Quantifying spatial heterogeneity and associated factors of Nanjing’s street green view index using urban vegetation structure and street view imagery

Abstract

Urban street greenery is commonly evaluated using either street-level visual indicators or remote-sensing-based vegetation indices. However, the relationship between pedestrian-visible greenery and overhead vegetation greenness remains insufficiently understood across seasons, vegetation compositions, and heterogeneous urban contexts. Taking Nanjing, China, as a case study, this study develops a citywide analytical framework that integrates Baidu Street View imagery, semantic segmentation, NDVI-based remote sensing, vegetation-composition information, and spatial analysis. A total of 118,830 valid panoramic street-view images were used to calculate the Green View Index (GVI) along Nanjing’s road network. The analysis examined the spatial distribution of GVI, its relationship with NDVI, seasonal variation in GVI–NDVI correspondence, vegetation-composition characteristics, road-class and functional-zone differences, and local clustering patterns. The results show that roadside visible greenery in Nanjing exhibits substantial spatial heterogeneity, with a citywide mean GVI of 0.273 and a standard deviation of 0.208. High-GVI areas were mainly concentrated around major ecological and waterfront landscapes, including the southern foothills of Purple Mountain and the Qinhuai River corridor, whereas low-GVI areas were more frequently observed in dense commercial centers, transportation hubs, and older industrial–residential mixed districts. GVI and NDVI showed a moderate positive correlation, indicating that street-level visible greenery and remotely sensed vegetation greenness are related but not interchangeable. Their relationship also varied seasonally: the correlation was weakest in spring, strengthened in summer, reached its highest level in autumn, and declined slightly in winter. Vegetation composition further contributed to differences in GVI–NDVI correspondence, as trees and grasslands dominated roadside green space while shrubs accounted for a much smaller proportion. Road-class and functional-zone analyses showed that sidewalks, internal roads, ecological zones, and public zones generally had higher GVI values, whereas traffic-priority roads and highly urbanized functional areas tended to show lower visible greening levels. These findings demonstrate the value of combining street-view-based and remote-sensing-based indicators in urban greenery assessment. They also highlight the need to consider seasonal context, vegetation composition, road hierarchy, and functional zoning in citywide evaluations of street greenery. The proposed framework can support more differentiated assessment and planning of visible street greenery in Nanjing and other spatially heterogeneous cities.

1 Introduction

Urban streets are among the most frequently used public spaces in everyday life and therefore play an important role in shaping urban environmental quality and residents’ subjective well-being (Marans, 2003). As a key component of the built environment, street greenery provides both ecological regulation and landscape benefits (Li et al., 2005; Lee and Maheswaran, 2011; Fruth et al., 2019; Ligtermoet et al., 2022). It can help mitigate the urban heat island effect (Wang and Akbari, 2016), improve air quality (Mandal et al., 2023), reduce traffic noise (Van Renterghem et al., 2023), and moderate the urban microclimate (Feng et al., 2022). From the perspective of everyday street users, greenery also contributes to thermal comfort, visual amenity, and more pleasant outdoor experiences, thereby supporting pedestrian well-being (Bowler et al., 2010; Carrus et al., 2015; Speak et al., 2020; Mendzina and Vugule, 2020). In increasingly dense urban environments, maintaining accessible green space has also become a planning challenge, as urban densification can threaten green-space provision and complicate conservation strategies (Haaland and van den Bosch, 2015). At the same time, human-scale measurements of daily accessed street greenery have become important for linking visible greening assessment with everyday exposure, spatial perception, and activity-related urban experience (Ye et al., 2019; Zheng et al., 2023).

Urban street greenery has traditionally been assessed using planimetric or land-use-based indicators, such as the Green Coverage Ratio (GCR) and the Green Land Ratio (GLR), which quantify vegetation cover or the amount of roadside green space from a top-down perspective (Susaki and Kubota, 2017; Pan and Kao, 2021; Zhu et al., 2023). These indicators are useful for describing the quantity and distribution of urban vegetation, but they do not directly capture the greenery that pedestrians perceive while moving through streets. To address this limitation, the Green View Index (GVI) was developed to measure the proportion of visible vegetation within the human field of view and has become an important indicator for evaluating street-level green exposure from a pedestrian perspective (Jing et al., 2024; Zhang and Zeng, 2024). This distinction between overhead vegetation measures and street-level visible greenery also provides the conceptual basis for comparing GVI with remote-sensing-based indicators such as the Normalized Difference Vegetation Index (NDVI).

The conceptual foundation of GVI can be traced to early landscape-perception studies that examined how visible greenery shapes human experience in urban environments. Aoki et al. (1985) and Aoki (1991) provided foundational explorations of greenery visibility and its perceptual significance in urban landscapes. Building on this perspective, Yang et al. (2009) formalized the Green View Index as a measure of visible urban greenery and showed that it captures street-level green exposure in ways that differ from conventional planimetric greening indicators. With the emergence of street-view platforms, Li et al. (2015) extended GVI measurement to large-scale urban assessment by combining Google Street View imagery with a modified GVI calculation framework. Subsequent studies have further established GVI as a pedestrian-centered complement to conventional urban greening indicators and one of the most widely used measures in street-view-based greenery research (Kumakoshi et al., 2020; Lu et al., 2023).

Recent advances in street-view imagery and computer vision have further enabled GVI to be estimated at a much larger scale and with greater analytical efficiency. In particular, semantic segmentation models can identify vegetation pixels in street-view images and distinguish them from non-vegetated elements, thereby supporting automated GVI extraction. Internet-crawling street-view images have been used to quantify green view values and examine their influencing factors at urban-agglomeration scales (Chen et al., 2020), while deep-learning approaches have enabled large-scale quantification of urban tree cover from street-level images (Cai et al., 2018), streetscape greenery assessment using Google Street View (Kameoka et al., 2022), and broader street-view-based urban environmental interpretation. Related semantic-segmentation approaches have also been used to calculate the Green View Index from street-view images (Xia et al., 2021a), estimate other street-level visual environmental indicators such as sky view factor (Xia et al., 2021b), and assess the structure of urban street greenery using panoramic street images (Zhang et al., 2023). More broadly, machine-learning and deep-learning approaches have also been applied to pixel-level classification of ground-based urban imagery, tree-crown and height extraction, and satellite-image-based urban green-space assessment, further supporting the use of image-based methods in urban greenery research (Qamar and Dobler, 2020; Hao et al., 2021; Alfarhood et al., 2025). More recent work has further explored the estimation of visible street greenery by integrating multisource remote-sensing imagery with deep learning (Ma et al., 2025). Together, these technological developments have made it possible to move beyond small-sample visual assessment toward citywide and fine-grained analyses of street greenery.

Beyond improvements in GVI measurement itself, a growing body of research has examined how street-level visible greenery relates to conventional remote-sensing-based vegetation indicators, particularly NDVI. Existing studies generally suggest that GVI and NDVI are positively associated but not interchangeable: NDVI reflects vegetation conditions from a top-down remote-sensing perspective, whereas GVI captures greenery visible to pedestrians at street level (Tong et al., 2020; Kumakoshi et al., 2020; Li et al., 2021; Sun et al., 2026). Recent work has also begun to show that street-visible greenery can vary across observation periods and seasonal contexts (Yu et al., 2022), and that street-view-derived and remotely sensed greenness provide complementary rather than identical information about urban green exposure (Sun et al., 2026). However, most existing analyses of GVI–NDVI correspondence have relied on static or single-period assessments. As a result, the seasonal dynamics of this cross-scale relationship remain insufficiently examined at the citywide scale.

The relationship between visible and remotely sensed greenness may also depend on the structural characteristics of urban vegetation. Previous studies have shown that street-level greenery is shaped not only by the total amount of vegetation, but also by road patterns, vegetation structures, vegetation cover, and broader landscape configurations (Tong et al., 2020; Zhu et al., 2023; Zhang and Zeng, 2024). These characteristics may influence whether vegetation that is prominent in the pedestrian visual field is equally represented by overhead remote-sensing indicators. Recent evidence further suggests that greenery types and built-environment conditions can modify the correspondence between street-view-based and aerial greenness assessments (Sun et al., 2026). Nevertheless, the role of urban vegetation composition in explaining differences in GVI–NDVI associations across heterogeneous street environments has not yet been sufficiently quantified.

In addition to cross-scale correspondence, a further body of scholarship has examined the spatial heterogeneity of GVI across different urban contexts. Existing studies have reported substantial differences in visible street greenery among administrative districts, functional zones, old and new urban areas, and cities with distinct landscape settings (Li et al., 2021; Chen et al., 2022; Lu et al., 2023; Zhang and Zeng, 2024; Gou et al., 2025). Spatial clustering approaches have also begun to be introduced into street-view-based greenery assessment. For example, Song et al. (2023) applied Getis–Ord Gi* analysis to identify hotspots and coldspots of visible greenery in residential communities in Beijing, demonstrating the value of spatial clustering for revealing localized disparities in urban greening. These findings indicate that GVI is sensitive to urban development intensity, land-use context, and broader landscape conditions. However, many existing studies have examined only individual spatial dimensions, selected subareas, or specific urban settings. Relatively few have developed a citywide analytical framework that jointly considers road hierarchy, functional zoning, and spatial clustering characteristics. This limitation is particularly relevant for large and spatially heterogeneous cities such as Nanjing, where mountains, rivers, dense built-up districts, and expanding peripheral areas coexist and may jointly shape the spatial organization of street greenery.

