wo-Depth models.

Prediction accuracy

Table 7 presents a comparative analysis of the Perceiver model and its ablation variant, in terms of prediction accuracy measured by the root mean square error (RMSE) of ROS scores across seven psychological dimensions (Q1Q7) and their average. The Perceiver consistently outperforms the ablation model across all dimensions, achieving a lower RMSE (0.29) in average ROS score, compared to 0.41 for Perceiver-wo-Depth. This table substantiates the efficacy of incorporating depth in the Perceiver structure, affirming its value for precise landscape evaluation in the context of psychological restoration.

Table 7.

Comparison of Perceiver and Perceiver-wo-Depth in accuracy.

Model	RMSE of ROS score
	Avg.	Q1	Q2	Q3	Q4	Q5	Q6	Q7
Perceiver	0.29	0.31	0.26	0.33	0.27	0.27	0.28	0.29
Perceiver-wo-Depth	0.41	0.44	0.33	0.43	0.42	0.45	0.40	0.42

Furthermore, Fig. 13 presents the prediction accuracy of the Perceiver across four groups of test images, categorized by their ground-truth average ROS score: [1, 2), [2, 3), [3, 4) and [4, 5]. For example, an image with an average ROS score of 2.6 belongs to the [2, 3) group. The left part (violin plot) presents RMSE distributions, while the right part (bar plot) shows per-question prediction errors within each group. Across all groups, the Perceiver consistently achieves lower RMSEs than the ablation, indicating more accurate and stable predictions. The difference is especially pronounced in the low-score groups, suggesting that the full model’s depth mechanism significantly improves predictions for moderately restorative landscapes. The bar plots provide a question-level breakdown of performance, where Perceiver outperforms its ablation across all seven dimensions. In high-score groups, prediction errors are generally below 0.25—most of which fall under 0.2—relative to ground-truth, demonstrating the Perceiver’s strong capability in estimating perceived restorativeness.

Fig. 13.

Prediction accuracy of both Perceiver and Perceiver-wo-Depth under four groups of ground-truth ROS scores. RMSE comparisons across score bins [1,2), [2,3), [3,4), and [4,5) were statistically evaluated using Wilcoxon signed-rank tests and two-way ANOVA, yielding model effect F-values ranging from 41.39 to 566.63 with p-values .

Evaluation of Miner

For each graph annotation, Miner iteratively learns an optimal weight assignment, as each graph possesses a distinct optimal graph mask. Miner operates 200 iterations, using the Adam optimizer with a learning rate of . During the learning process, the Perceiver model remains frozen, while Miner optimizes its node and edge masks.

According to the -thresholding strategy defined in Sect. "Principles underlying Miner", a perceptual graph can be extracted from the original full graph. Here, a larger results in the selection of more nodes, while a smaller allows a greater number of edges to be included in the sugbraph. At the extreme setting of and , the resulting perceptual subgraph is equivalent to the original full graph. In our experiments, we empirically found that setting and can produce a subgraph that achieves high representational capacity while maintaining structural simplicity; this can be observed in Fig. 14, where the cumulative distribution curves show that over 90% of perceptual subgraphs yield prediction errors (in MSE) below 0.1 across all seven dimensions of ROS score vector and their average.

Fig. 14.

Representational capacity of (0.25, 0.5)-thresholding perceptual subgraphs of all landscape images.

Results and analyses

The preceding evaluation demonstrates that Perceiver achieves high-accuracy ROS score prediction consistent with human ratings, and Miner effectively assigns importance weights and extracts perceptual subgraphs that it preserves representational fidelity with respect to restorative outcomes. In this section, we investigate which structural features of urban park landscape influence ROS scores, by analyzing the importance weights and perceptual subgraphs generated by Miner. The analysis unfolds across three levels: individual instance classes, triplet types, and perceptual subgraphs. This supports a progressively holistic understanding of how urban park landscape composition influences psychological restoration.

Does individual instance classes matter?

