Revealing the Relationship Between Urban Park Landscape Features and Visual Aesthetics by Deep Learning-Driven and Spatial Analysis

Abstract

Urban parks are an important component of public urban spaces, which directly impact the living experiences of residents and the urban image. High-quality urban parks are crucial for enhancing the well-being of residents. This study selected Fukuoka, Japan, as the study site. Five urban parks were chosen to evaluate landscape visual quality by using the Scenic Beauty Estimation (SBE) method. The Semantic Differential (SD) method was used to get sample subjective landscape features. Meanwhile, sample objective landscape features were obtained by using semantic segmentation techniques in deep learning and combined with spatial analysis to understand their distribution. A regression model was established, which used the SBE values as the dependent variable and subjective landscape features as the independent variables to analyze the relationship between urban park landscape visual quality and subjective landscape features. The regression analysis revealed that sense of layering, harmony, interestingness, sense of order, and vitality were the core factors influencing visual quality. All five features had a significant positive impact on landscape visual quality. The sense of order was the most influential factor, which would be the key to enhancing the landscape perception experience. Moreover, the XGBoost model and SHAP value from machine learning were used to reveal the nonlinear relationships and significant threshold effects between urban park visual quality and five objective landscape features: openness, greenness, enclosure, vegetation diversity, and Shannon–Wiener diversity index. This study showed that when openness exceeded 0.27, the positive effect was significant. The optimal threshold for the greenness was 0.38. Vegetation diversity and enclosure had to be below 0.82 and 0.58, respectively, to have a positive impact. Meanwhile, the positive influence of the Shannon–Wiener diversity index reached its maximum at a value of 1.37. This study not only establishes a systematic method for diagnosing landscape problems and evaluating landscape visual quality but also provides both theoretical support and practical guidance for urban park landscape optimization.

1. Introduction

In the context of rapid urbanization, it is estimated that by 2050, 70% of the world’s population will live in urban areas. At the same time, this trend has prompted a renewed examination of the key factors essential for creating livable and healthy urban environments [1]. Urban parks are an important component of urban green spaces, which not only provide residents with opportunities to connect with nature but also offer space to encourage physical activity [2,3,4,5] and other beneficial behaviors. The urban park landscape environment influences people’s perceptions [6] and feelings [7], which play a significant role in impacting human perception of urban spaces. It also affects individuals’ psychological states and behaviors. Moreover, high-quality landscape environments can effectively relieve stress and anxiety, thereby enhancing overall well-being [8]. Among various types of perception, visual perception accounts for approximately 80% of the information humans acquire. It plays a crucial role in impacting individuals’ experiences of natural environments [9]. Therefore, the urban park visual landscape design is important in enhancing resident environmental perception and psychological comfort.

With the growing public interest in the quality of the living environment, as an important component of urban environmental experience, the correlation between visual environment and landscape features has increasingly become a research hotspot [10,11,12]. In recent years, more studies have begun to explore the impact of visual quality on human [13,14], as well as the ways and evaluation mechanisms by which different landscape features affect visual quality [9,15,16]. Research in this field has gradually expanded traditional aesthetic evaluation to incorporate multidimensional perspectives, including psychological restoration [17], social identity [18], and ecological conservation [19].

Meyer Granbastien et al. [20] found that a high-quality urban park landscape environment can effectively promote psychological recovery, highlighting the regulatory role of landscape spatial structure on psychological effects. In discussions of visual landscape quality, landscape features play an important role as influencing factors. Wartmann et al. [21], based on a national survey in Switzerland, pointed out that resident evaluations of visual landscape quality were influenced by individual factors such as length of residence and degree of visual openness, as well as landscape features. In research on park landscape features, scholars have generally focused on how different spatial elements affect the visual experiences of users.

By integrating image recognition technologies to extract objective landscape features and further quantify features such as green view index, spatial openness, and vegetation diversity [22], it is possible to assess the visual quality of landscape environments and promote the vectorization and spatialization of visual quality evaluation [15,23]. At the same time, visual quality is also influenced by subjective perception. Studies have shown that people’s perception of landscapes is typically based on a comprehensive evaluation of features such as naturalness, sense of order, layering, and tranquility [24,25]. Therefore, understanding how different landscape features influence visual perception is important for optimizing landscape design and enhancing the quality of urban spaces. Some studies have assessed park quality by focusing solely on a single or a limited number of objective features, such as landscape elements and infrastructure. However, beyond these objective factors, subjective dimensions—such as human needs and user perceptions—also play a crucial role in defining park quality. In recent years, studies have emerged that combine subjective landscape features with objective landscape features [26,27], but their overall number remains limited. Further research is needed to fully reveal the integrated impact of landscape features on visual quality.

Now, lots of research methods are used for landscape visual evaluation, including the Scenic Beauty Estimation (SBE) method, BIB-LVJ, Analytic Hierarchy Process (AHP), and Psychophysiological Index (PPI) testing, among others [28]. Among these, the SBE method is widely regarded as the most accurate and rigorous methodology [29,30,31]. This method was proposed in 1976 by the American environmental psychologists Daniel and Boster [32]. Due to its scientific and flexible nature, the SBE method has been widely applied in studies of visual landscape evaluation. Xuehong Tan et al. [33] applied the SBE method to the aesthetic evaluation of plant landscapes, showing that it effectively quantifies public visual perception and preferences for different types of plant landscapes [34].

Based on an analysis of the factors influencing landscape perception, Tempesta evaluated the agricultural landscape of the Veneto Plain in Italy and developed a mathematical model for assessing landscape aesthetic quality [35]. The Semantic Differential method (SD method) is a psychological technique commonly used to measure subjective perceptions and attitudes toward specific objects. It was proposed by Osgood and his colleagues. In landscape feature research, the SD method is widely applied to assess subjective cognition and emotional responses to visual environments [36]. It offers advantages such as ease of use [37], a high degree of quantification, and strong comparability of results [38]. Sun [39] systematically evaluated the subjective perceptions of resident for visual landscape quality in urban waterfront parks using both the SBE and SD methods and quantified 22 subjective landscape features. Cao applied the SD method to assess the quality of urban waterfront landscapes from the perspective of users’ psychological perception [40]. It further confirms the objectivity and validity of the SD method in landscape evaluation. Overall, the greatest advantage of the SD method lies in its ability not only to study visual aesthetics but also to quantitatively analyze various types of sensory experiences beyond vision. This is difficult to achieve with other traditional methods [41].