Taken together, previous studies have established the value of GVI for pedestrian-oriented greenery assessment, clarified its complementarity with NDVI, and highlighted the importance of vegetation structure and urban spatial context. However, three interrelated gaps remain. First, the seasonal dynamics of the GVI–NDVI relationship have not been adequately examined at the citywide scale. Second, the contribution of urban vegetation composition to cross-scale differences between visible and remotely sensed greenness remains insufficiently quantified. Third, integrated analyses that connect road-type differentiation, functional-zone characteristics, and GVI spatial clustering are still limited, especially in cities with complex landscape structures and heterogeneous development patterns. These gaps concern not only whether street-level and remotely sensed greenness are related, but also when they correspond, why they diverge, and how such differences are organized across urban space.

To address these gaps, this study takes Nanjing, China, as a case study and develops a multidimensional assessment of urban street greenery. Specifically, this study: (1) maps the citywide spatial distribution and heterogeneity of GVI across the street network; (2) examines seasonal variation in the GVI–NDVI relationship using temporally matched multi-seasonal street-view and remote-sensing data; (3) evaluates how vegetation composition is associated with differences in GVI–NDVI correspondence; (4) compares GVI variation across road types and urban functional zones; and (5) identifies GVI hotspots and coldspots and interprets their spatial correspondence with functional-zone characteristics. By integrating street-view imagery, remote-sensing indicators, vegetation-composition information, and urban spatial differentiation, this study aims to provide a more coherent understanding of the spatial and seasonal patterns of urban street greenery and to support differentiated greening assessment and planning in complex urban environments.

2 Materials and methods

2.1 Study area

This study focuses on Nanjing, a major city in the Yangtze River Delta region of China (Figure 1). According to China’s climatic regionalization, Nanjing is located in the northern subtropical humid climate zone. The city is characterized by a typical humid northern subtropical climate, with distinct seasonal variation and abundant precipitation.

Figure 1

2.2 Methodology

2.2.1 Data collection

Road-network data were obtained from Geofabrik (https://download.geofabrik.de/), covering all roads within Nanjing. Functional-zone classification was based on Point of Interest (POI) data acquired through the AutoNavi Open Platform Web Service API (https://lbs.amap.com/api/webservice/guide/api-advanced/search), which was used to examine the relationship between visible street greenery and urban functional zones (Figure 2).

Figure 2

Street-view images were acquired from the Baidu Street View API. Sampling points were generated at 50 m intervals along the preprocessed road network and were then used to retrieve panoramic images through the API. It should be noted that the Baidu Street View API returns the most recent available panoramic imagery for each sampling location, rather than allowing complete retrieval of images from a predetermined year for all points. Therefore, the acquisition years of valid street-view samples were not uniform across the full dataset. This characteristic of the data source was considered in the subsequent seasonal matching procedure and is further discussed as a limitation of the GVI–NDVI comparison.

During data screening, a set of multidimensional exclusion criteria was applied to improve the reliability of GVI estimation. A total of 323,421 raw sampling points were collected through the Baidu Street View API. After screening, 118,830 valid panoramic images were retained, corresponding to a validity rate of 36.7% (Figure 3). The exclusion criteria were as follows: (1) missing or empty images, including sampling points without accessible street-view data; (2) images in which major sightlines were substantially obstructed by large vehicles, buildings, billboards, or temporary construction sites; (3) images with abnormal shooting angles, such as excessive tilt or rotation, incomplete horizontal field of view, or lenses directed vertically toward the sky or ground, which failed to represent actual street-view conditions; and (4) images with very low clarity or severe blurring, which prevented reliable identification and extraction of vegetation pixels.

Figure 3

All screening procedures were conducted automatically using image metadata and quality-detection algorithms, without manual selection based on land-use type, urban functional zone, or greening condition. Therefore, the screening process was not expected to introduce systematic bias across commercial districts, ecological zones, residential areas, or other functional zones.

Remote-sensing imagery was acquired from the Geospatial Data Cloud Platform (www.gscloud.cn), including Landsat 8 OLI_TIRS and Sentinel-2 MSI images from 2014 to 2022 across 44 temporal intervals. Before NDVI calculation, Sentinel-2 imagery was resampled to 30 m resolution in Google Earth Engine using bilinear interpolation to ensure consistency with the spatial resolution of Landsat 8 while preserving spectral continuity.

2.2.2 Street network preprocessing and sampling optimization

To improve the accuracy of subsequent GVI sampling, NDVI extraction, and spatial clustering analysis, road-network data obtained from Geofabrik were preprocessed and optimized in ArcGIS 10.8.1. First, topological corrections were performed to address data anomalies. For duplicated road segments, the “Delete Duplicates” tool was used to retain the segment with the most complete attribute information and remove redundant records, thereby avoiding repeated sampling. For self-intersecting nodes, the “Snap” tool was used to align intersecting nodes to a shared coordinate within a tolerance of 2 m. The “Merge” tool was then applied to integrate disconnected parts into continuous road segments and maintain consistent sampling intervals. For dangling nodes, local feature-point matching and adjacency verification were conducted to connect endpoints located within 5 m of valid junctions. Road segments shorter than 10 m were removed to improve network consistency.

Second, the street network was spatially calibrated to improve positional accuracy. Baidu Maps high-precision road-centerline data, with a positional accuracy of ±1 m, were used as the reference dataset. A total of 500 control points, including road intersections and landmarks, were systematically selected across the study area to align the original road network with the Baidu Maps reference data. The post-calibration accuracy was evaluated using 100 randomly selected road segments. The average positioning error was 3.2 m, and the maximum error was 4.8 m, meeting the predefined accuracy criterion of ≤5 m.

The calibrated and continuous road network was then used for GVI sampling at 50 m intervals. The positional accuracy of ≤5 m ensured reliable alignment between street-view images retrieved from the Baidu Street View API and their corresponding road segments. The connected road segments also provided a more accurate basis for buffer-zone generation, reducing potential NDVI extraction errors caused by fragmented road geometry. For spatial clustering analysis, the topological integrity of the road network helped ensure that the spatial weight matrix reflected adjacency relationships among sampling locations, thereby improving the reliability of Global Moran’s I and Getis–Ord Gi* analysis. Coordinates and road-class information were recorded for subsequent analysis.

2.2.3 GVI calculation

A ResNet50-based semantic segmentation model from the MIT SemSeg framework was used for pixel-level classification of street-view images. The model incorporates dilated convolutions and the PPM-deepsup module and was initialized with weights pre-trained on the ADE20K scene parsing benchmark, which includes 150 semantic categories for scene classification. It was then adapted to Nanjing street-view imagery through the local training and validation procedure described in Section 2.2.3.1 (Zhou et al., 2017).

The RGB calibration of vegetation-related categories and the subsequent GVI calculation procedure are described in detail in Section 2.2.3.2. After vegetation extraction, the effective pixel counts of trees, shrubs, and herbaceous plants were aggregated for each street-view sample. The Green View Index (GVI), defined as the proportion of visible vegetation pixels within the effective panoramic image area, was then calculated after excluding black borders and blurred regions.

Representative street-view scenes from different functional zones and road types were selected to construct a typical street-scene atlas for different GVI levels in Nanjing (Figure 4). For visual consistency, the representative scenes in Figure 4 were displayed in a uniformly cropped horizontal layout. This cropping was applied only for figure presentation and did not affect the original panoramic images used for semantic segmentation or GVI calculation. The atlas illustrates perceptual differences among streets with high, medium, and low greening levels and provides visual support for the subsequent quantitative analysis of GVI spatial distribution.

Figure 4

2.2.3.1 Model fine-tuning and validation configuration

To adapt the ADE20K-pretrained semantic segmentation model to the visual characteristics of Nanjing street-view imagery, a locally labeled dataset was constructed for model fine-tuning and validation. A random sample of 1,000 street-view images representing typical urban functional zones in Nanjing was manually labeled and divided into a fine-tuning set (80%, 800 images) and a validation set (20%, 200 images). The training hyperparameters followed the SemSeg framework (Zhou et al., 2017) and urban green-space segmentation benchmarks (Oliveira et al., 2019), including the Adam optimizer, an initial learning rate of 0.001, a batch size of 16, a weight decay of 5e-4, 5,000 training epochs, and an exponential learning-rate policy with a decay factor of 0.9. To improve model generalization, data augmentation techniques were applied, including random horizontal flipping (probability p=0.5) and random rotation (0–10°) (Oliveira et al., 2019). Standard metrics for evaluating green-space segmentation performance, namely Intersection over Union (IoU) and F1 score, were used (Zhao et al., 2017; Oliveira et al., 2019). The green-space IoU reached 0.81 (95% confidence interval: 0.78–0.83), and the green-space F1 score reached 0.85. Both metrics fall within the range reported in recent high-precision urban green-space identification studies (Zhu et al., 2025), supporting the reliability of the segmentation results.

To further assess the robustness of vegetation extraction beyond the original training–validation procedure, an additional season-stratified independent test set was constructed. Following the methodological rationale of strengthening model reliability through independent testing in recent street-view-based predictive studies (Xu et al., 2026), 150 street-view images were randomly selected from the full image dataset, including 38 images from spring, 37 from summer, 38 from autumn, and 37 from winter. Vegetation masks were prepared according to a unified annotation criterion focused on visible green vegetation components relevant to GVI calculation, including tree canopies, shrubs, hedges, grasses, and groundcover. These masks were visually reviewed before being compared with the vegetation masks extracted from the semantic segmentation outputs. Pixel-level agreement between the reviewed reference masks and the model-derived vegetation masks was evaluated using IoU, precision, recall, and F1 score. The overall micro-averaged IoU was 0.586, with a precision of 0.699, recall of 0.784, and F1 score of 0.739. Seasonal stratification showed stronger performance in summer (IoU = 0.742, F1 = 0.852) and autumn (IoU = 0.634, F1 = 0.776), moderate performance in spring (IoU = 0.488, F1 = 0.656), and lower performance in winter (IoU = 0.345, F1 = 0.513). The lower winter performance may reflect weaker vegetation chromaticity and more complex vegetation–background boundaries under leaf-off conditions. These supplementary results provide additional evidence for the reliability of vegetation extraction across seasonally heterogeneous street-view scenes, while also indicating that winter-related GVI results should be interpreted with appropriate caution. The season-stratified independent test-set evaluation results are summarized in Table 1.