To evaluate whether specific landscape instance classes are more influential for restorative outcomes, we analyze their average importance values across four ROS score groups, as illustrated in Fig. 15. Although elements such as tree, lawn, and leisure facility exhibit higher importance in high-scoring groups, they also appear prominently in lower-scoring ones. This overlap indicates that instance classes alone do not reliably differentiate restorative potential. For example, a tree beside a road may have different restorative effects than one near a fountain; a leisure facility near a natural element may be perceived differently than the same facility surrounded by civil infrastructure. Consequently, Fig. 15 highlights a limitation: relying solely on the importance of instance classes is insufficient for interpreting restorative potential. Substantial variance within groups, as indicated by error bars, further underscores the inconsistency of instance-level significance. In sum, meaningful interpretation requires integrating spatial structure and relational context beyond individual instance classes.

Fig. 15.

Importance distribution of all 18 Level-2 instance classes within images grouped by ground-truth ROS scores.

In Fig. 16, we investigate how Miner learns to assign importance weights to nodes and edges in order to identify their contribution to the ROS scores. Specifically, we track the variation of these importance weights throughout Miner’s iterative learning process for each input graph annotation. Figure 16 demonstrates this variation using two example graphs, highlighting Miner’s progressive refinements of assessment. As shown in the top-row subfigures of Fig.  16, throughout the learning process, Miner adjusts edge importance weights across a wide range, typically between 0.2 and 0.75, consistently for all four graphs representing different ROS score groups. In contrast, node importance weights exhibit a more constrained range of updates, typically within a bandwidth of approximately 0.3, indicating relatively stable refinement over iterations. Recall that in each learning iteration, Miner first applies the current graph mask to generate an updated weighted graph, which is then passed to the Perceiver for scoring. The predicted score is compared with that of the original graph to compute the loss for mask updating. The observed wide-range variation in edge importance weights indicates that Miner continuously and substantially adjusts edge relevance throughout learning iterations. This dynamic adjustment suggests that edges—each corresponding to a triplet composed of two instances and their spatial relationship—play a critical role in shaping ROS predictions. As a result, Miner explores a broader spectrum of edge weights to more effectively learn an optimal edge mask that aligns with the underlying restorative patterns in the data.

Fig. 16.

Illustration of variation in importance weights for four example graphs during Miner’s learning process. For clarity, only the weight distributions of ten nodes and edges, all of which are randomly chosen, are shown for each example graph.

Which individual triplet configurations are more contributive?

For a given triplet, we identify all matching edges across the weighted graphs output by Miner, and calculate the average of their importance weights. Figure 17 shows the distribution of average importance weights for all triplets within images from the two groups with high ROS scores. The curve displays a classic long-tail pattern: a small number of top-ranked triplets (the top 75) exhibit notably higher average importance weights (above 0.615), while the majority show moderate and gradually declining values. Interestingly, the proportion of top-ranked triplets (i.e., the minority) is very close to the Golden Ratio. The steep drop at the beginning highlights a concentrated set of highly influential triplets, suggesting that only a limited subset of spatial-semantic relationships plays a dominant role in influencing ROS scores. Beyond the top 75 triplets, the importance values of triplets flatten, showing a diminishing marginal contribution from lower-ranked triplets. This observation justifies the focus on top triplets in downstream analysis and interpretability studies. Table 8 lists the top-20 triplets within images with higher ROS scores; the triplet “sunny sky over trees” rank first, followed by the triplet “civilian architecture behind trees”. This reinforces the restorative benefit of sky visibility and spatial depth.

Fig. 17.

Average triplet importance (sorted in descending order) computed across all images in groups [3, 4) and [4, 5].

Table 8.

Top-20 triplets ranked by average importance weight.

