Due to the varying names of residential communities and the absence of applicable programming rules for extracting residential community areas, all the residential communities covered with BSV images in the study area were identified manually, and so were their inner road networks. A total length of 79.40 km road network was delineated in the study communities using the professional geographical information system software ArcGIS Desktop^® 10.1 (Environmental Systems Research Institute, Inc., Redlands, CA, USA). Considering the visible distance in the residential community, a minimum distance of 100 m was established between two adjacent sampling sites. A total of 7,940 sampling sites were thus randomly selected throughout the road networks. Subsequently, these sites were thoroughly examined and 245 candidates that were situated too close to other sites at the intersections were eliminated. Finally, a total of 7,695 sampling sites were selected and assigned sequential BSV-ID (0-7694) for investigation and analysis (Table 1). It should be noted that “greenery” is constantly used in this study to characterize the amount of green vegetation and to differentiate it from the “greenness” of the tasseled cap transition in remote sensing processing method.

To reduce the impact of fluctuations in road width on the assessment, this study used BSV images that were taken along the road direction. For instance, all the sampling sites distributed along the north-south (N-S) oriented roads would have BSV images heading northwards; BSV images facing east were chosen and acquired for the roads with an east-west orientation (E-W). Each available BSV image can be requested in an HTTP URL form using BSV Image API.¹ In this study, the obtained BSV images have a 160° horizontal coverage and 0° vertical coverage, which is similar to people’s optic angle and view.

Similar to other profile image systems, BSV imagery includes no NIR information but captures only R, G, and B band information. The spectral ranges of landscape features in the residential community, including sky, shade, pole, vehicle, building, and pavement, largely overlap with each other’s, which makes the images less useful as the source of information for vegetation classification. Thus, image classification methods that employ NIR band information or NIR-based vegetation indices may not be able to effectively identify vegetation in BSV images and distinguish vegetation from other artificial features.

This study compared the various combinations of the three band information (Table 2). The information of R, G, and B, and the sum, subtract, and divide of these three bands were computed respectively. The R, G, and B band values were normalized using Equation 1 to minimize the pixel value difference of the same commonly seen features. The value ranges of vegetation greenery and other features using the equations in Table 2 were compared to examine the separation possibility in the BSV images. Finally, a NVGI (Equation 2) in the line of ID 9 in Table 2 was proposed to detect and monitor the vegetation greenery in the BSV images.

where G, R, and B represent the normalized G, R, and B band values of the BSV images; G_pixel, R_pixel, and B_pixel are the corresponding band values; G_min, R_min, and B_min denote the minimum band values, and G_max, R_max, and B_max are the maximum band values.

Subsequent to the NVGI calculation, the grayscale images were obtained and the NVGI value ranges of green vegetation and other features can be identified. The green vegetation area in each image was then distinguished based on the segmentation method (AlSaeed et al., 2016), in which the pixels of each image are represented in L gray levels [1,2,…L], the number of pixels at level i is represented by n_i, and the total number of pixels using N = n₁+n₂+….n_L. The probability distribution is:

If the image is classified into two classes C₀ and C₁ by a threshold at level t, C₀ is the pixels with gray levels [1,2…,t] and C₁ is [t+1,….L]. The probability of class occurrence and class means are:

where ω₀ and μ₀, ω₁ and μ₁ are the zeroth and first-order cumulative moments of the histogram up to the t-th level and are defined as follows:

μ_T is the total mean level of the original image which is computed as:

The class variances and total variance are computed using the equations below:

To evaluate the threshold (at level t), the within-class variance (σw2) and between-class variance (σB2) is used as measures of class, separability, and their equations are as follows.

in which

As σB2is dependent on first order statistics, while σw2 depends on the second order statistics. Therefore, σB2is the simplest measure for t. Thus, η(t) is adopted as the criterion measure to select the best threshold value (t) which is determined by a sequential search using the following functions.

