Incorporating Street-View Imagery into Multi-Scale Spatial Analysis of Ride-Hailing Demand Based on Multi-Source Data

Abstract

The rapid expansion of ride-hailing services has profoundly impacted urban mobility and residents’ travel behavior. This study aims to precisely identify and quantify how the built environment and socioeconomic factors influence spatial variations in ride-hailing demand using multi-source data from Haikou, China. A multi-scale geographically weighted regression (MGWR) model is employed to address spatial scale heterogeneity. To more accurately capture environmental features around sampling points, the DeepLabv3+ model is used to segment street-level imagery, with extracted visual indicators integrated into the regression analysis. By combining multi-scale geospatial data and computer vision techniques, the study provides a refined understanding of the spatial dynamics between ride-hailing demand and urban form. The results indicate notable spatiotemporal imbalances in demand, with varying patterns across workdays and holidays. Key factors, such as distance to the city center, bus stop density, and street-level features like greenery and sidewalk proportions, exert significant but spatially varied impacts on demand. These findings offer actionable insights for urban transportation planning and the design of more adaptive mobility strategies in contemporary cities.

1. Introduction

In recent years, the transportation industry has experienced significant transformations driven by the widespread adoption and advancement of mobile and wireless communication technologies [1]. These innovations have led to the electrification of vehicles, the integration of intelligent systems, enhanced networking capabilities, and the rise of shared mobility options. Among these, ride-hailing services facilitated by smartphone applications developed by Transportation Network Companies (TNCs) have emerged as a dominant and flexible travel option for urban residents [2]. A notable example is Didi Chuxing, which, since its launch in December 2015, has achieved remarkable popularity in China, serving over 1.3 billion users across 60 cities and handling an average of more than 2.4 million daily orders [3]. This success underscores the effectiveness of shared mobility solutions in modern urban transportation.

Compared to conventional transportation modes such as public transit or taxis, ride-hailing offers several advantages, including greater convenience, personalized service options, real-time responsiveness, and cost transparency [4,5,6]. With intelligent matching algorithms and GPS-based tracking, ride-hailing platforms reduce passenger wait times and provide flexible routing options [7]. Passengers can further tailor their experiences by selecting vehicle types and ride conditions, such as requesting child seats or non-smoking environments [8,9]. The shared economy model inherent to ride-hailing improves vehicle utilization and incorporates green transportation options, such as electric vehicles, which help to alleviate traffic congestion and reduce environmental pollution [10,11,12]. Despite these benefits, the rapid surge in demand, especially during peak periods, has led to service supply shortages, increased wait times, and localized inefficiencies [13]. These challenges underscore the importance of identifying the key spatial and temporal drivers of ride-hailing demand. Empirical evidence suggests that ride-hailing demand is shaped by complex temporal and spatial dynamics [14,15]. Demand tends to rise during commuting hours and in high-density or commercial areas, while decreasing during late nights, holidays, and in peripheral zones. Seasonal and event-related factors, such as weather fluctuations, public holidays, and festivals, further contribute to demand variability.

Research indicates that ride-hailing usage patterns result from a combination of internal and external factors [16,17,18], including user preferences, sensitivity to convenience and cost, and the demand for personalized services [19,20]. While early studies focused on broad socioeconomic indicators to assess the development of ride-hailing services across different regions [21], recent research has increasingly adopted a finer spatial lens. Built-environment variables at the neighborhood- or census-tract level have gained prominence in understanding demand heterogeneity [22,23]. Variables such as population density, land-use diversity, road network structure, access to public transportation, the presence of commercial and service facilities, and traffic conditions have all been shown to affect ride-hailing patterns [24,25,26,27]. Areas characterized by limited public transportation access often exhibit higher dependence on ride-hailing services [28,29], while the spatial distribution and accessibility of transit stations play a critical role in determining the frequency and spatial distribution of ride requests [15,28,30]. Thus, spatial modeling of these built-environment features provides valuable insights for improving demand forecasting and optimizing mobility planning [31]. These insights support more informed planning decisions and enable efficient scheduling, planning, and resource allocation for ride-hailing services by providing a robust theoretical foundation.

Against this background, the present study investigates the spatial heterogeneity of ride-hailing demand in Haikou, China, by integrating multi-source data, including street-level imagery and spatial indicators. The research seeks to address three key questions: (1) Do built-environment factors influence ride-hailing demand consistently across different geographic areas? (2) How can traditional planning indicators, such as points of interest (POIs), population, or GDP, be refined to improve spatial resolution and granularity? (3) How can the effective range and thresholds of various built-environment elements be accurately determined? This study makes several significant contributions as follows: (1) expanding the analytical focus from spatial to spatiotemporal dimensions, capturing the dynamic interplay between time and geography in shaping demand; (2) incorporating street-view images and applying computer vision techniques to extract fine-grained environmental features, allowing for a more accurate representation of urban context; and (3) employing the multi-scale geographically weighted regression (MGWR) model to account for spatial non-stationarity, enabling each predictor to operate across distinct spatial scales.

The rest of the paper is structured as follows. Section 2 reviews the relevant studies. Section 3 describes the data, study variables, and methodologies. Section 4 presents the analyses and discusses the modeling results. Section 5 concludes this study and discusses future work.

