The Impact Mechanism of Urban Built Environment on Urban Greenways Based on Computer Vision

Abstract

With the development and widespread adoption of smart machines, researchers across various disciplines and fields are exploring the contributions of computers and intelligent machines to human science and society through interdisciplinary collaboration. In this study, we investigated the potential applications of artificial intelligence and multi-source big data in the selection and design of urban greenways, using the city of Nanjing as a case study. Utilizing computer vision technology and the DeepLabV3+ neural network model, we analyzed over 320,000 street view images and 530,000 fine-grained urban data points from Nanjing. We also trained the place space material quantification model using the Street Space Greening Structure (S.S.G.S) dataset. This dataset not only achieved high-precision semantic segmentation but also surpassed previous datasets in predicting greenery at the street level. The performance metrics for this model are as follows: MIoU is 0.6344, Recall is 0.7287, and Precision is 0.8074. Through Robust regression, we identified several micro and macro-level factors influencing the Panoramic View Green View Index (PVGVI). The results indicate that multiple factors have significant positive or negative effects on PVGVI. This research not only provides new decision-making tools for landscape architecture and urban planning but also opens new avenues for applying artificial intelligence in urban environmental studies.

1. Introduction

Urban greenways, as linear spaces that integrate natural and recreational functions, play a critical role in urban landscapes. These green corridors, often established along underutilized spaces such as defunct railways or riverbanks, are gaining increased attention in urban development due to their multifaceted benefits. Ecologically, greenways serve as essential corridors that support biodiversity, mitigate urban heat island effects, and improve air quality, thus enhancing overall urban sustainability. Socially, greenways provide vital public spaces that foster community interaction and social cohesion, offering residents places for recreation and leisure. The presence of greenways has been linked to higher physical activity levels among urban dwellers, contributing to improved physical and mental health by reducing risks associated with obesity, cardiovascular diseases, and stress. Economically, the proximity to well-designed greenways can enhance property values and contribute to the local economy by attracting tourism and promoting community events.

The planning and design of urban greenways necessitate a holistic approach, considering factors such as dimensions, materials, and aesthetics, alongside practical aspects like traffic and safety. These efforts must align with the unique geographic, climatic, and demographic profiles of urban areas to craft spaces that resonate with residents’ needs and aspirations.

From an ecological standpoint, greenways are instrumental in counteracting urban challenges like heat islands and air pollution, enhancing urban sustainability. Socially, they provide vital public spaces for recreation and community interaction, improving the overall quality of urban life. The environmental, social, and economic dividends of greenways, from ecosystem services to health benefits and property value appreciation, underscore their significance in urban development.

This study delves into the nuanced relationships between urban greenways and the built environment, employing computer vision and big data analytics to dissect these interactions across scales. By elucidating the influences of built structures on urban greening, the research aims to refine landscape architecture and urban planning practices, highlighting the potential of artificial intelligence systems in advancing urban environmental studies.

2. Related Works

2.1. Traditional Methods for Assessing the Quality of Greenery at the Urban Street Level

To enhance urban street greenery and its ecological benefits, a meticulous evaluation is necessary. This requires a detailed visual representation and quantification of its structural components. Traditional approaches to assess urban greenery often hinge on direct field investigations conducted by experts. However, such methods are susceptible to external influences that could skew data accuracy. These external influences include (i) the weather on the day of the on-site investigation; (ii) differences in image collection; (iii) the accuracy of the image acquisition location; (iv) accuracy of manual extraction, etc. Furthermore, conventional techniques are marked by their labor-intensive nature and inefficiency, requiring significant human and material investments. This becomes particularly challenging in large-scale, city-wide studies, where data collection often relies on volunteers. These volunteers may lack the specialized skills and knowledge needed, further complicating the accuracy and reliability of the collected information.

The advent of methodologies employing remote sensing satellite and aerial imagery for green space quantification has paved the way for advancements in this field [1], integrating proximate remote sensing visuals with conventional remote sensing strategies. Additionally, the utility of high-definition Light Detection and Ranging (LiDAR) technology in mapping the volume of urban greenery has been corroborated by research [2]. Nevertheless, LiDAR technology is currently encumbered by several impediments, including questions about its scalability, the cumbersome and costly nature of its sensors, intricate signal and data handling processes, and its vulnerability to atmospheric conditions such as precipitation, snow, and variations in lighting. Therefore, a new approach is needed to address its limitations.

To precisely gauge urban street greenery and understand its effects on human well-being and the ecological milieu, innovative methodologies are essential. Utilizing the DeepLabV3+ neural network for feature extraction, in conjunction with openly accessible panoramic street view imagery that mimics human perspective, emerges as a potent remedy to these challenges. This method offers urban developers and landscape architects richer insights, significantly enhancing the assessment and scrutiny of urban greenery. Such advancements hold the promise of shaping effective urban strategies to tackle environmental issues, thereby fostering healthier and more sustainable urban landscapes.

2.2. Evaluation of Urban Street-Level Greenery Using Street View Images

As urbanization and city planning progress rapidly, the human-centric aspect of urban spaces is garnering more focus. It is crucial for urban design and planning to center on the populace’s necessities and emotional responses, aiming to forge urban settings that are secure, agreeable, accessible, aesthetically pleasing, and sustainable. These environments should cater to a wide array of human needs, ultimately uplifting residents’ life quality and contentment. Consequently, from a scholarly standpoint, assessing the myriad facets of the urban landscape through a human-centered lens is becoming paramount.

