Rural Local Landscape Perception Evaluation: Integrating Street View Images and Machine Learning

Abstract

Rural landscape perception is of great significance in understanding the emotional connection between people and rural local environments. Seeking to rectify the problem of incomplete or biased results owing to the separation of objective and subjective landscape perception in previous studies, this study took the village of Chongming District in Shanghai, China, as an example and built an evaluation model that integrated four dimensions and 20 indicators of objective and subjective landscape perception, and used machine learning technology to analyze street view images. Subjective perception has been influenced by landscape color, style, and element perception. Notable spatial disparities have been observed in the distribution of rural landscape indicators across Chongming. This study refines key subjective and objective factors affecting rural landscape perception, and the model provides a new method for the perception evaluation of complex landscapes, providing a theoretical basis and practical reference for rural landscape planning.

1. Introduction

Owing to China’s rapid globalization and urbanization, most rural areas of the country continue to include farmland, homesteads, collective villages, and other traditional elements [1], while also integrating some urban functions and features. Over time, the original natural texture and features of the countryside are gradually being destroyed [2], and rural landscapes are increasingly displaying a mix of traditional and modern, Chinese and Western styles, which further aggravates their fragmentation and declining quality [3]. Rural landscapes are relatively stable environments that include characteristics of natural, cultural, and human activity formed through the dynamic coupling of built and natural environments, creating local characteristics unique to each area [4]. However, China’s current top-down rural planning and management mode is predominantly based on subjective decision-making by planners [5] and by a unified planning mode that lacks the incorporation of local rural landscape features or the quantitative evaluation of the subjective perceptions of rural landscape users [6]. This approach weakens rural landscape planning and renders the construction and implementation of high-quality human settlements difficult.

The concept of place [7] and the relationship between humans and the land are generally understood from the perspectives of human psychology, society, and culture. Local characteristics are those characteristics of a place that cannot be replicated elsewhere [8] and are distinctly regional, unique, and original to a place [9]. These characteristics comprise the geographical material environment and the spiritual and cultural environment formed by the interaction of humans and landscape [10]. Tuan contends that “place” is not merely a geographical entity but a “center of perceived value” [11]. By serving as a point of “pause”, the place transforms dynamic spatial dimensions into a stable source of meaning, facilitating the construction of individual identity through both material satisfaction and spiritual well-being.

In the 1960s, proponents of landscape perception theory argued that perception and experience of the surrounding environment are generated by individuals or groups through psychological and physiological processes, such as sensation, cognition, and emotion. Upon receiving landscape stimuli, individuals exhibit emotional responses, such as pleasure, comfort, or disgust, which guide the recognition and evaluation of the environment. These emotions are subsequently expressed through behavior. As a result, landscape perception has been categorized into objective and subjective dimensions [12]. Objective perception of landscape refers to an individual’s sensory perception of the physical characteristics of landscape, such as color, form, and texture [13]. Existing studies have primarily focused on the perception of landscape form and component elements, while landscape color and style have received limited attention [14]. Subjective perception of landscape refers to an individual’s understanding based on objective perception combined with personal experience, emotion, and cultural background. The two interwoven dimensions constitute landscape perception. Therefore, this study incorporated color and style perception into the objective dimension of local landscape perception. For sustainable and context-sensitive development, any evaluation of local landscape perception should integrate and interpret the relationship between objective and subjective dimensions.

Rural revitalization strategies and the reconstruction of urban-rural relations have rendered the perception of rural local landscape an important channel to enhance the sense of experience, subjective consciousness and satisfaction of people who live in or engage with rural landscapes [15], landscape character assessment [16], scenic beauty estimation [17], semantic differential [18], and other methods have been employed to capture participants’ multi-sensory experiences and psychological responses across visual, auditory, tactile, and related dimensions. It has gradually become a research hotspot in landscape science, ecology, sociology, human geography, and other fields, which has been applied to explore issues, such as collective emotion, cultural identity, rural homogenization, and the protection of landscape authenticity [19,20]. However, most related studies focus on analyzing a single aspect of the objective perception of the landscape; these studies often conduct analyses of objective elements such as rural green spaces, buildings, and roads in terms of pattern evolution, differentiating characteristics, and morphological changes [13,14]. In terms of subjective perception, the differences in people’s perception of beauty, comfort level, preferences, and other similar factors are primarily measured through threshold measurement, sensory scales, and the wearing of physiological measurement devices [14,21]. It can be seen from this that the objective and subjective perceptions of rural landscapes have generally not been integrated in research. This has resulted in it being difficult to comprehensively understand the differences and corresponding relationships between the physical spatial characteristics of the landscape and the real perceptions of villagers and tourists, which is not conducive to building a rural environment that satisfies its users. Therefore, this study innovatively constructs a landscape evaluation system that integrates both subjective and objective perceptions. To comprehensively capture objective perceptions, it selects colors, styles, and elements that are closely related to the objective perception of rural landscapes and creates a new method for landscape color perception evaluation and landscape style extraction. To capture subjective perceptions, it employs senses of beauty, tradition, belonging, attraction, vitality, and security, all of which have a high correlation with local and objective perception. In this way, this study enriches current understandings of objective and subjective perceptions of landscapes and their corresponding relationships, allowing for the formation of plans that can comprehensively guide the sustainable development of local landscapes and the inheritance of cultural context.