Rank	Triplet						Rank	Triplet
	Source instance		Relation type		Target instance			Source instance		Relation type		Target instance
1	C12	sunny	T4	above	C8	tree	11	C9	shrub	T7	adja. to	C8	tree
2	C4	civil. arch.	T6	behind	C8	tree	12	C12	sunny	T4	above	C16	road
3	C8	tree	T7	adja. to	C10	lawn	13	C9	shrub	T7	adja. to	C10	lawn
4	C12	sunny	T4	above	C10	lawn	14	C10	lawn	T3	below	C8	tree
5	C10	lawn	T7	adja. to	C8	tree	15	C12	sunny	T4	above	C9	shrub
6	C13	overcast	T4	above	C8	tree	16	C8	tree	T4	above	C10	lawn
7	C8	tree	T7	left of	C16	road	17	C8	tree	T7	adja. to	C9	shrub
8	C12	sunny	T4	above	C6	civil. arch.	18	C10	lawn	T7	adja. to	C16	road
9	C16	road	T7	adja. to	C10	lawn	19	C12	sunny	T4	above	C6	civi. faci.
10	C8	tree	T2	right of	C16	road	20	C8	tree	T7	adja. to	C16	road

To provide a more comprehensive and intuitive understanding, Fig. 18 visualizes the cumulative importance weights of all possible triplets, categorized by nine distinct spatial relation types, within landscape images with higher ROS scores. Each heatmap corresponds to a specific relation type (e.g., “above”, “behind”, “right of”), and quantifies the accumulated edge importance of instance pairs (e.g., skytree, buildingplant) under that relation. Darker shades indicate stronger contributions to predicted restorative potential, offering insights that go beyond individual element importance. Clearly, the distribution of high-contribution triplets is uneven across relation types. Relations such as “adjacent to” and “in front of” exhibit relatively dense distributions of high-importance triplets, suggesting that proximate and frontal spatial arrangements are particularly influential in shaping restorative perceptions. In contrast, relations such as “around”, “on”, and “left of” show relatively sparse distributions of impactful triplets, indicating a lesser role for lateral or ground-level proximities. Next, we compare pairs of spatially symmetric relations: “left of” vs. “right of”, “behind” vs. “in front of”, and “above” vs. “below”. The heatmaps for “left of” and “right of” show nearly identical triplet importance distributions, suggesting symmetry in their influence on restorative perception. A similar pattern is observed for “above” and “below”. Notably, a comparison between the heatmaps for “behind” and “in front of” reveals a strong asymmetry of triplets in spatial influence. This asymmetry underscores the directional sensitivity embedded in visual processing, particularly in how foreground elements contribute more strongly to perceived restorativeness.

Fig. 18.

Cumulative importance of instance-relation triplets across different relation types.

Structural insights from perceptual subgraphs

Building on the analysis of individual triplets, we next examine full graphs and perceptual subgraphs to understand how higher-order structural patterns (comprising interconnected elements and their spatial relations) contribute to restorative perception. This shift enables a more holistic assessment of landscape configurations beyond isolated instance pairs. Figure 19 presents 16 graph annotations and their corresponding perceptual subgraphs that are extracted by Miner using a threshold on 16 randomly selected images. Each subgraph yields a ROS prediction with an MSE error below 0.1, indicating high fidelity in representing its original graph in terms of restorative potential. The two figures provide valuable insights into what combinatorial patterns of relevant triples shape the restorative perception in urban park landscapes.

Fig. 19.

Illustration of original graph annotations and their corresponding -thresholded perceptual subgraphs, derived from 16 example images. For clarity, edge directions are omitted.

We first compare the full original graph annotations shown in the middle gray rows of Fig. 19. This can reflect topological features that help elucidate the factors influencing ROS scores. (1) Connectivity: Low-ROS graphs tend to exhibit sparser connectivity among landscape elements, whereas high-ROS graphs demonstrate denser, more cohesive structures characterized by uniformly connected nodes. This structural coherence is indicative of greater spatial integration and is consistent with the principle of coherent scene organization proposed in Attention Restoration Theory (ART). (2) Triplet balance: High-ROS graphs exhibit more balanced and meaningful spatial triplets, which reflect layered, naturalistic compositions that may evoke soft fascination and spatial depth; in contrast, low-ROS graphs more frequently involve redundant relations and the spatial semantics are less likely to invoke restorative responses. (3) Topological complexity: While high-ROS graphs are denser, they are not chaotic. They maintain a moderate level of structural complexity that aligns with theories of optimal stimulation—rich enough to sustain attention, but not overwhelming. Low-ROS graphs may be either overly simplistic (failing to engage the viewer) or separately clustered, which can disrupt perceptual fluency. (4) Topological shape: Low-ROS graphs are more linear or unbalanced, lacking a clear center of perceptual gravity, while high-ROS graphs often form radial or clustered shapes, indicating a strong center or focal region surrounded by supportive elements—a common motif in restorative natural scenes.