Understanding how residents perceive and experience the urban vegetation landscape at the street level is a key component in assessing urban greenery levels [3]. The stark contrast between street-level imagery and remote sensing visuals lies in the former’s ability to more directly mirror the human perception of their environment. With the growing availability of open-access, geotagged data sources, like Google Street View (GSV), Baidu Street View (BSV), and Tencent Street View (TSV), obtaining street-level imagery has become significantly more straightforward. This accessibility has empowered numerous researchers to employ various methodologies to quantify urban street greening, leveraging the rich data these street view platforms offer.

This investigation harnesses the potential of panoramic street view imagery to assess the vegetative framework of urban street greenery. These images yield an in-depth view of the greenery landscape at street level, shedding light on its influence on how residents perceive and interact with the urban milieu. The use of panoramic street view images for evaluating urban street greenery emerges as a compelling method to capture the human-centric view of the urban landscape, offering a more nuanced understanding of the interaction between urban dwellers and their green surroundings.

2.3. Introduction of Computer Vision to Help Assess Urban Street-Level Greenery

In the swiftly evolving realm of computing, significant strides in computer technology have brought artificial intelligence (AI) and computer vision to the forefront of public awareness. Predominantly applied in the domain of autonomous vehicles, these technologies have recently made their way into architecture, urban planning, and landscape architecture, addressing specific professional challenges. This includes the analysis and quantification of street-level imagery and the physical dimensions of urban spaces. Notable applications include the assessment of Sky View Factor (SVF) and the computation of the Green View Index (GVI) through GSV images. These methods provide insights into the proportion of greenery visible from pedestrian perspectives, enhancing our understanding of urban green spaces [4].

In 2009, Yang et al. introduced a novel approach for evaluating urban greenery, which integrated on-site assessments with manual photography to introduce the concept of the Green View Index (GVI) [5]. Despite its innovative approach, this method faced challenges such as heavy reliance on manual effort, increased uncertainties, and laborious procedures, potentially skewing data and leading to inaccurate outcomes. In a significant advancement in 2015, Li and colleagues presented a technique for urban streetscape analysis and GVI calculation using GSV imagery [6]. This technique employed pixel color recognition in Photoshop to identify greenery within images. However, the method was not without its flaws, still necessitating substantial manual input and struggling to precisely quantify urban street greenery due to the color overlap with non-vegetative green objects like signs. Additionally, the use of street view images can also analyze the impact of plants and humans on the perception of urban spaces [7,8].

In a pioneering study conducted in 2021, Xia and colleagues introduced a cutting-edge methodology for quantifying the green landscape index along urban thoroughfares, leveraging the semantic segmentation of streetscape imagery. This approach birthed the PVGVI, a novel metric designed to gauge the extent of greenery within visible street vistas. This technique effectively addresses the issue of the manual reliance inherent in traditional GVI assessments, significantly enhancing the computational speed. Furthermore, it introduces a fresh index for green landscapes, offering a more precise tool for evaluating urban greening efforts.

In summary, a series of prior investigations have been dedicated to the examination of urban greenery at the street level [9,10], alongside efforts to enumerate urban tree coverage at a city-wide scale [11,12,13,14]. Additionally, these studies have delved into the dimensions of urban street perception [15,16], contributing to a comprehensive understanding of urban green spaces and their impacts.

In essence, the advent of innovative data sources and technologies has unlocked new vistas for the detailed analysis of street-level greening as experienced in everyday urban life, heralding a burgeoning research frontier. While certain studies have ventured into utilizing machine learning algorithms for assessing visible street greening, the intricate interplay between the urban built environment and urban greenways remains largely untapped. Furthermore, studies that amalgamate large-scale evaluation with granular, human-scale spatial resolution are scant. It is imperative to dissect the influence of the urban built environment on urban greenways through a dual lens, examining both the macro and micro dimensions. From a macro perspective, in addition to considering ecology and economics, beyond society, it is necessary to incorporate considerations of urban spatial attributes into the urban built environment, while from a micro perspective, it is necessary to extract the types of urban street landscape interface elements from the human visual perspective of BSV. The approach of integrating vast datasets and high-resolution imagery to distill multi-scale, multifaceted factors holds promise for practical implementation by professionals in the field. These insights shed light on the nuanced ways in which the different scales and aspects of the urban built environment shape the urban green view index, enriching our comprehension of urban greening’s current landscape and trajectories. This not only augments our understanding of urban green dynamics but also serves as a crucial reference for future urban planning and ecological conservation efforts, setting robust groundwork for subsequent inquiries in urban studies.

3. Study Area and Data

3.1. Study Area

Nanjing, the capital of Jiangsu Province in eastern China, is a significant hub for research, education, and transportation, covering an area of 6587.02 square kilometers with an urban zone of 868.28 square kilometers. As of 2021, it has a population of 9.4234 million, with an urbanization rate of 86.9%. Located along the lower Yangtze River, its coordinates range from 31°14′ N to 32°37′ N latitude and 118°22′ E to 119°14′ E longitude. This study’s research area, Nanjing, is a key city in China’s eastern region, illustrated in Figure 1.

3.2. Panoramic Street View Image Acquisition

The role of streetscape imagery in urban science research is becoming increasingly pivotal, as these images afford researchers a more direct glimpse into the urban fabric and its myriad environmental components. This growing significance is attributed to the fact that street view photographs closely mirror the human vantage point, capturing the essence of urban landscapes as perceived by a pedestrian. This not only engenders a palpable sense of presence but also enriches the dataset with nuanced details. Beyond serving as navigational aids on mobile and desktop platforms, many mapping services extend their utility through the provision of Application Programming Interfaces (APIs) like those of Baidu Maps, Tencent Maps, and Google Maps. These APIs empower users to retrieve street-level imagery tagged with precise geographical coordinates, thereby facilitating a wide array of urban studies and applications.