Although traditional questionnaire surveys and on-the-spot interviews can quantitatively evaluate emotions, they are time-consuming, expensive to undertake, and are hampered by interviewees’ differing understandings of evaluation indicators, failing to establish generally recognized evaluation criteria, or to conduct large-scale fine-grained analysis. Machine learning technology offers an opportunity to explore the relationship between the physical characteristics of the landscape and landscape perception. Scholars have used large-scale visual models [14], convolutional neural networks [22], scene semantic segmentation [6], and other methods to explore the relationship between factors and perception in urban areas [23]. For example, Muhammad and Heung [22] utilized crowdsourced data to predict perceptions of safety and beauty, Naik et al. [24] used Place Pulse 1.0 to evaluate perceived safety. In general, machine learning technology is used in the single-aspect perception evaluation of small-scale urban landscapes, but it has not been widely used to study the complex content and corresponding relationships of combined subjective and objective landscape perception in large-scale rural landscapes.

To address the persistent disconnection between physical landscape features and human perceptual experiences in existing rural landscape studies—in which subjective and objective evaluations are often treated separately—this study proposes an integrated evaluation framework that bridges this gap. Specifically, we develop the Rural Local Landscape Perception Evaluation Model (RLLPEM), which systematically fuses subjective and objective perceptions to capture the multi-dimensional and spatially heterogeneous characteristics of rural environments. The Chongming District in Shanghai, China—a region undergoing rapid rural–urban transformation while simultaneously striving to preserve its ecological and cultural identity—was selected as the study area. Based on 360° panoramic images from Baidu Street View and supplemented by road, administrative, and perception-related datasets, this study extracts 20 indicators across four categories: landscape color (brightness, vividness, richness, complexity), landscape style (Chinese, European, modern, hybrid), landscape elements (e.g., green view index, road width, enclosures), and subjective perceptions (e.g., esthetics, tradition, safety). This study develops and fine-tunes several machine learning models, including Landscape Style Mask R-CNN, DeepLabv3-Plus, and VGG16 networks, to automatically perform large-scale semantic perception extraction. Pearson correlation coefficients are then used to analyze the relationship between objective features and subjective perceptions. This not only provides methodological tools for landscape perception analysis, but also offers new theoretical insights into the mechanisms by which rural environments evoke emotional, cultural, and psychological responses. These contributions support more human-centered, perception-informed rural planning strategies.

2. Research Design

2.1. Study Area

Chongming District, located in the north-eastern outskirts of Shanghai at the mouth of the Yangtze River, is the world’s largest estuarine alluvial island. It consists of Chongming, Changxing, and Hengsha islands, covering an administrative area of 2494.5 km², including 1413 km² of land. The district has nearly 20,000 rivers, 16 towns, and 2 townships. Among these, Xinhai Town and Dongping Town do not have administrative villages, and therefore are not included in this study. Within the scope of this study, there are a total of 269 administrative villages (Figure 1). Situated in the northern subtropical zone, it experiences a mild and humid climate, with an average annual temperature of 16.5 °C. The forest coverage rate is 23.2%, the highest among all Shanghai districts. Chongming is home to Chongming Island National Geopark and Dongping National Forest Park, and has been recognized as China’s longevity township, a national ecological demonstration zone, and one of the first national agricultural sustainable development experimental zones. The district’s unique island geography and rich natural ecological resources contribute to its distinct character.

The rapid urbanization of Chongming District has led to the blending of old and new functions, with mixed colors and architectural styles across its landscapes. As a result, the original island culture and the distinct texture of paddy fields are gradually fading, weakening, diminishing the villagers’ sense of identity, attraction, and beauty. This has made Chongming a typical example of the decline of local characteristics in Shanghai’s rural landscape. To address these challenges, The Master Plan for Land Use of Chongming District of Shanghai (2017–2035) emphasizes highlighting the features of a world-class ecological island, guided by “Chinese elements and island characteristics,” aiming to make it a key area for Shanghai to experience a sense of nostalgia. However, current rural development efforts have yielded limited results, and Chongming village has yet to establish a landscape pattern or overall image that reflects its ecological island identity. As a result, the sense of happiness and identity among residents remains underdeveloped. Therefore, it is crucial to leverage the perception characteristics of the Chongming District village more objectively and propose optimization strategies. This study also offers case studies and references for addressing similar rural landscape issues.

2.2. Data Sources

This study utilized 2024 Baidu Street View 360° panoramic images, Open-Street Map road data, administrative boundary data, and MIT Place Pulse data (

Table 1. Data and description.

Dataset	Source	Application Indicator
Baidu Street View 360° panoramic picture	Baidu Street View Application Programming Interface (https://lbsyun.baidu.com/ accessed on 10 January 2025), with a sampling period of 2024. Sampling points were set at 100 m intervals. Points without available pictures were skipped, and those with incomplete or inaccessible images were excluded. A total of 2700 valid sampling points and corresponding street view images were retained.	Landscape color: brightness, vividness, richness, complexity. Landscape style: European style, Chinese style, modern style, hybrid style. Landscape elements: green view index (GVI), sky open index (SOI), road width index (RWI), building view index (BVI), facility diversity index (FDI), enclosure integrity index (EII).
MIT Place Pulse	Papers with code (https://paperswithcode.com/dataset/place-pulse-2-0 accessed on 16 January 2025)	Subjective perception: esthetic, tradition, belonging, attractiveness, vitality, safety.
Road data	OpenStreetMap (https://www.openstreetmap.org/ accessed on 12 January 2025)	/
Administrative boundary data	Standard Map Service System of the Ministry of Natural Resources, Map No. GS (2024) 0650	/

).

