Exploring the Spatiotemporal Associations Between Ride-Hailing Demand, Visual Walkability, and the Built Environment: Evidence from Chengdu, China

Abstract

Ride-hailing services have reshaped urban commuting patterns, yet the spatiotemporal mechanisms linking built environment features to ride-hailing demand remain underexplored. Existing studies often overlook the joint effects of origin–destination visual walkability. This study integrates ride-hailing GPS trajectories and geospatial data to quantify mobility patterns and built-environment indicators in Chengdu, China. A dual analytical framework combining global regression and localized modeling was applied to disentangle spatial–temporal influences of urban form and socioeconomic factors. The results reveal that population density, floor–area ratio, and housing prices positively correlate with demand, while road density and distance to city center exhibit negative associations. Visual walkability metrics show divergent effects: psychological greenery and pavement visibility reduce ride-hailing usage, whereas outdoor enclosure enhances it. Temporal analysis identifies time-dependent impacts of built environment variables on main urban area travel. Housing price effects demonstrate spatial globality, while population density and city-center proximity exhibit geographically bounded correlations. Notably, improved visual walkability in specific zones reduces reliance on ride-hailing by facilitating sustainable alternatives. These findings provide empirical support for optimizing urban infrastructure and land-use policies to promote equitable mobility systems. The proposed methodology offers a replicable framework for assessing transportation–land-use interactions, informing targeted interventions to achieve metropolitan sustainability goals through coordinated spatial planning and pedestrian-centric design.

1. Introduction

Propelled by information and communication technology, the ride-hailing industry has undergone exponential growth in the last ten years. Taking China’s leading mobility service platform, Didi, as an example, its services have reached over 550 million registered users. Platform operational data indicate that the average daily order volume has surpassed 30 million, and the total daily mileage has reached 350 million kilometers, underscoring the pivotal role of ride-hailing platforms within urban transport networks [1]. The commercial success of ride-hailing platforms stems from their innovative service model, which has reconfigured traditional travel methods through mobile internet technology. Its fundamental merit is in offering customized transit solutions that provide prompt responsiveness and direct, door-to-door service cost-effectively [2,3]. Technically, such platforms leverage smart terminal applications to establish a supply–demand matching system that accurately connects drivers with passengers. Specifically, the system employs geographic information system (GIS) technology for real-time positioning, integrates intelligent algorithms for demand forecasting and resource allocation, and uses a floating pricing mechanism to dynamically adjust supply capacity, thereby ensuring balance between service supply and demand [4,5].

The rapid development of ride-hailing platforms has generated transformative effects on urban mobility networks marked by ambivalent consequences that reflect both innovation and disruption. From a positive perspective, these platforms have created innovative mobile travel solutions, allowing users to enjoy convenient services without bearing the costs of vehicle purchase and parking [2,6]. Studies show that this ride-sharing model helps optimize the distribution of transportation resources and improve road efficiency, which in turn alleviates urban traffic pressures and lowers carbon emission levels [7]. Additionally, its service network successfully reaches remote regions that conventional public transit finds difficult to serve, thus compensating for the shortcomings in public transportation coverage [8]. However, this innovative model has also brought challenges and controversies, which require further exploration and regulation. For example, the adoption rate of active travel modes (such as walking) is showing a downward trend [9].

Recent scholarship has increasingly analyzed the determinants of ride-hailing adoption through the lenses of sociodemographic characteristics and trip-specific motivations [10]. The built environment exerts structural influence on the formulation of urban development strategies and the design of transportation demand governance frameworks. The spatial behavior theoretical framework indicates that there is a bidirectional interaction mechanism between the urban built environment and human activities [11]. Analyzing the physical space dimension, the built environment—by combining spatial factors like land-use diversity, public service accessibility, and road network connectivity—forms the constraints and enabling factors that shape human spatial activities. On the other hand, the spatiotemporal distribution and group-level regularities of residents’ travel behaviors effectively mirror their functional requirements of the built environment.

In the context of the digital technology revolution, information and communication technology (ICT) demonstrates distinct contemporary characteristics in its influence on residents’ travel decision making. Notably, the evolution of ICT has not only reshaped the traditional notion of geographical distance, but it has also instigated a structural shift in individual life patterns by increasing the accessibility of virtual spaces [12]. Existing studies have verified that non-motorized travel modes like walking demonstrate marked benefits in improving urban ecological outcomes, safeguarding public health and safety, and enhancing energy efficiency, and their socioeconomic advantages are widely recognized within academia [13,14,15]. However, ride-hailing platforms’ economical door-to-door offerings could potentially trigger a mode substitution effect, particularly in short-range travel contexts, by competing with non-motorized transit options. Rayle et al. revealed based on field research data in the San Francisco area that about 40% of ride-hailing orders have essentially replaced travel needs that were originally planned to be completed through the public transportation system or walking [2]. Cross-city empirical research shows that the typical service radius is concentrated in a 3–6 km range, which overlaps significantly with traditional pedestrian advantage areas, confirming its crowding out effect on sustainable travel patterns [16].

Recent studies have drawn attention to a counterintuitive yet revealing trend: ride-hailing trips frequently originate or terminate in areas that are highly walkable and rich in urban amenities [17]. While such neighborhoods theoretically support active modes of travel, their dense, vibrant character often generates substantial short-distance travel demand, some of which is met by ride-hailing services rather than walking. This phenomenon suggests that traditional walkability measures—such as density, land-use mix, or Walk Score—may not fully capture the behavioral nuances influencing mode choice, particularly in high-activity urban centers where both walking and ride-hailing coexist. Emerging evidence suggests that visual walkability—defined as the perceived quality of the streetscape from a pedestrian’s perspective—plays a crucial role in shaping individuals’ travel behavior [18]. A visually unappealing environment (e.g., poor cleanliness, lack of greenery, or perceived insecurity) may deter walking even in areas with structurally walkable features. Conversely, visually pleasant, green, and safe streetscapes can actively encourage walking and suppress unnecessary ride-hailing demand.

Visual walkability refers to the visually perceptible qualities of the built environment that affect pedestrians’ comfort, safety, and willingness to walk, particularly from a street-level perspective. In this study, visual walkability is operationalized using four street-level indicators derived from deep learning on street-view imagery: Visual Greenery (visible vegetation), Visual Crowdedness (perceived spatial compression from buildings and vehicles), Visual Enclosure (degree of sky openness and spatial definition), and Visual Pavement (visible pedestrian infrastructure). These indicators reflect pedestrians’ sensory and emotional responses to their environment—factors that traditional spatial metrics overlook. Empirical studies have shown that incorporating visual walkability indicators can significantly enhance the explanatory power of models that analyze travel behavior patterns [19].

In summary, while empirical inquiry into ride-hailing platforms has expanded rapidly in recent years, critical knowledge gaps persist across multiple research domains that warrant systematic investigation. Generally speaking, the spatial dimension of ride-hailing has been under-researched. A limited body of research have attempted to describe the associations among ride-hailing trips, the built environment, and neighborhood characteristics [1,3,20]. However, previous studies have seldom considered the neighborhood characteristics at both the origins and destinations of trips. Crucially, no prior studies have examined the spatial distribution of ride-hailing demand along areas characterized by visual walkability attributes. Understanding the relationship between demand and the visual walkability properties of the built environment can help clarify whether ride-hailing services are used only when no other transportation options are available or if they continue to be utilized even when more sustainable alternatives exist.

We intend to tackle two research questions to bridge the current gap: (1) How can we develop a novel methodological framework to improve the precision of visual walkability characterization? (2) What is the relationship between the building environment at the starting and ending points of the trip and the demand for ride-hailing services? (3) How does ride-hailing demand change as visual walkability varies?