In this study, streetscape imagery was sourced from BSV panoramas, adhering to the methodology delineated by Li [17]. The initial phase entailed acquiring BSV panorama IDs through coordinate inputs. Subsequently, the panorama segments were retrieved from Google’s servers and, through automated software, assembled into comprehensive BSV panoramas. The study’s road network data was extracted from OpenStreetMap (OSM), with streetscape capture points strategically placed at 50 m intervals utilizing ArcGIS 10.8 software, culminating in a total of 14,279 sampling points across the study area. Continuous urban street view images were obtained after collecting the BSV of the sampling points. The street view imagery for each roadway within the study zone was amassed in alignment with the coordinates of these designated points. However, it is noteworthy that some sampling locations lacked corresponding imagery, necessitating their exclusion from the dataset. The coordinates of these street points facilitated the download of BSV panoramas through a custom Python script, enabling the comprehensive collection of metadata for all BSV panoramas encompassing the study locale.

3.3. S-G-S-S: Street Greening Spatial Structure Dataset

Cityscapes and ADE20K stand as significant large-scale street image repositories traditionally utilized for the appraisal of street-level green quality. Cityscapes, a relatively recent addition to the urban landscape datasets, serves as a benchmark for urban scene image segmentation models. This dataset, unveiled by Daimler AG, now operating under Mercedes-Benz R&D, during the CVPR 2016, is accessible to researchers affiliated with Daimler AG and the Technische Universität Darmstadt in Germany. Its compilation involved capturing stereo visual video sequences through binocular cameras across 50 cities within Germany and adjacent nations during the seasons of spring, summer, and fall. Cityscapes is distinguished by its inclusion of neighborhood scenes and meticulous pixel-level annotations for 5000 images, complemented by a more extensive collection of weakly annotated images. However, it is noteworthy that the dataset does not encompass extreme weather scenarios, such as rainfall and snowfall, which necessitates the adoption of specialized processing techniques and datasets for research endeavors [18].

The S-S-G-S dataset is a specialized resource designed for the evaluation of greening quality at the street level, drawing on a trio of data sources: the Cityscapes dataset, GSV panoramic images, and photographs taken manually. This dataset encompasses a quartet of scene categories—urban, rural, campus, and residential—each selected for their relevance to everyday life and the overarching aims of the research. Urban and rural areas are a broad category, while campuses and residences are subcategories of cities. Therefore, the clear distinction between urban and rural landscapes necessitates distinct datasets for each, due to their different compositions. Campus and residential settings, integral to urban fabric, offer a detailed look at both the interior and exterior aspects of such areas, thus enriching the urban scene spectrum further. This dataset is meticulously labeled to identify four feature types: background, trees, bushes, and grass, allowing for a nuanced analysis of greenery. It is organized into two segments: labeled data and the original images. The original images maintain a resolution of 1280 × 639 pixels, a DPI of 96, and a 24-bit depth, mirroring the specifications of the labeled images. Geographically focused on the Providence area, the dataset offers a localized yet diverse array of scenes for assessing street-level greening quality.

Large street image datasets such as Cityscapes boast considerable benefits, including an expansive sample size, heightened accuracy in data processing, and a diverse array of street images spanning various cities and scenes. Nonetheless, these datasets encounter limitations in variability, falling short in fully encapsulating the comprehensive structure of street greening. The S-S-G-S dataset addresses this shortcoming by incorporating a wide spectrum of scenes, specifically tailored for the evaluation of street-level greening quality. This focus ensures a more detailed and representative analysis of urban greenery, enhancing the depth and breadth of urban environmental research.

In 1981, the National Institute for Environmental Studies in Japan proposed a quantitative statistical analysis method to identify green spaces that influence positive subjective feelings through specific psychological changes. Subsequently, researchers at the institute formally introduced the concept of green visible value [19]. This index represents the percentage of green visible within the human field of vision, and this physical quantity can serve as a factor for evaluating landscape greening. Since then, many studies have utilized the GVI to assess the visibility of urban green spaces. This is calculated as the ratio of the total green area in photographs taken from four directions (front, back, left, right) on a street to the total area of the four photographs. Although the traditional GVI considers the horizontal view, it still overlooks some vegetation information. Therefore, in this study, we use 360° panoramic street view images as single images representing the overall visual environment of a specific location to evaluate the Panoramic View Green View Index (PVGVI) around the streets [3].

4. Methods

In this research, BSV were employed to capture the streetscapes within the designated study area (Figure 2). The underlying road network information was sourced from OpenStreetMap (OSM), and ArcGIS software facilitated the establishment of street view collection points at 50 m intervals, amassing a total of 80,249 points. Street view imagery for each designated location within the study zone was systematically gathered based on these coordinates. In instances where imagery was unavailable for a given point, that location was omitted from the collection process. Leveraging a Python script and adopting the methodology outlined in Li’s research, panoramic BSV imagery was procured from the Baidu Street View platform. This process entailed the initial input of coordinate data to retrieve the ID of the corresponding BSV panorama, followed by the download of panoramic image segments and their subsequent assembly into complete panoramic views directly from Baidu’s servers. The endeavor resulted in the acquisition of 320,996 street view images and 80,249 panoramic views within the study’s confines. Following a meticulous data cleansing process, which involved the removal of blank and non-viable images, the final tally stood at 308,996 street view images and 77,249 panoramic images. These panoramic captures were then processed through the DeepLabV3+ model, which had been previously trained to delineate micro-level built environmental features within the urban context. Concurrently, granular urban data was collected and categorized using the C.M.E.P.R. (C, commercial; M, medical; E, education; P, public service; R, recreation.) [20] classification framework to represent macro-level built environmental elements within the city.