2. Literature Review

2.1. Ride-Hailing Services and Data Sources

The sharing economy involves internet-based platforms that enable individuals to exchange capital, assets, and services, thereby making use of underutilized resources and shifting consumers from owning assets to sharing them [32]. Within this context, the transportation sector has been at the forefront of the sharing economy, introducing a range of technology-enabled mobility solutions such as car-sharing, bike-sharing, and scooter-sharing [33,34,35]. Among these, ride-hailing, a commercial form of car-sharing, separates vehicle usage from vehicle ownership [36], allowing individuals without a personal car to access on-demand driver and vehicle services through smartphone applications [37,38]. Studies indicate that the key factors attracting residents to ride-hailing services include lower costs, reasonable travel times, convenient payment, avoiding driving under the influence, and short waiting periods [36,37,38,39,40,41]. These advantages not only streamline travel decisions but also gradually influence travel behaviors, shaping trip frequency and type, destination choices, travel times, and the range and nature of activities [42,43,44,45]. Since ride-hailing services can pick up and drop off passengers at virtually any point in a city [46,47,48], they have a particularly strong connection to the built environment, offering new perspectives for examining interactions between urban spaces and travel behavior. However, many of these studies rely on data from specific regions or user surveys, which may limit the generalizability of their conclusions across different socio-spatial and cultural contexts.

Recently, the data availability for analyzing ride-hailing demand has expanded significantly, with cities such as Chengdu [49], New York [50], and London [51] making ride-hailing data publicly accessible. This openness holds substantial potential for advancing the understanding of spatiotemporal travel demand. Traditional surveys or official statistics are often used to explore factors related to the built environment [52,53], and ongoing developments in information technology have made data collection through web scraping more feasible [54]. POIs, serving as one representative form of open data, have gained popularity in land-use analysis and are considered a practical means of accurately identifying and interpreting the nature of specific areas [15,55,56]. Moreover, the broad use of social media platforms such as Twitter and Weibo offers an opportunity to examine the public sentiment and behavioral patterns related to ride-hailing services [57,58,59], supplementing or even replacing traditional surveys by providing real-time detailed insights.

2.2. Built Environment and Travel Behavior

Understanding ride-hailing demand patterns is critically important as it not only provides a scientific foundation for developing diverse operational strategies but also supports more effective urban and transportation planning [60]. In recent years, researchers have extensively investigated how the built environment influences ride-hailing demand, recognizing it as a primary driver of mobility needs [15,25,28,29,31,34,36,52,53,55,60]. Defined as a human-made space shaped by policies and human activities, the built environment consists of multiple interrelated spatial variables [61]. These variables are commonly grouped into three dimensions: density, diversity, and design. Density refers to the spatial concentration of economic activities in a given area, often measured by key indicators such as population density, employment density, or building density [62]. It reflects the scale of human and economic activities in a region and serves as a foundational metric for evaluating transportation demand. The diversity dimension considers the mix of land uses within a region, ranging from residential to commercial or industrial, which can shape both the types and patterns of travel [63]. The design dimension encompasses the features of the transportation network, including road connectivity, pedestrian facilities, and bicycle lane accessibility [64]. Well-designed networks can improve traffic efficiency and lower travel costs, indirectly affecting ride-hailing demand. Building on these fundamentals, researchers have further expanded the analytical framework for the built environment by incorporating transportation distance and demand management characteristics into a “5D” model [65,66]. This extension adds two key dimensions: transportation distance, which gauges how readily users can reach their destinations, and demand management, which includes transportation policies, parking facilities, and pricing incentives, factors that directly influence the frequency and rate of ride-hailing use.

Over the past decade, a substantial body of research has confirmed that the built environment has a broad and significant influence on travel behavior [67,68,69,70]. A number of these studies emphasize statistical associations without adequately addressing the potential for self-selection bias, where individuals may choose neighborhoods that align with their travel preferences, thereby complicating causal interpretation. Scholars adopt various perspectives and methods in their analyses. Some focus on geographic patterns, examining how spatial characteristics in different areas affect the levels of ride-hailing demand [60,71,72]. For example, city centers often exhibit higher demand due to dense populations and frequent commercial activities, while peripheral areas with less-developed transportation infrastructure may see lower demand. Other studies employ the concept of “islands”, highlighting areas where ride-hailing demand is heavily concentrated or minimal. By identifying these “demand islands”, researchers can reveal spatial distribution patterns and intrinsic characteristics of specific regions [73,74]. This approach helps to isolate the relationship between the built environment and ride-hailing demand, particularly in high-demand “hotspots” and low-demand “coldspots”. Meanwhile, temporal variation is also crucial. Differences in demand during peak and off-peak periods, as well as weekday versus weekend travel behaviors, illustrate how time heterogeneity significantly affects demand patterns [15,53,71,72]. Including the temporal dimension not only highlights how built-environment factors may shift in importance over time but also enables a more comprehensive understanding of demand dynamics.

Furthermore, incorporating socioeconomic variables broadens the examination of how the built environment relates to ride-hailing demand. Factors such as demographic structure (e.g., age and gender distribution), income levels, and education attainment help to explain regional differences in demand [24,49,71,75,76]. Areas with higher incomes may display greater acceptance of ride-hailing services, while lower-income areas might rely more on public transit. Regions with low population density but robust economic activity can exhibit unique demand patterns that interact with built-environment factors.