2.3. Research Methods

This section outlines the methodological framework adopted to construct the RLLPEM. To achieve an integrated evaluation of both the objective and subjective dimensions of rural landscapes, we developed a multi-model approach based on computer vision and deep learning techniques. Specifically, four categories of indicators were extracted: (1) landscape color features, including brightness, vividness, richness, and complexity, all derived from HSV color space analysis; (2) landscape style classification using Mask R-CNN, chosen for its ability to perform accurate instance segmentation and identify diverse architectural styles in complex rural scenes; (3) landscape elements segmented via DeepLabv3-Plus, which offers robust semantic segmentation of large-scale images, allowing for the precise quantification of green spaces, roads, the sky, buildings, and other components; and (4) subjective perception prediction based on VGG16, a well-established convolutional network known for its reliable visual feature extraction, particularly for perception-related tasks. These models were selected to complement each other, enabling high-resolution, large-scale, and automated perception evaluation that directly supports the study’s objective of bridging visual spatial characteristics with human emotional and cultural responses in rural environments.

2.3.1. Landscape Color Perception Method

Based on relevant studies [14], we selected four color indicators: brightness, saturation, richness, and complexity to analyze landscape visual perception color features.

Brightness refers to the color brightness features of an image. Each image was converted to the HSV (hue defines the type of color, saturation controls the purity, and value regulates the brightness) color space using OpenCV (Open Source Computer Vision Library). The average brightness (V channel) of all pixels in the image was then calculated as follows:

$$\mathrm{Brightness}={\displaystyle \frac{1}{N}}\sum _{i=1}^N V_i$$

(1)

where $V_i$ is the brightness value of pixel i, and N is the number of pixels in the image.

Vividness reflects purity of color: the more saturated the color, the higher the vividness. The saturation (S channel) of the HSV color space was used to represent vividness, and the mean saturation value of all pixels in the image was computed as follows:

$$\mathrm{Saturation}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{S}_i$$

(2)

where $S_i$ is the saturation value of pixel i.

Richness is the number of types of colors in the image. The greater the hue (H channel) change, the higher the richness of color. The range of hue change (that is, the difference between maximum and minimum values) was calculated as follows:

$$\mathrm{Hue} \mathrm{range}=\mathrm{m}\mathrm{a}\mathrm{x}\left(H\right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(H\right)$$

(3)

where H is the hue value of all pixels in the image.

Complexity is the quantitative analysis of color distribution in the image. The occurrence probability of each color was calculated based on the color histogram, and the uniformity of the color distribution was measured by the entropy value (Formula (4)). The higher the entropy, the wider the color distribution, and the lower the complexity. Conversely, the lower the entropy, the more concentrated the color distribution and the higher the complexity:

$$\mathrm{C}\mathrm{o}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{x}\mathrm{i}\mathrm{t}\mathrm{y}=-\sum _{i=1}^k p_i{\mathit{log}}_2\left(p_i\right)$$

(4)

where $p_i$ represents the occurrence probability of color i, and k is the total number of color categories.

2.3.2. Landscape Style Perception Method

Based on the Mask Region-based Convolutional Neural Network (Mask R-CNN) model, we carried out semantic segmentation and style recognition of architectural and landscape objects and adopted transfer learning strategies to reduce the burden of manually labeling landscape style datasets. Referring to relevant literature [23,24] and the local characteristics of Chongming’s rural landscape, the landscape styles are classified as follows: European style (featuring Western columns, decorative lines, European patterns, etc.), Chinese style (using white walls, blue-tiled sloping roofs, Chinese classical patterns, etc.), modern style (adopting simple shapes, glass curtain walls, steel frame structures, etc.), and mixed style (a fusion of the above styles). The specific methods are as follows (Figure 2):

(1)

Classification and annotation. The Labelbox platform was used to manually classify and label buildings and landscape pieces in street view images according to European, Chinese, modern, and hybrid styles. 70% of the dataset was used as the model training sample set, 20% was the verification set of the adjusted and optimized model, and 10% was the test set to evaluate the performance of the model.

(2)

Model fine-tuning. The existing Mask R-CNN model was used for transfer learning. The large-scale dataset pre-trained model was taken as the starting point, and the annotated street view image dataset was fine-tuned to construct Landscape Style Mask R-CNN (LSMask R-CNN). The key hyperparameters in the fine-tuning process are shown in

Table 2. Key hyperparameters in the fine-tuning process.

Parameter	Instructions	Fine-Tuned Value
Learning Rate	Control the step size of each parameter update: over-learning will lead to unstable training, under-learning will lead to slow convergence.	0.0001
Batch Size	The number of samples used during each training, smaller batch sizes are suitable for fine adjustment, larger batch sizes speed up training.	16
Epochs	The number of training rounds determines the total number of training times for the model.	25
Optimizer	Controlling the model weight updating mode and selecting a suitable optimizer speeds up the training process.	Adam
Freezing Layers	Freeze the early layers of the model to maintain its general feature extraction capabilities and fine-tune later layers only.	Freeze the first 12 layers and fine-tune the last 4 layers
Regularization	Techniques to prevent overfitting and help model generalization.	Dropout (0.5)

. The comparison of model accuracy before and after fine-tuning is shown in

Table 3. Comparison of model accuracy before and after fine-tuning.

Precision Index	Pre-Fine Tuning Model (%)	Fine-Tuned Model (%)
Accuracy	83.26	95.45
Precision	81.20	94.35
Recall	87.05	92.33
F1 Score	81.25	90.58

, indicating that the fine-tuned model was better able to learn landscape style characteristics.