This study innovatively establishes a multi-dimensional analytical framework for ride-hailing trip generation, focusing on examining the spatiotemporal associations between ride-hailing demand and built environment characteristics. Based on travel big data, the research team developed a built environment indicator system comprising 14 core dimensions, including population density, road density, floor–area ratio (FAR), housing prices, subway accessibility, land-use mix, POI density, and visual walkability index. The study develops a comprehensive visual walkability indicator system. Integrating visual walkability into ride-hailing research represents a novel and necessary extension to the analytical toolkit, providing new empirical perspectives for exploring how different types of streetscape visual environments are statistically associated with variations in ride-hailing usage, particularly in contexts where such services may coincide with or potentially displace walking and other sustainable travel modes. Methodologically, we overcome the limitations of traditional global regression models by incorporating the Geographically Weighted Regression (GWR) technique from spatial econometrics, constructing a spatiotemporal heterogeneity model to accurately capture the spatial differentiation patterns of the elasticities of explanatory variables.

The rest of this paper is structured as follows. Section 2 reviews the related literature. Section 3 describes the data and methods utilized in this study. Section 4 and Section 5 comprehensively present the results and discuss insights regarding the impact of built environment characteristics on ride-sharing services. Section 6 concludes with the key findings, provides policy proposals, and highlights potential research avenues.

2. Literature Review

This section establishes the conceptual and empirical foundations linking built environment characteristics, walkability, and ride-hailing demand. Section 2.1 outlines the well-established 5D framework and key spatial determinants of travel behavior. Section 2.2 explores the mode substitution effects between walking and ride-hailing, particularly in short-distance urban trips. Section 2.3 extends the discourse by introducing visual walkability as a perceptual construct, increasingly measurable through street-level imagery and deep learning techniques. Collectively, this literature review supports the study’s integrated analytical framework that combines conventional built environment metrics with visual walkability indicators to explain variations in ride-hailing demand.

2.1. Determinants of Travel Demand

We focus on the classic proposition concerning the interactive relationship between the built environment and travel behavior. Existing empirical analyses have demonstrated that the configuration of spatial elements has a lasting influence on residents’ transportation mode choices [21]. The 5D framework provides a comprehensive explanation of how the five dimensions—Density, Diversity, Design, Destination Accessibility, and Distance to Transit—influence travel decisions [11]. Yet, the latest analyses reveal substantial differences in the elasticity coefficients of these elements across various urban settings, highlighting the influence of spatiotemporal heterogeneity on the research conclusions.

Several studies in the existing literature employ various variables to characterize aspects of the built environment. Population density and the FAR are considered density variables and have been employed in travel behavior analyses [22,23]. It is widely acknowledged that enhancing density can diminish reliance on private cars and foster public transit and non-motorized modes; yet, two distinct viewpoints emerge regarding elasticity: one perspective holds that the very low elasticity implies that boosting density is not sufficiently effective to warrant density-centric policies [24], while the other contends that even a modest elasticity can significantly curb unsustainable travel because density increases are typically linked with improvements in other built environment measures (such as increased land-use mixing), ultimately producing a net sustainable effect [25].

Diversity has been extensively utilized and demonstrated to predict urban travel behavior [14,21,26,27]. Primarily, empirical research consistently demonstrates that elevated land-use mixing significantly shortens trip lengths while also reducing reliance on personal automobiles [28,29,30]. Notably, scholars have validated opposing causal pathways through spatial econometric models. Their studies demonstrate that high land-use mix areas may stimulate non-commuting travel demand, ultimately leading to cumulative increases in individual mobility distances, while the moderating effects on private vehicle usage exhibit significant spatial heterogeneity [31,32].

As carriers of urban functions, POI data serve as core explanatory variables for ride-hailing service demand prediction due to their distinctive travel demand characteristics [33]. Empirical studies utilizing multi-source data fusion have uncovered asymmetric facility impacts: food-service POIs show the strongest positive correlation with trip origin demand, while recreational facilities predominantly affect the spatial distribution of trip destinations [34,35]. Such discrepancies highlight the modulating effects of contextual spatial heterogeneity.

Road density is commonly considered a design-related factor and has been utilized in the analysis of travel behavior [27,36]. Existing studies demonstrate that this metric profoundly influences residents’ travel decisions through pathways that include spatial accessibility modulation, travel cost constraints, and transportation mode selection. High-density road density enhances spatial accessibility by reducing average trip distances, thereby promoting residents’ preference for non-motorized travel modes [11]. Distance to transit is suggested as a predictive factor in transportation generation models. Numerous empirical studies indicate that proximity to public transit exhibits a significant negative correlation with travel demand in predicting regional trip generation [37].

Housing price has been employed as a proxy variable for socioeconomic status [38]. Housing price research elucidates user stratification patterns in ride-sharing systems through the lens of socioeconomic disparities. Cross-national comparative studies reveal distinct consumption stratification in the Indian market: low-income groups struggle to afford ride-hailing costs due to service price elasticity constraints, while high-income groups use such services less frequently given greater availability of alternative transportation options [39]. In China’s urban transportation context, user behavioral decisions are primarily driven by dual factors of time cost constraints and travel budget limitations [40].

Despite the flourishing development of such research, there are still several areas that require more research. Firstly, due to the spatiotemporal uncertainty of ride hailing demand, there are contradictions in the conclusions about the impact of the built environment on ride-hailing travel at different times and spaces [41,42,43]. However, there are few studies that consider both temporal and spatial differences in methodology selection. Secondly, previous studies lacked synchronous analysis of the built environment characteristics of the departure and destination, resulting in the inability to reveal the spatial selection mechanism of the complete travel chain.

2.2. The Relationship Between Ride-Hailing and Walking

The digitalization process has profoundly restructured urban residents’ mobility decision-making mechanisms, with deep penetration of information and communication technologies leading to structural changes in travel patterns. As a core element of sustainable transportation systems, non-motorized travel methods possess multi-dimensional advantages in promoting public health, reducing energy consumption, and controlling environmental pollution. However, the substitution effect generated by optimizing service efficiency through mobile travel platforms presents new governance challenges.

Rayle et al. (2016), based on field survey data from the San Francisco area, revealed that approximately 40% of ride-hailing orders essentially replaced travel demand originally intended to be met by public transit or walking [2]. Intercity empirical studies have demonstrated that the typical service radius lies between 2 and 4 km [16], a spatial extent that substantially coincides with regions traditionally favorable for walking, thus substantiating its crowding-out impact on sustainable travel alternatives. Notably, the intensity of such substitution effects exhibits significant spatial differentiation characteristics, with its mechanism primarily constrained by three dimensional factors: platform technology accessibility, adaptation of urban morphological characteristics, and heterogeneity of user groups [6,44]. The environmental externality generated by this substitution effect cannot be ignored. Carbon emission model calculations show that if the current level of substitution is maintained, the target for peak traffic carbon emissions will face a 5% shortfall in achievement [45].

More concerning are the cascading effects on health: epidemiological evidence suggests that a reduction in average daily steps may elevate the risk of cardiovascular diseases and certain cancers [46], whereas high-density urban districts, thanks to better accessibility of walking facilities, exhibit lower health risks than suburban areas [47].

Therefore, understanding the connection between demand and the walkability attributes of the built environment can help clarify whether ride hailing services are being used without other transportation options or are still being used even when there are other more sustainable options available. The existing models only include macro indicators such as road network density and bus accessibility, and there are very few studies on the selection of built environment indicators that consider walkability [17], ignoring the potential impact of visual walkability factors such as street interface green visibility on travel mode selection.

2.3. Measurement of Visual Walkability

The objective walkability evaluation framework has undergone a three-stage evolution. Initially, as represented by the Walk Score, it utilizes GIS-based network analysis to compute the index-weighted distances of nine types of everyday facilities within a 15-min walk, thereby ensuring regional applicability but neglecting the influence of the physical street context. Subsequently, one study developed a four-dimensional spatial morphology framework. By employing parameters such as street network permeability (intersection density), functional mix (land-use entropy), spatial compactness (residential core density), and interface permeability (proportion of retail interfaces), they empirically demonstrated that every 10% increase in street interface permeability is associated with a 6.2% rise in the walking commuting rate [48]. Recent studies on segment-level walkability evaluation have considered factors such as the appropriateness of walkability and the features of street connectivity [48,49,50].