2.3. Spatial Analysis Methods

The previous studies on travel behavior commonly employed Ordinary Least Square (OLS) regression to investigate influencing factors [60]. Other methods, such as discrete choice models, spatial clustering, and machine learning algorithms, have also been used under various research contexts [77,78,79]. However, these approaches often assume that all the explanatory variables remain spatially stationary across the study area [80,81]. In reality, travel distribution varies across urban spaces and is influenced by built-environment characteristics, resulting in substantial spatial heterogeneity. To address this issue, empirical research has introduced local models that account for spatial non-stationarity, most notably the geographically weighted regression (GWR) model [60,72,82,83,84]. GWR uses a spatial weighting matrix that locally calibrates model parameters for each observation, capturing the spatial variations in the estimated coefficients. Recognizing that spatial heterogeneity can manifest at different scales, some scholars have proposed the semi-parametric geographically weighted regression (SGWR) model, which categorizes influencing scales into global or local [81,85,86]. Nevertheless, SGWR has certain limitations in refining spatial detail. More recently, the multi-scale geographically weighted regression (MGWR) model has gained traction [71,73,87]. Unlike traditional GWR, MGWR allocates a distinct bandwidth to each variable as an indicator of its spatial scale, enabling a better fit for scenarios where the underlying processes operate at multiple scales. Despite these advances, the MGWR model also has limitations, such as increased computational complexity and potential multicollinearity, which require careful parameter tuning and model diagnostics, which are often underreported in empirical applications.

2.4. Research Gaps

Despite the extensive literature on the relationship between ride-hailing and the built environment, several limitations remain that warrant further exploration. First, many studies still rely on conventional surveys or statistical data to represent built-environment variables, often defining the density of POIs as a proxy for specific land-use functions. This approach can introduce inaccuracies stemming from the data collection methods, varying levels of granularity, and imbalanced distributions of different POI categories [88]. Second, the temporal dimension, an essential factor influencing ride-hailing usage, has often been overlooked. Ride-hailing trips can vary significantly over time due to diverse passenger purposes (e.g., commuting versus non-commuting) and may exhibit uneven distributions across different periods of the day. To address these issues, the present study collects actual ride-hailing trip records, integrates street-view features extracted via image segmentation, and combines additional raster-based data on population, GDP, housing prices, and public transit accessibility. Recognizing the distinct usage patterns on workdays and holidays, as well as throughout the day, trips are categorized into five time periods, morning peak (7:00–10:00), daytime (10:00–17:00), evening peak (17:00–20:00), night (20:00–24:00), and overnight (24:00–7:00), to reflect varying trip purposes. Leveraging the MGWR model’s demonstrated strength in handling complex spatial contexts provides a robust methodological framework for a deeper analysis of ride-hailing behavior.

3. Data and Methodology

3.1. Study Area, Time Span, and Dependent Variables

This study focuses on Haikou, a major city in southeastern China. As the capital of Hainan Province, Haikou serves as a key economic, cultural, and transportation hub in the region. According to the 2017 governmental statistical report [89], the city’s annual GDP was approximately USD 20.585 billion, and it had a population of about 2.2721 million. The ride-hailing dataset used in this research originates from the Didi Chuxing platform and covers all travel orders in Haikou during 2017. Each record includes various fields, such as order ID, service type (e.g., express or premium), pick-up and drop-off coordinates, departure and arrival times, as well as the distance between origins and destinations, with all data referenced to the WGS84 coordinate system. Administratively, Haikou is divided into sub-districts and towns. Sub-districts are typically located in urban areas, whereas towns lie farther from the city center. To ensure the representativeness of the data, this study selected travel orders from October 2017 as a case sample, encompassing all 21 sub-districts in Haikou, primarily within Meilan District, covering approximately 19°56’–20°6’ N and 110°15’–110°25’ E. During data preprocessing, we excluded all trips whose pick-up or drop-off points were outside the study area, as well as trips lasting under five minutes or over 120 min, resulting in a final set of 1,732,629 valid records. To facilitate the integration of multiple data sources and enhance the spatial resolution of the analysis, the entire study area was divided into 611 grid cells of 500 m × 500 m each (Figure 1), with each cell roughly corresponding to one neighborhood [81,88].

In this study, ride-hailing usage was calculated separately for workdays and holidays, resulting in data from 17 workdays and 14 holidays. To further examine how built-environment factors influence demand across different times and dates, this study divides daily travel into five time slots: morning peak (7:00–10:00), daytime (10:00–17:00), evening peak (17:00–20:00), night (20:00–24:00), and overnight (24:00–7:00). Unlike previous research that aggregates demand to the daily level [60,90], we separately summarize average hourly demand for each time slot within each grid cell, yielding ten dependent variables in total.

3.2. Explanatory Variables

Building on the literature review, this study selects explanatory variables from three dimensions, socioeconomic and transportation accessibility, land-use characteristics, and streetscape features, corresponding to regional, neighborhood, and street-level scales.