(3)

Landscape style perception. The experiment used the trained LSMask R-CNN model to identify the style of Street View images at the sampling points, extract the landscape pixel region at each sampling point, and calculate the proportion of each style. This represented the visual proportion of different landscape styles at each sample point in the image. It was defined as Landscape Style View Index (LSVI) and calculated as follows:

$$LSV{I}_i={\displaystyle \frac{A_{t,i}}{A_{b,i}}}\times 100\%$$

(5)

where $A_{t,i}$ is the pixel area of a certain style of landscape at sampling point i and $A_{b,i}$ is the total pixel area of all landscapes at the sampling point.

Next, the landscape style visual rate in all Street View scenic spot images was aggregated experimentally, and the overall style ratio was calculated, classified into European, Chinese, modern, and hybrid style categories:

$$RLSVI=\left({\displaystyle \frac{\sum _{j=1}^n \sum _{i=1}^m A_{t,ij}}{\sum _{j=1}^n \sum _{i=1}^m A_{b,ij}}}\right)\times 100\%$$

(6)

where RLSVI is the Rural Landscape Style View Index, n is the total number of Street View image sampling points in the landscape, m is the number of different landscape styles identified at each sampling point, and $A_{t,ij}$ is the landscape pixel area of style i at sampling point j. $A_{b,ij}$ is the total pixel area of all landscapes at sampling point j.

(4)

Model validation. To ensure the robustness of style classification, the manually labeled dataset was randomly split into 70% training, 20% validation, and 10% testing sets. Model performance was evaluated using standard classification metrics, including accuracy, precision, recall, and F1-score (Table 3). The results show that the fine-tuned LSMask R-CNN significantly outperformed the pre-trained baseline model across all metrics, indicating its strong generalization ability and effectiveness in capturing diverse rural architectural styles.

2.3.3. Landscape Element Perception Method

The semantic segmentation model (DeepLabv3-Plus model) was used to pre-train weights, and graphics processing unit (GPU) acceleration technology was employed to process large-scale street view data. This model accurately segments the input Baidu API street view data into categories, such as green, sky, architecture, small objects, and other targets. This segmentation supports the calculation of various indices, including green view index (GVI), sky open index (SOI), road width index (RWI), building view index (BVI), facilities diversity index (FDI), and enclosure integrity index (EII) (Figure 3), as well as the landscape view index (LVI) for different landscape elements, as follows:

$$LV{I}_x={\displaystyle \frac{A_x}{A_{\mathit{total}}}}$$

(7)

where $LV{I}_x$ represents the apparent rate of landscape element x and $A_x$ is the pixel area of this kind of landscape element in the image (green pixel area is $A_g$, sky pixel area is $A_s$, building pixel area is $A_b$, road pixel area is $A_r$, landscape pixel area is $A_f$, and the pixel area of enclosed objects, such as buildings and facilities is $A_e$). $A_{\mathit{total}}$ is the total pixel area of a street view image.

2.3.4. Landscape Subjective Perception Method

The Visual Geometry Group 16 Network model convolutional neural network was adopted as the framework, and the MIT Place Pulse dataset was used to pre-train the landscape subjective perception model. The MIT dataset included esthetic, vitality, boredom, pleasure, safety, and wealth perception labels [6]. The experiment fine-tuned the model using 2700 labeled local street view images and incorporated pairwise comparison methods to simulate the public’s perception and evaluation of two street view images. Additionally, new tags for tradition and belonging were introduced. The experiment marked and predicted the results for the image’s esthetics, tradition, belonging, attractiveness, vitality, and safety (Figure 4).

(1)

Model fine-tuning. Using 2700 Street View images to complete feature extraction pre-training, the fully connected classification layer was adjusted to six label categories. The labels of esthetic, attractiveness, vitality, and safety in the MIT dataset were mapped to the corresponding perceptual labels, and traditions and belonging were also added. Next, the pre-trained feature extraction function was frozen, the classification layer was adjusted, and the model was optimized with a small learning rate to adapt to the new perceptual classification task. The accuracy, precision, recall, and F1 score of each perception index of the model after centralized fine-tuning were verified to range from 78.45% to 87.3% (

Table 4. The performance of the subjective perception index.

Perception Labeling	Accuracy (%)	Precision (%)	Recall (%)	F1 Score (%)
Esthetic	86.12	87.30	85.45	86.50
Tradition	78.95	78.45	80.12	79.05
Belonging	80.45	79.80	81.35	80.50
Attractiveness	83.01	83.60	81.70	82.80
Vitality	81.80	80.60	82.90	81.75
Safety	83.30	84.90	82.65	83.75

), indicating that the transfer learning and fine-tuning model had good applicability for subjective perception classification [20].

(2)

Perceptual label prediction. We conducted a perception label prediction of rural Street View images. The prediction results were distributed to villages according to latitude and longitude through the administrative boundary aggregation method, and the distribution proportion of each village in the six perception label categories was calculated. The perception proportion was calculated as follows:

$$P_i^j={\displaystyle \frac{\sum _{k=1}^{N_i} {\mathrm{Indicator}}_k^j}{N_i}}$$

(8)

where $P_i^j$ is the proportion of perception label j in village i, $N_i$ is the number of Street View images in the village, and ${\mathit{Indicator}}_k^j$ is the prediction probability of image k in label j.

(3)

Indicator visualization. According to the rural perception distribution results, the spatial distribution characteristics of the six rural perception indicator types were generated.