As discussed in Sect. "Generation of perceptual subgraph", focusing on perceptual subgraphs offers distinct advantages over analyzing full-scene graphs. We next examine the perceptual subgraphs acros varying ROS score groups, as visualized in the bottom light-green rows of Fig. 19. The comparative analysis reveals several key structural and semantic differences. (1) Instance diversity: High-ROS subgraphs typically feature a greater variety of natural elements (e.g., trees, lawns, flowers, water), which frequently appear as central nodes with high relative importance. In contrast, low-ROS subgraphs are often dominated by more built or inert elements (e.g., architectural features, sidewalk), displaying less diversity in instance classes. (2) Structural balance and symmetry: High-ROS subgraphs tend to display visually balanced or radially distributed structures, where nodes are evenly distributed and symmetrically connected around a central focal element. This configuration suggests a higher degree of perceptual coherence. Low-ROS subgraphs, by contrast, often present linear or sparse arrangements, often lacking a clear central element, which may lead to fragmented or less engaging visual experiences. (3) Relation richness: In high-ROS subgraphs, core nodes are typically linked to multiple peripheral nodes via semantically meaningful and spatially diverse relations (e.g., “adjacent to”, “above”, “in front of”). These layered relationships contribute to immersive and complex scene compositions. Low-ROS subgraphs, on the other hand, exhibit fewer and more homogeneous relations, indicative of simple and monotonous spatial structures. (4) Edge saliency: High-ROS subgraphs feature strong, consistently colored edges, denoting stable and confident attribution of spatial importance and suggesting well-defined structural hierarchies. In contrast, low-ROS subgraphs contain weaker, more diffusely colored edges, indicating lower structural cohesion and perceptual clarity.

These above observations are consistent with those observed in the full original graphs. Importantly, these findings indicate that restorative urban park landscapes are not defined solely by the presence of individual elements, but by the richness, coherence, and connectivity of their underlying structural configurations—features that are prominently revealed in perceptual subgraphs. The perceptual subgraph thus functions as a distilled yet faithful representation of the restorative core of a landscape, offering both psychological interpretability and structural fidelity.

To clarify which instance classes are more influential for restorative perception, we further analyze perceptual subgraphs extracted from high-ROS images using chord plots (Fig. 20). These plots visually summarize the occurrence frequency of each landscape element (instance) along with their typical relational pairings. As illustrated in Fig. 20, spatial relations such as “above”, “adjacent to”, “in front of” and “behind” show notably dense connections. This suggests that vertical layering, proximity, and foreground-background interactions among landscape elements are particularly significant factors contributing to psychological restoration. Several instance classes appear frequently and prominently across all relation types. Among the nine chord plots, instance classes such as C8/“tree” and C10/“lawn”, and C16/“road”, have dense interactions across multiple relation types. Furthermore, most of these instances are natural elements. Their frequent involvement and interaction indicate a central role in restorative compositions and offer direct insight for designers. Some interactions are strongly directional or asymmetric. For example, the “above” relation often flows from sky-related elements (e.g., sunny sky) toward vegetation or built elements, highlighting the importance of open vertical space in restorative perception. Similarly, relations like “behind” or “in front of” illustrate clear foreground-background arrangements. The spatial relation “around” appears notably sparse, implying this type of spatial arrangement rarely contributes significantly to perceived restorative quality, possibly due to its inherent ambiguity or limited perceptual saliency.

Fig. 20.

