| 摘要: |
| 在“以人为核心”的时代语境下，城市景观作为全面推进美丽中国建设、城市空间品质提升的重要抓手，缺乏智能化、精细化和主客观结合的感知方
法。首先，构建融合传统景观与认知科学的多学科理论框架，阐明景观感知的机制、过程和内容，建立主观与客观耦合的指标体系；其次，集成色彩分析、目标
检测、语义分割和深度预测4种计算机视觉方法，并应用人工智能视觉大模型构建可解释的视觉融合模型，对城市景观的色彩特征、景观要素和空间形态3类客观
感知进行精细化测度；再次，通过随机森林算法，捕捉客观感知与主观感知之间的非线性关系，实现对主观感知的精准预测；最后，在重庆市主城区开展实证研
究，验证了视觉融合模型的有效性与科学性 |
| 关键词: 风景园林 景观感知 人工智能 计算机视觉 重庆 |
| DOI：10.19775/j.cla.2025.03.0076 |
| 投稿时间：2024-04-07修订日期：2024-10-08 |
| 基金项目:国家重点研发计划项目(2023YFC3804803)；国家资助博士后研究人员计划(GZC20241084)；自然资源部城市国土资源监测与仿真重点实验室开放
基金资助课题(KF-2023-08-17)；广东省哲学社会科学规划项目(GD24DWQGL03) |
|
| Research on Fine-grained Perception Method of Urban Landscape Based on Visual Fusion Model: ACase Study of Chongqing's Main Urban Area |
| CAO Yuehao,,YANG Peifeng,,LI Minmin,,LI You*,,GUO Renzhong |
| Abstract: |
| In the context of the "human-centered" era, urban landscapes are key
instruments for advancing the construction of a Beautiful China and enhancing
the quality of urban spaces. However, current research methods for landscape
perception lack intelligence, precision, and integration of subjective and objective
dimensions. Specifically, three main limitations persist: 1) the absence of a
systematic theoretical framework to elucidate the mechanisms of landscape
perception, resulting in deficiencies in the breadth and accuracy of measurement
content; 2) reliance on labor-intensive annotation and model training processes,
leading to high costs in manpower, time, and computation; and 3) incomplete
coverage of all perceptual dimensions, requiring the integration of more diverse
computer vision models to align with the multifaceted nature of human visual
perception. To address these limitations, this study develops a multidisciplinary
theoretical framework that integrates traditional landscape theories with
cognitive science. By aligning sensory, perceptual, and cognitive processes
with the biological visual system's processing of sensory information, the
study elucidates the mechanisms, processes, and content of landscape
perception. A coupled subjective-objective indicator system is constructed,
encompassing four dimensions—color characteristics, landscape elements,
spatial morphology, and landscape imagery—with a total of 28 indicators. The
study then integrates four computer vision methods—color analysis, object
detection, semantic segmentation, and depth prediction—along with advanced
AI vision models, including SAM, Grounding DINO, and DINO v2, to construct
an interpretable visual fusion model. This model facilitates refined measurement
of three categories of objective perception: color characteristics, landscape
elements, and spatial morphology. To model subjective perception, a dataset
is created through online surveys, and multiple machine learning algorithms,
including Random Forest, GBRT and SVM, are applied to establish nonlinear
relationships between objective and subjective perceptions. Comparative
experiments identify Random Forest as the best-performing algorithm, enabling
accurate prediction of six types of landscape imagery: ecology, aesthetics,
culture, safety, attractiveness, and vitality. Using this theoretical framework and
visual fusion model, an empirical study is conducted in the main urban districts
of Chongqing. The findings validate the framework's scientific robustness and
effectiveness. Analysis of key landscape perception characteristics aligns with
traditional understandings of Chongqing's rich mountain and water resources,
multi-group spatial structure, and major urban green spaces and centers.
Optimization strategies for landscape perception are proposed, including
introducing locally distinctive color elements in public spaces, improving
pedestrian and vehicular layouts to enhance walkability and visual appeal, and
strengthening the continuity and interaction between foreground, midground,
and background views. The study also recommends optimizing nodal designs
for various landscape scenarios, leveraging Chongqing's unique terrain to create
distinctive spatial experiences, and restoring and preserving historic buildings
to enhance cultural appeal. Critical factors influencing landscape perception are
identified. Beyond the single dimension of landscape elements emphasized in
prior studies, color characteristics and spatial morphology are shown to have
significant impacts. For example, contrast, color diversity, and color complexity
exhibit strong positive correlations with most landscape imagery dimensions.
Midground and average depth correlate positively with ecological perception,
while background and ultra-background views are negatively correlated with
ecological perception and all subjective perceptions. This study contributes two
key innovations. First, it refines the landscape perception process by clarifying
the correspondence between physiological and psychological processes and
the content of landscape perception, enriching theoretical foundations. Second,
it constructs a visual fusion model that integrates advanced computer vision
and machine learning techniques, enhancing the precision and generalizability
of perception methods. These advancements strengthen planners' ability to
perceive complex urban scenarios, fostering harmonious coexistence between
humans and urban environments and driving the overall improvement of the built
environment's quality. Looking to the future, this study highlights the need to
integrate sensory information from additional dimensions and employ intelligent
analytical techniques to produce more comprehensive and scientific perception
results. Such advancements will provide robust support for constructing green,
intelligent ecological civilizations. |
| Key words: landscape architecture landscape perception artificial intelligence computer vision Chongqing |