2.2.3.2 Vegetation RGB calibration and GVI calculation

RGB reference values for common vegetation types, including trees, shrubs, and herbaceous plants, were calibrated using 300 panoramic photographs selected from the full road network of Nanjing through a stratified sampling strategy.

To clarify the sampling logic and ensure reproducibility, the 300 vegetation RGB calibration samples were selected through a single multi-criteria stratified sampling procedure rather than through two independent sampling schemes. During this process, ecological or vegetation context and urban functional-zone distribution were considered simultaneously. Therefore, the two classifications describe different dimensions of the same sample set rather than separate sampling operations.

From an ecological and vegetation perspective, the samples were distributed across three representative landscape contexts in Nanjing: (1) the native forest area of Purple Mountain, which is predominantly composed of indigenous tree species; (2) urban commercial–residential areas characterized by cultivated landscape vegetation and common street trees, such as plane trees, cherry-plum, and boxwood; and (3) riverbank and suburban zones containing mixed vegetation types, including aquatic plants, grasses, and shrubs. Approximately 100 samples were selected from each context to capture variation in dominant vegetation types and spectral characteristics.

At the same time, the same 300 samples were proportionally allocated across six urban functional zones, including ecological, public-service, industrial, commercial, residential, and mixed-use areas, with approximately 50 samples per zone. This allocation was intended to ensure that the calibration dataset reflected the spatial composition of the full set of 118,830 valid street-view samples. Within each functional zone, the distribution of samples across different road classes, including arterial, secondary, and local roads, was also examined to reduce sampling bias.

These criteria were applied jointly within the same sampling procedure, meaning that each selected sample belonged to both one ecological context and one functional zone. This multi-criteria stratified sampling approach was designed to improve vegetation-type coverage, spatial representativeness, and methodological reproducibility.

The K-means clustering algorithm was then used to calibrate RGB baseline values for the three vegetation types. It should be clarified that this K-means procedure was not applied directly to raw street-view images to distinguish vegetation from all urban objects. Instead, it was conducted after semantic segmentation, which had already performed pixel-level scene classification and largely separated vegetation-related pixels from non-vegetation elements. Therefore, visually green non-vegetation objects, such as billboards, building façades, or vehicles, were primarily addressed during the semantic segmentation stage rather than by the subsequent K-means calibration.

The number of clusters was set to three, corresponding to trees, shrubs, and herbaceous plants. Euclidean distance was used as the similarity metric, and K-means++ initialization was adopted to reduce random initialization bias. The convergence threshold was set to 1 × 10⁻⁶, with a maximum of 300 iterations. The green band (G) was selected as the clustering feature dimension for vegetation RGB calibration because it provides a strong discriminative response among the vegetation samples considered in this study. The reliability of the overall vegetation-extraction pipeline, including possible residual ambiguity in complex street scenes, was further examined through the independent test-set evaluation described in Section 2.2.3.1.

To verify representativeness, homogeneity tests were conducted by comparing vegetation RGB features from the 300 calibration samples with those from the remaining 118,530 valid samples. The intra-class variance of RGB values for similar vegetation types across the city remained below 0.02 after normalization to the 0–1 range, and the calibration accuracy of the 300 samples in representing citywide vegetation RGB characteristics reached 94.7%. These results indicate that the selected samples adequately captured the variation in roadside vegetation RGB features in Nanjing.

A sensitivity analysis was also performed to evaluate the influence of calibration sample size on GVI computation accuracy. GVI values derived from the 300 stratified samples were compared with those obtained from 1,000 and 2,000 randomly selected samples. The results showed a Mean Absolute Error (MAE) of 0.012 and a Root Mean Square Error (RMSE) of 0.018, indicating only minor differences among the sample-size settings. This suggests that 300 calibration samples were sufficient for reliable GVI estimation in this study, and that further increasing the sample size did not substantially improve computational accuracy.

Using the statistical distribution of the 118,830 valid GVI samples and the five-tier classification system for green visibility proposed by Orihara (2006) and summarized by Xiao et al. (2018), GVI values were classified into five levels using the Jenks natural breaks method. The thresholds were defined as follows: no greening (GVI < 0.05), low greening (0.05 ≤ GVI < 0.15), moderate greening (0.15 ≤ GVI < 0.25), greening (0.25 ≤ GVI < 0.35), and high greening (GVI ≥ 0.35). These thresholds are consistent with the perceptual evaluation ranges reported in previous studies (Orihara, 2006; Xiao et al., 2018).

In particular, GVI = 0.15 represents the lower boundary of the “moderate greening” level in Orihara’s perceptual classification, corresponding to the point at which urban greenery becomes visibly present in the human visual field. Therefore, this value was adopted as the minimum benchmark for identifying basic street-greening conditions in this study. By combining an established perceptual classification framework with the statistical distribution of citywide samples, this classification supports both methodological consistency with previous research and practical applicability to large-scale semantic-segmentation-based GVI estimation.

2.2.4 NDVI calculation

The NDVI values for road segments were derived from Landsat 8 OLI_TIRS and Sentinel-2 MSI remote-sensing imagery acquired from the Geospatial Data Cloud Platform for the period 2014–2022, comprising 44 temporal intervals. To ensure consistency in spatial resolution, Sentinel-2 imagery, originally available at 10 m resolution, was resampled to 30 m in Google Earth Engine using bilinear interpolation before NDVI calculation. This procedure aligned the spatial resolution of Sentinel-2 imagery with that of Landsat 8 data.

NDVI was then calculated for each image using the following formula, producing 30 m × 30 m NDVI raster datasets covering the entire study area:

where NIR denotes reflectance in the near-infrared band and R denotes reflectance in the red band.

Road boundary polygons were generated from the preprocessed road network described in Section 2.2.2 by creating 10 m buffer zones on both sides of each road centerline. To ensure spatial consistency between the road polygons and the NDVI raster datasets, all spatial data were projected to WGS84/UTM Zone 50N and aligned using the ArcGIS “Snap Raster” procedure. The mean NDVI value for each road segment was then extracted using the “Zonal Statistics as Table” tool in ArcGIS, with the road boundary polygon used as the statistical unit and the overlapping NDVI raster cells used for value aggregation.

Figure 5 illustrates the zonal statistics procedure. The road boundary polygon identifies the raster cells spatially associated with each road segment, and the mean value of these overlapping cells was used as the NDVI indicator for the corresponding road segment.

Figure 5

2.2.5 Statistical examination of the seasonal GVI–NDVI correlation

Given that GVI and NDVI characterize urban greenery from different observational perspectives—pedestrian-level visible greenness and top-down vegetation coverage, respectively—their relationship was examined as a supplementary comparison and consistency analysis rather than as direct mutual validation. Correlation analysis was conducted for each season using matched GVI–NDVI samples to evaluate the association between street-level visible greenness and remotely sensed vegetation coverage. Spearman’s rank correlation coefficients were calculated to assess the strength of the monotonic association between the two indicators, with statistical significance evaluated at α=0.05. To examine whether the observed GVI–NDVI relationship remained stable across different analytical contexts rather than depending solely on a pooled result, the association was further assessed through stratified analyses by season, administrative district, and functional zone. These stratified analyses served as consistency checks for the cross-scale comparison between pedestrian-level GVI and remotely sensed NDVI.

To improve spatiotemporal comparability, GVI and NDVI were matched at both temporal and spatial levels before correlation analysis. Because the Baidu Street View API provides the most recent available imagery rather than a complete year-specific archive for all sampling points, strict one-to-one matching by identical acquisition year and date could not be achieved across the full dataset. Therefore, this study adopted a seasonal-window-based matching strategy. Temporally, each valid street-view sampling point was paired with NDVI imagery acquired within the corresponding 30-day seasonal window of the street-view acquisition period, so that both indicators represented vegetation conditions during a comparable phenological phase. Spatially, all datasets were projected to WGS84/UTM Zone 50N. Each GVI sampling point was then aligned with the corresponding 30 m NDVI raster cell using the ArcGIS “Snap Raster” procedure, with the spatial offset controlled within 5 m. Samples lacking valid street-view imagery, temporally corresponding NDVI observations, or valid NDVI raster values were excluded from the matched dataset.

Due to the limited availability of summer street-view data, a post hoc statistical power analysis was conducted to assess whether the available sample size was sufficient to detect the observed relationship. Power was estimated using Fisher’s Z transformation of the observed rank correlation coefficient, with Spearman’s ρ=0.5399, n=2,725, and α=0.05 under a two-tailed test. The resulting statistical power exceeded 0.99, indicating that the available summer sample size provided sufficient statistical sensitivity to detect the observed GVI–NDVI association. Therefore, the significant summer correlation is unlikely to be explained solely by insufficient sample size or random sampling error, although potential bias related to sample representativeness and image availability should still be considered.

The Kruskal–Wallis H test was further applied to evaluate whether GVI–NDVI correlation strengths differed significantly across seasons, thereby assessing seasonal variation in the cross-scale relationship between the two indicators.

2.2.6 Spatial cluster analysis of road GVI

The Getis–Ord Gi* statistic (Getis and Ord, 1992), a widely used local spatial autocorrelation measure, was applied to identify statistically significant hotspots and coldspots of GVI across Nanjing’s road network.