The culmination of this study involved the application of a Robust regression model, integrating the two distinct data sets to elucidate the dynamic interplay between the urban built environment and the proliferation of urban greenways.

4.1. DeepLabv3+ Neural Network Model and Model Prediction

In this investigation, the selection was made to employ a semantic segmentation network grounded on the DeepLabV3+ neural network architecture, as innovated by Chen and colleagues in 2018 [21]. This model stands out for its blend of precision and speed, setting it apart from conventional models prevalent in the field.

DeepLabV3+ distinguishes itself with superior capabilities for interpreting urban landscapes compared to other leading deep learning frameworks. This advantage stems from the model’s inherent design focus on urban scene analysis, enabling it to adeptly identify the greening structures within urban streetscapes. Utilizing DeepLabV3 as its encoder, the model employs Atrous Convolution to generate feature representations of varied dimensions. It incorporates the Atrous Spatial Pyramid Pooling (ASPP) strategy, allowing for multi-scale feature extraction by sampling at multiple rates for effective upsampling. Moreover, DeepLabV3+ enhances the delineation of boundary details through its cascade decoder mechanism. To streamline the model and boost both accuracy and processing velocity, deep separable convolution is implemented, reducing the overall parameter count. This combination of advanced techniques underpins the model’s effectiveness in urban scene segmentation, as highlighted in the research by Chen et al. [21]. In this study, we used a system bit of Windows 11, Python version 3.8, PyTorch version 1.7.0, Cuda 11, Torchvision 0.8.0. Pytorch is currently the most popular deep learning tool, which facilitates the reproduction and dissemination of our work.

We tested the model’s MIoU using the VOC dataset and found that it had a 9.62% and 7.73% improvement compared to SegNet and PSPNet, respectively [22]. The VOC dataset is a dataset format, which is combined with a world-class computer vision model, PASCAL VOC challenge (The PASCAL Visual Object Classes). It mainly includes the following categories: object classification, object detection, object segmentation and action classification. The dataset created and used in this study are both in VOC format, which means that any model trained with VOC can use the dataset we created.

MIoU is a standard metric for semantic segmentation and is used in almost all segmentation related papers. MIoU is the intersection over union ratio (IOU) between the true label and the predicted result. Then, the average IOU of all categories is calculated as follows. According to the confusion matrix MIoU formula, it can be equivalent to:

$$MIoU=\frac{1}{k+1}{\displaystyle \sum }_{i=0}^k\frac{TP}{FN+FP+TP}$$

(1)

Figure 3 shows the operation process of the DeepLabV3+ neural network. In this study, we used this model to extract micro urban built-up environmental factors and calculated the percentages of these factors.

A total of 77,249 panoramic street view images were uploaded into the trained DeepLabV3+ neural network model for processing. The model executed predictions on each image by navigating through the street view image folder via folder traversal. It classified the pixels within each panoramic street view image according to predefined semantic labels set during the training phase of the neural network model. The process involved calculating the percentage of various semantic labels within the panoramic street view images, in addition to determining the PVGVI and SVF. The results from this comprehensive traversal were then compiled into a CSV file. This step involved linking the latitude and longitude coordinates of the street view points with the derived analytical outcomes, which were subsequently visualized using ArcGIS software. The outcomes of these predictions are presented in

Table 1. Partial prediction results of DeepLabV3 + neural network model.

Image ID	X	Y	PVGVI	SVF	…
0	118.7793	32.06055	0.067337	0.36426	…
1	118.7796	32.06054	0.161056	0.277854	…
10	118.7831	32.06027	0.249094	0.179226	…
100	118.7809	32.06701	0.100215	0.240295	…
1000	118.8376	32.05576	0.608715	0.183073	…
10,000	118.7585	32.04672	0.033413	0.320545	…
10,001	118.793	32.05983	0.298045	0.271198	…
10,002	118.7929	32.05988	0.326236	0.243465	…
10,003	118.8133	32.07326	0.213156	0.152733	…
…	…	…	…	…	…

, and Figure 4 displays the visualized prediction results, offering a graphical representation of the analyzed data.

4.2. Obtaining Fine-Grained Urban Data

This study utilizes Python with the GeoPandas library and the Amap API to comprehensively collect urban Points of Interest (POI). The workflow begins with setting up an environment via Anaconda, followed by GeoPandas installation using pip. After acquiring an Amap API key, we define the coordinates for the bounding rectangle around Nanjing and employ a quadtree index to divide this area into smaller segments, each within the POI retrieval limits.

Each segment is then processed to gather detailed urban data, resulting in a collection of 585,006 data points. Data cleansing via pandas removes irrelevant and duplicate entries, yielding 534,606 valid POIs across five Nanjing districts: Gulou, Jianye, Qixia, Qinhuai, and Xuanwu. The cleaned data is summarized in Figure 5, providing a foundation for subsequent analysis.