3.2.1. Socioeconomic and Transportation Accessibility Variables

The initial set of socioeconomic and transportation accessibility variables includes gross resident product, population, housing, road mileage, number of bus stops, and distance to the city center. Gross resident product data are derived from actual production values adjusted by nighttime light data [91]. Population data come from the open-source WorldPop database (https://www.worldpop.org/ (accessed on 20 February 2025)) and have been calibrated based on China’s Seventh National Population Census. Housing information, such as prices, floor area, construction year, addresses, and names, is obtained from Beike (https://sh.ke.com/ (accessed on 20 February 2025)), one of China’s largest and most reputable online real estate agencies. From 9280 records of second-hand housing transactions, 6260 records within the study area were selected, and these were used to calculate the average housing price in each neighborhood. Road mileage data are collected from OpenStreetMap (OSM) (https://www.openstreetmap.org/ (accessed on 20 February 2025)), including major roads, primary roads, secondary roads, and urban side streets. Since Haikou’s subway system is still in the planning stage, buses currently serve as the city’s primary form of public transit, leading this study to select the number of bus stops as an indicator of local bus accessibility. All of these variables are aggregated at the grid-cell level. Distance to the city center is measured as the geodesic distance from each grid cell’s centroid to the Hainan Center (110.35522° E, 20.01773° N) [92].

3.2.2. Land-Use Variables

Land-use variables are obtained through Baidu’s open platform, including different POI categories, such as metro stations, tertiary hospitals, large shopping centers, parks, schools, and commercial districts (http://lbsyun.baidu.com/ (accessed on 20 February 2025)). POIs are generally regarded as key indicators of urban land-use functions and thus represent built-environment diversity [73]. This study classifies the collected POI data into 14 categories to capture the functional distribution of different urban areas. However, because the quantity of POIs varies considerably across categories, simply counting the number of POIs could lead to significant errors. Instead, this study uses Enrichment Factors ($EF$s) within each grid cell rather than the density of POIs to represent local land-use characteristics. $EF$ serves as a normalized measure that mitigates the imbalance in the distribution of POIs among categories, thus more accurately reflecting real land-use functions [88]. The $EF$ value is calculated by comparing its local proportion to the overall proportion, as shown in Equation (1).

$$E{F}_j={\displaystyle \frac{N_j^i/N_j}{N^i/N}}$$

(1)

where i refers to ith POI category, j refers to the jth analysis grid, and N refers to the total number of POIs.

3.2.3. Streetscape Feature Variables

Street-level variables are derived from semantic segmentation of street-view images to capture features such as roads, buildings, and greenery. These features effectively reflect street design. Semantic segmentation partitions an image into non-overlapping meaningful subregions where pixels within the same region share certain similarities. This study employs the DeepLabV3+ neural network model [93], an improved form of a deep convolutional neural network (DCNN). With its encoder–decoder structure, DeepLabV3+ gradually captures semantic information and high-level features. In the encoder, the model integrates a large number of dilated convolutions to expand the receptive field without losing critical information, while the decoder fuses both low-level and high-level features to improve boundary accuracy [94,95].

We adopt ResNet-101 as the backbone network for DeepLabV3+ and train the model on the public Cityscapes dataset, which is a large-scale open dataset for urban spatial scene understanding. Model parameters are updated through backpropagation in each iteration, defined as processing one batch of data. After each epoch, model performance is evaluated on the validation set. If the evaluation metric improves, the current model parameters are saved to replace the previously stored version. In our experiment, the learning rate is set to 0.1, the maximum number of epochs is 30,000, and the batch size is 8. Under the optimal configuration, DeepLabV3+ achieves strong segmentation performance on the validation set, with an overall accuracy of 95.96%, mean accuracy of 84.73%, and mean intersection over union (MIoU) of 77.36%. Once trained, the model segments each street-view image into 19 semantic categories, including road, sidewalk, building, wall, fence, pole, traffic light, traffic sign, vegetation, terrain, sky, person, rider, car, truck, bus, train, motorcycle, and bicycle. The model training and testing are conducted on a server equipped with one NVIDIA GeForce RTX 4090 GPU (NVIDIA Corporation, Santa Clara, CA, USA) and anIntel Core i9-13900K CPU (Intel Corporation, Santa Clara, CA, USA).

The application of DeepLabV3+ in this context assumes that the target street-view images are semantically and visually similar to the urban street scenes in the Cityscapes dataset used for training. It also presumes that objects in street scenes exhibit spatial coherence and that the image quality provides sufficient visual cues for accurate pixelwise classification. To simulate how residents perceive the urban streetscape, street-view images were acquired via the Baidu Street View API. In each grid cell, four sampling points were placed at 125-m intervals and shifted to the nearest roads, ensuring the images captured relevant street sections. Moreover, because a single image cannot provide a complete panoramic view of each sampling point, images were taken from four angles (0°, 90°, 180°, and 270°), each with a 90° field of view. Due to data gaps in certain angles, a total of 5915 images were finally obtained. These images were segmented using the trained DeepLabV3+ model, generating 19 streetscape variables. This study focuses on six categories that are most closely related to ride-hailing demand: buildings, greenery, sky, roads, traffic control facilities (traffic lights and traffic signs), and sidewalks. The average proportion of each category within each grid cell is calculated to represent how street-level environmental features may influence travel demand.

3.2.4. Variable Description and Classification

In summary, this study incorporates ten dependent variables and three categories of explanatory variables.

Table 1. Definitions and categories of study variables.