With the age of the big data era, traditional data collection methods, which often rely on extensive field surveys, are gradually becoming obsolete due to their time-consuming nature, high labor demands, and costly implementation. The rapid development of remote sensing technology, computer vision, and deep learning algorithms, along with the widespread accessibility of image data resources such as Google Maps [42] and Baidu Maps [43], has opened a new way to overcome the limitations of traditional methods [44]. For example, Xia et al. [45] used panoramic images to assess the visual accessibility of streetside greenery in Mizushima, Japan.

At the same time, with the emergence of advanced deep learning algorithms such as SegNet and DeepLab [46], the automatic recognition of landscape elements has become more efficient and accurate [47]. Using high-resolution imagery, Sun et al. [48] successfully identified and analyzed various streetscape elements in Los Angeles County. Similarly, Lyu et al. [49] employed semantic segmentation techniques based on deep learning to extract the Green View Index (GVI) and analyzed the relationship between perceived greenery and visual comfort (VICO) along coastal streets in Qingdao. These technological advancements have brought unprecedented operability and precision to the analysis of large-scale park landscape features.

Existing studies have shown that landscape features such as greenery and openness often have nonlinear effects on visual quality [50,51]. However, many studies still rely on linear models for analysis, making it difficult to effectively capture the complex nonlinear relationships and potential interaction mechanisms between landscape features and visual quality. In addition, quantitative research on threshold effects remains limited, and there is a lack of interpretable nonlinear modeling approaches to uncover the mechanisms by which key variables influence visual perception. Therefore, to more accurately reflect the mechanisms of visual perception, it is necessary to introduce nonlinear models that can capture complex relationships and potential interaction effects among variables. With the advancement of artificial intelligence, new perspectives and tools have been introduced into urban spatial research. Machine learning methods—such as Gradient Boosting Decision Trees (GBDTs), the ensemble learning algorithm XGBoost, Random Forests (RFs), and Multilayer Perceptrons (MLPs) [52,53,54]—are increasingly applied in urban studies due to their strong data processing capabilities and high predictive accuracy.

Among these methods, the XGBoost (Extreme Gradient Boosting) model has shown superior fitting ability and predictive performance in handling high-dimensional, nonlinear, and multivariable interaction structures [55,56,57]. To enhance model interpretability, many studies have further employed SHAP (SHapley Additive exPlanations) to interpret model outputs, thereby revealing the importance of various landscape elements in visual quality perception as well as identifying threshold effects [58,59].

The main research questions addressed in this study are the following: 1. Among the subjective landscape features, which factor has the greatest impact on visual quality? 2. What is the mechanism through which subjective landscape features influence the visual quality of park landscapes? 3. Among the objective landscape features, which factor has the greatest impact on visual quality? 4. What are the nonlinear relationships and threshold effects between objective landscape features and the visual quality of urban parks?

To address the above research questions, this study takes urban parks in Fukuoka City as the research site to explore the complex relationship between landscape features and visual quality. The SBE method is used to obtain visual quality ratings of park landscapes. Subjective and objective landscape features are extracted using the SD method and deep learning-based semantic segmentation techniques. In terms of analytical methods, a linear regression model was first employed to examine the impact mechanism of subjective landscape features on visual quality. Additionally, machine learning techniques were introduced. The XGBoost model combined with the SHAP (SHapley Additive exPlanations) interpreter reveal the nonlinear influence patterns and threshold effects of objective landscape features on visual quality. The complementary analysis of subjective and objective features provides a new methodological framework and support for gaining a deeper understanding of the relationship between landscape features and visual quality in urban parks. This study integrates subjective evaluations with objective computational methods to propose a comprehensive landscape assessment framework, enriching theoretical research on landscape quality and visual aesthetics. By introducing deep learning and interpretable machine learning techniques into landscape studies, it enhances the accuracy and interpretability of landscape feature identification and analysis, providing a scientific basis for the design, renewal, and management of urban park landscapes.

2. Materials and Methods

2.1. Study Site and Data

Fukuoka is in the northern part of Kyushu Island, Japan, along the Hakata Bay. It is the largest city in Kyushu and has a unique geographical environment and urban structure [60,61]. Its favorable climatic conditions have contributed to the development of numerous urban parks in the city.

This study selected five representative urban parks located in the central area of Fukuoka City as research sites (Figure 1): Ohori Park, Maizuru Park, Nishi Park, Suzaki Park, and Island City Central Park. These parks differ in terms of their geographical locations, spatial structures, vegetation compositions, and landscape characteristics, providing both representativeness and contrast. This is beneficial for a deeper investigation into the relationship between the landscape visual quality of urban parks and features.

In recent years, the rise of big data on the Internet has provided abundant data resources for deep learning applications. Emerging online images supplied by providers such as Google, Baidu, and Tencent have been widely used in many fields such as architecture and environmental studies [62,63]. Among them, Google Maps is one of the largest online mapping platforms [64], allowing users to access static landscape images from various perspectives. OpenStreetMap (OSM) is a free, editable, and open-source global geographic information database [65].

In this study, the basic road network data of the five urban parks were first obtained from OSM. Using the “Construct Points” function in GIS, a total of 231 sample points were generated at intervals of 50 m. To accurately simulate human visual experience and closely approximate the pedestrian perspective, the pitch for the image was set to 0°, and the headings were set to 0°, 90°, 180°, and 270° (Figure 2). Python version 3.9 was used to obtain four landscape images in the horizontal direction for each sample point (

Table A1. The complete set of parameters required to obtain street view images from different angles.