Parameter settings were defined according to the spatial characteristics of the data. A fixed-distance-band approach was used to construct the spatial weight matrix, reducing potential bias caused by uneven point distribution. The distance threshold was set to 300 m, taking into account the 50 m street-view sampling interval and the average density of Nanjing’s road network. This setting ensured that each sampling point was evaluated in relation to a local neighborhood of nearby points while retaining regional interpretability. Row standardization was applied to reduce bias associated with differences in neighbor counts. Two-tailed significance tests were conducted at the 90%, 95%, and 99% confidence levels, and clustering intensity was categorized into Gi_Bin values of ±1, ± 2, and ±3; Gi_Bin = 0 indicates no statistically significant spatial clustering. The analysis was implemented in ArcGIS 10.8.1 using the Hot Spot Analysis (Getis–Ord Gi*) tool, with GVI sample points as input features and GVI values as the analysis field. For each sampling point, the output included Gi_Bin, GiZScore, and GiPValue, which were used to classify locations as significant hotspots, significant coldspots, or non-significant areas for subsequent spatial overlay analysis with functional zones.

3 Results

3.1 Spatial distribution

GVI values across Nanjing showed clear spatial heterogeneity, with a citywide mean of 0.273 and a standard deviation of 0.208, reflecting marked variation in roadside visible greenness among urban areas (Figure 6).

Figure 6

Figure 7 shows the spatial concentration of high- and low-GVI areas across Nanjing’s road network. Kernel density estimation was used to generate a heat map of GVI concentration, and the values were classified into five categories using the Jenks natural breaks method, consistent with Figure 6. High-GVI clusters were mainly concentrated in the southern foothills of Purple Mountain, along the Qinhuai River waterfront corridor, and within Xianlin University Town, forming several contiguous high-value zones. In contrast, low-GVI clusters were concentrated in the Xinjiekou commercial core, the Nanjing South Railway Station transportation hub, and mixed industrial–residential areas in the old city, showing a more dispersed multi-point pattern. This spatial pattern is broadly consistent with the GVI distribution shown in Figure 6 and further highlights areas of local concentration.

Figure 7

The observed spatial pattern indicates that natural landscape contexts and urban functional layouts are closely related to the spatial variability of GVI in Nanjing. High-GVI clusters were mainly observed near ecological and waterfront landscapes, including the southern foothills of Purple Mountain and the Qinhuai River corridor. Low-GVI areas were more frequently concentrated in highly developed commercial, transportation, and mixed-use urban districts, indicating a clear spatial contrast between landscape-oriented and highly urbanized street environments.

3.1.1 GVI zones

High-GVI areas (GVI > 0.25) accounted for 45.3% of all samples, with values ranging from 0.250 to 0.999 and a mean of 0.454. These road segments were mainly concentrated in four types of urban landscape settings. First, road sections along the southern foothills of Purple Mountain, such as Lingyuan Road and Zhongshanmen Avenue, were characterized by continuous woodland edges and tree-lined corridors, with GVI values commonly ranging from 0.35 to 0.50. The highest GVI values, approaching 0.999, were observed in highly enclosed shaded corridor landscapes. Second, the upper Qinhuai River waterfront corridor, including Qingliangmen Street and Hanzhongmen Street, showed vertically layered roadside greening composed of waterfront green spaces, street trees, and riparian vegetation, with GVI values generally ranging from 0.28 to 0.40. Third, auxiliary arterial roads within the core residential area of Hexi New City, such as Jiangdong Middle Road and Mengdu Avenue, formed relatively continuous green interfaces through roadside parks and community green spaces, with GVI values of approximately 0.25–0.35. Fourth, central Xianlin University Town, including Wenyuan Road and Xuehai Road, combined campus greenery, community green spaces, street trees, and lawns, with GVI values of approximately 0.25–0.32.

Low-GVI areas (GVI < 0.08) accounted for 17.1% of all samples, with values ranging from 0.000 to 0.080 and a mean of 0.038. These road segments showed clear spatial concentrations in areas with limited roadside vegetation and intensive built-up conditions. First, the southern section of the Xinjiekou–Fuzimiao commercial core, including Zhongshan South Road and Hongwu Road, was characterized by high-density commercial buildings and compact street interfaces that restricted visible greenery. Second, the area east of the Nanjing South Railway Station–Jiangning Dongshan transportation hub, including Yonglan Road and Shuanglong Avenue, was dominated by intensive traffic functions and motor-vehicle-oriented road cross-sections, with GVI values generally ranging from 0.005 to 0.080. Third, older industrial–residential mixed areas, such as Yuhuatai District and eastern Qinhuai District, including Daming Road and Kazi Men Street, exhibited limited green infrastructure and discontinuous roadside vegetation, with many road segments showing GVI values between 0.002 and 0.070.

3.1.2 Spatial clustering of GVI

Hotspot and coldspot analysis further revealed the local clustering pattern of GVI (Figure 8). Hotspots at the 95% confidence level (Gi_Bin = 2, 0.01 ≤ GiPValue < 0.05) were mainly concentrated in ecological and public-service areas. These hotspots were associated with road segments adjacent to core green spaces, especially the inner ring of the Purple Mountain Scenic Area, as well as roads near schools, hospitals, parks, and university campuses. Coldspots at the 99% confidence level (Gi_Bin = −3, GiPValue < 0.01) accounted for approximately 4.3% of the analyzed units and were mainly located along Software Avenue, Yongxin Avenue, and adjacent areas. These coldspots were primarily associated with industrial clusters and densely built residential districts, with a mean GVI of 0.228. Coldspots at the 95% confidence level (Gi_Bin = −2, 0.01 ≤ GiPValue < 0.05) accounted for approximately 47.5% and included areas such as Beijing East Road and Fuchunjiang East Street. These areas mainly corresponded to mixed commercial–residential zones, where GVI values ranged from 0.15 to 0.31. Non-significant areas (Gi_Bin = 0) accounted for more than 48% of the analyzed units, indicating that no statistically significant local clustering of GVI was detected in these locations.

Figure 8

3.2 Multidimensional GVI variations

3.2.1 Administrative district variations

GVI values varied clearly among administrative districts in Nanjing. Xuanwu District recorded the highest mean GVI (0.342), followed by Gaochun District (0.305), Gulou District (0.298), Qixia District (0.293), and Jianye District (0.292). These higher-ranking districts were generally associated with road environments adjacent to major ecological landscapes, relatively open peripheral settings, university-town areas, or newly developed urban districts with visible roadside greening.

To further characterize district-level variation in visible greening, this study calculated the street-greening compliance rate, defined as the proportion of road samples with GVI ≥ 0.15 within each administrative district. This threshold corresponds to the lower boundary of the “moderate greening” level adopted in this study. Based on this criterion, Gaochun District achieved the highest compliance rate (80.3%), followed by Xuanwu District (72.2%), Qixia District (71.4%), Jianye District (70.4%), and Jiangning District (70.1%). In contrast, Qinhuai District (58.2%), Liuhe District (60.3%), and Pukou District (61.2%) showed lower compliance rates, with a larger share of road segments remaining below the basic visible-greening benchmark.

Mean GVI and compliance rate reflect different aspects of district-level visible-greening performance. Mean GVI describes the overall intensity of visible greenery, whereas the compliance rate indicates the proportion of road segments that reached the greening threshold used in this study. For example, Pukou District showed a relatively low mean GVI (0.251) but a compliance rate of 61.2%, suggesting that many road segments reached the basic threshold even though the overall greening intensity remained comparatively modest. By contrast, Gaochun District combined a high compliance rate (80.3%) with a relatively high mean GVI (0.305), reflecting both broader threshold attainment and stronger overall visible greenery.

The distribution of higher-GVI road segments also differed among districts. Gaochun District had the highest proportion of road samples with GVI ≥ 0.25 (56.6%), followed by Xuanwu District (54.8%), Qixia District (51.6%), Jianye District (50.9%), and Gulou District (49.0%). In comparison, Liuhe District (36.5%), Lishui District (39.7%), Qinhuai District (39.5%), and Pukou District (39.9%) had lower proportions of high-GVI road segments. These differences indicate that administrative districts varied not only in overall greening averages, but also in the prevalence of road sections with relatively high levels of visible roadside greenery.

Sample sizes also varied across administrative districts in accordance with differences in road-network extent and the availability of valid street-view samples. Jiangning District (n = 24,581), Pukou District (n = 17,518), and Qixia District (n = 17,545) contained the largest numbers of samples, whereas Gaochun District (n = 3,763) and Xuanwu District (n = 6,837) had smaller sample sizes. These differences mainly reflect variation in district size, road-network extent, and valid image availability. Overall, the administrative-district comparison revealed marked variation in roadside visible greening across Nanjing (Figure 9).

Figure 9

3.2.2 Functional zoning differences

GVI values varied clearly across the six urban functional zones. Mean GVI values decreased in the following order: ecological zones (0.412) > public zones (0.299) > industrial zones (0.272) > commercial zones (0.242) > other areas (0.234) > residential zones (0.231).

Ecological zones showed the highest mean GVI (0.412 ± 0.275, n = 336). Public zones ranked second (0.299 ± 0.217, n = 23,554), followed by industrial zones, which recorded an intermediate mean GVI of 0.272 ± 0.206 (n = 80,581).

Commercial zones (0.242 ± 0.184, n = 1,274), other areas (0.234 ± 0.192, n = 11,029), and residential zones (0.231 ± 0.193, n = 785) showed comparatively lower mean GVI values. Among these categories, residential zones recorded the lowest mean GVI, followed closely by other areas and commercial zones.

Overlaying GVI hotspots and coldspots with functional zones further revealed the spatial differentiation of street greening (Figure 10). Hotspots were more frequently associated with ecological and public zones, especially road segments adjacent to major green landscapes, campuses, parks, and waterfront corridors. In contrast, coldspots were more commonly observed in commercial, industrial, and older mixed-use areas with a more compact urban form and weaker roadside vegetation continuity.