(4)

Model validation. For the subjective perception classification task, 2700 Street View images were divided into 70% training, 15% validation, and 15% testing subsets. Model performance was evaluated using accuracy, precision, recall, F1-score, and R² across six perception labels (Table 4). The results indicate that the model has a generally stable predictive performance, with relatively lower scores for tradition and vitality, likely due to subjective variability and limited training samples in those categories.

2.4. Technical Process

This study investigates “data acquisition and preprocessing—landscape objective perception analysis—landscape subjective perception analysis—result visualization and correlation analysis” (Figure 5).

(1)

We collected 360° Street View image data from Chongming District in Shanghai, and performed unified coding, formatting, and data cleaning to ensure consistent data quality. Four color indicators were extracted, and four style classification labels were created to establish an objective database of perceptual indicators.

(2)

Key parameters of streetscape images were extracted, and distribution maps of generated landscape styles were identified, alongside the extraction of six landscape element indicators for quality checking and spatial aggregation.

(3)

By combining the MIT Place Pulse dataset with Chongming Rural Streetscape data, experiments were conducted to train six types of subjective perception models and evaluate their performance.

(4)

The experiment generated a spatial distribution map of landscape perception, performed correlation analysis between objective perception and subjective perceptions, and synthesized the results to propose optimization suggestions.

3. Experimental Process and Analysis of Results

3.1. Landscape Color Perception Analysis

Analysis of street view image data revealed the characteristic results and distributions of color indicators (

Table 5. Analysis results of landscape color perception.

Index	Max	Min	Mean	Histogram
Brightness	215.83	146.29	190.96
Vividness	127.71	25.91	85.96
Richness	179	177.52	178.98
Complexity	20.58	16.16	19.23

and Figure 6). The mean values of the indicators in the study area (Table 5) were relatively high, with histograms predominantly skewed to the right. This finding suggests that brightness, vividness, richness, and complexity are generally elevated. Significant differences between the maximum and minimum values for brightness and vividness (69.54 and 101.8, respectively), indicate that, while lighting conditions and greenery vividness in Chongming’s rural areas are typically excellent, there is a notable disparity in color brightness and vividness between highly developed rural areas and those dominated by traditional agriculture. The range in richness and complexity (1.48 to 4.47) was relatively narrow, suggesting that recent efforts to enhance rural landscapes have enriched plant colors and ornaments. However, the lack of comprehensive planning has led to overly complex landscape colors in many rural areas. Only a few villages, such as Xin’an Village in Sanxing Town and Jingting Village in Xinhe Town, which feature extensive paddy fields, forests, and ecological protection zones, exhibit more monotonous colors.

In summary, under the influence of urban development in central Shanghai, Chongming’s rural areas—featuring a mix of farmlands, forests, and wetlands—have incorporated a variety of building facade colors. The recent rise in agricultural tourism has further contributed to elevated color indicator levels across the rural landscape. In contrast, construction in conservation-focused areas has progressed slowly, with landscapes largely composed of traditional white and gray structures, greenery, and water channels, resulting in limited color variation and richness. Overall, the findings indicate a lack of comprehensive color planning in Chongming’s rural development. Moreover, rural landscape plans have not provided village-specific implementation guidelines. Consequently, independently executed renovations have resulted in a pronounced spatial imbalance in color-related indicators across the region.

3.2. Landscape Style Perception Analysis

Table 6. Analysis results of landscape style perception.

Index	Max (%)	Min (%)	Mean (%)	Histogram
European style	17.02	12.5	14.99
Chinese style	24.91	20.39	22.53
Modern style	16.25	13.29	14.94
Hybrid style	50.49	44.46	47.54

and Figure 7 present the measurement outcomes and distribution patterns of European, Chinese, modern, and mixed landscape styles in Chongming’s rural areas. The mean values for each style indicator are relatively high, and the histograms are predominantly right-skewed, suggesting a broad distribution of these styles (Table 6). The differences between maximum and minimum values are relatively small—4.41%, 4.52%, 2.96%, and 6.03%, respectively—indicating limited regional variation. The spatial distribution of styles is highly dispersed among villages (Figure 7). Influenced by urban immigrant culture and Western lifestyles, recent developments in Chongming have included the construction of European-style residences and homestays, alongside preserved traditional Jiangnan water town architecture. Additionally, rural office areas and industrial parks feature modern, minimalist buildings with glass curtain walls. A variety of stylistic elements have been integrated into existing structures and landscapes by residents, resulting in multifunctional hybrid-style environments that reflect diverse esthetic preferences.

In summary, influenced by ongoing globalization and modernization, Shanghai’s culturally integrated environment has significantly impacted Chongming’s rural landscape. The traditional layout—characterized by white-walled, black-tiled buildings and narrow streets—has been increasingly replaced or supplemented by European, modern, and hybrid styles that reflect the evolving esthetic preferences of residents. This stylistic diversification highlights the absence of a unified rural esthetic strategy, resulting in a visually eclectic yet broadly inclusive landscape.

3.3. Landscape Element Perception Analysis

Semantic segmentation and analysis of street view images enabled the extraction of six key landscape element indicators (

Table 7. Analysis results of landscape element perception (n = 2700).