Building pedestrian-friendly communities is fundamentally an extension in practice of human-centered urban planning ideals. The core challenge facing current research is that traditional methods struggle to effectively capture the complex mapping between the built environment and pedestrians’ emotional spatial experiences. Existing research paradigms encounter methodological dilemmas on three levels: The prevailing approaches primarily rely on questionnaires, in-depth interviews, spatial audits, and behavioral observations [51,52]; although these methods can capture subjective spatial perception data, they are affected by both selection bias and social desirability bias. Traditional spatial audit tools (such as the MAPS system) can systematically assess physical street features; however, their metrics rely excessively on morphological parameters (e.g., facade continuity and street width-to-height ratio), thereby neglecting neuro-perceptual dimensions (e.g., visual complexity and environmental stress responses). Survey-based sampling approaches encounter limitations regarding both the spatial extent and the continuity over time. Even though this approach is convenient, it might produce misleading results, as it does not account for or control the potential influence of socioeconomic factors [53].

As sensor technology advances, the street view analysis framework has been steadily refined. In its early academic phase, colored perspective images, aerial photography, and LiDAR point clouds were frequently used as media to simulate pedestrian viewpoints for studies in street vegetation visualization. As technology iterates, street view imagery has increasingly served as a key data source for decoding the visual features of urban spaces [54], with its evolution displaying distinct phase-wise characteristics. Existing research has confirmed that street view parsing technology based on image segmentation can effectively quantify the dimensions of urban spatial perception [55,56,57,58]. Notably, the Zeng team (2018) innovatively combined a sky region recognition algorithm with the SVF (Sky View Factor) computation model, successfully establishing an urban spatial openness evaluation system based on street-view imagery [58]. In the field of vegetation research, this technology effectively addresses the shortcomings of traditional survey methods—such as sample bias and insufficient timeliness [59,60]—marking an important shift in data collection from manual measurement to intelligent recognition.

Although street-view databases have provided a new avenue for large-scale walking environment assessments [57], current methodologies still exhibit significant limitations: firstly, the insufficient standardization of data processing workflows reduces the comparability of results; secondly, most studies focus on a single environmental factor (such as the green-view index or sky visibility), lacking a coordinated analysis of multi-dimensional indicators [56]. Given the increasing leisure demands of urban residents, proposing an effective method to approximate visual walkability is of significant importance both theoretically and practically.

3. Data and Methodology

3.1. Data and Study Area

We utilized a de-identified travel dataset, accessible to academia via the Didi Chuxing platform, collected throughout November 2016. The dataset’s raw fields comprise driver anonymized IDs, unique journey identifiers, departure and arrival timestamps, and pick-up/drop-off geographic coordinates (with a geocoding accuracy of up to five decimal places), forming the basis for high-resolution spatiotemporal analysis of travel behavior.

Analysis of spatial distribution characteristics indicates that travel demand exhibits a significant core–periphery spatial differentiation pattern: the vast majority of ride-hailing orders are concentrated within the core urban area inside the Third Ring Road. Based on principles of data integrity and spatial heterogeneity control, this study focused on analyzing the 200-square-kilometer core urban area enclosed by the Third Ring Road (Figure 1). The delineation of this study area followed two criteria: first, the built environment element database (including land-use attributes, point of interest distribution density, and real estate valuation levels) had complete coverage in the core area; second, the sample region encompassed high-frequency ride-hailing zones in Chengdu, effectively representing the main characteristics of urban traffic activities.

3.2. Dependent Variables

This study employed a multi-dimensional time segmentation approach to analyze transportation modes. As shown in Figure 2, data visualization shows the demand characteristics of ride-hailing trips across different time periods.

The division of time periods was based on the integration of multiple data sources, including the commuting report released by the Didi platform [61], the urban rail transit timetable [61], and empirical survey data from the study area. Based on traffic flow fluctuation characteristics, the periods were primarily divided into four typical segments: morning peak, evening peak, leisure, and night. Based on traffic flow fluctuation characteristics, the periods were primarily divided into four typical segments: morning peak, evening peak, leisure, and night. On weekdays, the morning peak was defined as 07:00–10:00 from Monday to Friday, while the evening peak was defined as 17:00–20:00 from Monday to Friday [62]. Considering that some private enterprises implement flexible work schedules, which may induce commuting demand on Saturdays, this study specifically classified Sundays as non-working day leisure periods. The night period was determined to be 23:00–06:00, considering infrastructure constraints including the discontinuation of rail transit services and the curtailed operational hours of regular buses. In terms of modeling, urban communities were selected as the basic unit of analysis, and spatiotemporal panel data were constructed by calculating the connection frequencies for each time period. Specifically, the number of boardings and alighting during the morning peak, evening peak, late-night, and leisure periods divided by the community area served as a core explanatory variable and was incorporated into the regression analysis framework of the trip generation model.

3.3. Independent Variables

The construction of the variable system in this study followed the widely used “5D” theoretical framework in the field of transportation demand analysis [11,21]. We have selected variables that may be related to travel demand: population density, road density, FAR, housing prices, transportation availability, land-use mix, the number of points of interest, and visual walkability. It is particularly noteworthy that the evaluation of visual walkability was implemented using deep learning techniques on street-view images; this method employs convolutional neural networks to extract visual features of street facades and establishes a quantitative model correlating these features with walking comfort.

3.3.1. “5D” Variables

Given the absence of official population census data at the urban level in Chinese cities, this study utilized the 2019 100 m × 100 m gridded population spatial differentiation data provided by the WorldPop spatial demography platform (http://www.worldpop.org (accessed on 1 October 2020)). The dataset was generated by fusing diverse data sources such as remote sensing imagery, nighttime light intensity, and ground survey data and by utilizing machine learning algorithms for spatially discrete modeling of population distribution, effectively mitigating the scale dependency problem associated with traditional administrative unit-based population statistics. Compared to conventional census data, this grid-based population density estimation method exhibits significant advantages in terms of spatial resolution and timeliness [63], making it particularly suitable for fine-scale studies of highly dynamic urban systems.

In this study, we measured regional road network density using a spatial heterogeneity analysis approach. This was achieved by computing the ratio of the total mileage of roads within each unit (measured in kilometers) to the area of the unit (in square kilometers), forming a road density indicator expressed in km/km². The fundamental road network data was obtained from the open-source geographic data platform OpenStreetMap (https://www.openstreetmap.org/ (accessed on 15 October 2020)).

This study employed the FAR as a core indicator for characterizing urban spatial morphology, with its calculation based on a comprehensive three-dimensional building database covering Chengdu. This database fully includes the geometric characteristics, footprint area, and height parameters of all constructed buildings. The data were sourced from the 2018 urban building census project conducted by the Chengdu Municipal Planning and Natural Resources Bureau and have undergone topological verification as well as attribute integrity checks.

In response to the widespread challenge of limited micro-level socioeconomic data in Chinese cities, this study, based on the housing price–income elasticity theory [64], innovatively constructed a spatial proxy variable system centered on the transaction unit price of second-hand residences. Using the Amap API, second-hand housing was geocoded to form a spatial point database in the WGS84 coordinate system. Relying on the Scrapy 2.6 framework, a targeted web crawler system was developed to collect second-hand housing transaction data listed on Beike (https://bj.ke.com/ (accessed on 20 October 2020)) from 2015 to 2016, focusing on key fields such as property characteristics, historical transaction prices, and neighborhood environmental parameters. During data cleaning, Tukey’s fences method was employed to eliminate price outliers, and spatial imputation was conducted via the k-nearest neighbors’ algorithm, ultimately establishing a residential price characteristic database with spatiotemporal attributes.

Distance to public transit constitutes a critical determinant in urban mobility decisions. This study focused on the 2016 rail transit network (containing 3 lines and 38 stations), constructing a network distance-based accessibility evaluation model by integrating geographic data with the ArcGIS 10.8 platform: Actual route distances from community unit centroids to the nearest stations were computed using the Network Analyst toolkit.