Parameter Name	Description
fovy	Vertical field of view; higher values represent a wider viewing angle
quality	Image quality; range typically from 0 to 100, higher values give clearer images
heading	Horizontal rotation angle (0–360°); defines viewing direction (0 = North, 90 = East, etc.)
pitch	Vertical angle (−90 to 90); 0 represents a horizontal view
width	Image width in pixels
height	Image height in pixels
lng	Longitude
lat	Latitude

in Appendix A). Additionally, the “Timeline” function of Google Maps was employed to adjust the imagery, ensuring that the collected landscape images were primarily from the summer season, thereby effectively decreasing the influence of seasonal variations on the research results.

2.2. Research Framework

In this study, the objective landscape feature data had a large sample size, broad coverage, and rich variable dimensions, making it suitable for modeling with nonlinear approaches. This allows for a more effective exploration of the complex relationships between landscape elements and visual quality, as well as the potential nonlinear threshold effects among variables. In contrast, subjective landscape features were obtained through questionnaire surveys and SD scales. In this study, representative and typical sample points were selected for manual scoring. A multiple linear regression model was employed to quantitatively model the relationship between subjective landscape features and visual quality. Both models in this study construct a systematic and comprehensive evaluation framework for the visual quality of urban parks. The experimental workflow is shown in Figure 3.

2.3. Landscape Visual Quality Based on the SBE Method

Previous studies have shown that public aesthetic preferences can be well represented by student samples [66]. Due to their high task completion efficiency and strong participation willingness, students have been widely adopted as survey participants in related research [67,68]. Moreover, students from different academic backgrounds exhibit similar preferences for landscape visual quality (LVQ) [69]. In this study, the sample consisted of 89 students from various backgrounds, including architecture (37 students), landscape architecture (29 students), and environmental design (23 students). Students with various academic backgrounds exhibit varied perceptual focuses: architecture students tend to pay more attention to spatial structure, landscape architecture students are more sensitive to natural elements and ecological perception, while environmental design students place greater emphasis on color, form, and overall visual harmony. The participation of students from diverse disciplines contributes to a more comprehensive representation of public perceptions and preferences regarding landscape visual quality from multiple perspectives. During the evaluation process, identical slide shows were presented to each one. Before the experiment, all participants took up training to ensure they had a basic understanding of the assessment tasks (Appendix B).

Subsequently, the observers independently completed paper-based evaluation questionnaires based on their personal preferences, with communication strictly prohibited around the process. To help establish a unified evaluation standard among participants, a video was shown before the formal evaluation, presenting typical samples and background information of urban parks in Fukuoka City. During the evaluation phase, each landscape image was displayed for 15 s before advancing to the next, and participants rated the images using a seven-point Likert scale (1 = “strongly dislike” to 7 = “strongly like”). In total, 6 invalid questionnaires were excluded, and 85 valid questionnaires were retained for data analysis. Each participant’s scores were standardized into z-scores using a general formula (Equation (1)), which was subsequently used to calculate the final SBE value for each sample photo.

$$Z_{ij}={\displaystyle \frac{R_{ij}-{\overline{R}}_i}{S_i}}$$

(1)

$$Z_j={\displaystyle \frac{\sum Z_{ij}}{N_j}}$$

(2)

In the formulas, $Z_{ij}$ represents the standardized evaluation score given by the evaluator $i$ for the photo of the sample point $j$; $R_{ij}$ represents the actual evaluation score given by the evaluator $i$ for the photo of the sample point $j$; ${\overline{R}}_i$ represents the average actual evaluation score given by the evaluator $i$ for all sample point photos; $S_i$ represents the standard deviation of the actual evaluation scores given by the evaluator $i$ for all sample point photos; $Z_j$ represents the mean standardized evaluation score for the sample point j photo across all evaluators; $N_j$ denotes the number of evaluators who rated the photo of the sample point $j$.

2.4. SD Method and Subjective Landscape Features

The SD method was used in this study to measure subjective landscape features. Additionally, the Delphi method was used to consult experts in the field of landscape architecture, and based on the current conditions of urban parks in Fukuoka City, eight subjective landscape features were ultimately selected as evaluation indicators (

Table 1. Summary table of subjective landscape features.

Evaluated Item	Adjectives in Pairs	Description
Interestingness	Boring–interesting	It reflects whether the image can capture attention, spark imagination, or leave a strong impression.
Tidiness	Dirty–tidy	It reflects whether the image appears clean and gives a sense of comfort.
Tranquility	Noisy–tranquil	It reflects whether the visual elements create a quiet and peaceful atmosphere.
Vitality	Unvital–vital	It refers to whether the image conveys a sense of liveliness and visual energy.
Sense of layering	Unclear layering–clear layering	It refers to the clarity with which visual elements are separated in spatial depth.
Sense of order	Disorderly–orderly	It refers to whether the elements in the image are arranged in an orderly manner, following visual logic or compositional rules.
Color richness	Monotonous–colorful	It reflects whether the image has visual appeal, emotional expression, and aesthetic value.
Harmony	Inharmonious–harmonious	It refers to the overall coordination and unity in composition, color, and element arrangement within the image.

). Because the landscape quality needs detailed analysis and assessment, and to enhance operational feasibility, the number of sample images was reasonably controlled [70]. Experts selected 80 representative sample points based on differences in landscape quality (high, medium, low), including 30 from Maizuru Park, 20 from Ohori Park, and 10 each from Nishi Park, Suzaki Park, and Island City Central Park.

Typically, 20 to 50 assessors with relevant expertise are required for SD evaluations; in this study, 35 participants were collected, including 15 graduate students and 20 undergraduate students majoring in landscape architecture. During the assessment, adjectives were used according to a 5-point Likert scale. For example, the rating for “interestingness” ranged from 1 (“boring”) to 5 (“interesting”), with higher scores indicating that the perceived characteristics of the landscape were closer to those described by the adjective on the right side. After the evaluation, 30 valid questionnaires were collected and used for data analysis.