Considering σT2 is not a function of threshold (t), the optimal threshold should be the value that maximizes the between-class variance [σB2⁢(t)]. Thus the optimal threshold value (t*) can be computed with the equation below.

The support vector machine (SVM) classification method was used as a comparison to distinguish the green vegetation pixels based on the spectral information of the BSV images, which is a typical conventional classification approach for multi-spectral images (Ding et al., 2016). To maximize the distance between the hyperplane and the nearest pixel in each of the two groups, it divides pixels into two groups using hyper planes (Vapnik, 1999). Up until a pair of pixels has the closest distance in various classes, the SVM classification operations are repeated. The key issue that makes SVM difficult to automate is that it is a type of supervised classification approach that requires the user to designate classes of interest and train the classifier before the classification (Woodard, 1992; Samiappan and Moorhead, 2015). In this study, the kernelType was RBF, the gamma was set to 0.5, and the cost was 10. Another 200 randomly chosen BSV images were visually interpreted in this study to create a set of reference data for training the SVM classifier (Lillesand et al., 2015).

Two hundred new BSV images were used to create the verification dataset, which served as a means of validating the identification findings obtained using the NVGI-based and SVM methods. These BSV images’ green vegetation was manually portrayed and utilized as a reference. The confusion matrix (user’s, producer’s, and overall accuracy) and the Kappa coefficient were used to evaluate the correctness of the results generated by the NVGI and SVM methods. When compared to the total number of pixels, the user’s accuracy indicates how many pixels accurately reflect a given class; the producer’s accuracy indicates how many portions of a given class were correctly represented in the categorized image (Tung and LeDrew, 1988). The proportion of correctly categorized pixels to all pixels is known as the overall accuracy. Along with the overall consistency between the reference and classified images, the Kappa coefficient also considers incorrectly classified pixels from the error matrix (Rosenfield and Fitzpatrick-Lins, 1986).

where r represents the number of rows and columns in the confusion matrix, N is the total number of elements, X_ii is the elements in row i and column i, X_i+ is the marginal total of row i, and X_+i is the marginal total of column i.

The green view index (Deng et al., 2021) was used here to quantitatively describe and compare the green vegetation distribution pattern of each sampling site. The green view index is defined as the proportion of green vegetation pixels in the BSV image, its computation equation is shown below.

where N_veg represents the pixel amount of green vegetation in each BSV image, it can be identified using the NVGI method and counted automatically; N_total means the total pixels of each BSV image, which is a constant value in this study (1,521,216 pixels). After the procedure of green view index computation, the sampling sites were further classified into four different levels based on the green view index computation results and the gradation system of extant studies (Aoki, 1987; Yu et al., 2019) with an interval of 10% (Table 3).

Moran’s I and Getis-Ord GI* analysis were employed to analyze the spatial distribution patterns of the green view index. The global Moran’s I is a measure of the overall clustering of the spatial data (Moran, 1950). Compared with the global spatial autocorrelation, the local Moran index calculates the spatial correlation degree between each spatial object in the analysis region and its neighboring objects, analyzes the local characteristic differences in the distribution of spatial objects, and reflects the spatial heterogeneity and instability in the local region (Anselin, 1995). They are defined as:

where I is the Global Moran’s I measuring global autocorrelation, and I_i is local; N is the number of spatial units indexed by i and j; x is the variable of interest; x¯is the mean of x; w_ij is a matrix of spatial weights with zeroes on the diagonal; and W is the sum of all w_ij.

The hot spot analysis tool calculates the Getis-Ord Gi* statistic for each feature in a dataset. It is an effective means to explore the characteristics of local spatial clustering distribution, which can distinguish the degree of clustering of variable spatial distribution by cold and hot spots (Getis and Ord, 2010). The Gi* statistic returned for each feature in the dataset is a z-score. For statistically significant positive z-scores, the larger the z-score is, the more intense the clustering of high values (hot spots); the lower the z-score is less than 0, the more intense the clustering of low values (cold spots).