The locational entropy of 2016 POI data within Chengdu’s Third Ring Road was employed to characterize functional diversity [65]. POI data sourced from map operators encompass 11 primary categories (corporate enterprises, residential areas, catering services, accommodation services, tourism attractions, cultural/recreational facilities, shopping venues, transportation infrastructure, research/education institutions, government agencies, and public services), with 30 subcategories. Higher entropy values indicate more balanced distribution of functional facilities, while lower values correspond to reduced functional mix.

$$MixUse{d}_k=-{\displaystyle \frac{{\displaystyle \sum _{i=1}^M{P}_{k,i}ln{P}_{ki}}}{lnM}}$$

(1)

where $P_{ki}$ denotes the percentage contribution of POI type i within spatial unit k relative to the unit’s total POI count, and M signifies the categorical variety of POIs in the spatial unit.

Our dataset incorporates quantitative measures of three POI categories—retail establishments, food/beverage outlets, and recreational venues—systematically retrieved via AMAP’s application programming interface. Empirical studies confirm that social interactions and recreational engagements represent primary trip motivations underlying ride-hailing service utilization [35,40].

3.3.2. Visual Walkability Variables

We utilized Baidu Map Street View (BMSV) API to download street-view imagery. An automated scraper was engineered to systematically capture street-level imagery using geospatial coordinates (latitude and longitude) of target locations. The crawler incorporated URL parameters defining directional and angular specifications transmitted via standard HTTP requests to obtain BMSV images for targeted street positions. The visual parameters were calibrated to emulate human-scale perception, with Field of View (FOV) fixed at 60° and pitch angle standardized at 22.5° for pedestrian perspective simulation. Omnidirectional coverage (360° horizontal span) was achieved through quad-directional image capture at 0°, 180°, 270°, and 360° azimuths per sampling location.

Street geometry parameters were computed through ArcGIS 10.2 after coordinate system transformation to Baidu Coordinate System, followed by attribute table exportation for spatial position extraction. The analytical framework generated 85,281 geospatial sampling points. SegNet’s architecture consists of sequential encoder layers and corresponding decoder layers (Figure 3), followed by a pixel-wise soft-max classification layer. This SegNet implementation was built upon the Caffe deep learning library. SegNet is based on the VGG16 network, incorporating convolutional layers, pooling layers, and ReLU activation functions for feature extraction. The decoder mirrors the encoder’s structure, leveraging saved pooling indices during upsampling to recover precise spatial information and minimize edge artifacts. A Softmax-activated classification layer outputs per-pixel class probabilities, ultimately yielding the semantic segmentation map. For training optimization, cross-entropy loss is employed to update parameters via backpropagation. Data augmentation techniques like rotation and flipping enhance the model’s generalization capability. Adaptive learning rate strategies (e.g., Adam optimizer) are implemented, with batch size configured according to GPU memory constraints. Categories with weak visual distinguishability are merged to reduce noise interference. Isolated erroneous pixels in segmentation results are removed through neighboring pixel analysis.

Based on the aforementioned expertise and theories, we proposed a method to measure four subindices related to visual walkability and calculate the visual walkability for each street location. The four subindices (Visual Greenery, Visual Crowdedness, Visual Enclosure, Visual Pavement) were constructed based on coverage calculations of distinct street features in street-view imagery (

Table 1. Definition and formula of visual walkability indicators.

Indicators	Definition	Formula
Visual Greenery	Visual Greenery is the street vegetation that pedestrians can see.	Visual greenery = ${\displaystyle \frac{\mathrm{green}\mathrm{pixels}}{\mathrm{total}\mathrm{image}\mathrm{pixels}}}$
Visual Crowdedness	Visual Crowdedness measures spatial oppression through the proportion of buildings and motor vehicles.	Visual crowdedness = ${\displaystyle \frac{\mathrm{building}\mathrm{pixels}+\mathrm{vehicle}\mathrm{pixels}}{\mathrm{total}\mathrm{image}\mathrm{pixels}}}$
Visual Enclosure	Visual Enclosure assesses visual openness through sky openness and interface permeability, which means the ratio of vertical objects to horizontal features.	Visual enclosure = ${\displaystyle \frac{\mathrm{building}\mathrm{pixels}+\mathrm{tree}\mathrm{pixels}}{\mathrm{pavement}\mathrm{pixels}+\mathrm{road}\mathrm{pixels}}}$
Visual Pavement	Visual Pavement determines pedestrian safety through sidewalk proportion.	Visual pavement = ${\displaystyle \frac{\mathrm{sidewalk}\mathrm{pixels}}{\mathrm{total}\mathrm{image}\mathrm{pixels}}}$

).

3.4. Models

Conventional trip generation frameworks have predominantly employed linear regression methodologies to predict aggregate trip frequencies, with these models historically correlating land-use variables and socioeconomic indices to trip volumes within discrete spatial units and temporal intervals. Nevertheless, these models fail to consider the spatial interdependencies. We estimated OLS and GWR models separately for morning peak, evening peak, late-nightm and leisure trip purposes. Conventional trip generation models apply global linear regression (OLS), which assumes constant spatial relationships between variables but fails to capture spatial heterogeneity. The incorporation of spatial weight matrices (e.g., Gaussian kernel functions) in GWR enables geographically varying coefficients that better capture localized effects of built environment factors on trip production. To statistically verify the existence of spatial autocorrelation in the dataset, the global Moran’s I index was computed. The spatial clustering of trip volumes in adjacent areas demonstrates the necessity of GWR modeling. Rigorous mitigation of local multi-collinearity concerns enhanced the statistical robustness and theoretical plausibility of GWR parameter estimates. Given the homogeneous spatial distribution of observational units (grid cells), a fixed-bandwidth Gaussian kernel was implemented, featuring distance-decaying weights.

$$w_{ij}=exp\left(-{\displaystyle \frac{d_{ij}^2}{{\theta }^2}}\right)$$

(2)

Here, $w_{ij}$ denotes the spatial weight assigned to observation j when estimating parameters for location i, $d_{ij}$ represents the Euclidean distance between locations i and j, and $\theta$ signifies the predetermined bandwidth parameter.

The optimal bandwidth $\theta$ was selected based on the AICc criterion (corrected Akaike Information Criterion), balancing model goodness-of-fit with complexity. AICc demonstrates superior performance over cross-validation (CV) in small-sample contexts, substantially reducing computational demands while maintaining selection reliability. Spatial diagnostics, including LISA cluster maps and Geary’s C index, were applied to residual analysis, confirming the successful elimination of residual spatial autocorrelation and validating the model’s spatial heterogeneity characterization capability.

4. Results

Descriptive statistics for neighborhoods inside Chengdu’s Third Ring Road are summarized in

Table 2. Community area statistics.

		Min	Max	Mean	Std. Dev.
Trips in morning peak hours/community area	average pick-ups (sum = 1,678,964)	92.12	37,627	4866.56	6100.813
	average drop-offs (sum = 1,598,715)	38.34	40,956	4633.96	5314.522
Trips in afternoon peak hours/community area	average pick-ups (sum = 263,961.3)	0	6047.9	765.1	1005.995
	average drop-offs (sum = 227,851.3)	0	5522.4	660.4	638.7626
Trips in night hours/community area	average pick-ups (sum = 3,329,599)	69.38	46,844	9651.01	9152.022
	average drop-offs (sum = 3,617,335)	41.83	103,365	10,485.03	14,742.28
Trips in recreation hours/community area	average pick-ups (sum = 1,545,129)	45.8	31,870	4478.6	4669.945
	average drop-offs (sum = 1,527,366)	43.22	37,686	4427.15	5093.238
Population density (residents/km2)		3996.2	62,731	25,476.24	8786.411
Road density (km/km2)		0	37.181	9.652022	4.984533
FAR		0	13.462	2.625983	1.617472
Housing price by cell (CNY/m2)		0	43,105	15,863.12	6071.479
Average counts of entertainments		0	268.53	36.26	34.61367
Average counts of restaurants		0	739.23	160.97	129.4699
Average counts of shopping malls		0	90.921	11.168	9.458874
Distance to city center (m)		220	9686	4740	2053.016
Distance to subway station (m)		74.89	3279.6	759.03	532.7586
Land-use mix entropy		2 × 10 −7	0.9872	0.828126	0.145013
Visual Greenery		0.0607	0.5333	0.22761	0.076622
Visual Crowdedness		0.0108	0.1343	0.05492	0.015487
Visual Enclosure		0.7052	176.45	5.1521	13.14099
Visual Pavement		0.024	0.5809	0.27674	0.096921

. Table 2 illustrates the generated travel spatial patterns during different time periods on weekdays. The trip counts generated within Chengdu’s Third Ring Road during nighttime period surpassed those observed in other temporal intervals. During morning/evening peak hours and leisure periods, pick-ups dominated drop-offs, whereas nighttime hours exhibited higher drop-off frequencies compared to pick-up activities.