2.5. Semantic Segmentation and Objective Landscape Features

With the rapid development of artificial intelligence technologies such as deep learning, the limitations of traditional image processing in accuracy and efficiency have been effectively addressed and improved [70]. Semantic segmentation technology enables the detailed classification of regions extracted from images based on landscape features [71]. This is difficult to achieve using traditional image processing methods. In addition to identifying and extracting green spaces, it can also accurately distinguish areas such as the sky, buildings, and water, thereby providing a basis for the evaluation of landscape quality. Common semantic segmentation models include FCN, SegNet, U-Net, and the DeepLab series [72]; all of these improve segmentation accuracy and the restoration of spatial details through different structural designs. In this study, DeepLabV3+ was used to divide landscape elements within images. DeepLabV3+, proposed by the Google group, is an advanced semantic segmentation model designed for pixel-level object recognition in images [73]. It is currently one of the most widely used and effective models in the field of image segmentation.

DeepLabV3+ was used for semantic segmentation of landscape elements such as trees, buildings, fences, roads, and sidewalks, based on the ADE20K dataset [74]. Based on previous studies, five key objective landscape features were identified as shown in

Table 2. Objective landscape features of the five parks.

Feature	Formula	Expression	Definition
Openness	$O_i={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n{S}_i}\left\{i\in \left(1,2,\mathrm{\dots },n\right)\right\}$	$S_i$ denotes the proportion of sky pixels.	It refers to the ratio of sky pixels to the total number of pixels.
Greenness	$G_i={\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n{T}_i}+{\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n{P}_i+\frac{1}{n}{\displaystyle {\sum }_{i=1}^{n}G{1}_i}}$	$T_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{tree}\mathrm{pixels},P_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{plant}\mathrm{pixels},\mathrm{and}G{1}_i$ denotes the proportion of glass pixels.	It refers to the ratio of plant, tree, and grass pixels to the total number of pixels.
Enclosure	$E_i={\displaystyle \frac{\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{B}_i}+\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{T}_i}+\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{P}_i}}{{\displaystyle \frac{1}{n}{\sum }_{i=1}^n{R}_i+\frac{1}{n}{\displaystyle {\sum }_{i=1}^{n}G{1}_i}+{\displaystyle \frac{1}{n}{\sum }_{i=1}^n{W}_i}}}}$	$B_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{building}\mathrm{pixels},T_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{tree}\mathrm{pixels},P_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{plant}\mathrm{pixels},R_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{road}\mathrm{pixels},W_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{water}\mathrm{pixels},\mathrm{and}G{1}_i$ denotes the proportion of glass pixels.	It refers to the ratio of vertical to horizontal surfaces in the park.
Vegetation diversity	$V{D}_i=1-\left[{\left({\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n{T}_i}\right)}^2+{\left({\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n{P}_i}\right)}^2+{\left({\displaystyle \frac{1}{n}}{\displaystyle {\sum }_{i=1}^n{G}_i}\right)}^2\right]$	$T_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{tree}\mathrm{pixels},P_i$$\mathrm{denotes}\mathrm{the}\mathrm{proportion}\mathrm{of}\mathrm{plant}\mathrm{pixels},\mathrm{and}G{1}_i$ denotes the proportion of glass pixels.	Vegetation diversity indicates the variety and composition of plant elements present in each landscape environment.
Landscape diversity index	$SHD{I}_i=-{\displaystyle {\sum }_{i,j=1}^s(P_{i,j}\times {\mathrm{InP}}_{i,j})}$	$P_{i,j}$ denotes the proportion of pixels occupied by the j-th landscape element in image i.	It refers to the diversity of various elements within the landscape image.

, all derived from the segmented landscape elements.

2.6. Statistical and Machine Learning Techniques

XGBoost is an efficient, flexible, and scalable Gradient Boosted Decision Tree (GBDT) algorithm, commonly used for classification, regression, and ranking tasks [75]. SHAP is a model interpretation method based on the concept of Shapley values from game theory, used to measure the “contribution” of each feature to the model’s prediction results [76]. In this study, SPSS version 25.0 software was used to perform Pearson correlation analysis between the visual quality of typical SD sample images and subjective landscape features data. Subjective landscape features with significant correlations were identified and retained. Taking landscape visual quality as the dependent variable and the subjective landscape features as independent variables, a landscape quality evaluation model for urban parks was constructed through linear regression analysis. Meanwhile, during the analysis process, the XGBoost model in machine learning was applied to establish the relationship between landscape visual quality and five objective landscape features. SHAP values were calculated to deeply explore the nonlinear effects of objective landscape features on landscape visual quality.

3. Results

3.1. Evaluation of SBE

By comparing the mean values of the SBE score of the parks (

Table 3. Scenic beauty estimation score of the five parks.

	Mean Score	Standard Deviation	Proportion of High Values	Proportion of Low Values
Maizuru Park	−0.053	0.917	0.300	0.388
Ohori Park	0.049	1.004	0.526	0.193
Nishi Park	−0.203	1.049	0.323	0.452
Suzaki Park	−0.042	0.949	0.161	0.323
Island City Central Park	0.028	0.986	0.281	0.344

), it can be seen that SBE Ohori Park (0.049) > SBE Island City Central Park (0.028) > SBE Susaki Park (−0.042) > SBE Maizuru Park (−0.053) > SBE Nishi Park (−0.203). Ohori Park showed significantly higher landscape visual quality than the others among the five parks. It meant greater attractiveness in landscape design and the arrangement of natural elements. This suggests that Ohori Park can provide residents with a better visual experience. In contrast, Nishi Park had the lowest score for visual quality. It may be caused by its relatively confusing landscape layout and conflicting vegetation combinations that fail to give a pleasant visual impression.

In terms of the proportion of low-quality samples, Nishi Park showed the highest proportion at 45.2%, indicating a higher level of negative perception. This reflected the presence of many sheltering elements, such as dense vegetation and enclosing walls. Horizontal sightlines of pedestrians were blocked, making it difficult to form a good vista extension and spatial accessibility. In contrast, Ohori Park had the lowest proportion of low-quality samples at only 19.3%, further confirming its better overall evaluation. Additionally, Maizuru Park and Island City Central Park showed relatively similar proportions of high- and low-quality samples, indicating a moderately low evaluation trend.