Figure 10

Considerable within-zone variation was also observed. For example, ecological and public zones contained both high-GVI road segments and areas with weaker visible greening, indicating that functional-zone categories do not fully capture local variation in street greenery. Similar internal differences were also observed in commercial and residential zones.

Overall, the functional-zone comparison showed clear differences in roadside visible greening across urban land-use contexts. Ecological and public zones generally recorded higher GVI values, whereas commercial, residential, and mixed-use settings tended to show lower visible greening levels.

3.2.3 Differences across road classes

GVI values differed clearly across road classes in Nanjing. Lower values were generally observed on traffic-priority roads, whereas higher values appeared on road types with weaker traffic dominance or greater pedestrian accessibility. Elevated roads and expressways recorded the lowest mean GVI (0.1829 ± 0.1519, n = 22,114), followed by urban arterial roads (0.2610 ± 0.1859, n = 35,566) and collector roads (0.2906 ± 0.2003, n = 23,406). Road types with lower traffic intensity or greater pedestrian accessibility showed higher mean GVI values, including urban local roads (0.3158 ± 0.2288, n = 20,757), internal roads (0.3585 ± 0.2554, n = 9,277), sidewalks (0.3658 ± 0.2408, n = 1,346), and suburban rural roads (0.3543 ± 0.2724, n = 229). Bicycle lanes recorded a moderate mean GVI of 0.3012 ± 0.2248 (n = 516). The “Other” category (0.6132 ± 0.2195, n = 5) was excluded from substantive interpretation because of its very small sample size. Overall, the road-class comparison revealed a clear gradient in visible greenery, with lower GVI values on traffic-priority roads and higher values on more pedestrian-oriented or lower-intensity road types.

3.3 Comparative analysis of GVI and NDVI

Within the matched GVI–NDVI dataset (n=117,546), the mean GVI was 0.2744, while the mean NDVI was 0.1451. The two indices showed a moderate positive correlation (Spearman’s ρ=0.5730, p<0.001), indicating a measurable correspondence between pedestrian-visible greenery and remotely sensed vegetation cover, while also suggesting that the two indicators capture different aspects of urban greenness (Figure 11). To further examine their overall relationship, scatter plots and linear regression analysis were used to assess the linear association between GVI and NDVI (Figure 12). The fitted regression equation was y=0.21304x+0.08652, with a coefficient of determination of R² = 0.29. This result shows that NDVI and GVI are positively related but only partially overlapping, reflecting the distinction between top-down vegetation coverage and street-level visible greenery. The stability of this cross-scale relationship was further examined across functional zones, administrative districts, and seasons in the following analyses.

Figure 11

Figure 12

3.3.1 Functional zone differences in GVI–NDVI correlation

GVI–NDVI correlations differed clearly among functional zones, and all associations were statistically significant (p<0.001). As shown in Figure 13, ecological zones exhibited the strongest correlation between GVI and NDVI (Spearman’s ρ=0.6262, n=336). Industrial zones followed closely (ρ=0.5946, n=80,578), while public zones also showed a relatively strong association (ρ=0.5790, n=23,554). Overall, the correlation was stronger in ecological, industrial, and public zones than in the remaining functional categories.

Figure 13

Commercial zones showed a moderate correlation (ρ=0.5421, n=1,267), while other areas exhibited a slightly weaker but still significant association (ρ=0.5223, n=11,026). Residential zones showed the weakest correlation among the six functional categories (ρ=0.4823, n=785). These results indicate that the strength of the GVI–NDVI relationship varied substantially across urban functional contexts.

3.3.2 Variations in GVI–NDVI consistency across administrative districts

The correlation between GVI and NDVI also varied clearly across administrative districts. Among all districts, Xuanwu District showed the strongest association (Spearman’s ρ=0.6699, n=6,837), followed by Yuhuatai District (ρ=0.6498, n=10,452), Gulou District (ρ=0.6416, n=6,768), and Jianye District (ρ=0.6361, n=7,890); all four coefficients exceeded 0.63 (Figure 14). These results indicate relatively strong GVI–NDVI correspondence in several central urban districts.

Figure 14

Several districts exhibited moderate correlations, including Qinhuai District (ρ=0.5952, n=6,960), Qixia District (ρ=0.5818, n=17,532), Pukou District (ρ=0.5705, n=17,518), and Jiangning District (ρ=0.5630, n=24,581). These districts occupied an intermediate position in the overall distribution of district-level GVI–NDVI correlations.

Relatively weaker correlations were observed in the outer suburban districts. Liuhe District (ρ=0.4728, n=10,393) and Lishui District (ρ=0.4720, n=4,852) showed similar moderate-to-low associations, while Gaochun District recorded the weakest correlation (ρ=0.3911, n=3,763). Overall, the district-level results revealed substantial spatial variation in the correspondence between street-level visible greenery and remotely sensed vegetation coverage.

3.3.3 Vegetation composition characteristics

Roadside vegetation in Nanjing was mainly composed of trees and grasslands, while shrubs accounted for a much smaller share. Trees represented 47.09% of the total green-space area, grasslands accounted for 46.15%, and shrubs accounted for 6.32%. These vegetation types differ in vertical structure and spatial coverage, which may partly explain their different contributions to GVI and NDVI.

Trees contributed to both vertical and horizontal greening patterns (Figure 15). As the largest vegetation category, trees can increase street-level GVI by forming visible canopy cover within the pedestrian field of view. Their relatively large biomass and canopy extent may also contribute to NDVI values captured from overhead remote-sensing imagery. The widespread presence of trees therefore helps explain the positive correspondence observed between GVI and NDVI.

Figure 15

Shrubs accounted for only 6.32% of the total green-space area (Figure 16). Although their overall proportion was limited, shrubs may contribute to street-level visible greenery by adding lower- and mid-level vegetation layers, while also contributing to surface vegetation cover captured by NDVI. However, their relatively small share suggests that shrubs played a more limited role than trees and grasslands in shaping the overall GVI–NDVI relationship.

Figure 16

Grasslands accounted for approximately 46.15% of the total green-space area (Figure 17). Their broad horizontal distribution may contribute substantially to NDVI because low-lying vegetation is readily captured in overhead remote-sensing imagery. However, grasslands contribute less to vertical green visibility from the pedestrian perspective than trees or shrubs. In grassland-dominated areas, relatively high NDVI values may therefore coexist with comparatively lower GVI values, which may weaken the correspondence between the two indices.

Figure 17

This study did not conduct a comprehensive quantitative assessment of how specific vegetation-structure parameters, such as plant height, diameter at breast height, and crown spread, influence GVI. This limitation was mainly due to the lack of detailed field measurements and a finer classification of roadside vegetation structure. Future research could incorporate these variables to further clarify how vegetation form and composition shape variations in GVI and its relationship with NDVI.

3.3.4 Seasonal trends of average NDVI and GVI

The seasonal means and standard deviations of roadside GVI and NDVI in Nanjing were calculated using 56,562 spring samples, 2,725 summer samples, 49,024 autumn samples, and 9,235 winter samples (Table 2). The seasonal variation patterns of the two indices are further illustrated in Figure 18.

Figure 18

Both GVI and NDVI reached their highest mean values in summer, accompanied by relatively large standard deviations. This pattern indicates that roadside greenery was most pronounced during the summer growing season, when evergreen vegetation and the fully developed foliage of deciduous species contributed to higher greenness levels. The larger standard deviations also suggest greater variation in roadside vegetation conditions during this period.

In winter, both indices declined to their lowest mean values and showed smaller standard deviations. This seasonal reduction is consistent with the leaf-off period of deciduous vegetation, during which visible and remotely sensed greenness are mainly sustained by evergreen species. The lower standard deviations indicate a more uniform greenness pattern across road samples in winter.

Across all seasons, the mean GVI remained higher than the mean NDVI. This difference reflects the distinct measurement perspectives of the two indices: GVI captures greenery visible within the pedestrian field of view, whereas NDVI measures vegetation greenness from an overhead remote-sensing perspective. The contrast between the two indices was particularly evident in autumn.

3.3.5 Seasonal variation in GVI–NDVI correlation coefficient

The seasonal comparison further showed that the difference between mean GVI and mean NDVI varied across seasons (Figure 19). The GVI–NDVI correlation also varied seasonally, with a lower coefficient in spring, a stronger association in summer, the highest correlation in autumn, and a slight decline in winter (Figure 20).

Figure 19

Figure 20

To examine seasonal differences in the GVI–NDVI relationship, the study period was divided into four climatic seasons according to Nanjing’s subtropical monsoon climate: spring (March–May), summer (June–August), autumn (September–November), and winter (December–February of the following year).

As described in Section 2.2.5, GVI and NDVI samples were aligned through a seasonal-window-based matching procedure and a unified projected coordinate system before seasonal correlation analysis. This strategy improved temporal comparability and spatial consistency among the matched samples.

The correlation between GVI and NDVI showed clear seasonal variation (Figure 20). Spring exhibited a moderate association (Spearman’s ρ=0.4982, p<0.001). The correlation strengthened in summer (ρ=0.5399, p<0.001) and reached its highest level in autumn (ρ=0.5672, p<0.001). In winter, the association declined slightly (ρ=0.5245, p<0.001), although it remained stronger than that observed in spring. Overall, the correspondence between street-level visible greenery and remotely sensed vegetation greenness was strongest in autumn and weakest in spring.

The Kruskal–Wallis H test confirmed statistically significant seasonal differences in GVI–NDVI correlation coefficients (H = 28.64, p<0.001). This result indicates that the strength of the GVI–NDVI relationship varied significantly across seasons rather than remaining constant throughout the year.