The Getis-Ord local statistic is given as:

where x_j is the attribute value for feature j, w_i,j is the spatial weight between feature i and j, n is the total number of features.

The original value ranges of the 200 BSV images were shown in Figure 2A. Spectral ranges of vegetation greenery and other common features, including shade, poles, vehicles, buildings, and pavement, are largely overlapped, which reduces the value of the images as a source of data for identifying vegetation. After the normalization process, the features can be separated into two groups, as shown in Figure 2B, the pixel values of vegetation greenery and green shade were different from those of other features. The NVGI computation results were obtained using Equation 2 and shown in Figure 2C. The NVGI value range of green vegetation in the study area is 1.06–1.32, which is different from that of the sky, shade, pole, vehicles, building facade, and pavement. Based on the NVGI differences, the green vegetation pixels in the BSV images can be identified.

The NVGI method was applied to the 7,695 BSV images, and the findings were saved in the GIS database for subsequent study. In two randomly chosen BSV images, shown in Figure 3, the vegetation greenery was extracted using the NVGI, SVM, and manual identification methods to display the extraction findings. White was used to symbolize the green vegetation, while black was used to represent the other features. Green vegetation in the BSV images may be distinguished in the NVGI-based method with high accuracy when compared to the SVM and the reference (the manual method). When employing the NVGI approach, for example, it is possible to identify artificial green features as non-vegetation, such as the vehicles (Figure 3b). However, the SVM-derived results incorrectly classify the green vegetation with other features (Figures 3c, g).

To further verify the accuracies of the NVGI-based and the SVM method, another 200 BSV images were randomly selected from the BSV image dataset. All the images were manually classified by using the polygon partitioning function in the ArcGIS Desktop^® 10.1 (Environmental Systems Research Institute, Inc., Redlands, CA, USA). The classification accuracies of these images were computed, and the green vegetation identification results (dependent variable) with the manual depiction ones (independent variable) were compared. As shown in Table 4, the NVGI method improved the accuracy of vegetation greenery identification by 18% (Kappa of 0.85 vs. 0.72) and 20% (Overall Accuracy of 0.91 vs. 0.76), compared to the SVM methods (Table 4), which is more suited for vegetation greenery identification in the street view images. The results using NVGI-based and SVM methods were then compared with the manually depicted maps. In the fitting results (Figure 4), the red dashed lines are the fitting ones between the independent and dependent variables. The regression line of the NVGI-based method and manual results is close to the diagonal line, and Pearson’s r = 0.993, with p = 0.000, shows that the green vegetation classification made using the NVGI method was closely correlated with those of the reference classification (Figure 4A); the SVM extracted green view indexes were overestimated compared to the manual ones (Figure 4B). The vegetation greenery index values computed using the NVGI-based method were consistent with the artificial refinement, which can be used in further spatial analysis.

The green view index of each sampling site was computed and shown in Figure 5. The dot color represented the percentage of green vegetation in each sampling site, in which red and orange colors indicated low and relatively low green vegetation in the study area, respectively, and green and blue color dots represented the moderate and high green vegetation distribution. It can be found that the sample sites were not evenly distributed within green index grades (Table 5). Sampling sites classified as Grade II and III accounted for more than 60% of the total sites, which indicated that the communities in the study area were not green enough for residents’ sensory and psychological health. Among them, 2,440 sampling sites (31.70%) were classified as Grade II, which had relatively low proportions (10–20%) of vegetation greenery in these BSV images, the residents in the communities may have an impression of relatively low greenery and high man-influenced landscape; 2,469 sites (32.09%) were classified as Grade III, i.e., moderate greenery with 20–30% portions in the BSV images, in which the artificial environment is natural due to the relatedly high proportion of greenery (Table 5). In addition, 584 sites (75.90%) were found to have small proportions of green vegetation (0–10%) in the images, in which 211 sampling sites were found with no green vegetation planted (0%), residents standing or walking past these sites will have a strong perspective of the man-made landscape. A total of 28.62% (2,202) sites had a high percentage of vegetation coverage, showing strong visual and spiritual pleasure.