Categories	Variable	Description
Socioeconomic and Transportation Accessibility Variables	GDP	GDP Density
	Population	Population Density
	Housing Price	Average housing price of the community
	City Center	Distance from the center point to the city center
	Road Mileage	Total mileage of arterial, major, minor roads, and urban branch roads
	Bus Stops	Number of bus stops
Land-Use Variables	Healthcare	$EF$ value for general hospitals, specialty hospitals, clinics, etc.
	Companies	$EF$ value for enterprises, companies, factories, etc.
	Education	$EF$ value for universities, high schools, libraries, etc.
	Shopping	$EF$ value for shopping centers, department stores, etc.
	Catering	$EF$ value for restaurants, snack bars, coffee houses, etc.
	Living	$EF$ value for post offices, courier logistics, laundromats, etc.
	Sports	$EF$ value for fitness centers, stadiums, swimming pools, etc.
	Auto Repair	$EF$ value for car sales, maintenance, car rentals, etc.
	Hotels	$EF$ value for star-rated hotels, chain hotels, aparthotels, etc.
	Tourism	$EF$ value for parks, zoos, botanical gardens, museums, etc.
	Entertainment	$EF$ value for cinemas, game centers, dance halls, KTV, etc.
	Finance	$EF$ value for banks, ATMs, financial companies, etc.
	Residential	$EF$ value for office buildings, residential areas, communities, etc.
	Transportation	$EF$ value for airports, train stations, bus stops, parking lots, etc.
Streetscape Feature Variables	Buildings	Proportion of buildings in streetscape images
	Greenery	Proportion of vegetation/green coverage in streetscape images
	Sky	Proportion of sky area in streetscape images
	Roads	Proportion of visible roads in streetscape images
	Traffic Control	Proportion of traffic lights, traffic signs, guardrails, etc.
	Sidewalks	Proportion of sidewalks in streetscape images

summarizes the definitions and categories of all variables considered in the empirical analysis.

3.3. Methodology

3.3.1. Variable Selection

Linear Correlation

Initially, 26 candidate variables were selected to explore the factors influencing ride-hailing service demand (Table 1). An important preprocessing step before model building was to remove variables that were highly linearly correlated with others to avoid overfitting. To achieve this, we employed bidirectional stepwise regression, which balances model explanation and complexity by combining forward selection and backward elimination strategies [96]. In this method, forward selection tests the significance of each unselected variable and adds those that meet the significance threshold, while backward elimination re-evaluates the significance of variables already in the model and removes those that no longer meet the criteria. By dynamically adjusting the variable set at each step, this method selects key variables from the 26 candidates, ensuring that the final model includes only the statistically significant variables. Additionally, Variance Inflation Factor ($VIF$) analysis was performed to reduce the effects of multicollinearity. Multicollinearity occurs when independent variables are strongly correlated with one another, potentially distorting the interpretation of other variables’ significance. The $VIF$ is used to measure the severity of multicollinearity, with higher values indicating stronger correlations with other variables. The formula for $VIF$ is calculated as in Equation (2):

$$VI{F}_k={\displaystyle \frac{1}{1-R_k^2}}$$

(2)

where $R_k^2$ is the regression coefficient of variable k on the remaining variables. The larger value of $VI{F}_i$ indicates the higher degree of correlation between variable i and the other variables. According to existing research, a $VIF$ value greater than 10 indicates significant multicollinearity, while a value below 5 suggests minimal concerns and is considered within a safe threshold.

Spatial Autocorrelation

In addition to linear correlation checks, another key issue in variable selection is the consideration of spatial autocorrelation, which is often overlooked in global models. Due to differences in spatial distribution of various functional zones within a city, human activities tend to exhibit significant spatial autocorrelation, meaning activity patterns in neighboring areas may show similarity or clustering effects. To assess spatial autocorrelation, this study applies global Moran’s I index. As a widely used method for detecting spatial autocorrelation, it helps to identify spatial dependencies and clustering phenomena across the study area. The formula for Moran’s I is calculated as follows in Equation (3):

$$I_k={\displaystyle \frac{G}{{\sum }_{m=1}^G{\sum }_{n=1}^G{w}_k^{m,n}}}\ast {\displaystyle \frac{{\sum }_{m=1}^G{\sum }_{n=1}^G{w}_k^{m,n}\left(x_k^m-\overline{x_k}\right)\left(x_k^n-\overline{x_k}\right)}{{\sum }_{m=1}^G{\left(x_k^m-\overline{x_k}\right)}^2}},\left(m\ne n\right)$$

(3)

where $I_k$ is the Moran’s I index of variable k, and $x_k^m$ and $x_k^n$ are the values of variable k at grid m and n; there are a total of G grids; $\overline{x_k}$ is the mean value of variable k at all grids, $w_k^{m,n}$ is the spatial weight between grids m and n, and I takes the value between $[-1,1]$.

Higher positive Moran’s I index values mean that close observations tend to have similar attribute values, while distant observations have different attribute values, which indicates spatial aggregation. However, a negative index indicates spatial dispersion, while an index near zero indicates a spatially random distribution. The null hypothesis of Moran’s I is that the variable is randomly distributed without spatial correlation; i.e., $I=0$. The Z-score is usually adopted as the significance index of the Moran index test, as shown in Equation (4):

$$Z\left(I\right)={\displaystyle \frac{I-E\left(I\right)}{\sqrt{Var\left(I\right)}}}$$

(4)

where $E\left(I\right)$ and $Var\left(I\right)$ are the expectation and the standard deviation of the Moran’s I statistic, respectively. The significance level in this study is set as $p<0.05$, so the hypothesis was rejected when $\left|Z\right|>1.96(p<0.05)$, indicating significant spatial autocorrelation.

