Speaker
Description
Green perception, as a crucial indicator for measuring urban greening levels and the visibility of neighborhood green spaces, is widely applied in research on built environment quality, resident satisfaction, social equity, attention restoration, and more. It holds significant importance in urban planning and public policy. With the rapid development of big data, machine learning, and artificial intelligence, utilizing Street View Imagery (SVI) and computer vision to measure green perception has become an important approach. However, existing methods have three limitations: (1) They predominantly rely on a single Green View Index (GVI) metric, overlooking the multidimensionality and complexity of green perception; (2) They focus on extracting 2D features from SVI, neglecting the 3D spatial characteristics of urban landscapes, thus failing to accurately capture the impact of green landscapes at different distances on subjective perception; (3) In subjective perception assessments, participants typically use computer screens for simulated experiences, lacking immersion and realism.
To address these gaps, this study proposes a green perception measurement method that employs mixed reality (MR) and large visual models (LVMs) to collect and integrate multidimensional information. Specifically: (1) Based on an in-depth exploration of landscape perception mechanisms, this study establishes a multidimensional green perception indicator system integrating both objective and subjective perceptions. The objective perception indicators encompass 3 dimensions—color features, landscape elements, and spatial forms—totaling 30 indicators. Subjective perception includes 6 indicators: green coverage, vegetation diversity, natural ambiance, relaxation, attractiveness, and environmental quality. (2) Utilizing SVI from Baidu Map and state-of-the-art LVMs including SAM, DINOv2, Depth Anything, this study measures the objective perception of green landscapes. These models not only extract semantic information of green landscapes from SVI but also estimate the spatial depth of landscape scenes, achieving the integration of color features, landscape elements, and spatial forms. Furthermore, pre-trained models are employed to perform computer vision tasks under zero-shot conditions, significantly reducing computational complexity and training time. (3) Volunteers were recruited to measure subjective perception data using both computer screens and MR devices. The relationship between objective and subjective perceptions was established through 3 machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost).
In an empirical study conducted in Shenzhen, China, experimental results indicate that: (1) the proposed method achieved higher prediction accuracy for greening degree, vegetation diversity, and natural ambiance compared to relaxation, attractiveness, and environmental quality; (2) RF consistently outperformed SVM and XGBoost in predictive performance across different perception dimensions; and (3) models trained using the MR-based perception method exhibited significantly higher accuracy than those based on computer screens. This verifies that MR devices offer greater advantages over computer screens in terms of immersion, realism, and user experience in urban built environment audits. These findings validate that the green perception measurement method integrating mixed reality and large vision models outperforms traditional methods in terms of perception dimensions and accuracy. Specifically, in urban built environment audits, MR devices offer significant advantages over computer screens in immersion, realism, and experiential quality. Furthermore, the study generated green perception distribution maps, analyzed the spatial characteristics of Shenzhen’s green landscapes, and identified key factors influencing green perception.
This study proposes an innovative approach to measuring urban green perception by integrating mixed reality and large vision models, effectively overcoming the limitations of traditional methods. The approach not only enriches the theoretical dimensions of urban green perception measurement but also enhances measurement accuracy from a technical standpoint. It holds significant application value in urban built environment auditing, assessment, and planning, contributing to the development of greener, healthier, and more livable cities.
| Keywords | Urban Green Perception;Mixed Reality;Large Vision Models;Machine Learning;Street View Imagery |
|---|---|
| Best Congress Paper Award | No |