According to the standard deviation data, although Ohori Park had the highest visual quality score, it had the relatively highest standard deviation. It indicated that the different opinions in the public perception of Ohori Park led to inconsistent visual quality assessments. Nishi Park had the lowest mean visual quality but also exhibited a high standard deviation. It suggested different opinions, with some individuals giving positive evaluations to Nishi Park. Island City Central Park showed both a moderate average value and standard deviation, indicating they had consistent evaluations among respondents. In contrast, Maizuru Park and Suzaki Park had both low visual quality scores and low standard deviations, reflecting generally unfavorable and similar perceptions, with a consistent evaluation trend.

3.2. Evaluation of SD and Spatial Analysis

By using Excel to calculate the landscape feature values for each sample, the average feature values for each park, and the overall feature values for all 924 samples, a comparison of the landscape features across the 924 samples was conducted. In terms of the overall subjective evaluation across parks, the ranking of landscape features was as follows: Vitality (3.591) > Sense of layering (3.303) > Color richness (3.297) > Tidiness (3.241) > Interestingness (3.238) > Sense of order (3.153) > Tranquility (3.106) > Harmony (3.100).

By comparing the mean values of subjective landscape features for the five parks with the overall mean values as shown in

Table 4. Mean values of subjective landscape features for the five parks and the overall mean.

	All-Park Mean Score	Maizuru Park	Ohori Park	Nishi Park	Suzaki Park	Island City Central Park	Standard Deviation
Vitality	3.591	3.417	3.875	4.050	3.150	3.525	1.139
Sense of layering	3.303	3.450	3.400	3.275	2.525	3.475	1.356
Color richness	3.297	3.258	3.550	3.200	3.075	3.225	1.114
Tidiness	3.241	3.158	3.563	3.575	2.625	3.125	1.215
Interestingness	3.238	3.200	3.738	3.150	2.225	3.450	1.345
Sense of order	3.153	3.383	3.088	3.025	2.775	3.100	1.283
Tranquility	3.106	3.083	3.538	3.100	2.450	2.975	1.347
Harmony	3.100	3.150	3.263	3.300	2.450	3.075	1.343

, the result can be seen: Ohori Park has higher scores than the overall sample average in seven subjective landscape features, namely vitality, sense of layering, color richness, tidiness, natural interestingness, tranquility, and harmony, with tranquility being significantly higher than the sample mean. In Maizuru Park, the sense of layering, sense of order, and harmony were all higher than the overall sample averages. Island City Central Park had only two subjective landscape features—sense of layering and interestingness—that exceeded the overall sample mean. In Suzaki Park, only the sense of layering was higher than the overall sample mean.

By comparing the subjective landscape features of each park, the following conclusions can be seen: In Maizuru Park, the sense of layering (3.450) was the highest score, indicating a clear spatial structure and well-organized landscape layout. In contrast, vitality (3.417) was relatively low, suggesting that there is still a need for improvement in vegetation conditions within the park. Ohori Park performed notably well in terms of interestingness (3.738) and tranquility (3.538), highlighting its strengths in creating a natural atmosphere and a serene environment. However, the sense of order (3.088) was the lowest-scoring feature, possibly reflecting a weaker sense of spatial organization in certain areas.

In Nishi Park, vitality (4.050) received the highest score, indicating good vegetation conditions and a strong sense of liveliness. The sense of order (3.025) was the lowest, suggesting that improvements are needed in plant composition and spatial design. Suzaki Park showed generally lower scores across all features, but vitality (3.150) performed relatively better. Interestingness (2.225) and harmony (2.450) had the lowest scores, reflecting weaknesses in naturalness and landscape integration. In Island City Central Park, interestingness (3.450) and sense of layering (3.475) had higher scores, highlighting a comfortable environment and rich spatial layering. However, tranquility (2.975) had the lowest score, possibly due to the impact of high pedestrian density or functional disturbances.

By comparing the standard deviations of the eight subjective landscape features, it was found that vitality and color richness both had relatively high overall mean values and low standard deviations. This indicates that most parks performed well in terms of vegetation vigor and ecological conditions, with rich color presentation and relatively consistent visual perceptions. In contrast, both harmony and tranquility showed relatively low overall mean values and low standard deviations, suggesting that landscape composition across the parks generally lacked coordination, with some disorganized plant arrangements. The standard deviations for sense of layering and interestingness were relatively high, with sense of layering showing the highest variability. Sense of layering was the highest value, indicating that some of the park landscapes are organized, and some are more chaotic. The values fluctuate widely.

3.3. Objective Landscape Features Results and Spatial Analysis

Table 5. Objective landscape features of the five parks and their overall mean values.

	Openness	Greenness	Enclosure	Vegetation Diversity	Shannon–Wiener Diversity Index
All-park mean score	0.213	0.472	3.731	0.829	1.042
Maizuru Park	0.176	0.457	3.336	0.845	1.021
Ohori Park	0.250	0.410	3.944	0.839	1.033
Nishi Park	0.123	0.668	7.591	0.660	0.916
Suzaki Park	0.166	0.571	3.660	0.799	1.191
Island City Central Park	0.334	0.401	2.137	0.904	1.035

shows the objective landscape features of the five parks and the overall mean values. The table shows that Island City Central Park had the highest openness value (0.334), indicating the strongest sense of spatial openness. Nishi Park had the lowest openness value (0.123), reflecting a more enclosed space. Island City Central Park had lots of lawns combined with waterside vegetation, creating a natural and open spatial environment. Additionally, the park is surrounded by fewer buildings, providing a wide field of vision. In contrast, Nishi Park contains abundant greenery, but the dense vegetation along both sides of the pathways blocks the sky, resulting in a lower degree of openness.

In terms of greenness, Nishi Park showed the highest value (0.668), indicating a significant level of vegetation coverage. Island City Central Park had the lowest green view index (0.401), reflecting relatively less greenery. From the enclosure, Nishi Park showed a much higher enclosure value (7.591) compared to the other parks, primarily due to its dense vegetation and strongly enclosed spatial layout. In contrast, Island City Central Park showed the lowest enclosure value (2.137), corresponding to its open spatial structure. The landscape of Island City Central Park is characterized by extensive lawns combined with low-growing plants and open boundaries, creating a transparent space with a strong sense of spatial extension. The dense planting within Nishi Park, especially the continuous green walls formed along both sides of the pathways, results in a high sense of enclosure.