3.3.2. Global Models

As discussed in the literature review, several spatial regression models have been applied to study the relationship between ride-hailing demand and influencing factors, including Ordinary Least Squares (OLSs), Spatial Error Model (SEM), and Spatial Lag Model (SLM). These models share the common feature that their regression coefficients are fixed globally, meaning that all observations (or spatial units) share the same coefficient. In other words, these models assume that the relationship between the dependent and independent variables is consistent across the study area, with the influence of the independent variables on the dependent variable being the same regardless of location.

Ordinary Least Squares

In the OLS model, regression parameters are estimated using the least-squares method. OLS assumes that the relationship between independent variables and the dependent variable is uniform across the study area, and the regression coefficients represent the average effect across the entire region [60], which can be denoted in Equation (5):

$$y={\beta }_0+X\beta +\epsilon$$

(5)

where y is the vector of the dependent variable, X is a matrix of independent variables, ${\beta }_0$ represents the intercept of the regression model, $\beta$ presents the vector of the regression coefficients, and $\epsilon$ is the vector of random errors.

However, since OLS does not account for the influence of neighboring regions, when spatial autocorrelation is present in the data, OLS may lead to inaccurate estimates. Therefore, in addition to the OLS model, this study applies SEM and SLM to account for spatial dependencies and improve the accuracy of the estimates.

Spatial Error Model

SEM assumes that error terms exhibit spatial dependence, meaning that the error term of one observation can influence the error terms of neighboring observations [28]. To reflect this, SEM introduces spatially autocorrelated error terms into the regression model, which helps to capture the spatial correlations between the errors, thus improving the estimation precision. The basic form of SEM can be expressed in Equation (6):

$$y={\beta }_0+X\beta +\mu ,\mu =\lambda W\mu +\epsilon$$

(6)

where $\mu$ is the vector of spatially correlated error terms, referring to those similar urban environmental characteristics hidden between adjacent neighborhoods, in which $\lambda$ is the spatial error coefficient and W is the spatial weight matrix obtained based on queen contiguity.

Spatial Lag Model

In contrast to SEM, SLM assumes that the impact of independent variables on an observation has a lag effect. Specifically, the independent variables at a given observation not only influence that observation’s dependent variable but may also affect neighboring observations. To model this, SLM introduces a spatially lagged dependent variable, linking the dependent variable to neighboring areas. This method effectively captures spatial dependencies, which can be denoted as in Equation (7):

$$y={\beta }_0+\rho Wy+X\beta +ϵ$$

(7)

where $\rho$ is the spatial lag coefficient. This formulation allows the model to capture the effect of spatial autocorrelation by directly including the influence of neighboring values of the dependent variable on each observation, with the coefficient $\rho$ quantifying the strength and direction of this spatial dependence.

3.3.3. Local Models

Global models provide a uniform estimate across the study area, ignoring potential local variations. However, cities typically have different functional features across regions, and the relationship between dependent and independent variables may change depending on spatial location. This phenomenon, known as spatial heterogeneity, contradicts the assumptions of global models. To more accurately capture these local features, we employ local regression models that allow parameters to vary spatially, thereby identifying spatial differences.

GWR Model

To overcome the limitation of global models, where regression coefficients only represent the global average, the Geographic Weighted Regression (GWR) model assigns geographic weights (typically distance-based) to each observation [72]. This allows GWR to estimate local regression coefficients that reflect the characteristics of the specific observation and its neighboring areas, thus more precisely capturing the local distribution of variables [88]. The GWR takes the following form as in Equation (8):

$$y(u,v)={\beta }_0(u,v)+\sum _k{\beta }_k(u,v)x_k(u,v)+\epsilon (u,v)$$

(8)

where $y(u,v)$ represents the dependent variable at the shperical coordinates $(u,v)$, ${\beta }_0(u,v)$ serves as the constant term, $x_k(u,v)$ denotes the explanatory variable k with ${\beta }_k(u,v)$ as the corresponding local regression coefficient for each variable, and $\epsilon (u,v)$ the error terms at the point. The local coefficients $\beta (u,v)$ are estimated by applying a weighted-least-squares procedure at each location as in Equation (9):

$$\widehat{\beta }(u,v)={\left[X{(u,v)}^{T}W(u,v)X(u,v)\right]}^{-1}X{(u,v)}^{T}W(u,v)y$$

(9)

where $X(u,v)$ is the local design matrix with each row corresponding to an observation’s explanatory variables, $W(u,v)$ is a diagonal matrix containing spatial weights $w^h(u,v)$ for each observation, and y is the vector of observed dependent variable values. This weighted-least-squares solution ensures that observations closer to the location $(u,v)$ have a higher influence on the parameter estimates, and the estimation begins with the minimization of the local weighted sum of squared residuals. The weights $w^h(u,v)$ determine the influence of each observation on the local regression. Using an adaptive bisquare kernel, the weight for the hth observation is defined as in Equation (10):

$$w^h(u,v)=\left\{\begin{array}{cc}{\left[1-{\left({\displaystyle \frac{d((u,v),(u^h,v^h))}{b{w}^{\ast }}}\right)}^2\right]}^2,\hfill & d((u,v),(u^h,v^h))\le b{w}^{\ast },\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(10)

where $d((u,v),(u_h,v_h))$ represents the distance between the target location $(u,v)$ and the hth observation’s location $(u^h,v^h)$, and $b{w}^{\ast }$ is the optimal bandwidth that controls the spatial extent over which data are used to estimate local parameters. The optimal bandwidth $b{w}^{\ast }$ is typically determined by minimizing the corrected Akaike Information Criterion (AICc) using the Golden Section Search.