The validity and accuracy of global OLS and local GWR models were assessed using key model performance indicators across different time periods.

Table 3. Morning pick-up model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−1.033	−0.155	0.367	0.278	−1.8 × 10−16
Population density	−0.18	0.219	0.773	0.247	0.1913 **
Road density	−0.172	−0.059	0.074	0.061	−0.07469 .
FAR	0.025	0.224	0.422	0.123	0.2949 ***
Average housing price	−0.035	0.217	0.362	0.117	0.1462 ***
Average counts of entertainment	−0.182	−0.059	0.066	0.064	−0.05239
Average counts of restaurants	−0.075	0.162	0.474	0.12	0.116 .
Average counts of shopping malls	−0.036	0.096	0.301	0.092	0.1397 **
Distance to city center	−0.676	−0.349	0.023	0.134	−0.314 ***
Distance to subway station	−0.188	0.025	0.215	0.087	0.04355
Land-use mix entropy	−0.543	−0.056	0.292	0.191	−0.2085 ***
Visual Greenery	−0.25	−0.093	0.032	0.063	−0.0947 *
Visual Crowdedness	−0.099	0.042	0.14	0.062	0.03279
Visual Enclosure	−0.006	0.215	0.632	0.166	0.1051 **
Visual Pavement	−0.218	−0.11	−0.013	0.05	−0.147 **
Adj.R2	0.753				0.629
AICc	−17,654.541				−12,670.456
Moran’s I					0.431 ***
Bandwidth					162

N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05, ^. p < 0.1.

,

Table 4. Morning drop-off model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−0.646	−0.096	0.353	0.231	−3.1 × 10−16
Population density	−0.151	0.165	0.481	0.173	0.145 .
Road density	−0.185	−0.055	0.042	0.054	−0.06795
FAR	−0.047	0.129	0.365	0.131	0.2096 **
Average housing price	−0.019	0.258	0.495	0.147	0.1597 ***
Average counts of entertainment	−0.207	−0.089	0.135	0.075	−0.06291
Average counts of restaurants	0.038	0.233	0.573	0.12	0.2091 **
Average counts of shopping malls	−0.015	0.09	0.266	0.081	0.0954 .
Distance to city center	−0.74	−0.355	0.157	0.196	−0.3309 ***
Distance to subway station	−0.151	0.04	0.236	0.079	0.05926
Land-use mix entropy	−0.553	−0.036	0.258	0.203	−0.1646 **
Visual Greenery	−0.3	−0.108	0.034	0.083	−0.09361 *
Visual Crowdedness	−0.151	0.026	0.12	0.075	0.01884
Visual Enclosure	−0.064	0.107	0.39	0.114	0.08026 *
Visual Pavement	−0.198	−0.07	0.046	0.058	−0.1234 *
Adj.R2	0.623				0.521
AICc	−23,901.615				−18,456.219
Moran’s I					0.528 ***
Bandwidth					185

N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05, ^. p < 0.1.

,

Table 5. Evening pick-up model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−0.25	0.03	0.251	0.134	2.162 × 10−16
Population density	−0.316	0.066	0.39	0.213	0.06701
Road density	−0.107	−0.032	0.073	0.042	−0.03154
FAR	−0.188	0.074	0.367	0.15	0.1147
Average housing price	−0.002	0.198	0.454	0.127	0.1425 **
Average counts of entertainment	0.069	0.255	0.417	0.111	0.2781 ***
Average counts of restaurants	−0.022	0.13	0.334	0.094	0.1157
Average counts of retails	−0.179	−0.083	0.025	0.062	−0.06012
Distance to city center	−1.044	−0.401	0.119	0.287	−0.3371 **
Distance to subway station	−0.073	0.082	0.19	0.047	0.06258
Land-use mix entropy	−0.612	−0.133	0.191	0.208	−0.159 **
Visual Greenery	−0.243	−0.088	0.08	0.083	−0.0696
Visual Crowdedness	−0.113	0.035	0.156	0.076	0.0348
Visual Enclosure	0.004	0.25	0.474	0.149	0.08475 *
Visual Pavement	−0.104	0.017	0.139	0.062	−0.02996
Adj.R2	0.721				0.674
AICc	−16,549.231				−14,320.636
Moran’s I					0.396 ***
Bandwidth					207

Note: N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05.

,

Table 6. Evening drop-off model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−0.602	0.078	0.546	375	1.021 × 10−17
Population density	−0.475	0.172	0.89	0.005839	0.226 **
Road density	−0.255	−0.061	0.076	5.787	−0.09337 *
FAR	−0.062	0.141	0.469	28.14	0.1828 *
Average housing price	−0.208	−0.061	0.248	0.00503	−0.03487
Average counts of entertainment	−0.213	−0.056	0.094	1.123	−0.06585
Average counts of restaurants	−0.014	0.129	0.517	0.3525	0.148 *
Average counts of retails	−0.06	0.078	0.287	3.681	0.2035 ***
Distance to city center	−1.114	−0.253	0.623	0.03124	−0.2222 *
Distance to subway station	−0.178	0.042	0.314	0.05676	0.03738
Land-use mix entropy	−0.35	0.11	0.756	254.8	−0.03498
Visual Greenery	−0.294	−0.066	0.187	388.5	−0.07183
Visual Crowdedness	−0.196	0.037	0.219	1785	0.04493
Visual Enclosure	−0.041	0.314	1.089	2.025	0.08798 *
Visual Pavement	−0.273	−0.074	0.077	337.9	−0.1368 **
Adj.R2	0.751				0.722
AICc	−14,503.853				−11,923.350
Moran’s I					0.568 ***
Bandwidth					149

Note: N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05.

,

Table 7. Nighttime pick-up model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−1.464	−0.181	0.52	0.41	1.68 × 10−17
Population density	−0.268	0.144	0.757	0.243	0.2091 **
Road density	−0.123	−0.017	0.097	0.054	−0.07127 .
FAR	−0.07	0.132	0.428	0.149	0.2224 ***
Average housing price	−0.147	0.02	0.281	0.122	0.02056
Avg. counts of entertainment	−0.214	−0.067	0.105	0.077	−0.1045 *
Avg. counts of restaurants	−0.024	0.152	0.536	0.123	0.1407 *
Avg. counts of retails	−0.087	0.094	0.263	0.088	0.1232 *
Distance to city center	−0.921	−0.43	−0.161	0.173	−0.3955 ***
Distance to subway station	−0.303	0.013	0.22	0.114	0.06498
Land-use mix entropy	−0.367	0.06	0.831	0.273	−0.08533 .
Visual Greenery	−0.175	−0.053	0.138	0.072	−0.07819 .
Visual Crowdedness	−0.07	0.048	0.182	0.059	0.04743
Visual Enclosure	−0.133	0.026	0.396	0.144	0.04817
Visual Pavement	−0.251	−0.091	0.013	0.062	−0.1703 ***
Adj.R2	0.875				0.811
AICc	−24,011.650				−21,789.121
Moran’s I					0.620 ***
Bandwidth					145

N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05, ^. p < 0.1.