The vegetation greenery among the sixteen districts varied and showed different spatial distribution patterns, which can be classified into three categories (Table 6): (1) In CP, SJS, FS, HD, DX, and TZ, the prevalent green view index grades were Grade IV (green view index ≥30%), and Grade III (green view index ∈[20%, 30%), Table 6), residents living in these districts may encounter more visible green vegetation when they stand or walk along the inner roads of the communities. (2) In CY, DC, FT, HR, MTG, PG, SY, and XC, the main grades of the green view index were Grade III [green view index ∈(20%, 30%)] and Grade I [green view index ∈(10%, 20%)], the residential communities in these eight districts appeared less green than those in the aforementioned six districts (Figure 5). (3) As for the left two districts (MY and YQ), the dominant grades were Grade I (green view index less than 10%), which meant the residential communities in these two districts lacked green vegetation.

The Moran’s I value of all green view index values based on the NVGI method in the study area (0.476, p = 0.000) demonstrated that the green view index values have positive spatial autocorrelation. Local Moran’s I analysis showed that High-High and Low-Low cluster types prevailed in the study area (Figure 6 and Table 7). Among the 467 sample sites in SJS, 260 exhibited a high-high cluster, accounting for 55.7%, which is the highest percentage in the high-high cluster type. A high-high cluster was present in 557 sample sites in CP, accounting for 52.4%; which was also present in FS and DX, accounting for 32.8 and 32.1%, respectively. In terms of the low-low cluster type, 213 of 266 sample sites in Miyun showed a low-low cluster, accounting for 80.1%. HR and CY followed, with 68 and 59.6%. Five hundred eighty-five of 1028 sample sites in FT showed a low-low cluster, accounting for 56.9%. In addition, the percentage of PG low-low clustering sites was 56.4% (Table 7).

The Getis-Ord GI* analysis was used to detect cold and hot spots of green view index values in the study area. Figure 6b showed whether the spatial clustering of the sampling sites was significant and, if so, at what level (0.01, 0.05, and 0.1 levels). The spatial weight matrix was calculated based on the Euclidean distance between sampling sites, the distance threshold was 485.447 m. The spatial heterogeneity analysis found that the core region of the study area (DC and XC), the core area of the suburban district, such as HR, MY, PG, and MTG, presented low-value clustering (cold spots, 3,205 sampling sites). Conversely, the peripheral zone (e.g., CP and SJS) surrounding the core area mainly presents high-value clustering (hot spots, 2,304 sites).

Baidu street view imagery uses people’s profile optic angle and view, which is different from those of conventional satellite RS images, to capture and record images. Consequently, BSV images are unable to capture green spaces such as green roofs that cannot be observed or detected from a road angle (Yu et al., 2022). However, BSV images enable it to be used to measure vegetation greenery from the viewpoint of people who are standing or walking on the roads, it can be more useful in residential community planning and management. During the image acquisition process, BSV images captured in alignment with the direction of a road’s heading were selected for analyzing vegetation’s greenery. This approach to image selection can enhance the effectiveness and accuracy of vegetation greenery analysis when dealing with varying road widths (Han et al., 2013; Chen et al., 2020). Previous studies employed multiple images (e.g., 18 images: 6 horizontal directions and 3 aspects in each direction; or 12 images: 4 horizontal directions and 3 aspects in each direction) to mitigate the impact of road width on vegetation greenery at a specific location of interest (Li et al., 2015; Chen et al., 2019; Yu et al., 2019). The extensive requirements for image or data can result in time-consuming and labor-intensive analysis of vegetation greenery. In contrast, the proposed method only requires BSV images facing the direction of the road while minimizing potential errors due to variations in road width.