In terms of vegetation diversity, Island City Central Park showed the highest value (0.904), indicating a rich variety of plant species. Nishi Park showed the lowest diversity (0.660), reflecting a relatively single plant composition. Island City Central Park has a diverse range of plant species with well-defined structural layers, and a variety of plants contribute to its high landscape vegetation diversity. In contrast, although Nishi Park has relatively high vegetation coverage, its plant species are single and poorly layered, resulting in lower vegetation diversity.

In terms of the Shannon–Wiener Diversity Index, Suzaki Park had the highest value (1.191) and Nishi Park showed the lowest value (0.916). The high value in Suzaki Park is largely attributed to its well-organized spatial function and rich landscape design, which includes a variety of plant compositions, water features, landscape structures, and paving elements. The diverse combinations of these elements enhance the visual layer and overall landscape quality. In contrast, Nishi Park lacks diversified landscape compositions and distinct landscape nodes.

Table 6. Coefficients of variation for the objective landscape features of the five parks.

	Openness	Greenness	Enclosure	Vegetation Diversity	Shannon–Wiener Diversity Index
Maizuru Park	0.710	0.444	1.190	0.145	0.960
Ohori Park	0.604	0.426	0.121	0.156	0.252
Nishi Park	0.870	0.278	0.716	0.269	0.301
Suzaki Park	0.653	0.327	0.955	0.147	0.186
Island City Central Park	0.350	0.425	1.299	0.085	0.171

and Figure 4 show the coefficients of variation for the objective landscape features of the five parks. Combined with the mean value comparisons discussed above, Nishi Park had the highest greenness among the five parks, and the coefficient of variation for this indicator was the lowest (0.278). This suggests that Nishi Park had excellent performance in terms of greenery coverage, with a relatively uniform distribution of greenery in the park, leading to a consistently high level of visual greening across space.

In Island City Central Park, the vegetation diversity was the highest among the five parks, and the coefficient of variation for this indicator was also the lowest (0.171). This indicates that the park not only had a rich variety of plant species and a complex ecological structure but also exhibited a relatively balanced spatial distribution of vegetation, with stable vegetation diversity across the park. At the same time, Nishi Park had the lowest enclosure value among the five parks, and its coefficient variation for enclosure was also the lowest (0.716), suggesting that the strong sense of spatial enclosure and privacy is consistently distributed throughout the park. It lacks significant enclosure changes.

In addition, several indicators exhibited relatively high coefficients of variation across the five parks, reflecting significant spatial differences. In Maizuru Park, both the openness (0.710) and the Shannon–Wiener Diversity Index (0.960) showed high coefficients of variation, indicating big fluctuations in spatial openness and ecological diversity. This suggests that there were significant differences in spatial visibility within the park. Meanwhile, the species distribution was uneven.

In Ohori Park, the greenness was the highest coefficient of variation (0.426). It shows the highly uneven spatial distribution of the park in terms of the degree of greening, and there are large differences in green plant cover in different areas of the park. Nishi Park showed the highest coefficient of variation in Vegetation diversity (0.269), while Island City Central Park recorded the highest coefficient of variation in enclosure (1.299) among the five parks.

3.4. Visual Landscape Quality Modeling Through Urban Informatics Tools

3.4.1. Impact of Subjective Landscape Features on Visual Landscape Quality

After standardizing the landscape visual quality scores, correlation analysis was conducted between park landscape visual quality and eight subjective landscape features, such as vitality, harmony, sense of order, and sense of layering, using SPSS version 25.0 statistical software. According to the correlation analysis results (see Figure 5), all eight indicators showed a highly significant relationship with landscape visual quality (p < 0.05) and were positively correlated. Among them, interestingness, tidiness, harmony, sense of layering, and sense of order exhibited stronger correlations.

3.4.2. Construction of the Evaluation Model

A multiple linear regression analysis was conducted with landscape visual quality as the dependent variable and eight subjective landscape features as independent variables. The results are shown in

Table 7. Model summary.

Model Summary					ANOVA
R	R Square	Adjusted R Square	Std. Error of the Estimate	Durbin–Watson	F	Sig.
0.794	0.631	0.625	1.093	2.107	107.401	0.000

and

Table 8. Model regression coefficients.

Subjective Landscape Features	Unstandardized Coefficients (B)		Standardized Coefficients (Beta)	t	Sig.	VIF
	B	Standard Error	Beta
(Constant)	−0.73	0.232		−3.147	0.002
Sense of order	0.381	0.072	0.275	5.286	0	2.308
Harmony	0.266	0.072	0.201	3.69	0	2.534
Interestingness	0.224	0.061	0.17	3.691	0	1.797
Sense of layering	0.281	0.071	0.215	3.975	0	2.489
Vitality	0.206	0.06	0.132	3.423	0.001	1.265

. According to the regression results, the correlation coefficient (R) was 0.794, and the R squared was 0.625, indicating a good model fit. The F-value for the model significance test was 107.401 with a significance level of p = 0.000 (p < 0.005), meaning that the model is statistically significant. The variance inflation factors (VIF) for the five selected independent variables were all less than 10, suggesting that there is no severe multicollinearity among the subjective landscape features, and the linear relationship between the dependent and independent variables is significant. The final regression equation describing the relationship between landscape visual quality and subjective landscape features was as follows:

$$Y=-0.73+0.381X_1+0.266X_2+0.224X_3+0.281X_4+0.206X_5$$

(3)

where X₁ = Sense of order, X₂ = Harmony, X₃ = Interestingness, X₄ = Sense of layering, and X₅ = Vitality.