4.3. Data Processing and Robust Regression Model

4.3.1. Classification of Environmental Factors for Urban Construction

In the study, Python-based web scraping was used to collect urban data, and the DeepLabV3+ computer vision model was utilized to quantify the physical attributes of spaces within the research area, resulting in the calculation of the PVGVI for all streets. Upon improving and creating C.M.E.P.R based on Ye Yu’s 5D model, 25 urban built environment impact factors were identified, including grass, roads, buildings, people, walls, sidewalks, sky visibility, food services, companies, shopping services, financial and insurance services, commercial residential areas, car sales, car services, motorcycle services, accommodation services, medical services, education and culture services, scenic spots, public facilities, transportation services, life services, passage facilities, government institutions, and community and sports leisure services.

These 25 impact factors are broadly divided into 18 macro-level urban built environment impact factors and 7 micro-level urban built environment impact factors (Figure 6). The macro-level factors are the fine-grained urban data (POI) collected through Python scraping, including food services, companies, shopping services, financial and insurance services, commercial residential areas, car sales, car services, motorcycle services, accommodation services, medical services, education and culture services, scenic spots, public facilities, transportation services, life services, passage facilities, government institutions, and community and sports leisure services. The micro-level factors result from the semantic segmentation by the DeepLabV3+ model, trained on the S.S.G.S (Street Space Greening Structure) dataset, and include grass, roads, buildings, people, walls, sidewalks, and SVF.

In this research, the evaluation of macro-level urban environmental factors and urban physical characteristics focuses primarily on the diversity and transportation aspects within the 5D model framework. This approach utilizes Points of Interest (POI) data to measure the diversity and transportation conditions of urban environments. POI data, being fine-grained, more comprehensively reflects the accurate information of urban land use. The POI data used in this study were all downloaded from the Amap (Gaode) API, and were reclassified according to basic urban functions and cleansed of irrelevant, duplicate, and empty data, resulting in 534,606 valid entity POIs. These valid entity POIs were then categorized using the C.M.E.P.R classification [20] method (Figure 7).

The C.M.E.P.R classification, a method for categorizing urban points of interest based on urban built characteristics, allocated the 534,606 valid entities into 18 macro-level urban built environment impact factors according to their actual functions, categorized by urban functions such as commercial, medical, educational, public services, and entertainment. The categorized valid entity POIs were mapped onto the research area’s grid, and the entropy score of POI data in each grid was calculated to determine the diversity, employing the following formula:

$$Mix Index=-{\displaystyle \sum }_{i=1}^n{p}_{i}ln{p}_i$$

(2)

In the formula, $p_i$ represents the proportion of the $i^{th}$ type of POI, and $n$ is the total number of all types of POIs within the grid. This approach allows for a better understanding of the impact relationship between the urban built environment and urban greenways.

At the micro level, urban built environment data is obtained using the DeepLabV3+ neural network model, which has been trained to perform image semantic segmentation on 77,249 panoramic street view images within the study area. This process extracts micro-level urban built environment characteristics. The outputs include the PVGVI and 7 categories of micro-level urban built environment impact factors: grass, roads, buildings, walls, people, sidewalks, and SVF. These elements are used to measure and evaluate urban greenways, in conjunction with macro-level indicators (Figure 8).

4.3.2. Urban Data Processing

Data normalization is a critical preprocessing step in artificial intelligence and deep learning to ensure uniform data scales, preventing the overshadowing of smaller values. Our study utilizes Min-Max Scaling (MMS) for this purpose, a technique that linearly transforms data to a [0, 1] range, enhancing model training and performance, using the following formula:

$$MMS=\frac{X-X_{Min}}{X_{Max}-X_{Min}}$$

(3)

Figure 9 indicates that there are no outliers in the current data, allowing for direct descriptive analysis based on the mean values. The process of observing the research data and determining whether the data distribution conforms to a normal distribution is known as normality testing. The study employs the Jarque–Bera test (J-B test) for this purpose. The results of the J-B test indicate that the research data conforms to a normal distribution.

After the initial processing and normality testing of the data, it is necessary to test the model’s data for outliers. Once it is confirmed that there are no outliers, regression analysis can be conducted on the data. Typically, the testing for outliers in a model includes the following points: (1) the verification of the data’s multicollinearity; (2) the verification of the data’s heteroscedasticity. If the data do not have multicollinearity, the OLS regression model can be used for regression. If there is multicollinearity in the data, ridge regression should be used for regression analysis. After testing for multicollinearity in the data, it is necessary to test for heteroscedasticity in the data. If heteroscedasticity is present, Robust regression (Huber method) should be used for regression analysis.

The study uses the Variance Inflation Factor (VIF) to measure the degree of correlation among several independent variables. In a regression model, the VIF for each variable depends on the level of correlation between that variable and the other independent variables. The formula for calculating VIF is as follows:

$$\mathrm{Variance} \mathrm{Inflation} \mathrm{Factor}=\frac{1}{1-R_i^2}$$

(4)

In the formula, $R_i^2$ refers to the R-squared value obtained by regressing the $i^{th}$ variable against all the other variables in the model, without including a constant term.

Typically, if the VIF of an independent variable exceeds 5, it indicates that the variable has a certain level of autocorrelation, which might minimally impact the results. However, if the VIF of an independent variable exceeds 10, it suggests the presence of high multicollinearity. In cases of high multicollinearity, it may be necessary to consider removing some variables or employing other methods to address the issue. Figure 10 presents the VIF test results for the data.

The test results indicate that the VIF value for roads in the model is greater than 5 but less than 10. This suggests that there may be some collinearity issues present, but they are not significant enough to affect subsequent regression analyses. Based on this preliminary assessment, the data can be included in the regression model. Figure 11 and Figure 12 show the visualization of VIF results.