3.3.6 Seasonal limitations

Despite the seasonal analysis described above, the dataset was characterized by an imbalance in seasonal sample sizes. In particular, the number of summer street-view images was relatively limited (n=2,725), accounting for only about 4.8% of the spring sample size (n=56,562) and remaining substantially smaller than the sample sizes available for the other seasons. This imbalance was not artificially introduced during data processing but instead resulted from the retrieval logic of the Street View API, which returns the most recent available imagery for each sampling point rather than imagery from a predefined year or season. In the Nanjing region, recent summer imagery was unavailable or had not been updated for many road segments, leading to a smaller number of valid summer samples.

In addition, the available street-view images were not uniformly acquired within a single year. Potential interannual changes in the urban environment, such as road reconstruction, redevelopment of surrounding parcels, changes in roadside vegetation management, and broader urban renewal processes, may therefore introduce additional uncertainty into the interpretation of the GVI–NDVI relationship. Although the seasonal-window matching strategy improved temporal comparability between GVI and NDVI, it could not fully eliminate the influence of cross-year changes in the urban landscape. This issue should be considered when interpreting the seasonal correlation results and warrants the use of more temporally synchronized street-view archives in future research.

The limited number of summer samples may introduce potential bias into the interpretation of the summer GVI–NDVI correlation. However, a post hoc statistical power analysis (see Section 2.2.5) showed that the available summer sample size still provided sufficient statistical sensitivity to detect the observed correlation (power>0.99). Therefore, the significant summer correlation is unlikely to be explained solely by insufficient sample size or random sampling error, although potential bias related to sample representativeness and image availability should still be considered.

It should also be noted that, despite the smaller summer sample size, the summer observations retained a broadly comparable spatial distribution across urban functional zones relative to the other seasonal subsets. Their proportional distribution did not show an obvious spatial imbalance in relation to the overall seasonal sampling framework. Accordingly, the moderate summer correlation should be interpreted with caution because of sample-size imbalance, but it is unlikely to be driven solely by a severe spatial concentration of summer observations.

In addition, the season-stratified independent validation showed that vegetation-extraction performance was lower in winter than in summer and autumn, consistent with the greater difficulty of identifying vegetation under leaf-off conditions. Accordingly, interpretations of winter GVI results should be made with appropriate caution.

3.4 Quantitative differences in GVI across road classes and functional zones

Descriptive statistics and nonparametric tests were used to examine GVI differences across road classes and functional zones. This analysis focused on whether visible street greenery varied significantly among different urban spatial categories and on the magnitude of these intergroup differences.

3.4.1 Differences in GVI across road classes

GVI values varied clearly across road classes, showing a general gradient from lower values on traffic-priority roads to higher values on road types with weaker traffic dominance or greater pedestrian accessibility (Table 3).

Elevated roads and expressways recorded the lowest mean GVI (0.1829 ± 0.1519, n = 22,114), followed by urban arterial roads (0.2610 ± 0.1859, n = 35,566). Mean GVI values increased further among urban collector roads (0.2906 ± 0.2003, n = 23,406) and urban local roads (0.3158 ± 0.2288, n = 20,757). This pattern indicates a clear road-class gradient in visible street greenery.

Non-motorized and low-traffic road types showed comparatively higher GVI values. Bicycle lanes recorded a mean GVI of 0.3012 ± 0.2248 (n = 516), suburban rural roads recorded 0.3543 ± 0.2724 (n = 229), internal roads recorded 0.3585 ± 0.2554 (n = 9,277), and sidewalks showed the highest GVI among the major analyzable categories at 0.3658 ± 0.2408 (n = 1,346). The “Other” category contained only five samples (0.6132 ± 0.2195) and was therefore excluded from substantive interpretation because of its very small sample size.

3.4.2 Differences in GVI across functional zones

GVI values also varied substantially across the six urban functional zones (Table 4). The distribution of samples across functional zones reflects the broader urban spatial structure of Nanjing, while the descriptive statistics reveal clear differences in visible street greenery among these categories.

Ecological zones recorded the highest mean GVI (0.4119 ± 0.2750, n = 336), followed by public zones (0.2991 ± 0.2165, n = 23,554). Industrial zones displayed an intermediate GVI level (0.2718 ± 0.2062, n = 80,581).

Commercial zones showed a lower mean GVI (0.2418 ± 0.1842, n = 1,274). Other areas (0.2337 ± 0.1923, n = 11,029) and residential zones (0.2309 ± 0.1930, n = 785) recorded the lowest mean GVI values among the six functional categories.

Overall, the functional-zone comparison revealed clear differentiation in roadside visible greening. Ecological and public zones recorded relatively high GVI values, industrial zones occupied an intermediate position, and commercial, residential, and other mixed-use settings showed comparatively lower mean GVI levels.

3.4.3 Statistical significance tests for intergroup differences

Given the 50 m interval sampling design, neighboring road sampling points may not be fully statistically independent, as roadside visible greenness can exhibit local spatial dependence and clustering. Therefore, the Kruskal–Wallis H tests in this section were used as supplementary nonparametric evidence of intergroup differentiation rather than as the sole basis for substantive interpretation. The significance results were interpreted together with descriptive statistics, mapped spatial patterns, and the spatial clustering analysis presented in Section 2.2.6.

The Kruskal–Wallis H tests indicated statistically significant differences in GVI values across the nine road-class categories (H = 1286.74, df=8, p<0.001) and across the six functional-zone categories (H = 102.35, df=5, p<0.001). To avoid relying exclusively on p-values in a large spatial sample, effect sizes were additionally estimated using η²H derived from the Kruskal–Wallis statistics. The estimated effect size was η²H=0.011 for road classes and η²H=0.001 for functional zones.

These results indicate that the observed group differences were statistically detectable across the full dataset, but their effect magnitudes should be interpreted cautiously, particularly in light of the large sample size and the potential spatial dependence among neighboring road points. Accordingly, nominal significance levels alone were not treated as evidence of strong substantive effects. Instead, the interpretation of road-class and functional-zone differences was based on the combined consideration of statistical testing, effect-size estimates, descriptive GVI patterns, and mapped spatial variation.

4 Discussion

This study examined the spatial differentiation of road-level Green View Index (GVI) in Nanjing and its relationships with NDVI, vegetation composition, road classes, and urban functional zones. By integrating citywide street-view assessment with remote-sensing and spatial-context analyses, the study provides a more comprehensive understanding of how pedestrian-visible greenery varies across urban space and how it corresponds with broader vegetation conditions. The results show that the spatial clustering of GVI is associated with both natural landscape settings and urban functional configurations. They also extend current urban greening assessments by addressing citywide spatial variation, seasonal dynamics, and vegetation-composition differences within a unified analytical framework.

The relationship between the Green View Index (GVI) and the Normalized Difference Vegetation Index (NDVI) has attracted increasing attention in urban greening research. Existing studies have commonly treated NDVI as a remotely sensed indicator of vegetation coverage and GVI as a pedestrian-oriented measure of visible greenery (Zhang and Zeng, 2024). However, previous research has largely focused on urban core areas (Li et al., 2021; Chen et al., 2022) or relied on static observations (Tong et al., 2020; Zhang and Zeng, 2024), leaving citywide and spatially heterogeneous greening patterns less fully examined. In this study, the citywide mean GVI of Nanjing was 0.273, with a standard deviation of 0.208, indicating marked spatial variation in roadside visible greenery. High-GVI areas were concentrated around major ecological and waterfront landscapes, particularly the southern foothills of Purple Mountain and the Qinhuai River corridor, whereas low-GVI areas were mainly distributed in dense commercial centers, transportation hubs, and older industrial–residential mixed districts. At the administrative-district level, Xuanwu District recorded the highest mean GVI (0.342), while Gaochun District also showed relatively high visible greening (0.305). These findings extend previous localized studies by showing that high-GVI areas are spatially clustered around major ecological and waterfront landscape systems at the citywide scale (Li et al., 2021).

The seasonal analysis extends previous static assessments by showing that the correspondence between GVI and NDVI is not temporally uniform. The Spearman correlation was moderate in spring (ρ = 0.4982), strengthened in summer (ρ = 0.5399), reached its highest level in autumn (ρ = 0.5672), and declined slightly in winter (ρ = 0.5245). This seasonal pattern may reflect phenological changes under Nanjing’s humid northern subtropical climate. In spring, transitional leaf emergence may weaken the consistency between street-visible greenery and remotely sensed vegetation greenness; in autumn, relatively stable canopy structures before leaf fall may strengthen their correspondence; and in winter, leaf-off conditions may produce a more complex relationship between the two indicators. These findings complement prior evidence of a positive GVI–NDVI association (Zhang and Zeng, 2024) by demonstrating that the strength of this relationship varies across seasonal contexts. They also highlight the importance of incorporating temporal comparability into cross-scale urban greening assessment.

Vegetation composition provides a further perspective for interpreting differences in the observed GVI–NDVI relationship. Roadside vegetation in Nanjing was dominated by trees (47.09%) and grasslands (46.15%), while shrubs accounted for only 6.32%. Trees are likely to contribute to both indices because their canopy structure increases visible greenery within the pedestrian field of view while also producing substantial vegetation signals in overhead remote-sensing imagery. By contrast, the limited proportion of shrubs may reduce the contribution of lower- and middle-layer vegetation to the visual structure of streets, potentially weakening the consistency between ground-level and remotely sensed greenness in some areas. This finding extends the qualitative observation by Tong et al. (2020) that vegetation structure matters for GVI–NDVI correspondence by providing quantitative evidence on the relative composition of roadside vegetation types in Nanjing.