,

Table 8. Nighttime drop-off model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−0.842	−0.127	0.285	0.228	−2 × 10−17
Population density	−0.183	0.143	0.445	0.158	0.1007
Road density	−0.127	−0.04	0.074	0.044	−0.05766
FAR	0.019	0.188	0.366	0.109	0.2694 ***
Average housing price	−0.003	0.198	0.337	0.1	0.1416 **
Avg. counts of entertainment	−0.146	−0.054	0.116	0.063	−0.03186
Avg. counts of restaurants	−0.031	0.124	0.366	0.088	0.09014
Avg. counts of retails	0.002	0.148	0.404	0.11	0.1861 ***
Distance to city center	−0.727	−0.427	−0.2	0.119	−0.3775 ***
Distance to subway station	−0.279	−0.03	0.154	0.088	0.01201
Land-use mix entropy	−0.473	−0.113	0.21	0.161	−0.2254 ***
Visual Greenery	−0.21	−0.089	0.005	0.041	−0.08202 .
Visual Crowdedness	−0.138	0.012	0.163	0.081	0.01187
Visual Enclosure	0.007	0.341	0.665	0.185	0.1186 **
Visual Pavement	−0.259	−0.146	−0.032	0.061	−0.1673 ***
Adj.R2	0.710				0.682
AICc	−22,067.144				−20,137.250
Moran’s I					0.597 ***
Bandwidth					149

N = 354; Significance levels: *** p < 0.001, ** p < 0.01, ^. p < 0.1.

,

Table 9. Casual time pick-up model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−0.652	−0.091	0.486	0.279	1.83 × 10−16
Population density	−0.148	0.231	0.665	0.219	0.2136 **
Road density	−0.227	−0.085	0.061	0.074	−0.0884 *
FAR	0.019	0.194	0.439	0.133	0.246 ***
Average housing price	−0.059	0.187	0.458	0.140	0.105 *
Avg. counts of entertainment	−0.201	−0.073	0.111	0.075	−0.05974
Avg. counts of restaurants	−0.020	0.215	0.556	0.127	0.1822 **
Avg. counts of retails	−0.047	0.060	0.197	0.062	0.08243
Distance to city center	−0.724	−0.249	0.575	0.241	−0.2689 **
Distance to subway station	−0.092	0.078	0.279	0.085	0.07149
Land-use mix entropy	−0.520	0.020	0.355	0.222	−0.128 *
Visual Greenery	−0.290	−0.128	0.021	0.079	−0.1165 **
Visual Crowdedness	−0.123	0.040	0.124	0.064	0.03445
Visual Enclosure	−0.012	0.109	0.458	0.135	0.06823 .
Visual Pavement	−0.185	−0.064	0.057	0.054	−0.1186 *
Adj.R2	0.742				0.711
AICc	−21,003.167				−19,082.249
Moran’s I					0.467 ***
Bandwidth					181

N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05, ^. p < 0.1.

and

Table 10. Casual time drop-off model.

	GWR				OLS
Variables	Min	Median	Max	St. Dev.	Coef.
Intercept	−0.592	−0.076	0.421	0.240	2.91 × 10−16
Population density	−0.158	0.145	0.377	0.145	0.1306 .
Road density	−0.193	−0.079	0.050	0.060	−0.08376 .
FAR	0.043	0.205	0.390	0.112	0.2517 ***
Average housing price	−0.028	0.219	0.435	0.137	0.1228 **
Avg. counts of entertainment	−0.199	−0.097	0.130	0.073	−0.07013
Avg. counts of restaurants	0.026	0.203	0.476	0.101	0.1758 *
Avg. counts of retails	0.016	0.128	0.284	0.071	0.1387 **
Distance to city center	−0.689	−0.320	0.239	0.185	−0.310 **
Distance to subway station	−0.095	0.055	0.287	0.080	0.05592
Land-use mix entropy	−0.471	−0.047	0.288	0.195	−0.1536 **
Visual Greenery	−0.315	−0.121	0.000	0.072	−0.1082 *
Visual Crowdedness	−0.191	0.006	0.134	0.087	−0.00071
Visual Enclosure	−0.005	0.241	0.527	0.143	0.08437 *
Visual Pavement	−0.172	−0.065	0.054	0.050	−0.1153*
Adj.R2	0.701				0.654
AICc	−17,981.326				−15,022.956
Moran’s I					0.511 ***
Bandwidth					149

N = 354; Significance levels: *** p < 0.001, ** p < 0.01, * p < 0.05, ^. p < 0.1.

summarize the Adjusted R², AICc, and Moran’s I values for both models. Across all temporal segments—morning peak, evening peak, nighttime, and casual hours—the GWR models consistently outperformed the OLS models in terms of model fit and spatial accuracy. In all cases, the GWR model not only improved the explanatory power of the model but also effectively mitigated spatial autocorrelation in residuals. These findings verify the spatial non-stationarity of explanatory variables and confirm the superiority of GWR in capturing localized patterns in ride-hailing demand.

4.1. Morning Peak Hours Model

Table 3 and Table 4 display the results of both the Geographically Weighted Regression (GWR) and Ordinary Least Squares (OLS) models applied to pick-up and drop-off patterns during the morning rush hour. For the GWR model, the spatial variability in the local coefficients of each independent variable is reported, including their minimum, median, maximum, and standard deviation values. In contrast, the OLS model provides global coefficient estimates along with their statistical significance.

The OLS model indicates that population density, FAR, average housing price, average counts of shopping mall and Visual Enclosure are all positively correlated with pick-up/drop-off activities during morning peak hours. Distance to city center, land-use mix entropy, Visual Greenery, and Visual Pavement show negative correlations with pick-up/drop-off activities during morning peak hours. Average counts of entertainment, distance to subway station, and Visual Crowdedness reveal no statistically significant associations with pick-up/drop-off dynamics. Population density, FAR, and distance to city center show larger coefficients for pick-ups than for drop-offs. In contrast, land-use mix, average housing price, and Visual Pavement exhibit larger coefficients for drop-offs.

In the GWR model, the magnitude of median coefficient estimates is similar to those in the global model for statistically significant coefficients. For most of our GWR coefficients, the estimated standard deviations exceed the median values. This indicates substantial spatial variability, suggesting that the effects of explanatory variables differ greatly across locations, thus reinforcing the necessity of localized modeling [35,66]. For pickups, FAR, average housing price, average restaurant count, average shopping malls count, and Visual Enclosure show median estimates exceeding standard deviations. For drop-offs, average housing price, average counts of entertainment, and average shopping malls count have median estimates exceeding the standard deviations.

4.2. Evening Peak Hours Model

Table 5 and Table 6 present the outcomes of the GWR and OLS models applied to evening rush hour pick-up/drop-off dynamics. The results are generally similar, but we find positive and significant coefficients for pick-ups regarding housing price, average recreational facility count, and Visual Enclosure. Pick-ups show negative and significant coefficients for distance to city center and land-use mix entropy. Drop-off models reveal statistically significant positive parameter estimates for population density, FAR, average counts of restaurants, average counts of shopping malls, and Visual Enclosure. The road density, distance to urban center, and visual walkway ratio exhibit negative and significant coefficients. In the GWR model, the magnitude of median coefficients is generally similar to those in the global model.

4.3. Nighttime Model

Table 7 and Table 8 present the outcomes of the GWR and OLS models applied to night hour pick-up/drop-off dynamics. All variables except housing price, distance to nearest subway station, Visual Crowdedness, and Visual Enclosure show statistical significance for pick-ups. Road density, average entertainment counts, distance to city center, land-use mix entropy, Visual Greenery, and Visual Pavement exhibit negative correlations with pick-ups. Population density, FAR, average restaurant counts, and average shopping malls counts show positive correlations with pick-ups. FAR, average housing price, average shopping malls counts, and Visual Enclosure show positive correlations with drop-offs. Distance to city center, land-use mix entropy, Visual Greenery, and Visual Pavement exhibit negative correlations with drop-offs. In the GWR model, the median coefficients of statistically significant variables are comparable; land-use mix entropy may show slight deviations in pickup parameters. Visual Enclosure may display larger deviations in drop-off aspects.