The study employs the White test and the Breusch–Pagan (B-P) test to assess the data for heteroscedasticity. The null hypothesis for these tests is that the model does not exhibit heteroscedasticity. A p-value less than 0.05 indicates the presence of heteroscedasticity, leading to the rejection of the null hypothesis; conversely, it suggests the acceptance of the null hypothesis, meaning the model does not have heteroscedasticity issues. All data were imported into Pycharm, integrated into the model, and calculated for all sample data, with results shown in

Table 2. Data heteroscedasticity test results.

White Heteroscedasticity Test		Breusch–Pagan Heteroscedasticity Test
χ2	p	χ2	p
8152.290	0.000	2089.960	0.000

. The null hypothesis assumes no heteroscedasticity in the model, and both tests in the table reject the null hypothesis (p < 0.05), indicating the presence of heteroscedasticity in the model.

4.3.3. Robust Model Construction and Data Analysis

Upon testing the data, it was found that there was no multicollinearity, eliminating the need for ridge regression or stepwise regression. However, based on the results of the White test and the Breusch–Pagan (B-P) test, heteroscedasticity was present in the data. Consequently, the conventional linear regression and Ordinary Least Squares (OLS) methods may lead to unstable regression outcomes and biased results due to the heteroscedasticity in the data. Therefore, in selecting a regression model, the first step is to address the data’s heteroscedasticity, leading to the choice of using Robust regression (Huber method) for data analysis. The study employs Python to construct the Robust regression, where the Robust regression algorithms from the statsmodels library can be directly utilized. The results of the regression are presented in Figure 13 and Figure 14.

5. Results

5.1. Micro Variable Robust Regression Results

According to the Robust regression analysis results for micro variables (Figure 9), it is shown that in the micro variables, Grass (regression coefficient = 0.688, p = 0.000 < 0.01), Road (regression coefficient = 0.119, p = 0.000 < 0.01), Person (regression coefficient = 0.616, p = 0.000 < 0.01), and Sidewalk (regression coefficient = 0.615, p = 0.000 < 0.01) have a positive impact on PVGVI. The p-values of all four micro variables are less than 0.01 (p = 0.000 < 0.01), proving a significant influence of these variables. Among them, Grass (regression coefficient = 0.688), Person (regression coefficient = 0.616), and Sidewalk (regression coefficient = 0.615) have relatively high regression coefficients, greater than 0.6, indicating that a unit change in these micro variables leads to a larger average change in the value of PVGVI.

Conversely, Building (regression coefficient = −0.066, p = 0.000 < 0.01), Wall (regression coefficient = −0.008, p = 0.000 < 0.01), and SVF (regression coefficient = −0.046, p = 0.000 < 0.01) have a negative impact on PVGVI. The p-values of these three micro variables are less than 0.01 (p = 0.000 < 0.01), indicating a significant negative influence. However, the regression coefficients for Building (−0.066), Wall (−0.008), and SVF (−0.046) are less than 0.6, suggesting that a unit change in these micro variables leads to a smaller average change in the value of PVGVI. This result may seem counterintuitive.

To explain this, two hypotheses are proposed regarding the effect of reduced sky visibility on PVGVI. Hypothesis 1 suggests that tall buildings and structures may block the view of the sky, reducing sky visibility without necessarily decreasing PVGVI, as tall plants and buildings could both obstruct the sky. Hypothesis 2 proposes that dense, tall vegetation might obstruct both the sky and buildings, implying that at very high PVGVI values (exceeding the normal 30%), vegetation covers both the sky and the buildings, reducing their proportions. Given these hypotheses and the large size of the data sample, the negative impact of buildings and sky visibility on PVGVI, along with the small regression coefficients despite p-values being less than 0.01, can be attributed to these factors. Figure 15 visualizes the micro variable Robust regression results.

5.2. Macro Variable Robust Regression Results

In the Robust regression analysis results for macro variables (Figure 10), catering services (regression coefficient = 0.061, p < 0.01), scenic spots (regression coefficient = 0.023, p < 0.01), transportation facilities services (regression coefficient = 0.688, p < 0.01), science, education, and cultural services (regression coefficient = 0.016, p < 0.01), motorcycle services (regression coefficient = 0.195, p < 0.01), car sales (regression coefficient = 0.030, p < 0.01), life services (regression coefficient = 0.733, p < 0.01), passage facilities (regression coefficient = 0.385, p < 0.01), and medical insurance services (regression coefficient = 0.018, p < 0.01) are among the nine macro variables that have a positive impact on the PVGVI. The p-values of all nine macro variables are less than 0.01, proving a significant influence on PVGVI. Particularly, transportation facilities services (regression coefficient = 0.688) and life services (regression coefficient = 0.733) have relatively high regression coefficients, greater than 0.6, indicating that a higher presence of life service facilities correlates with greater green coverage on city streets.

Conversely, public facilities (regression coefficient = −0.046, p < 0.01), companies (regression coefficient = −0.347, p < 0.01), shopping services (regression coefficient = −0.223, p < 0.01), financial insurance services (regression coefficient = −0.059, p < 0.01), commercial residential (regression coefficient = −0.014, p < 0.01), indoor facilities (regression coefficient = −0.102, p < 0.01), leisure service facilities (regression coefficient = −0.012, p < 0.01), government institutions and social organizations (regression coefficient = −0.012, p < 0.01), and accommodation services (regression coefficient = −0.069, p < 0.01) are nine micro variables that negatively impact PVGVI. The p-values of these nine micro variables are also less than 0.01, indicating a significant negative relationship with PVGVI. Figure 16 visualizes the micro variable Robust regression results.