Frequencies and interaction of instance class in perceptual graphs across ROS score groups [3,4) and [4, 5].

We analyze the centrality of each instance class in high-score perceptual subgraphs, using three standard graph-theoretic measures: degree centrality, closeness centrality, and eigenvector centrality. Degree centrality quantifies the number of direct connections a node has, reflecting its local connectivity. Closeness centrality captures how near a node is to all other nodes in the graph based on the shortest path lengths, indicating its overall positional centrality. Eigenvector centrality measures a node’s influence by accounting for the importance of its neighbors, thereby capturing its role within the broader graph structure. Several notable patterns emerge from Fig. 21. Instance classes such as C7/“leisure facility”, C8/“tree”, C9/“shrub”, C14/“exercising person”, and C16/“road” exhibit high degree and closeness centralities, indicating these landscape instances frequently form directed connections to many others and serve as key hubs for shortest paths. Their topological prominence suggests a central role in organizing spatial interactions within restorative environments. High eigenvector centrality values observed for classes such as C6/“civilian facility”, C7/“leisure facility”, C12/“sunny sky” and C14/“exercising person” emphasize their structural significance, reflecting strong connections with other highly influential instances in the subgraph. Classes including C3/“fountain”, C15/“sitting person”, C17/“pet”, and C18/“wild animal” demonstrate consistently low values across all three centrality measures. This suggests their limited involvement in the critical restorative interactions.

Fig. 21.

Centralities of the instance classes within the perceptual subgraphs of ROS score groups [3, 4) and [4, 5].

Discussion

Achievements

In this paper, we present a novel graph learning-based framework, which can systematically explore the structural relationships between urban park landscapes and psychological restoration. This methodological innovation addresses critical limitations in previous approaches that pay less attention to structural interrelationships. In the proposed framework, Perceiver leverages graph convolutional neural networks and depth extraction to accurately predict restorative potential, paving a machine learning way to achieve a human-aligned perception in psychological restoration. On the other hand, Miner effectively interprets Perceiver’s predictions, by explicitly identifying influential nodes and edges, effectively bridging computational models with psychological insights. Moreover, the perceptual subgraphs produced by Miner provide simplified yet representative structural insights, reflecting the theoretical proposition that spatial structures of landscape elements influence human psychological restoration. This interpretability is crucial for actionable urban park landscape design, enabling stakeholders to make evidence-based decisions about spatial arrangements and element combinations. Moreover, it facilitates the design of environments that are better aligned with mental health needs, thereby helping to mitigate urban stress and promote broader community well-being.

In addition to proposing a novel framework for understanding influential elements and spatial structures, our study successfully leverages LLMs to facilitate dataset annotation. Specifically, by designing a targeted prompting strategy, we leverage the reasoning and contextual understanding capabilities of LLM to interpret spatial relationships among pre-segmented landscape elements. Compared to traditional annotation approaches, our LLM-assisted method considerably reduces human labor and time cost, while maintaining semantic richness and structural consistency across the dataset. This not only streamlines large-scale dataset development but also exemplifies the practical synergy between vision-language models and graph-based spatial representation, offering a scalable and generalizable pathway for future research.

Our findings suggest several practical strategies for configuring restorative park environments. First, designers should favor compositions where vegetation frames, rather than blocks, mid- and long-range views, balancing refuge and prospect. Second, integrating water features, continuous canopy edges and gently curved paths appears especially beneficial, whereas excessive hard infrastructure or visual clutter may weaken restorative effects. These scene-level patterns can be translated into concrete guidelines for siting vegetation, water and amenities in future park designs.