4.4. Casual Time Model

Table 9 and Table 10 present the outcomes of the GWR and OLS models applied to casual time pick-up/drop-off dynamics. Population density, FAR, average housing price, restaurant counts, and Visual Enclosure show significant positive correlations with pick-ups. Road density, distance to city center, land-use mix entropy, Visual Greenery, and Visual Pavement width exhibit significant negative correlations with pick-ups. Drop-offs show similar patterns to pick-ups, except average shopping malls counts are positively correlated with drop-offs, while average entertainment counts show no significant relationship. For covariates exhibiting significance in both the global and GWR specifications, the median parameter estimates demonstrate approximate alignment, suggesting baseline spatial stationarity in central tendencies. However, the pronounced dispersion of the GWR coefficients—with standard deviations systematically surpassing median magnitudes—underscores pervasive spatial non-stationarity in elasticity structures. For pick-ups, road density, distance to city center, land-use mix entropy, Visual Greenery, and Visual Pavement show standard deviations estimates exceeding the median. For drop-offs, road density, distance to city center, land-use mix entropy, Visual Greenery, and Visual Pavement have standard deviations estimates exceeding the median.

5. Discussion

This study addresses a research gap by jointly considering origin–destination neighborhood features and incorporating visual walkability into ride-hailing demand analysis. By constructing a novel indicator system and applying both the OLS and GWR models, it uncovers spatial–temporal variations in the elasticity of built environment effects.

The global regression model validates statistically significant associations between Didi Chuxing trip generation and predominant urban built environment determinants. Although the GWR results reveal spatial variations in these effects, the median values of the local coefficients are basically consistent with the global model outcomes in direction and magnitude. The analytical sequence proceeds from synthesizing global regression findings to interrogating spatial non-stationarity through GWR modeling.

5.1. OLS

The global model reveals that higher population density is generally associated with increased DiDi frequency. Areas exhibiting these characteristics are predominantly located in central business districts (e.g., Tianfu Square), traditional residential neighborhoods (e.g., communities around Kuanzhai Alley), and established commercial zones (e.g., Chunxi Road). First, areas with higher population density typically exhibit greater travel demand. The concentration of population and elevated transportation needs make ride-hailing services a convenient solution, naturally increasing pick-up/drop-off activity at service locations [35]. Furthermore, high-density areas often experience public transit congestion or accessibility limitations, particularly during peak hours. When conventional public transportation becomes inadequate in meeting mobility demands, ride-hailing services function as complementary or alternative mobility options [9], thereby increasing service adoption rates.

In urban areas with higher road density, residents have more diverse transportation options [67,68]. Beyond ride-hailing services, residents can choose from various transportation modes, including metro, buses, taxis, or bike-sharing systems. The operational efficiency and spatial accessibility of these complementary modes facilitates demand dispersion across transportation networks, thereby diminishing the necessity for ride-hailing platform dependency.

High-FAR zones (e.g., commercial hubs, office clusters, and large residential complexes) generally exhibit dense populations and frequent economic activities, generating substantial travel demand [35,69]. High-density zones like Chunxi Road and the Financial District naturally require more ride-hailing service points to accommodate commuting and shopping needs due to concentrated pedestrian flows and high trip frequency.

Residents in high housing price areas possess stronger economic capacity: Such areas generally indicate higher regional economic development levels and elevated household incomes. These residents exhibit greater willingness to pay for convenient and comfortable mobility solutions [70]. The operational flexibility and temporal efficiency of ride-hailing platforms, particularly in circumventing peak-hour congestion in mass transit systems, drive preferential modal selection that elevates service patronage within these urban zones.

Regions exhibiting elevated densities of commercial establishments, dining venues, and recreational amenities demonstrate statistically significant associations with ride-hailing trip generation across multiple models. This empirical pattern indicates that ride-hailing services predominantly facilitate social, recreational, and entertainment-oriented mobility patterns [33,71,72].

Urban cores typically concentrate commercial, office, and residential zones, featuring abundant workplaces, business facilities, recreational venues, and dining districts. High-density travel demand in city centers naturally increases ride-hailing pick-up points and passenger volumes [73,74]; well-developed transportation infrastructure in CBDs attracts ride-hailing drivers due to high demand and frequent order dispatch rates. This further enhances ride-hailing service availability and utilization at pick-up/drop-off locations.

Areas with high land-use mix are typically multi-functional zones (e.g., integrated residential, commercial, and office spaces), creating diversified travel demands [40]. These mixed-use environments enable residents to fulfill daily activity requirements (retail consumption, professional engagements, gastronomic needs, etc.) through pedestrian-scale accessibility. The spatial efficiency of mixed-use development reduces dependency on motorized transport, with active mobility (walking/cycling) and mass transit systems absorbing most trip demand despite elevated urban density [35,69]. But there are also studies showing that land-use mix has a significant positive impact on ride-hailing demand [41,73].

Studies indicate that the green-view index (GVI)—the visible green coverage from human eye level—significantly promotes active transportation modes like walking and cycling [75,76]. For example, street GVI positively correlates with walking duration, particularly in areas with harmonious street height-to-width ratios and high pedestrian concentration [77]. The findings imply that visual greenery exposure modifies mobility decisions through enhanced aesthetic and microclimatic conditions [78], inducing behavioral shifts toward non-motorized transport for trip distances under 2 km and ultimately mitigating reliance on motorized transport systems.

Outdoor enclosure generally indicates dense building clusters and elevated population density. Regions with greater outdoor enclosure typically contain more commercial and office districts. These areas experience heavy pedestrian flows and substantial travel demand, making ride-hailing services a natural first-choice transportation mode for residents and commuters due to their speed and convenience.

Visual Crowdedness is not significantly correlated with the boarding and alighting points of ride-hailing services at different time periods, indicating that the level of street chaos is not related to the volume of ride-hailing services.

Areas with wider pedestrian walkways generally indicate pedestrian-friendly environments and improved walking conditions [79], allowing residents to conveniently access surrounding destinations or transportation hubs on foot. When walkways are spacious and comfortable, more people are likely to choose walking over ride-hailing services, leading to reduced usage at online car-hailing pickup points. Enhanced pedestrian accessibility between neighborhood zones reduces dependence on ride-hailing services for short-distance mobility, ultimately diminishing passenger throughput at corresponding pickup nodes.

5.2. GWR

The results of the GWR model are shown in the Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10 and Figure 11. The GWR model reveals that the influence of population density on ride-hailing demand is spatially heterogeneous, with stronger positive effects concentrated in urban cores such as Tianfu Square and Chunxi Road. These findings align with prior research emphasizing the role of population agglomeration and commercial clustering in shaping mobility patterns [3,69,80]. Notably, the effect is most pronounced during evening peak hours, indicating intensified demand pressure. This localized amplification complements the OLS model’s global results by highlighting spatial variations in the strength of association.

Compared to the OLS model, which indicates that higher FAR areas—such as commercial and high-density residential zones—generate greater ride-hailing demand due to economic activity and pedestrian intensity, the GWR model uncovers spatial heterogeneity in this relationship. Specifically, FAR demonstrates stronger positive effects in the northwest, with gradually diminishing influence toward the southeast. These results differ from Wang and Noland’s (2021) findings that elasticity peaks in central zones (e.g., Chunxi Road) [35], yet they align with prior studies highlighting diminishing marginal returns in saturated areas. A possible explanation is that our model integrates visual walkability and employs neighborhood-scale spatial units, which may modulate the local influence of FAR and produce distinctive elasticity patterns compared to global models.

Consistent with the global OLS results, the GWR model shows predominantly positive coefficients for housing price, affirming its strong association with ride-hailing demand. However, spatial heterogeneity emerges: Central zones exhibit the highest elasticity, which weakens toward peripheral areas—especially in the east. Some studies emphasize that the relationship between housing prices and ride-hailing usage is influenced by local environmental factors, such as traffic network density [40,74]. Central districts’ high-density, small-block configurations generate fragmented mobility needs that are well served by ride-hailing’s operational flexibility. Peripheral low-density, large-block layouts create dispersed trip endpoints, reducing ride-hailing service efficiency and suppressing demand.