Index	Max (%)	Min (%)	Mean (%)	Histogram
GVI	57.62	9.09	27.16
SOI	54.23	18.39	37.74
BVI	32.12	7.02	19.3
RWI	15.74	3.65	10.5
FDI	3.84	0.13	0.88
EII	13.06	1.16	4.24

, Figure 8), revealing distinct spatial distribution patterns across villages. The average values of these indicators are relatively low, indicating a considerable disparity compared to their respective maximum values (Table 7). Apart from the centrally positioned sky openness index (SOI), the remaining indicators are predominantly left-skewed in the histogram, suggesting limited visibility of landscape elements throughout rural Chongming. Due to widespread agricultural land, the SOI is slightly elevated. Notably, large variations were observed between the maximum and minimum values across indices (GVI: 48.53%, SOI: 35.84%, BVI: 25.1%, RWI: 12.09%, FDI: 3.71%, EII: 11.9%). Figure 8 further illustrates the uneven spatial distribution. For example, forest coverage rates vary significantly, from 31.11% in Xin’an Village (Sanxing Town) to 43.98% in Beigang Village (Xianghua Town), and up to 56.9% in Lvgang Village (Lvhua Town). Additionally, the GVI is generally higher in areas with the extensive forests and ecological reserves. In contrast, resettlement villages, such as Jingting (Xinhe Town), Huimin (Shuxin Town), and Hongtian (Chenjia Town), which have undergone concentrated relocation, have experienced significant reductions in green space due to the construction of numerous new buildings. Therefore, greening efforts have lagged, leading to abandoned homesteads, diminished vegetation coverage, and comparatively low GVI values.

In summary, rapid urbanization in China has led to the conversion of substantial farmland and forest land in Chongming’s rural areas into construction land, significantly increasing building density. However, the development of green infrastructure, rural roads, service amenities, and public facilities has not kept pace. Consequently, the visibility of key landscape elements remains limited. In the absence of comprehensive and systematic planning, landscape improvements have been carried out independently by individual villages, resulting in pronounced disparities in element distribution across the region.

3.4. Landscape Subjective Perception Analysis

Convolutional neural networks (VGG16) were employed as a machine learning framework, with transfer learning and fine-tuning applied to train and predict perceptions of esthetics, tradition, belonging, attractiveness, vitality, and safety. Model performance was evaluated using accuracy, F1 score, precision, recall, and R². The R² values for each perception index were as follows: esthetics (0.63), tradition (0.17), belonging (0.52), attractiveness (0.25), vitality (0.24), and safety (0.67) (Figure 9). Higher R² values indicate greater model stability, stronger data fitting, and improved prediction accuracy for landscape perception.

The trained model was utilized to analyze landscape perception in street view images of Chongming District, with perception data aggregated at the village level to map the spatial distribution of subjective perceptions (Figure 10 and

Table 8. Analysis results of landscape subjective perception.

Index	Max	Min	Mean	Histogram
Esthetic	0.44	0.03	0.06
Tradition	0.44	0.10	0.17
Belonging	0.65	0.13	0.24
Attractiveness	0.28	0.14	0.23
Vitality	0.32	0.18	0.22
Safety	0.54	0.03	0.07

). As shown in Table 8, mean values for each indicator range from 0.07 to 0.24, and the histograms are slightly left-skewed, reflecting generally low levels of subjective perception across the district. The difference between maximum and minimum values (0.14–0.52) is substantial, indicating pronounced spatial variation in perceived landscape qualities. Analysis of Figure 10 reveals that rural traditions, sense of belonging, and attractiveness are notably higher in northern villages, such as Sanxing Town, Miao Town, Xianghua Town, and Xincun Township. In contrast, esthetic perceptions are relatively elevated only in select areas of Chengqiao, Xinhe, and Chenjia towns. Security and transportation-related perceptions are generally weak across the district, with exceptions in a few villages in Chengqiao Town, Xincun Township, Changxing Town, and Hengsha Township, where public service and safety infrastructure are comparatively more developed.

In conclusion, the overall level of subjective perception of Chongming’s rural landscape remains low and unevenly distributed. This is largely attributed to the recent emphasis on constructing select “beautiful villages” and demonstration sites, such as Yingdong, Yuyuan, and Xin’an, which have seen notable improvements. However, most rural planning efforts have prioritized building facades and main roads, while overlooking the preservation of folk traditions and the integration of local cultural symbols. This has led to diminished cultural identity, reduced attractiveness, and a weakened sense of belonging among villagers. Furthermore, Chongming’s geographic isolation—requiring over an hour’s drive and a cross-sea bridge to access—continues to impede the inflow of resources and expertise, ultimately constraining rural landscape development and contributing to the region’s overall esthetic quality.

3.5. Correlation Analysis

We established a correlation matrix to analyze the nonlinear relationship between objective perception and subjective perception (Figure 11).

(1)

Esthetic perception demonstrates a strong positive correlation with hybrid landscape styles and RWI. In recent years, Chongming has integrated modern design elements into traditional rural landscapes, enabling a simultaneous appreciation of cultural heritage and contemporary innovation. Many rural villages have also expanded main roads and enhanced green roadside layers, contributing to continuous and visually appealing landscape corridors. Conversely, esthetic perception shows a strong negative correlation with brightness and EII. Chongming’s characteristic low-brightness palette—dominated by green vegetation, blue waterways, and gray-tiled traditional architecture—is often disrupted by high-brightness features, such as neon signage and glass curtain walls, which clash with the natural rural setting. Additionally, dense building clusters and high enclosure walls obstruct scenic views, fragment public spaces, reduce spatial transparency, and exacerbate discomfort during the humid summer season by intensifying a sense of confinement and visual dissonance.