Limitations

Our framework is currently limited to the domain of visual perception. Other sensory modalities—such as auditory and olfactory stimuli, which are known to influence restorative experiences in urban environments^72–74—have not been yet incorporated. This study is based on isolated landscape images of urban parks, which do not capture the spatiotemporal continuity of real-world environments as experienced by urban dwellers. To more robustly evaluate the framework, future studies should incorporate landscape imagery with spatiotemporal correlations that reflect the scene continuity from the viewer’s perspective. Additionally, we primarily focus on the classes of landscape elements and their spatial relationships, with less attention given to their proportional presence within the viewer’s visual field, a factor shown in prior studies to influence restorative perception^36,38,51. Our framework is not in conflict with such findings; rather, it is complementary. Integrating the proportional presence of landscape elements does not require architectural modifications to our framework. Such proportional information can be quantified through semantic segmentation and encoded as an additional node feature for graph annotations.

The developed dataset ParkScape-PR provides a diverse collection of urban park landscape images, and our results on this dataset demonstrate a powerful proof-of-concept for the machine learning framework. However, the current dataset lacks cultural and seasonal diversity and a broad subject pool, which may constrain its ability to capture global variations in urban park design and restorative experiences. Future expansion of the dataset to include parks and participants from different cultural and geographic contexts can address these limitations without requiring any modification to the framework itself. Another limitation of this study is the absence of a formal integration between the observed patterns and established psychological theories in visual attention and restoration. Our findings intuitively align with concepts such as focal attention, spatial coherence, and element saliency; however, we have not yet established a rigorous mapping to these theoretical constructs. Addressing this gap will be a valued avenue for future research, which could strengthen the theoretical grounding of our framework and enhance its interpretability and practical applicability in urban park landscape design.

Theoretical implications

Our framework does more than predict restorative potential; it offers a new language to formalize and extend the theoretical premises of ART and SRT. By distilling landscapes into networks of spatial relations (i.e., the perceptual graphs), we shift the theoretical focus from restorative elements to restorative configurations, suggesting that psychological efficacy is encoded in a scene’s underlying spatial syntax.

First, we provide a structural operationalization of key ART constructs. The higher graph connectivity and cohesion characteristic of high-ROS scene quantitatively define what constitutes “coherent organization”—a property hitherto described qualitatively. Furthermore, the predominance of relations like “above” and “in front of” in influential triplets delineates the layered depth that materially creates the sense of ART’s extent, essential for fascination. Crucially, our model reveals that coherence is not merely visual homogeneity but a specific topological property of the perceptual graph.

Second, our approach refines the prospect-refuge theory from a configurational typology to a relational one. High-importance subgraphs often encapsulate a refuge node (e.g., a tree or pavilion) spatially linked via “in front of” or “adjacent to” relations to a prospect node (e.g., a lawn or water body). This suggests the theory’s restorative effect arises not from the mere co-presence of enclosed and open spaces, but from their specific, legible spatial linkage as captured in the graph structure.

Ultimately, the most significant theoretical implication is the primacy of relations over elements. An element’s restorative contribution is conditional on its relational context—“a tree behind a building” differs fundamentally from “a tree in front of a lake”, a distinction our triplet analysis captures but element-centric approaches miss. Thus, our graph-based formalism moves the field toward a more predictive, mechanistic understanding of restoration, where theories can be expressed and tested as hypotheses about spatial graph topology.

Future research grounded in this study

In the future, our framework can be extended in several directions to deepen the understanding of restorative urban landscapes and enhance its practicality and potential for integration with emerging smart city applications.

Keeping human in the loop. Collecting diverse urban landscape data is critical to understand the the relationship between urban landscapes and their restorative benefits. A promising extension for our framework is to deploy it on smartphones, leveraging the mobile crowd sensing (MCS) paradigm^51,75 to enable urban-scale data collection. With MCS, park visitors—or urban citizens more broaderly—act as active participants, using their smartphone to capture landscape settings they are experiencing. Modern smartphones are typically equipped with camera (visual sensor), microphone (acoustic sensor), and GPS unit (location sensor), enabling the collection of rich, multimodal data. In addition, users can annotate their experiences with self-reported emotions or reflections (via text or voice input), providing valuable ground-truth labels for restorative assessment. The MCS-powered data collection keeps human in the loop and thus, could allow us to capture fine-grained dynamics of human-environment interactions across time and space. Our framework is lightweight and efficient, implemented as a compact software module of approximately 1.5 MB. On devices with moderate computational resources, Perceiver generates ROS predictions in ~0.5 seconds per image, while Miner constructs perceptual subgraphs in ~2 seconds. This computational efficiency makes the framework well-suited for smartphone deployment, enabling real-time, user-driven, and scalable collection of restorative landscape data.