Consistent with OLS findings that highlight the significance of recreational amenities, the GWR model further reveals spatial heterogeneity in this relationship during evening peak hours. The number of leisure facilities is positively associated with ride-hailing demand [34,81], but the effect weakens from south to north. This gradient reflects differentiated urban structures and user profiles. Southern core districts, featuring high-end malls like SKP and Intime in99, host a dense concentration of financial and IT professionals with strong consumption preferences and high ride-hailing adoption. These areas support seamless post-work transitions from offices to entertainment venues, triggering spatially coupled mobility patterns. In contrast, northern neighborhoods—characterized by wholesale markets and aging residential blocks—lack vibrant leisure infrastructure, resulting in insufficient demand nucleation. The observed pattern deviates from previous studies, which found stronger effects near airports or evenly across urban cores, possibly due to intercity differences in urban form and socioeconomic composition [34].

Echoing the OLS results, the GWR model confirms a positive but spatially heterogeneous relationship between restaurant density and ride-hailing demand. The effect is strongest in the urban core and weakens toward peripheral areas. Temporal differences are also evident: Restaurant counts are positively associated with pick-ups during both morning and night hours, though the morning result contrasts with earlier findings using the same dataset [35]—likely due to different definitions of peak periods. During recreational hours, both pick-ups and drop-offs exhibit significant positive correlations with restaurant density, suggesting that dining remains a major driver of discretionary travel. In suburban zones, fragmented land use and modal competition may suppress this effect.

The relationship between ride-hailing demand and the distribution of shopping facilities varies by time of day and spatial context. Different from the OLS results, pick-up demand during the evening peak hours is negatively associated with the average number of shopping malls, suggesting that shopping-oriented trips may not coincide with traditional commuting peaks [35]. In contrast, during other time periods, shopping facility density positively influences ride-hailing demand, a finding that aligns with those observed in earlier studies [34,81]. Spatially, the strength of this influence exhibits a declining gradient from southeast to northwest. In southeastern Chengdu—particularly within the Third Ring Road—large-scale transit-oriented development (TOD) commercial complexes such as Wanda Plaza Phase II attract over 100,000 daily visitors and generate consistent, all-day travel demand. By comparison, the northwestern periphery is dominated by wholesale markets and warehouse-style hypermarkets with limited retail diversity, resulting in a lower frequency of travel demand and reduced spatial impact of shopping amenities on ride-hailing activity.

While the OLS model confirms a general negative relationship, the GWR model uncovers spatial heterogeneity in how the distance from the city center affects ride-hailing demand. A negative effect dominates in southeastern peripheral areas, likely due to long-distance commuters preferring cost-effective public transit. In contrast, a positive effect appears near the northern Second Ring Road, particularly around intercity transit hubs, where ride-hailing supplements first- and last-mile connectivity. These spatial variations align with prior studies and highlight demand nuances across urban subregions [69,74].

Unlike the uniformly negative association identified in the OLS model, the GWR results reveal spatial heterogeneity in the relationship between land-use mix entropy and ride-hailing demand, with significant positive correlations in some neighborhoods and negative correlations in others—a pattern partially consistent with findings from earlier studies [34,69,74]. The number of neighborhoods with significant associations is highest for pick-ups during the evening peak hours and recreational periods, indicating that land-use diversity may play a more prominent role during non-work and discretionary travel times. In other time periods, the spatial distribution of significant effects remains relatively consistent. Notably, in the southwestern region of Chengdu, land-use mix entropy is negatively associated with both pick-up and drop-off densities. This area, typified by clear functional zoning—such as the Financial City in the High-Tech Zone—is characterized by highly concentrated and predictable commuting patterns.

In suburban areas, the green-view index (GVI) exerts a significant negative effect on the density of ride-hailing pick-up and drop-off points during weekends. One possible explanation is that greenery-rich environments are generally more walkable [78,82], and on weekends, when residents have more flexible schedules, short-distance ride-hailing trips are more likely to be substituted by walking. Additionally, street-level imagery with high vegetation coverage is often concentrated in peripheral or suburban zones, where ride-hailing service availability is relatively low, leading to longer wait times and lower service efficiency [19]. Moreover, green spaces such as parks and open lawns are frequently disconnected from major arterial roads, making it more difficult for drivers to access these locations promptly, thereby reducing the attractiveness of ride-hailing as a travel option in such areas [19]. Notably, during other time periods, the influence of the green-view index on ride-hailing demand is largely insignificant across most neighborhoods. This may be because residents face tighter time constraints during morning and evening peak hours, rendering walkability less influential in shaping mode choice. At night, concerns about personal safety may further discourage walking, diminishing the relative impact of green streetscapes on mobility decisions.

The enclosure positively affects the density of boarding and alighting points for ride hailing services [19], with the impact decreasing from southwest to northeast. The traditional pattern of narrow roads and dense networks is preserved within the southwest Third Ring Road, where commercial and residential functions are highly mixed along the street, giving rise to the demand for all-weather ride-hailing services. The northeast Third Ring Road adopts modern planning with wide roads and streets, large building boundaries, wide roads, and clear functional zoning, resulting in linear travel demand. Shared electric bicycles have a high penetration rate in wide non-motorized lanes, and short-distance travel is diverted.

During nighttime hours, visual pavement shows a negative association with ride-hailing demand in certain communities (same as the OLS result), with stronger effects observed in peripheral areas compared to the urban core. This spatial pattern may reflect the reduced attractiveness of ride-hailing in peripheral zones, where service availability is lower and wait times are longer [83]. In such contexts, even though sidewalk quality may be relatively poor, residents may still resort to walking—despite safety concerns—due to limited alternatives and cost considerations. By contrast, in central districts where ride-hailing services are more accessible and response times shorter, pedestrian infrastructure plays a less decisive role in shaping mode choice. A possible explanation for this spatial heterogeneity lies in the constrained transportation options in peripheral areas, which amplify the behavioral sensitivity to built environment features such as visual pavement, thereby lowering the likelihood of choosing ride-hailing over walking for short nighttime trips.

6. Conclusions

In this study, we used Didi data integrated with multiple building environment variables to investigate the correlation between spatial characteristics and ride-hailing in Chengdu, China. This study introduced visual walkability into ride-hailing research, offering a novel analytical dimension. By constructing a new indicator system and modeling spatial–temporal variation, it reveals how built environment features influence ride-hailing demand and potential substitution of sustainable modes.

The global model shows that population density, road density, FAR, housing prices, and proximity to the city center are significantly positively correlated with Didi demand at all time periods. Restaurants and entertainment facilities have a greater impacts on taxi demand during peak and late-night hours. The GWR model indicates that population density, FAR, and distance from the city center are global driving factors, while entertainment facilities, road density, and spatial perception variables exhibit temporal specificity. In terms of visual walkability, the Visual Enclosure significantly increases the demand for shared transportation, while Visual Greenery and Visual Pavement exhibit local inhibitory effects. The impact of Visual Crowdedness is not significant.

Population density and land-use data rely on multi-source estimation, which may introduce errors. The research scope is limited to within the Third Ring Road of Chengdu and may overlook the ride-hailing mode characteristics of suburban areas or emerging development zones. This study contributes to the understanding of spatial associations between ride-hailing demand and visual walkability. However, due to methodological limitations, causality cannot be inferred.

Policymakers can benefit from their understanding of ride-hailing modes. They can optimize ride-hailing service points in high-density areas during morning rush hour, such as the Financial City, to reduce congestion and strengthen roadside parking management in late-night dining and entertainment areas. They can also encourage carpooling through subsidies or priority lane policies, improve land-use diversity, reduce the demand for short-distance taxis, and promote walking and public transportation. Combining Didi’s big data, they can establish a dynamic demand response system to optimize transportation capacity allocation and suggest strengthening enclosed space design in commercial/transportation hubs to enhance parking demand. Policymakers can optimize pavement design in areas such as main roads to reduce negative impacts on vehicle parking, as well as promote collaborative planning to increase greenery and transportation services outside residential areas.

Future research can integrate real-time traffic, weather, and social data to deepen the analysis of travel purposes such as commuting, entertainment, and medical treatment. We can expand our work to the suburbs of Chengdu (such as Pidu District and Tianfu New Area) and analyze the differences in ride-hailing behavior between urban and rural areas. Furthermore, we can track the actual traffic reduction and emission reduction effects of policies such as carpooling incentives and shuttle area settings (refer to Didi’s carbon reduction data in Chengdu).