(2)

Tradition exhibits a strong positive correlation with Chinese-style architecture and BVI. Traditional villages in Chongming are often organized around familial settlements, characterized by architectural layouts that create continuous and enclosed street and alley interfaces. For example, Caopeng Village (Sanxing Town) integrates features, such as moon gates and curved corridors from Jiangnan garden design, enriching the spatial narrative with classical Chinese esthetic elements. Conversely, tradition shows a strong negative correlation with European-style architecture and RWI. Traditional Chongming architecture primarily employs a muted palette of black, white, and gray, with elements including pitched roofs and wooden lattice windows contributing to the iconic “white walls and black tiles” appearance of the Jiangnan region. However, the introduction of European-style features, such as Roman columns and iron railings, disrupts this traditional esthetic. Additionally, the traditional layout of the rural road—narrow paths interwoven with residential plots—typically maintains a height-to-width ratio of 1:0.8 to 1:1.2, offering an intimate and human-scale environment. Road widening efforts that expand these paths beyond 5 m lower this ratio to below 1:0.5, diminishing the enclosed feeling of traditional streets and alleys.

(3)

Belonging is strongly positively correlated with Chinese-style architecture and GVI. The visual identity of Jiangnan water towns—characterized by white walls, gray roofs, traditional floral windows, and wooden frames—harmonizes with surrounding natural features, such as farmlands and rivers, fostering a cohesive and historically resonant landscape. For example, the boxwood nurseries in Yuanyi Village (Gangyan Town), blue-tiled buildings, reinforce cultural memory and enhance residents’ sense of identity. Long-term engagement in agricultural activities further deepens attachment to the rural ecosystem, including farmlands, vegetable plots, and trees. Conversely, belonging is negatively correlated with European-style and SVI. Architectural elements, such as Roman column facade and red pitched roofs—exemplified by French-style homestays in Xianqiao Village, Shuxin Town—conflict with the traditional visual language of coexisting houses and fields. Additionally, modern village developments that lack mature tree coverage tend to increase sky visibility, thereby diminishing enclosure and weakening the villagers’ emotional connection to the local landscape.

(4)

Attractiveness exhibits a strong positive correlation with hybrid landscape styles and GVI. The integration of natural environments with modern amenities allows rural areas to retain their green esthetic while catering to urban recreational demands. For example, Yingdong Village in Chenjia Town has successfully blended agricultural activities with diverse tourism functions, such as vacation homestays, exhibition halls, and restaurants, forming a multifunctional rural tourism destination. In contrast, attractiveness is negatively correlated with color complexity and BVI. While Chongming’s traditional dwellings typically feature simple light gray tones, the excessive use of more saturated colors, such as beige, brick red, and brown, undermines overall visual harmony. Furthermore, the construction of buildings exceeding three stories disrupts the traditionally horizontal village skyline and pastoral character, making the environment appear more urbanized and less appealing to those seeking rural charm.

(5)

Vitality is positively correlated with color richness and FDI. The introduction of vibrant vegetation and art sculptures significantly enhances spatial liveliness in traditional villages. For instance, Qianwei Village in Shuxin Town has developed the “Flower Island on the Sea” resort area, incorporating seasonal blooms, interactive workshops, and flower corridors to attract tourists for leisure and hands-on experiences. The integration of planting, processing, and tourism forms a complete “planting + processing + experience” industry chain, injecting sustained vitality into the rural setting. Conversely, vitality is negatively correlated with BVI. In areas where building density exceeds 35%, the traditional hierarchical pattern of “house-field-forest” is disrupted. This not only reduces the availability of public activity spaces but also diminishes long-standing ecological interactions between villagers and their environment.

(6)

Safety was found to be positively correlated with hybrid styles and RWI, indicating that wider roads and the integration of multiple architectural styles are perceived as safer by villagers accustomed to modern lifestyles. In contrast, a negative correlation with color complexity suggests that overly varied and uncoordinated colors reduce visual order, leading to a decreased sense of comfort and security in rural environments.

4. Discussion

4.1. Development Proposal

(1)

To address the issue of relatively high comprehensive landscape color indices in Chongming, a rural color database should be established through oblique photography and remote sensing image collection. It should then be integrated into the national land spatial planning system to achieve the visualization of color approval. Additionally, in conjunction with the Five Beauties Co-construction Implementation Plan of Chongming District, the specific technical parameters should be further refined to complete the color spectrum system planning for each village, including the “primary color + secondary color + accent color” scheme [23]. Finally, a control model that combines rigid control (strict enforcement of planning) with flexible guidance (dynamic adjustment) should be implemented.

(2)

In response to the issue of mixed and scattered landscape styles in Chongming, based on the Shanghai Suburban Rural Landscape Planning and Design Guidelines, which propose shaping the “Shanghai-style Jiangnan” landscape as the foundation, a specialized landscape style plan was developed. This involved retaining several valuable and relatively intact Chinese-style neighborhoods and buildings, gradually expanding to surrounding areas, and establishing protection control zones (i.e., reinforcing Chinese-style elements), partial renovation zones (incorporating multiple styles in smaller proportions), and moderate renewal zones (allowing the independent construction of other styles in specific areas) [24]. Different styles are reflected through modifications to site elements such as color, materials, style, and scale, including white-walled, black-tiled Chinese-style buildings, Chinese-style window lattices, and the use of blue tiles and blue bricks.