5.3. The Impact Mechanism of Urban Built Environment Factors on Urban Greenways

Robust regression analysis, conducted simultaneously at both macro and micro levels, reveals significant relationships between a series of micro and macro variables and the PVGVI, uncovering a complex interplay of influences. At the micro level, grass, roads, people, and sidewalks significantly positively affect PVGVI, suggesting that an increase in these elements leads to a higher PVGVI. Conversely, buildings, walls, and sky visibility negatively impact PVGVI, indicating that a reduction in these factors leads to a lower PVGVI.

At the macro level, variables such as catering services, scenic spots, transportation facilities services, science, education, and cultural services, motorcycle services, car sales, life services, passage facilities, and medical insurance services all have a significant positive impact on PVGVI, implying that an abundance of these elements correlates with a higher PVGVI. On the other hand, public facilities, companies, shopping services, financial insurance services, commercial residential, indoor facilities, leisure service facilities, government institutions and social organizations, and accommodation services all negatively affect PVGVI, suggesting that a reduction in these factors leads to a lower PVGVI. Event activities do not influence PVGVI.

This study highlights the significant associations between various micro and macro variables and urban PVGVI, shedding light on intricate mechanisms of influence. At the micro level, grass, people, sidewalks, and roads positively influence PVGVI, reflecting the ecological and aesthetic value of green spaces and the active human engagement in a harmoniously designed urban environment, enhancing PVGVI. However, buildings, walls, and sky visibility have a negative effect on PVGVI, likely due to their constraining effect on green space and ecological diversity.

At the macro level, variables like life services, transportation facilities services, and motorcycle services positively impact PVGVI, possibly due to the economic and human activity flourishing, highlighting urban greening and ecological construction. Conversely, public facilities, companies, and shopping services have a negative impact, possibly due to the sacrifice of green spaces in the pursuit of economic benefits, overlooking environmental protection.

From an interdisciplinary perspective, these findings could result from multiple factors’ combined effects. Ecologically, biodiversity and the richness of green spaces are key to enhancing PVGVI. From an urban planning standpoint, human-centric design and rational spatial layout balance buildings, green spaces, and activity areas, provide more green vistas. Economically, while the prosperity of economic activities might bring attention and investment to the environment, excessive commercialization and pursuit of economic interests could lead to a reduction in green spaces and ecological degradation. These insights reveal how micro and macro factors of the urban built environment impact urban PVGVI from various dimensions, offering a comprehensive and in-depth understanding, valuable for grasping the current state and trends of urban greening and informing future urban planning and environmental protection.

6. Discussion

When discussing the impact of micro and macro factors on the PVGVI, it is essential to delve into how these factors, through complex interactions and mechanisms, shape the green characteristics of urban environments across different levels, dimensions, and domains. The positive influence of micro factors like grass, people, and sidewalks reveals how a city’s ecological environment and human activities intertwine and co-evolve at a granular level, creating vibrant and human-centric urban spaces. Grass and other greenery bring biodiversity and ecological balance to the city, absorbing carbon dioxide and releasing oxygen through photosynthesis, improving air quality, and providing a comfortable living environment for urban residents. The presence of greenery also enriches the urban landscape, enhances city aesthetics and residential satisfaction, and fosters cultural and psychological development.

The significant positive impacts of grass (regression coefficient = 0.688) and people (regression coefficient = 0.616) merit further exploration. The integration of greenery and human activity not only creates a harmonious ecological and cultural environment but also fosters communication and understanding between people and nature, nurturing environmental awareness and ecological ethics. The influence of sidewalks (regression coefficient = 0.615) reflects the important role of public spaces and transportation infrastructure in urban environmental construction. A well-developed network of sidewalks encourages walking and non-motorized travel, reducing car usage, further lowering carbon emissions and environmental pollution, and supporting urban sustainable development. However, the negative impacts of buildings (regression coefficient = −0.066), walls (regression coefficient = −0.008), and sky visibility (regression coefficient = −0.046) expose deeper issues in urban development. The presence of buildings and walls may compress green and open spaces, limiting the city’s ecological potential and environmental quality. Reduced sky visibility, possibly due to excessive building density and height, leads to spatial enclosure and insufficient sunlight, affecting residents’ quality of life and psychological well-being. This spatial and environmental oppression not only restricts urban ecological and cultural development but also diminishes human freedom and happiness. Therefore, achieving a harmonious coexistence of architecture, space, and ecology in densely populated and competitive urban environments poses a significant challenge for current urban planning and design.

Exploring macro-level factors, such as catering services (regression coefficient = 0.061) and life services (regression coefficient = 0.733), the positive correlations reflect the interactive relationship between urban economic activities and the environment. On one hand, economically prosperous and service-rich areas often attract more population and investment, promoting environmental protection and greening initiatives. On the other hand, a high-quality environment enhances the attractiveness and competitiveness of an area, further driving economic growth and social progress. This positive cycle between the economy and the environment constructs a sustainable urban ecosystem. However, the negative correlations of factors like companies (regression coefficient = −0.347) and shopping services (regression coefficient = −0.223) also reveal potential conflicts between economic interests and environmental values. In the pursuit of profit and efficiency, the environment is often neglected and sacrificed, leading to the overconsumption of resources and unsustainable ecological practices. Addressing these conflicts and challenges requires comprehensive and in-depth discussions across policy, theory, and practice. Governments and society should strengthen environmental protection laws and regulations, increase the cost of environmental pollution, and guide businesses and individuals towards green and sustainable development paths. The academic and research community should deepen theoretical studies of urban environmental systems, revealing their inherent mechanisms and patterns to provide scientific foundations and methodological support for urban planning and management. Practitioners in urban planning and design should explore and innovate green and human-centric urban forms and spatial models to achieve harmony between humans and nature, and between the economy and the environment.