MGWR Model

It is important to note that, in GWR, all the explanatory variables are assigned the same bandwidth parameter, meaning that the spatial influence range is kept consistent across all variables in the modeling process [84]. However, this assumption may not be valid as different variables might have varying spatial effects. To address this limitation, we use the Multi-Scale Geographic Weighted Regression (MGWR) model, which allows different bandwidths for each explanatory variable to better capture the spatial heterogeneity of variable impacts [75,87], and the model can be mathematically expressed as in Equation (11):

$$y(u,v)={\beta }_0(u,v)+\sum _k{\beta }_k^{b{w}_k^{\ast }}(u,v)x_k(u,v)+\epsilon (u,v)$$

(11)

where ${\beta }_{b{w}_k^{\ast }}^k(u,v)$ is the regression coefficient of the explanatory variable k at location $(u,v)$ based on a specific bandwidth $b{w}_k^{\ast }$, which is used in determining the regression coefficient of the explanatory variable k. In the GWR framework, the local coefficients are estimated using a locally weighted-least-squares approach, and the weights for all the explanatory variables are derived from a single bandwidth. In contrast, each explanatory variable in MGWR model is permitted its own optimal bandwidth $b{w}_k^{\ast }$ and constructs the variable-specific weight matrices. These individual weight matrices are then combined to form the overal weight matrix $W(u,v)$ used in the local parameter estimation as in Equation (9). For the hth observation at location, the weight function of the explanatory variable k in MGWR is defined as in Equation (12):

$$w_k^h(u,v)=\left\{\begin{array}{cc}{\left[1-{\left({\displaystyle \frac{d((u,v),(u^h,v^h))}{b{w}_k^{\ast }}}\right)}^2\right]}^2,\hfill & d((u,v),(u^h,v^h))\le b{w}_k^{\ast },\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(12)

where the optimal bandwidth $b{w}_k^{\ast }$ for the variable k is determined using the Golden Section Search, just as in GWR, to minimize the $AIC{c}_k\left(b{w}_k\right)$.

To operationalize the MGWR estimation, we employ a back-fitting procedure in which the optimal bandwidth for each explanatory variable is iteratively updated using Golden Section Search to minimize the corrected $AIC{c}_k$, as in GWR. This process is summarized in Algorithm 1:

Algorithm 1: Back-Fitting Procedure for MGWR Bandwidth Optimization

This method allows each variable to have a distinct spatial influence range, adapting more effectively to the spatial scale of relationships.

4. Results

4.1. Temporal and Spatial Pattern of Ride-Hailing Services in Haikou’s Urban Area

This section analyzes the temporal variation and spatial distribution of ride-hailing demand in Haikou. Figure 2 presents (a) the average number of trips and (b) the average travel duration of ride-hailing in Haikou. The x-axis, shared by both panels, represents the time of day. A distinct morning peak occurs between 7:00 and 8:00, followed by the highest daily volume between 17:00 and 18:00. The morning peak is more pronounced on workdays, whereas holidays tend to show higher average hourly trip volumes during non-peak periods. In contrast, workday peaks are associated with longer travel durations, forming a clear bimodal pattern around 20 min, indicative of longer commutes and potential traffic congestion.

Figure 3 presents the spatial distribution of ride-hailing demand in Haikou’s urban area across various time periods, revealing distinct workday and holiday patterns. Overall, the downtown core consistently registers higher volumes of orders compared with the peripheral and southern districts, where demand remains relatively modest.

On workdays, morning trips predominantly originate in residential neighborhoods and flow toward the central employment zones, especially along major arteries and within the urban core. As midday approaches, the focal points of activity shift toward shopping districts, schools, hospitals, and large commercial centers (Wanxiang City, No. 1 High School, and the Hainan Provincial Maternal and Child Health Hospital), signaling an increase in errands and service-related trips. Moving into the evening peak, commuting demand intensifies and gradually extends southeast, yet the core area retains notably high activity levels. Following this peak, demand subsides substantially; however, select central districts and transport hubs continue to see moderate traffic late into the night. By contrast, holidays exhibit a milder and later onset of the morning peak, with a more dispersed pattern of midday and evening trips. During these periods, popular shopping areas, tourist attractions, and recreational venues experience substantial increases in demand (the Hunan Provincial Museum and Fengxiang Wetland Park), forming a multi-centered spatial structure, and the evening peak often persists into the late-night hours, underscoring an elevated level of nightlife activity.

4.2. Variable Selection Results

In the model construction process, a combination of bidirectional stepwise regression and VIF tests were used to dynamically optimize the significance of the variables and address multicollinearity issues [96]. For all 10 dependent variables across the different time periods, significant variables were selected, as shown in

Table 2. Selected significant variables through bidirectional stepwise regression and VIF tests.