The results of the regression analysis indicated that all five subjective landscape features had significant positive effects on landscape visual quality. Among them, the sense of order had the highest regression coefficient (0.381), suggesting that it had the most significant influence on visual quality. As the organizational foundation of landscape space, the sense of order enhances spatial legibility and stability through structured and orderly layouts. It serves as a critical factor in improving visual appeal and cognitive comfort. The sense of layering (0.281) also contributed positively to visual quality by enriching spatial diversity through the effective organization of foreground, mid-range, and long-range views elements, thereby enhancing the overall spatial experience.

Harmony (0.266) shows consistency among landscape elements in terms of style, material, and proportion, contributing to the creation of a uniform overall image. The positive effect of interestingness (0.224) indicates that the inclusion of wild natural elements plays an important role in enhancing landscape attractiveness and natural perception, strengthening emotional and nature-friendly experiences. Vitality (0.206) is primarily reflected by dynamic natural elements such as plant growth, flowing water, and biological activity, injecting ecological liveliness into the landscape. In summary, all five subjective landscape features positively influence landscape visual quality through different mechanisms, providing multi-dimensional perceptual optimization directions for landscape design.

3.5. Correlation Analysis Between Visual Landscape Quality and Objective Landscape Features

3.5.1. The Relative Importance of Objective Landscape Features to Visual Landscape Quality

Figure 6 shows the importance of ranking and the swarm plot of objective landscape features about landscape visual quality. The importance ranking chart visually presents the relative importance of objective landscape features in influencing landscape visual quality. The swarm plot reflects the distribution and spread of SHAP values for each feature in the model. It displays the significance of each feature in contributing to the prediction. Each point represents a sample for the analyzed feature, with its horizontal position indicating the SHAP value. According to the color used in the figure, red represents points with greater influence, while blue represents points with lesser influence. Therefore, the horizontal location of the colored points also allows us to see whether their impact on prediction is positive or negative.

The research findings show significant differences in the importance of various objective landscape features in shaping landscape visual quality (Figure 7). Among these features, openness showed the greatest influence, followed by the Shannon–Wiener Diversity Index and enclosure. These three factors contribute over 65% to the overall effect, highlighting their role as the core determinants of visual quality in urban landscapes. These three factors contribute over 65% to the overall effect, indicating their role as the core determinants of visual quality in urban landscapes. Openness directly affects the transparency of sightlines and spatial perception, thereby playing a critical role in the overall evaluation of the visual experience. Meanwhile, enclosure and the Shannon–Wiener Diversity Index enhance visual layering and spatial richness, improving both the aesthetic appeal and the experiential interest of the landscape. In contrast, greenness showed the lowest contribution, which may be attributed to its limited ability to reflect the quality of greenery, as it primarily measures the quantity rather than the qualitative aspects of vegetation. Consequently, it may not fully capture the comprehensive perceptual experience of the landscape.

3.5.2. The Nonlinear Effects of Objective Landscape Features on Visual Landscape Quality

Figure 8 shows the nonlinear effects and threshold of objective landscape features on landscape visual quality. Openness exhibits an approximately linear increasing trend, with its influence on visual quality turning positive when the openness exceeds 0.27. This indicates that higher levels of openness help to enhance spatial transparency and a sense of openness. After reaching a certain threshold, a wide field of view not only reduces the feeling of oppression but also improves visual extension, thereby becoming a positive factor contributing to visual quality.

Greenness exhibited an approximately inverted U-shaped curve. When greenness is in the range of 0 to 0.48, it has a positive effect on landscape visual quality, reaching its maximum positive impact at a value of 0.38. This result suggests that at this range, moderate levels of green coverage can enhance the sense of ecological quality, visual comfort, and natural affinity, thereby improving the aesthetic experience of the landscape. However, as green coverage continues to increase beyond 0.48, the local impact of the green view index on the SHAP value turns negative.

Enclosure showed an approximately linear decreasing trend. When enclosure is in the range of 0 to 0.42, it exerts a positive effect on landscape visual quality. The positive effect of SHAP decreases with increasing enclosure. This indicates that a moderate degree of enclosure provides a sense of encompassing feeling and security from a visual perspective. However, when the enclosure degree exceeds approximately 0.42, its effect on the SHAP value turns negative, suggesting that excessive enclosure can decrease visual quality and lead to a bad visual experience.

Vegetation diversity also exhibited an approximately linear decreasing trend. Within the range close to 0 to 0.82, the SHAP values are generally positive but gradually decline. As the vegetation diversity approaches 0.82, its influence on the SHAP value shifts to negative. This suggests that high vegetation diversity may lead to an overly mixed plant composition and confused spatial layout, which can disrupt the visual arrangement of the space and negatively impact the overall landscape visual quality.

The Shannon–Wiener Diversity Index exhibited an approximately inverted V-shaped curve, showing an initial increase followed by a decline. When the Shannon–Wiener Diversity Index approaches 0.97, its impact on the SHAP value shifts from negative to positive, reaching the maximum positive effect around 1.37. Beyond this point, the positive impact gradually decreases and turns negative when the diversity index reaches approximately 1.6. This pattern suggests that a moderate level of landscape element diversity enhances spatial complexity and visual attractiveness, thereby increasing people’s interest and engagement with the landscape.

4. Discussion

4.1. The Impact of Subjective Landscape Features on Visual Landscape Quality

This study shows that a sense of order, harmony, interestingness, layering, and vitality all have a positive influence on the visual quality of landscapes. Among these, a sense of order was the most influential factor, playing an important role in shaping visual quality. Previous studies have shown that in large open spaces, a well-established sense of order is essential for organizing pedestrian flow and enhancing spatial experience. Well-arranged greenery can improve the perceived orderliness of a space, thereby enhancing its visual quality. At the same time, a harmonious landscape helps prevent visual overload and confusion, improving overall aesthetic appeal and spatial legibility [77,78]. Landscapes with a strong sense of order are often characterized by clear pathways [79], well-defined nodes [80], and distinct functions, which can reduce spatial disorientation, increase usability, and provide pedestrians with a sense of psychological safety [81].