Research has shown that improvements in GVI can improve the psychological and physical health of urban residents [23]. Moreover, the presence of vegetation typically enhances people’s aesthetic evaluation of urban landscapes [24,25]. Whether people can appreciate green landscapes seems to affect their postoperative recovery and enhance their rehabilitation potential [26]. Street greening also provides a warm environment for students and teachers, and encourages outdoor activities [27]. Therefore, the findings of this study have significant implications for urban planning policy and practical applications. The robust positive correlation between urban greenery and factors such as transportation facilities, life services, and catering services suggests that urban planners and policymakers should prioritize integrating green spaces with essential urban services. This integration not only enhances the aesthetic and environmental quality of urban areas but also promotes economic vitality and social well-being.

Policy initiatives should focus on incentivizing green infrastructure development within commercial and residential projects. For instance, zoning laws could be revised to mandate a minimum percentage of green coverage in new developments, especially in densely built environments. Additionally, the establishment of green corridors and urban greenways should be strategically aligned with key public service areas to maximize their accessibility and benefits to urban residents. Practically, urban planners can utilize the insights from this study to implement targeted greening strategies in areas identified as having lower PVGVI scores. By leveraging computer vision and big data analytics, cities can continuously monitor and assess the effectiveness of these strategies, ensuring that urban greening efforts are both efficient and adaptive to changing urban dynamics. Furthermore, community engagement programs can be designed to foster public participation in maintaining and expanding urban green spaces, thereby enhancing the overall success of these initiatives. Overall, the integration of advanced technologies and interdisciplinary approaches in urban planning can pave the way for more sustainable, livable, and resilient cities. The study underscores the need for a holistic approach that considers ecological, social, and economic dimensions in urban development policies and practices.

This study, by analyzing the impact of variables at micro and macro levels on PVGVI, unveils the multidimensionality and complexity of urban development and environmental construction. These findings offer valuable insights and lessons, aiding a more comprehensive and profound understanding of the composition and evolution of urban environments, and providing scientific evidence and strategic guidance for future urban greening and sustainable development.

7. Conclusions

7.1. Analysis of Multilevel Robust Regression Results

In summarizing our exploration into the determinants of the PVGVI, we underscore the importance of synthesizing insights from both granular and broader perspectives to inform sustainable urban development and environmental conservation. Our investigation elucidates the intricate interplay between natural ecosystems, human endeavors, economic functions, and societal frameworks, weaving a complex tapestry of urban environmental dynamics.

Micro-level positive influencers, such as vegetation density (evidenced by a regression coefficient of 0.688 for grass) and human presence (regression coefficient = 0.616), alongside urban infrastructure like sidewalks (regression coefficient = 0.615), underscore the symbiosis between urban biodiversity and cultural vitality. These elements are pivotal in fostering an urban ecological equilibrium and enriching urban living standards. Conversely, the deleterious effects of built structures and limited sky exposure spotlight the ecological and psychological strains imposed by rampant urbanization and densification, advocating for more sustainable urban planning and architectural strategies.

At the macro level, the positive association of service sectors like dining (regression coefficient = 0.061) and lifestyle services (regression coefficient = 0.733) with PVGVI underscores the mutual reinforcement of economic vibrancy and environmental stewardship. Prosperous regions with a wealth of services tend to prioritize and invest in green initiatives, which in turn, can spur regional economic growth and societal advancement. Nonetheless, the adverse implications of certain macro-level variables, such as corporate and retail services, signal the persistent tug-of-war between economic expansion and ecological preservation, urging a more conscientious integration of environmental ethics and sustainable practices in the pursuit of prosperity.

7.2. Research Limitations

In this investigation, while elucidating the PVGVI’s dependency on diverse factors, we must concede certain constraints that might curtail the breadth and fidelity of our findings. Primarily, the dependency on readily accessible datasets might have led to the exclusion of pivotal variables, potentially skewing the outcomes. Moreover, the analytical methods, particularly linear regression, may not fully encapsulate the intricate, non-linear interplays within the data, possibly inducing biases. Furthermore, the granularity of our factor analysis could be refined, as current representations might be overly reductive, impacting the nuance of our interpretations. Additionally, the study’s contextual framework might impinge on its broader applicability, given the variability across different geographical and cultural landscapes. Importantly, there is an avenue for methodological and theoretical enrichment, particularly through the integration of more sophisticated statistical models and machine learning algorithms, to unravel more complex interactions. Despite these limitations, our study marks a step forward in understanding sustainable urban environmental development, underscoring the necessity for ongoing enhancement in research methodologies and theoretical constructs.

7.3. Future Applicability

Our findings offer insights and directions for future urban environmental research. Firstly, it is crucial to delve into the interplay between micro and macro factors to grasp the complexity and diversity of urban environments more thoroughly. Future studies should incorporate a broader array of variables, including climate change impacts, social networks, and population dynamics, to enhance the richness and breadth of urban environmental research. Furthermore, there is a pressing need to develop and apply novel theoretical frameworks and methodologies for a more precise and effective analysis of urban environmental changes and their implications. Such advancements will provide valuable, innovative strategies for sustainable urban development and management. By examining the multifaceted influences at both micro and macro levels, this study lays a solid foundation for understanding and improving urban environments, contributing significantly to the promotion of green development and the construction of ecological civilizations in urban settings.