Clear differences in GVI distribution were also observed across road classes and urban functional zones. Previous studies have mainly examined GVI differentiation at the regional or functional-zone level (Li et al., 2015), while the association between road-class characteristics and visible street greenery has received comparatively less attention. In this study, traffic-priority road types generally showed lower GVI values, whereas road types with weaker traffic dominance or greater pedestrian accessibility tended to exhibit higher levels of visible greenery. Elevated roads and expressways recorded the lowest mean GVI, followed by urban arterial and collector roads, while urban local roads, internal roads, and sidewalks showed comparatively stronger visible greening. This pattern may reflect differences in road function, spatial organization, and opportunities for roadside vegetation arrangement. For example, pedestrian-oriented streets and internal residential roads often provide more continuous planting interfaces, whereas high-capacity traffic corridors may offer fewer opportunities for visually rich roadside greenery. These findings provide empirical support for considering road-class differentiation in street-greening assessment and planning.

The overlay analysis of GVI spatial clusters and functional zones further showed that hotspots were mainly associated with ecological and public zones, whereas coldspots were more frequently concentrated in commercial districts, industrial areas, and older residential neighborhoods. At the same time, notable variation was observed within the same functional-zone category. For example, only the inner ring of the Purple Mountain Scenic Area formed a pronounced hotspot within ecological zones, while in public zones, areas adjacent to schools tended to be hotspots and those near transportation hubs tended to be coldspots. These within-zone differences suggest that broad land-use categories alone do not fully explain the spatial differentiation of GVI; local construction context and specific spatial functions may also play important roles. This finding extends Gou et al.’s discussion of GVI spatial distributions in mountainous urban areas by highlighting the importance of within-zone heterogeneity in visible greenery assessment (Gou et al., 2025).

Although this study addresses several gaps in existing research, some limitations remain.

• Although the street-view dataset covered the full study area, the number of valid summer samples was relatively limited (n = 2,725). This imbalance may affect the precision of the summer GVI–NDVI association analysis and should be considered when interpreting seasonal comparison results, particularly those related to the vegetation growth period.

• This study did not incorporate the moderating influence of human activity intensity on GVI perception. Because GVI represents street-level visible greenery from a human-eye-level perspective, variations in activity intensity, user presence, and perceptual context may influence how greening quality is experienced and interpreted in practice.

• This study did not conduct cost–benefit analyses of different greening configuration models. As a result, the findings cannot directly provide quantitative economic evidence for prioritizing specific urban greening strategies or evaluating their implementation feasibility.

• Owing to the lack of field-measured vegetation data, this study focused on the proportional composition of major vegetation types—trees, shrubs, and herbaceous plants—and their relationship with GVI and GVI–NDVI correspondence. It did not examine how more detailed structural parameters, such as vegetation height, diameter at breast height, crown spread, or specific planting combinations, influence GVI. This limitation constrains a more comprehensive understanding of how fine-scale vegetation structure shapes visible street greenery.

• This study identified spatial associations between natural landscape endowments, urban functional layouts, and the distribution of GVI, but it did not quantify their relative contributions. The absence of high-resolution natural-environment datasets—such as detailed topographic information, boundaries of mountain–water ecological zones, and natural vegetation distribution data—limited the ability to estimate the independent and interactive effects of these factors. Future research could integrate multi-source geospatial datasets with quantitative spatial models to more precisely evaluate how natural terrain and urban functional structure jointly shape GVI heterogeneity.

• This study examined the macro-level patterns of administrative districts and broad functional zones, but it did not quantify the relationships between GVI and more refined functional indicators, such as land-use mix, Point of Interest (POI) density, or commercial-sector composition. Nor did it examine the micro-scale processes underlying within-zone differences in GVI. Addressing these issues will require finer spatial-unit definitions and the integration of more detailed multi-source urban data in future research.

Future research should first increase the availability and temporal balance of summer street-view samples and integrate higher-resolution remote-sensing data to improve the spatiotemporal precision of seasonal greening analysis. Second, incorporating human-perception and activity-related data, such as questionnaire surveys, pedestrian observations, or behavioral trajectory records, could support the development of a dual-perspective greening assessment framework that combines vegetation characteristics with human-use intensity. Third, systematic cost–benefit analyses of alternative greening configurations would help identify strategies that balance ecological performance, visual quality, and implementation feasibility. Finally, integrating detailed terrain, ecological-boundary, and natural-vegetation datasets with spatial quantitative models could further clarify the relative contributions of natural landscape conditions and urban functional structure to GVI heterogeneity.

5 Conclusions

This study developed a citywide analytical framework integrating street-view imagery, NDVI-based remote-sensing data, and spatial analysis to examine the differentiation of road-level Green View Index (GVI) in Nanjing, its relationship with NDVI, and its links to vegetation composition, road classes, and urban functional zones. The results reveal substantial spatial heterogeneity in roadside visible greenery across the city. High-GVI areas were mainly concentrated around major ecological and waterfront landscapes, such as Purple Mountain and the Qinhuai River corridor, whereas lower-GVI areas were more frequently observed in highly urbanized commercial, transportation, and mixed-use districts. The GVI–NDVI relationship showed clear seasonal variation: the correlation was weakest in spring, strengthened in summer, reached its highest level in autumn, and declined slightly in winter. Vegetation composition was also related to differences in the GVI–NDVI relationship, while road-class and functional-zone comparisons further revealed clear variation in GVI across different urban spatial contexts. Sidewalks and internal roads recorded comparatively high GVI values, and ecological and public zones were more frequently associated with higher levels of visible street greenery.

These findings extend existing urban greening research in three respects. First, they provide a citywide assessment of street-level visible greenery in a spatially heterogeneous urban environment. Second, they show that the correspondence between GVI and NDVI is not temporally uniform, highlighting the importance of seasonal context in cross-scale greenness assessment. Third, they demonstrate the relevance of vegetation composition, road hierarchy, and functional zoning for understanding variations in urban street greenery. Taken together, the study supports a more differentiated approach to street-greening assessment and may inform the planning and optimization of visible greenery across Nanjing’s road network and in other cities with similarly complex urban landscape structures.

Several limitations should be acknowledged. The relatively small number of valid summer street-view samples may affect the precision of seasonal GVI–NDVI comparison. The analysis did not incorporate the moderating influence of human activity intensity on GVI perception, which may lead to differences between objectively measured visible greenery and actual experiential needs. In addition, the study did not quantify the cost–benefit performance of alternative greening configurations, limiting the direct applicability of the findings to implementation-oriented planning decisions.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

References

• 1

  AlfarhoodM.AlahmadA.AlalwanA.AlkulaibF. (2025). Leveraging satellite imagery and machine learning for urban green space assessment: a case study from Riyadh City. Sustainability17, 6118. doi: 10.3390/su17136118

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AokiY. (1991). Evaluation methods for landscapes with greenery. Landscape Res.16, 3–6. doi: 10.1080/01426399108706344

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AokiY.YasuokaY.NaitoM. (1985). Assessing the impression of street-side greenery. Landscape Res.10, 9–13. doi: 10.1080/01426398508706131

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BowlerD. E.Buyung-AliL.KnightT. M.PullinA. S. (2010). Urban greening to cool towns and cities: a systematic review of the empirical evidence. Landscape Urban Plann.97, 147–155. doi: 10.1016/j.landurbplan.2010.05.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  CaiB. Y.LiX.SeiferlingI.RattiC. (2018). “ Treepedia 2.0: applying deep learning for large-scale quantification of urban tree cover”, in: 2018 IEEE International Congress on Big Data (BigData Congress), (San Francisco, CA: IEEE). 49–56. doi: 10.1109/BigDataCongress.2018.00014

  • CrossRef
  • Google Scholar

• 6

  CarrusG.ScopellitiM.LafortezzaR.ColangeloG.FerriniF.SalbitanoF.et al. (2015). Go greener, feel better? The positive effects of biodiversity on the well-being of individuals visiting urban and peri-urban green areas. Landscape Urban Plann.134, 221–228. doi: 10.1016/j.landurbplan.2014.10.022

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  ChenJ.ZhouC.LiF. (2020). Quantifying the green view indicator for assessing urban greening quality: an analysis based on Internet-crawling street view data. Ecol. Indic.113, 106192. doi: 10.1016/j.ecolind.2020.106192

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  ChenY.ZhangQ.DengZ.FanX.XuZ.KangX.et al. (2022). Research on green view index of urban roads based on street view image recognition: a case study of Changsha downtown areas. Sustainability14, 16063. doi: 10.3390/su142316063

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  FengL.MiX.YuanD. (2022). Optimal planning of urban greening system in response to urban microenvironments in a high-density city using genetic algorithm: a case study of Tianjin. Sustain. Cities Soc87, 104244. doi: 10.1016/j.scs.2022.104244

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  FruthE.KvistadM.MarshallJ.PfeiferL.RauL.SagebielJ.et al. (2019). Economic valuation of street-level urban greening: a case study from an evolving mixed-use area in Berlin. Land Use Policy89, 104237. doi: 10.1016/j.landusepol.2019.104237

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  GetisA.OrdJ. K. (1992). The analysis of spatial association by use of distance statistics. Geogr. Anal.24, 189–206. doi: 10.1007/978-3-642-01976-0_10

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12

  GouA.WangX.WangJ.WangC.TanG. (2025). Spatial pattern and heterogeneity of green view index in mountainous cities: a case study of Yuzhong district, Chongqing, China. Sci. Rep.15, 12576. doi: 10.1038/s41598-025-97946-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 13

  HaalandC.van den BoschC. K. (2015). Challenges and strategies for urban green-space planning in cities undergoing densification: a review. Urban Forestry Urban Greening14, 760–771. doi: 10.1016/j.ufug.2015.07.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  HaoZ.LinL.PostC. J.MikhailovaE. A.LiM.ChenY.et al. (2021). Automated tree-crown and height detection in a young forest plantation using mask region-based convolutional neural network (Mask R-CNN). ISPRS J. Photogramm. Remote Sens.178, 112–123. doi: 10.1016/j.isprsjprs.2021.06.003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 15