Research has confirmed that aesthetic experience, which comes from the integration of various visual elements such as vegetation composition, site facilities, plant arrangements, and color schemes, is a source of visual pleasure [82]. Studies have emphasized that rich and engaging landscapes can quickly capture the attention of viewers, stimulate positive emotions, and enhance attractiveness and memorability [83,84]. Moreover, a well-designed spatial hierarchy can guide visual flow [85] and expand the dimensionality of a landscape [86]. In urban parks and open spaces, layered configurations of greenery and structures not only enhance aesthetic appreciation but also significantly improve the sense of safety and approachability of the environment. This also enhances exploration and an experiential nature [87,88].

4.2. The Impact of Objective Landscape Features on Visual Landscape Quality

This study found that the objective landscape features of urban parks significantly influence perceived visual quality, with clear nonlinear and threshold effects. Specifically, when the Shannon–Wiener Diversity Index is the range of 0.97 to 1.6, an increase in these elements leads to enhanced visual quality as perceived by individuals, consistent with the findings of studies [89,90]. However, when the Shannon–Wiener Diversity Index exceeds this threshold, the impact on visual quality becomes negative. The results suggest that within this optimal range, a moderate increase in landscape elements [90] positively contributes to visual quality. Yet, exceeding this threshold, overly artificial interventions can disrupt the natural aesthetic of the environment [91], diminishing the perceived visual quality of the landscape [92,93].

In terms of openness, this study has shown that as the proportion of greenery and other landscape elements decreases, the proportion of visible sky increases. When the sky proportion exceeds 0.27, it has a positive impact on visual quality. Moreover, the perceived visual quality continues to improve as the sky proportion increases. In previous studies, the sky has been considered an important component of nature [94]. When the proportion of sky is below 0.27, it often indicates a weaker sense of naturalness. In contrast, higher sky openness represents a more authentic natural environment [95,96], offering a more comfortable and pleasant visual experience.

The results for enclosure indicated that when it was at the threshold range of 0 to 4.2, the greenery, buildings, and horizontal interfaces in the park space achieved a visual balance. It provides the public with a more relaxed and comfortable psychological perception [12]. However, once this balance is disrupted, the positive effect gradually diminishes and eventually turns negative, failing to offer a pleasant visual experience. This finding is consistent with previous research results [97].

In the study of greenness, it was found that although greenness is commonly regarded as an important indicator for measuring the level of greenery in urban landscape evaluation [98], its role in enhancing overall visual quality is relatively limited. In this study, when the proportion of greenness is at the threshold range of 0 to 0.48, the amount of greenery effectively enhances visual comfort and the perception of naturalness, thereby improving the visual experience [99]. However, when this threshold is exceeded, an excessively high greenness may lead to issues such as landscape monotony and spatial enclosure, which in turn suppress visual quality [100,101]. Compared to a single green quantity metric, considering factors such as plant diversity, vertical greening structures, the configuration of greenery, and the degree of integration with other landscape elements is more conducive to improving the comprehensive perception of landscape visual quality.

In terms of vegetation diversity, diverse vegetation combinations can create rich visual layers [102] and enhance the sense of naturalness and ecological value. It also increases the spatial interest and aesthetic appeal of a landscape. This study identified a threshold range for such influence: when vegetation diversity was below 0.82, it significantly enhanced visual quality [103]. However, when this threshold was exceeded, an overly high number of plant species may lead to visual confusion, negatively affecting visual quality. This finding suggests that in practical landscape design and renovation, emphasis should be placed on the diversity of plant composition rather than simply pursuing the amount or coverage of greenery.

4.3. Limitation

This study has several limitations. First, the samples we used were taken from five large urban parks in Fukuoka City. Future research could include parks from multiple cities to enhance generalizability. Secondly, the target population of this study could be expanded to explore the perceptual differences in landscape visual quality among different groups. Lastly, emerging technologies such as eye-tracking and EEG could be introduced in future research to more objectively uncover the underlying connections between landscape visual quality, human cognition, and emotional responses.

5. Conclusions

This study was in the five city parks of Fukuoka, Japan, specifically Ohori Park, Maizuru Park, Nishi Park, Suzaki Park, and Island City Central Park. In this paper, five objective landscape features and eight subjective landscape features were calculated using SBE, SD, and semantic segmentation to analyze the quality of landscape visual quality in Fukuoka City urban parks. We established a linear regression model for visual quality and subjective landscape features and analyzed the nonlinear relationship between objective landscape features and landscape visual quality, in addition to the threshold effect.

Specifically, among subjective landscape features, the visual quality of urban parks is most affected by five features: a sense of order, harmony, interestingness, a sense of layering, and vitality. According to the regression coefficient, all of them have a significant positive influence on the visual quality of the landscape, among which the sense of order has the most significant influence on the visual quality of the landscape, which is a key factor in constructing a high-quality landscape perceptual experience. The threshold for openness in objective landscape features was greater than 0.27, and exceeding or falling below this threshold range hurt visual quality, while the optimal threshold for greenness was 0.38, and an increase in greenness could effectively enhance visual quality. However, above this value, the positive effect gradually diminishes or even turns negative. Enclosure has shown the same effect with an optimal threshold of 0–4.2, as well as vegetation diversity with an optimal threshold of 0–0.82, below which the visual quality is negatively affected. The optimal threshold for the Shannon–Wiener diversity index is 1.37, which was the highest point of positive effect on visual quality.

In the design of urban parks and public spaces, it is recommended that the visual proportion between sky and tree be carefully balanced in landscape design, with the sky proportion maintained above 0.27 to ensure an appropriate sense of openness. When arranging vegetation, it should consider its density to create a moderate sense of enclosure, enhancing feelings of safety and intimacy within the space. The variety of vegetation should be neither overly simplistic nor excessively complex; keeping it below 0.82 can help improve the visual experience. The diversity of landscape elements is ideally close to 1.37. In addition, design should emphasize subjective perceptual qualities such as a sense of order, harmony, and interestingness by optimizing spatial structure, coordinating colors, and integrating natural elements. The design of urban parks should follow a human-centered approach and serve multiple functions within the broader urban public space system. It should aim to achieve a harmonious integration of ecological performance, aesthetic value, and perceptual experience, thereby enhancing the overall quality and efficiency of urban spaces.