TL;DR
This paper introduces a novel framework, RLLPEM, that integrates objective landscape features (color, style, elements) with subjective human perceptions using machine learning and street view imagery. This approach provides a more comprehensive understanding of rural landscape perception, offering a valuable tool for informed rural planning and development.
What Problem Does This Paper Address?
Rural landscapes are undergoing rapid transformation due to urbanization, leading to a blend of traditional and modern elements that can fragment their unique character and diminish their quality. Existing studies often separate objective landscape analysis from subjective human perception, leading to incomplete or biased results. This disconnect makes it difficult to understand how physical landscape characteristics translate into real perceptions for residents and visitors, hindering the creation of truly satisfying rural environments. The paper aims to bridge this gap by developing a unified evaluation system that considers both objective and subjective dimensions of rural landscape perception.
Methods
The study proposes a multi-model approach combining computer vision and deep learning techniques. The core of their methodology involves:
- Objective Perception Extraction:
- Landscape Color: Analyzed using HSV color space to extract brightness, saturation (vividness), hue range (richness), and color distribution complexity.
- Landscape Style: Identified using Mask R-CNN for semantic segmentation and style recognition, classifying styles into European, Chinese, modern, and hybrid.
- Landscape Elements: Segmented using DeepLabv3-Plus to quantify elements like 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).
- Subjective Perception Extraction:
- A VGG16 convolutional neural network, pre-trained on the MIT Place Pulse dataset, was fine-tuned to predict subjective perceptions of esthetics, tradition, belonging, attractiveness, vitality, and safety.
- Integration and Analysis:
- The RLLPEM framework integrates these objective and subjective indicators.
- Pearson correlation coefficients were used to analyze the relationships between objective features and subjective perceptions.
- Spatial distribution maps were generated to visualize perception patterns.
Datasets
- Baidu Street View 360° panoramic images: Collected from 2024, with sampling points at 100m intervals, forming the primary data source for visual analysis.
- MIT Place Pulse dataset: Used for pre-training the subjective perception model, containing labels for esthetic, vitality, boredom, pleasure, safety, and wealth.
- Open-Street Map road data: Used for landscape element extraction.
- Administrative boundary data: Used for spatial aggregation of perception results.
The study does not explicitly introduce a new benchmark dataset but rather utilizes existing ones and augments them with local data for a specific region.
Key Insights and Ideas
- Bridging the Objective-Subjective Divide: The most significant contribution is the successful integration of objective landscape features with subjective human perceptions into a single, coherent evaluation framework. This moves beyond siloed analyses.
- Machine Learning for Scalable Perception Analysis: The use of deep learning models (Mask R-CNN, DeepLabv3-Plus, VGG16) allows for large-scale, automated, and high-resolution analysis of street view imagery, overcoming the limitations of traditional manual surveys.
- Chongming as a Case Study: The study highlights Chongming District as a representative example of rural areas facing rapid urbanization and the challenge of preserving cultural identity while integrating modern development.
- Spatial Heterogeneity is Key: The research emphasizes that rural landscape perception is not uniform; significant spatial disparities exist in both objective features and subjective responses across different villages.
- Complex Interplay of Factors: The correlation analysis reveals nuanced relationships, for instance, how hybrid landscape styles and wider roads can positively influence esthetic perception, while overly saturated colors or dense building clusters can detract from it.
Key Results
- Landscape Color: Generally elevated brightness, vividness, richness, and complexity were observed, with notable disparities between developed and traditional agricultural areas.
- Landscape Style: A balanced distribution of European, Chinese, modern, and hybrid styles was found, reflecting a blend of influences and resident preferences, though lacking a unified strategy.
- Landscape Elements: Indicators like GVI, SOI, BVI, RWI, FDI, and EII showed complex and uneven spatial distributions, often limited by development and agricultural land use.
- Subjective Perception: Overall subjective perception levels were generally low and unevenly distributed across Chongming, with tradition and belonging being higher in some northern villages, while esthetics varied regionally. Security and transportation perceptions were consistently weak.
- Correlation Analysis:
- Esthetic perception positively correlated with hybrid styles and RWI, negatively with brightness and EII.
- Tradition strongly correlated with Chinese-style architecture and BVI, negatively with European style and RWI.
- Belonging positively correlated with Chinese-style architecture and GVI, negatively with European style and SVI.
- Attractiveness positively correlated with hybrid styles and GVI, negatively with color complexity and BVI.
- Vitality positively correlated with color richness and FDI, negatively with BVI.
- Safety positively correlated with hybrid styles and RWI, negatively with color complexity.
Limitations and Future Work
- Data Bias: The reliance on Baidu Street View data may introduce bias in less developed areas or regions with sparse data coverage. The study did not account for individual factors like cultural background, age, or gender.
- Machine Learning Complexity: The complexity of the machine learning models used might limit their adoption by researchers without extensive technical backgrounds.
- Future Work:
- Future research should integrate field studies, satellite remote sensing, and UAV aerial photography for more comprehensive data collection.
- Developing simplified models or user-friendly analytical tools is suggested for broader accessibility.
- Multi-region research is recommended to optimize methods and improve model universality.
- Further investigation into the impact of individual cultural backgrounds, age, and gender on perception is needed.
Summary updated