  JingX.LiZ.ChenH.ZhangC. (2024). Is what we see always real? A comparative study of two-dimensional and three-dimensional urban green spaces: the case of Shenzhen’s Central District. Forests15, 983. doi: 10.3390/f15060983

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16

  KameokaT.UchidaA.SasakiY.IseT. (2022). Assessing streetscape greenery with deep neural network using Google Street View. Breed. Sci.72, 107–114. doi: 10.1270/jsbbs.21073

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  KumakoshiY.ChanS. Y.KoizumiH.LiX.YoshimuraY. (2020). Standardized Green View Index and quantification of different metrics of urban green vegetation. Sustainability12, 7434. doi: 10.3390/su12187434

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  LeeA. C.MaheswaranR. (2011). The health benefits of urban green spaces: a review of the evidence. J. Public Health33, 212–222. doi: 10.1093/pubmed/fdq068

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  LiF.WangR.PaulussenJ.LiuX. (2005). Comprehensive concept planning of urban greening based on ecological principles: a case study in Beijing, China. Landscape Urban Plann.72, 325–336. doi: 10.1016/j.landurbplan.2004.04.002

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  LiT.ZhengX.WuJ.ZhangY.FuX.DengH. (2021). Spatial relationship between green view index and normalized differential vegetation index within the Sixth Ring Road of Beijing. Urban Forestry Urban Greening62, 127153. doi: 10.1016/j.ufug.2021.127153

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  LiX.ZhangC.LiW.RicardR.MengQ.ZhangW. (2015). Assessing street-level urban greenery using Google Street View and a modified green view index. Urban Forestry Urban Greening14, 675–685. doi: 10.1016/j.ufug.2015.06.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  LigtermoetE.RamalhoC. E.FoellmerJ.PauliN. (2022). Greening urban road verges highlights diverse views of multiple stakeholders on ecosystem service provision, challenges and preferred form. Urban Forestry Urban Greening74, 127625. doi: 10.1016/j.ufug.2022.127625

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  LuY.FerrantiE. J. S.ChapmanL.PfrangC. (2023). Assessing urban greenery by harvesting street view data: A review. Urban Forestry Urban Greening83, 127917. doi: 10.1016/j.ufug.2023.127917

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  MaY.ChenP.QinY.YangZ.LiS. (2025). From sky to ground: monitoring visible street greenery via multisource remote sensing imagery with deep learning. Urban Forestry Urban Greening109, 128866. doi: 10.1016/j.ufug.2025.128866

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  MandalM.PopekR.PrzybyszA.RoyA.DasS.SarkarA. (2023). Breathing fresh air in the city: implementing avenue trees as a sustainable solution to reduce particulate pollution in urban agglomerations. Plants12, 1545. doi: 10.3390/plants12071545

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  MaransR. W. (2003). Understanding environmental quality through quality of life studies: the 2001 DAS and its use of subjective and objective indicators. Landscape Urban Plann.65, 73–83. doi: 10.1016/s0169-2046(02)00239-6

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  MendzinaE.VuguleK. (2020). Importance and planning of pedestrian streets in urban environment. Landscape Architecture Art.16, 80–86. doi: 10.22616/j.landarchart.2020.16.08

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  OliveiraA. A.HirataN. S.HirataR. (2019). Greenery segmentation in urban images by deep learning. arXiv preprint arXiv:1912.06199. doi: 10.48550/arXiv.1912.06199

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  OriharaN. (2006). Study on the evaluation of green landscapes: consideration of green evaluation methods for forming favorable landscapes [緑景観の評価に関する研究―良好な景観形成に向けた緑の評価手法に関する考察] (in Japanese). IBEC: Building Environ. Energy Conserv. Inf.27, 32–35.

  • Google Scholar
  • View reference in article

• 30

  PanT.-C.KaoJ.-J. (2021). Comparison of indices for evaluating building green values based on greenhouse gas emission reductions. Ecol. Indic.122, 107228. doi: 10.1016/j.ecolind.2020.107228

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  QamarF.DoblerG. (2020). Pixel-wise classification of high-resolution ground-based urban hyperspectral images with convolutional neural networks. Remote Sens.12, 2540. doi: 10.3390/rs12162540

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SongJ.ZhuX.YuX. (2023). Monitoring and mapping vegetation greenery in residential communities using street view images and a Normalized Vegetation Greenery Index: a case study in Beijing, China. Front. For. Global Change6, 1071569. doi: 10.3389/ffgc.2023.1071569

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  SpeakA.MontagnaniL.WellsteinC.ZerbeS. (2020). The influence of tree traits on urban ground surface shade cooling. Landscape Urban Plann.197, 103748. doi: 10.1016/j.landurbplan.2020.103748

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  SunS.HussA.KarssenbergD.SchmitzO.JiangY.Probst-HenschN.et al. (2026). Street view versus remote sensing greenery – comparison of two exposure metrics across urban-rural settings. Geo-spatial Inf. Sci. doi: 10.1080/10095020.2026.2619315. Advance online publication.

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  SusakiJ.KubotaS. (2017). Automatic assessment of green space ratio in urban areas from mobile scanning data. Remote Sens.9, 215. doi: 10.3390/rs9030215

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36

  TongM.SheJ.TanJ.LiM.GeR.GaoY. (2020). Evaluating street greenery by multiple indicators using street-level imagery and satellite images: a case study in Nanjing, China. Forests11, 1347. doi: 10.3390/f11121347

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  Van RenterghemT.VermandereE.LauwereysM. (2023). Road traffic noise annoyance mitigation by green window view: optimizing green quantity and quality. Urban Forestry Urban Greening88, 128072. doi: 10.1016/j.ufug.2023.128072

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38

  WangY.AkbariH. (2016). The effects of street tree planting on Urban Heat Island mitigation in Montreal. Sustain. Cities Soc27, 122–128. doi: 10.1016/j.scs.2016.04.013

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  XiaY.YabukiN.FukudaT. (2021a). Development of a system for assessing the quality of urban street-level greenery using street view images and deep learning. Urban Forestry Urban Greening59, 126995. doi: 10.1016/j.ufug.2021.126995

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40

  XiaY.YabukiN.FukudaT. (2021b). Sky view factor estimation from street view images based on semantic segmentation. Urban Clim.40, 100999. doi: 10.1016/j.uclim.2021.100999

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41

  XiaoX.WeiY.LiM. (2018). Measurement methods and evaluation applications of urban green view index in Japanese cities [日本城市绿视率计量方法与评价应用] (in Chinese). Urban Plann. Int.33, 98–103. doi: 10.22217/upi.2015.547

  • CrossRef
  • Google Scholar

• 42

  XuC. F.LiY. H.ZhuZ. B.ZouZ.GengX.HuY. (2026). Same streets, different contexts: personality-based differences in cycling willingness revealed from objective and subjective perspectives. ISPRS Int. J. Geo-Inf.15, 179. doi: 10.3390/ijgi15040179

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43

  YangJ.ZhaoL.McBrideJ.GongP. (2009). Can you see green? Assessing the visibility of urban forests in cities. Landscape Urban Plann.91, 97–104. doi: 10.1016/j.landurbplan.2008.12.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44

  YeY.RichardsD.LuY.SongX.ZhuangY.ZengW.et al. (2019). Measuring daily accessed street greenery: a human-scale approach for informing better urban planning practices. Landscape Urban Plann.191, 103434. doi: 10.1016/j.landurbplan.2018.08.028

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  YuX.HerY.HuoW.ChenG.QiW. (2022). Spatio-temporal monitoring of urban street-side vegetation greenery using Baidu Street View images. Urban Forestry Urban Greening73, 127617. doi: 10.1016/j.ufug.2022.127617

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46

  ZhangL.WangL.WuJ.LiP.DongJ.WangT. (2023). Decoding urban green spaces: deep learning and Google Street View measure greening structures. Urban Forestry Urban Greening87, 128028. doi: 10.1016/j.ufug.2023.128028

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  ZhangW.ZengH. (2024). Spatial differentiation characteristics and influencing factors of the green view index in urban areas based on street view images: a case study of Futian District, Shenzhen, China. Urban Forestry Urban Greening93, 128219. doi: 10.1016/j.ufug.2024.128219

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48

  ZhaoH.ShiJ.QiX.WangX.JiaJ. (2017). “ Pyramid scene parsing network”, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (Honolulu, HI: IEEE). 6230–6239. doi: 10.1109/CVPR.2017.660

  • CrossRef
  • Google Scholar

• 49

  ZhengL.ZhaoY.DuanR.YangW.WangZ.SuJ. (2023). The influence path of community green exposure index on activity behavior under multi-dimensional spatial perception. Front. Public Health11, 1243838. doi: 10.3389/fpubh.2023.1243838

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50

  ZhouB.ZhaoH.PuigX.FidlerS.BarriusoA.TorralbaA. (2017). “ Scene parsing through ADE20K dataset”, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (Honolulu, HI: IEEE). 5122–5130. doi: 10.1109/CVPR.2017.544

  • CrossRef
  • Google Scholar

• 51

  ZhuJ.HuangY.CaoZ.ZhangY.DingY.DuJ. (2025). Evaluating urban greenery through the front-facing street view imagery: insights from a Nanjing case study. ISPRS Int. J. Geo-Inf.14, 287. doi: 10.3390/ijgi14080287

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  ZhuH.NanX.YangF.BaoZ. (2023). Utilizing the green view index to improve the urban street greenery index system: a statistical study using road patterns and vegetation structures as entry points. Landscape Urban Plann.237, 104780. doi: 10.1016/j.landurbplan.2023.104780

  • Pubmed Abstract
  • CrossRef
  • Google Scholar