(3)

To address the issues of low view index for landscape elements and significant regional differences, firstly, the layout is organized as “one ring, multiple belts, and multiple points,” with parks and green spaces serving as green points, roadside green belts as green lines, and large parks as green areas. This approach strengthens the development of rural industrial zones, street parks, and roadside and vertical greening. Second, alley widths are controlled within 8–12 m, and dilapidated pergolas and shade structures are removed to enhance the openness of the sky in the area. Moreover, the rural building density is controlled below 35%, with public activity spaces increased through the integration of local culture and commercial space development. Then, rural main roads are equipped with dedicated non-motorized vehicle lanes, and secondary roads are provided with sidewalks no less than 1.5 m wide. In addition, a specialized facility planning scheme is compiled for the entire district with a 500 m facility service radius, and functional small-scale structures with local characteristics are systematically added. Finally, the use of alleys and public spaces remove overly oppressive and obstructive walls, as well as dilapidated facilities in streets, alleys, and public spaces, thereby balancing the enclosure and openness of the site [4,6].

(4)

To resolve the problem of low subjective perception results and significant disparities in their distribution across various landscape dimensions, it is recommended that each village adopt outstanding measures from beautiful rural villages and rural revitalization demonstration village initiatives to beautify the water, forest, farmland, and residential spaces of the rural area. The historical culture, intangible cultural heritage, and folk customs of each township should also be explored and then translated into landscape elements with a traditional feel to stimulate villagers’ sense of identity and belonging. Local cultural and tourism resources should be combined, with modern living, recreational functions and facilities, and diverse cultural activities added to enhance the attractiveness and vitality of resources. Security systems and transportation system planning should also be improved to enhance environmental comfort and safety [6,16].

(5)

In response to the complex positive and negative relationships between the objective and subjective perceptions of landscape elements, extensive field research should first be conducted to extract the objective texture characteristics and landscape cultural symbols (such as architectural colors, styles, and facility designs) of each village, establishing a quantifiable morphological gene bank [23]. Second, for different groups such as tourists and residents, perception databases should be designed, perception mapping tables should be established, thresholds for beauty and belonging should be quantified, and village-level perception archives that dynamically link characteristic symbols with psychological perception indicators should be created. Finally, the aforementioned specialized plans for rural landscape color, style, greening, and small-scale features should be integrated to establish a perception-based intelligent control system that integrates landscape color, style, elements, and the subjective “six senses.” For example, after inputting the database and archives of Chongming Horticultural Village, a color spectrum for nursery auxiliary colors (light green) and accent colors (light brown) could be generated, and reference materials could be provided for signage and lighting fixtures with boxwood-shaped designs [24].

4.2. Limitations

(1)

This research, while novel, is not without limitations. Although the coverage of Baidu Street View image data is relatively comprehensive in urban and well-developed areas, the analysis may be biased in less developed areas or areas where Street View data are lacking. Meanwhile, the study does not consider the impact of individual cultural background, age, gender, and other factors on perception results. Future research should combine field research, satellite remote sensing, and UAV aerial photography, 3D Gauss [25], diffusion modeling [26], eye tracking [27], and other technologies to collect multi-type and high-precision data that capture the subtle perceptual differences in individual esthetics, emotions, and other aspects, to improve the timeliness of data and the comprehensiveness of research results.

(2)

Although the machine learning used in this study can provide high-precision analytical results, its complexity may limit its popularization by researchers with non-technical backgrounds. In the future, a simplified model structure or user-friendly analytical tools should be developed for use by researchers lacking computer science and machine learning expertise. In the later stage, multi-region research can be carried out to optimize the methods and indicators and to improve the universality of the model.

5. Conclusions

Rural landscape perception reflects the characteristics of the landscape, while also reflecting the impact of the landscape on the subject’s physiology, psychology, and response to the environment. In light of the limitations of single research content and neglect of the relationship between subjective and objective data in current landscape perception studies, this study took Chongming District in Shanghai as the research object and constructed a new rural landscape subjective and objective integrated place perception evaluation model (RLLPEM). This was based on open-source street view image data and provided a new approach to the comprehensive evaluation of visual and psychological perception of landscape on a macro scale. The results are as follows:

(1)

The RLLPEM, which covers 20 subjective and objective integration indicators across four dimensions, can comprehensively extract landscape issues in rural areas from the perspective of both residents and visitors, comprehensively improve the accuracy of perception analysis in complex rural scenarios, promote the widespread application of landscape perception results in rural planning practices, help guide rural planners in creating environments that satisfy the public and bring them physical and mental pleasure, and assist decision-makers in making wise decisions [16].

(2)

The RLLPEM was applied to evaluate the rural landscape of Chongming, Shanghai. The results indicated that color perception exhibits distinct regional heterogeneity alongside certain universal patterns. Landscape style perception was found to be generally balanced yet spatially dispersed. Visual landscape elements displayed notably complex distribution characteristics. Moreover, subjective perception showed considerable regional variation, with consistently low scores across all evaluated indices. Faced with the significant differences in landscape perception content across different rural areas, systematic specialized planning will be required in the future to coordinate and balance the proportions of various elements [6].

(3)

The correlation analysis between objective perception and subjective perception of landscape shows that 14 objective perception indicators of landscape color, style and elements have nonlinear positive and negative differential effects on six subjective perception elements of landscape, consistent with the findings of other relevant studies [6,20]. This reveals the significant effect of the objective perception of rural landscapes on subjective perception [6].

Overall, this study presents a novel integrative framework that advances both the theoretical understanding and methodological practice of rural landscape perception. By linking objective spatial features with subjective human responses through machine learning and street view imagery, the RLLPEM enables scalable and multidimensional evaluations of rural environments. This approach enriches perception theory by clarifying how spatial forms and visual elements shape cultural identity and emotional experience, while also providing actionable insights for more context-sensitive, perception-informed rural planning.