Using the grid-overlay method, the initial urban street green vegetation studies manually identified and measured vegetation greenery from profile images (e.g., Shafer et al., 1969; Nordh et al., 2009). Multi-spectral classification techniques were also developed in conjunction with the creation of RS images to increase the effectiveness and precision of identifying green vegetation from datasets of profile images (Li et al., 2017; Lu, 2019). However, the direct application of the conventional supervised approaches (such as SVM in this study) results in misclassification of vegetation greenery with other common features in the BSV images due to the absence of the NIR band and the spectral overlap of R, G, and B bands. In contrast, the NVGI methodology efficiently extracts more valuable spectral information from the BSV images in comparison to the traditional SVM method, improving the recognition accuracy of green vegetation.

Unlike previous studies that monitored urban street-side greenery using satellite RS images, such as Landsat and Sentinel series data, to compare the extracted green view index results with the typical NDVI (Normalized Difference Vegetation Index), this study did not utilize such methods due to the narrower roads within the residential communities under study, which have a width of less than 4 m. This is in contrast to trunk roads which can be as wide as 30 m, secondary trunk roads at 12 m and side roads at 8 m. Unfortunately, commonly used Landsat-8/9 and Sentinel-2A/B satellite imagery that is available for free download have a pixel size ranging from 10 to 60 m, rendering them unsuitable for conducting comparisons due to their low resolution.

Population growth in cities combined with urban planning policies of densification may drive the conversion of vegetation-planted areas into residential land (Kabisch and Haase, 2014; Weber et al., 2017). The findings of this study indicated that more than 60% of the study sites had low green view indexes. According to the data from the Bureau of Statistics, these residential communities were constructed between 1995 and 2010. Due to their construction period and design philosophy, the issue of massive car ownership was not fully taken into consideration (Frank et al., 2010; Xiao et al., 2021). In many regions of the study area, property management skills and levels were also insufficient. Consequently, private buildings have replaced the community’s open green space (Li and Zhao, 2017; Mandeli, 2019). These factors contribute to the comparatively low levels of greenery in the study area. To achieve better community management and promote residential greenery equality, it is essential to conduct timely and accurate community greenery monitoring.

The NVGI method was proposed to conduct vegetation extraction, whereby the R, G, and B band values were initially normalized to minimize the pixel value discrepancies of identical features in the BSV images. Subsequently, the NVGI was employed to amplify differences between distinct features and identify greenery within the BSV images, as depicted in Figure 2. It allows planners to identify the name and location of residential communities that contain few green areas (especially the zero-value areas). By analyzing the results, these areas can be identified and improved through various measures such as clearing privately occupied areas and implementing structural systems that facilitate the growth of plants like trees, shrubs, and lawns on vertical surfaces (Ghazalli et al., 2019; Chan et al., 2021). Currently, the BSV images have covered 424 cities in China. The proposed framework in this study can be directly applicable to multi-year BSV images, facilitating landscape planners’ evaluation of urban vegetation greenery and its variations. Furthermore, the proposed framework can be utilized in other cities and countries that have street view images (e.g., Google street view image, Tencent street view image, etc.) to explore the greenery characteristics of residential areas.

Although the proposed method enables monitoring of spatial variations in vegetation greenery within residential communities and improve the accuracy of greenery classification and evaluation, this study is not without limitations. Firstly, due to the prolonged impact of the COVID-19 pandemic in China, data on in situ community characteristics such as the facade of the building structure and resident opinions regarding green view index grading were temporarily unavailable for collection. This issue will be addressed in a future study through the implementation of additional interviews and research. Furthermore, the utilization of random sampling to establish study sites may have an impact on the accuracy of Getis-Ord Gi* spatial analysis due to variations in resident age grades and building construction dates among communities. More comprehensive investigations should be conducted to compare the outcomes obtained from different spatial sampling modes. Additionally, in this study, trees and grasses were not distinguished separately in BSV images due to their similar impacts on human pleasure and shared spectral change patterns within the study area. Future research will focus on examining their variations during other seasons and discussing their identification and correlation with NVGI individually.