基于图像深度学习的高密度城区小型绿地恢复性潜能测度——以重庆渝中区106个样本 为例

Restorative Potential Measure of Small Green Space in High-density Urban Areas Based on the ImageDeep Learning: A Case Study of 106 Samples in Yuzhong District, Chongqing

关键词:  风景园林  小型绿地  高密度城区  参数化  恢复性潜能

Key words:  landscape architecture  small green space  high-density urban area  parameterization  restorative potential

摘要:

从小型绿地恢复性环境体系构成及其物化要素解译出发，提取表征空间环境的共性特征因子，遴选出典型高密度区域重庆渝中区106个小型绿地研究 样本，利用图像深度学习对2 190张实景图像进行参数化计算，并招募60名被试对象开展实景图像的恢复性潜能分组实验。结果表明：1)渝中区小型绿地的基础 性构成以绿视率、乔灌木、天空开敞度、硬质铺装场地和建筑界面为主，休息和娱乐服务设施占比较低，环境维护度、环境静谧性和空间安全性品质特征较好； 2)2组被试对象的恢复性潜能评价结果趋向一致，不同要素类型及其规模大小对应不同的要素恢复性潜能差异，从吸引性、远离性、延展性和相容性4个维度建立 的要素恢复性潜能模型具有较好的拟合度。利用图像深度学习方法建立了一套高精度的小型绿地参数化方法，在此基础上实现了实景图像的恢复性潜能测度，可 为同类研究提供参考与借鉴。

Abstract:

Green spaces have significant potential to enhance public health and serve as valuable restorative environmental resources in urban areas, offering broad benefits at a low cost. However, densely populated cities often struggle with the limited availability of green spaces, and large-scale green space development faces strict supply constraints. As a result, smaller green spaces that can be flexibly arranged have emerged as a crucial form of green space development in these high-density areas. Their advantages include a high quantity, widespread distribution, easy accessibility for residents, and the potential for renewal and transformation, making them vital resources for improving health in urban settings. Current research on green restorative environment is more inclined to explore its restorative benefit results, causality, and mechanism, focusing on single factor analysis at the overall level, but less from the perspective of internal composition and different elements, unable to evaluate the difference in restorative potential of different elements. In addition, the lack of large samples and refined quantitative measurement methods for real sites, especially the lack of reports on small green Spaces in high-density urban areas, restricts the discussion of their causal mechanism or revealing mechanism and hinders the promotion of restorative environmental design and spatial optimization of small green spaces. Therefore, it is necessary to further reveal the internal system composition of small green restorative environment, and analyze the difference of restorative potential of different elements through large samples and refined quantitative measurement. The key lies in how to understand its internal composition structure, extract common feature factor measurement labels from many constituent elements, and then achieve large sample and high-precision characterization through parametric analysis. Therefore, based on the interpretation of the restorative environment composition and its physical elements of small green space, the common characteristic factors were extracted, and 106 small green space research samples were selected to parametrically measure. Then the images in the hierarchical catalog of each feature factor type were divided equidistantly according to the statistical results, and 22×3 images were sampled from the high, middle, and low value domains of each feature factor, and 60 subjects were recruited to carry out the grouping experiment of restorative potential of real images. Meanwhile, 60 subjects were recruited to carry out experiments on the real image restorative potential. The results: 1) The basic composition of small green space in Yuzhong District was dominated by green vision rate, trees and shrubs, sky openness, hard pavement site and building interface, the service facilities accounted for a low proportion, and the characteristics of environmental maintenance, quietness, and safety quality were better; 2) The two group evaluation results of restorative potential tend to be same, different element types and sizes correspond to different restorative potential, and the model established from the four dimensions of attraction, distance, extensibility and compatibility had a fit. A set of small green space parameterization methods suitable for large samples and high precision are established by using the image deep learning method, the restorative potential measurement of real scene images is realized based on these, which can be used as a reference for similar research.