Integrating more human senses. Another promising and feasible extension is to incorporate auditory perception into our framework. Auditory stimuli, such as birdsong, flowing water, or the rustling of leaves, enhance restorative experiences in natural environments⁷⁶. These sounds can be passively detected via the smartphone’s microphone or actively reported by users through text input. Such integration would involve real-time audio data collection and processing via machine learning models capable of sound classification. The resulting sound labels could be appended as attributes to existing instances or introduced as additional instance classes within graph annotations. This enhancement does not require fundamental changes to our framework. Instead, it can be implemented through dataset augmentation or multimodal feature fusion, enabling a seamless extension of Perceiver and Miner to incorporate auditory perception.

Enhancing interpretation through interaction. While our framework achieves high predictive accuracy and generates representative perceptual subgraphs, interpretability remains essential—particularly in urban planning contexts where stakeholders require clear, actionable insights. To this end, our framework could be extended with interactive features that allow users to intuitively explore why specific landscape elements or spatial configurations contribute to psychological restoration. Such interaction would facilitate connections between graph-derived visual patterns and underlying cognitive mechanisms from environmental psychology, supporting both psychologists and urban designers. Users could dynamically adjust threshold parameters in real time to observe how these variations affect perceptual subgraph construction and the resulting restoration outcomes. Additionally, the framework could incorporate predictive simulations to assess the potential restorative benefits of hypothetical or what-if landscape modifications. These capabilities would support informed, data-driven decision making grounded in interpretable and scenario-based feedback.

AI-generated designs of urban park landscapes. One promising direction for future research lies in leveraging the findings of this study to support generative design in urban park landscape planning. By identifying spatial structures associated with higher restorative potential and representing them through graph-based encodings, our framework offers a transferable and interpretable design language. Urban park planners and designers could use these high-ROS graphical representations as templates or constraints within AI-generated content (AIGC) tools for landscape synthesization or simulation⁷⁷. This extension would enable rapid visual prototyping and feedback, allowing stakeholders to explore, evaluate, and iterate on restorative design concepts more efficiently and with data-driven guidance.

Conclusion

This study introduces a graph-based computational framework for modeling the psychologically restorative potential of urban park landscapes. By combining perceptual encoding and structural reasoning, the proposed framework not only predicts restorative outcomes with high accuracy but also provides interpretable, structural explanations grounded in environmental psychology. The introduction of the ParkScape-PR dataset, along with a scalable LLM-assisted annotation pipeline, advances methodological capabilities for analyzing landscape imagery at the level of human visual experience. Our findings highlight the critical role of spatial structure—beyond the mere presence of natural elements—in shaping restorative perception. Importantly, the graph representation methodology developed in this study offer a transferable, interpretable format that can inform evidence-based and data-driven urban park design. By identifying structural patterns linked to high restorative value, this study lays the groundwork for generative design approaches that merge environmental cognition with computational creativity. More broadly, this study not only offers a new lens for understanding the relationship between urban park landscape and mental well-being. It also contributes to a growing interdisciplinary convergence between computational modeling, environmental psychology, and human-centered design—an intersection that promises to reshape how we envision, evaluate, and construct restorative urban spaces in increasingly complex cities.

Author contributions

Y.Z. and G.S. conceived the original idea and overall concept. Y.Z. and Y.L. led the manuscript writing. Y.Z., Y.L. and Y.Y. developed the ParkScape-PR dataset and led the model implementation and experiments. Y.J. and Y.Z. designed the restorative benefit questionnaire and coordinated the survey study.

Funding

This work was supported by the Fundamental Research Funds for the Central Universities of China (No. BFUKF202528).

Data availability

The datasets generated in this study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.