Variable	Workday										Holiday
	Morning Peak		Morning Day		Evening Peak		Evening		Overnight		Morning Peak		Morning Day		Evening Peak		Evening		Overnight
	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF	Coeff.	VIF
Population	—	—	—	—	—	—	—	—	—	—	0.404 **	1.482	0.570 **	1.458	—	—	—	—	—	—
Housing Price	—	—	−0.503 *	1.114	—	—	−0.391	1.115	−0.130 **	1.124	—	—	−0.480 *	1.119	—	—	—	—	−0.163 **	1.124
City Center	−0.853 ***	1.341	−0.926 ***	1.401	−1.257 ***	1.250	−1.322 ***	1.313	−0.191 ***	1.395	−0.615 ***	1.357	−1.041 ***	1.421	−1.793 ***	1.341	−1.659 ***	1.340	−0.248 ***	1.395
Road Mileage	0.648 **	1.999	—	—	—	—	—	—	0.214 ***	2.031	—	—	—	—	—	—	—	—	0.280 ***	2.031
Bus Stops	1.663 ***	1.906	3.191 ***	1.560	3.589 ***	1.652	3.025 ***	1.652	0.458 ***	1.834	1.667 ***	1.469	3.312 ***	1.447	4.334 ***	1.652	3.686 ***	1.447	0.569 ***	1.834
Finance	—	—	0.868 ***	1.229	1.210 ***	1.223	0.994 ***	1.244	0.142 **	1.203	—	—	0.517 **	1.202	0.902 **	1.220	1.122 ***	1.163	0.192 **	1.203
Auto Repair	—	—	−0.455 *	1.032	−0.553 *	1.032	−0.524 *	1.034	−0.093 *	1.030	—	—	−0.441 *	1.033	−0.592 *	1.032	−0.508	1.025	−0.120 *	1.030
Buildings	1.935 ***	2.026	1.928 ***	1.707	—	—	—	—	0.461 ***	1.671	1.678 ***	1.817	1.547 ***	1.889	1.100 *	2.604	1.098 **	1.635	0.473 ***	1.671
Greenery	0.499 *	1.917	0.452	1.578	−1.407 ***	2.863	−1.074 **	2.870	—	—	0.381 **	1.594	—	—	—	—	—	—	—	—
Sky	−1.283 ***	1.540	−1.375 ***	1.363	−4.046 ***	3.121	−2.967 ***	3.165	−0.327 ***	1.404	−0.618 ***	1.259	−1.083 ***	1.221	−2.728 ***	3.621	−1.103 ***	1.219	−0.402 ***	1.404
Roads	—	—	—	—	5.248 ***	7.905	3.606 ***	8.034	—	—	—	—	—	—	2.120 **	6.748	—	—	—	—
Traffic Control	0.541 **	1.420	0.953 ***	1.408	0.892 **	1.451	0.716 **	1.451	0.130 *	1.421	—	—	0.597 **	1.415	0.847 *	1.423	0.820 **	1.407	0.150 *	1.421
Sidewalks	0.498 *	2.319	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—

*** Significant at 0.001 level; ** Significant at 0.01 level; * Significant at 0.05 level.

. Most of the retained explanatory variables were statistically significant, and all the VIF values were below 10, with the majority even below 5, indicating low multicollinearity between the variables and ensuring the robustness and interpretability of the model [60]. The selected variables were then input into subsequent models, ensuring comparability across different model results. The analysis results show that there are significant differences in user travel demand across different time periods, and the significant explanatory variables associated with travel demand also vary by time period. Overall, the selected significant variables effectively reflect the user demand for ride-hailing services across the ten periods.

By analyzing the relationship between travel demand and variables using linear regression models, the results showed that the number of significant variables on workdays was eight in the morning peak, nine during the day, eight in the evening peak, nine at night, and nine during the overnight period. During holidays, the numbers of significant variables were six, nine, eight, seven, and nine, respectively. Specifically, in the morning peak on workdays, road mileage, the number of bus stations, the proportion of buildings, greenery, traffic control facilities, and sidewalks derived from street-view images showed a significant positive correlation with the use of ride-hailing services, while the distance to the city center and the proportion of sky showed a significant negative correlation. This suggests that the demand for travel during the morning peak is concentrated in areas with higher traffic convenience, larger building densities, and better greenery and walking facilities, which aligns with commuting-focused travel characteristics. Over time, the correlation with road mileage, bus stations, and traffic control facilities gradually strengthened, and the regression coefficient for financial-related places became significant, positively correlating with ride-hailing service usage. Meanwhile, the correlation with distance to the city center, the number of auto repair shops, greenery ratio, and sky ratio weakened, with the direction of some variable correlations even changing (e.g., the greenery ratio shifted from a positive to a negative correlation). During the evening peak, the concentration of traffic facilities and the regression coefficient for financial-related places peaked, indicating that, at this time, travel demand is more concentrated in the central urban area. Work-related traffic facilities and places have a significant positive impact on ride-hailing service use, while greenery and sky ratios significantly suppress travel demand. During the night and overnight periods, the correlation with traffic facilities and financial-related places began to weaken, while the correlation with the number of auto repair shops, greenery ratio, and sky ratio gradually strengthened, with a reduced demand for commuting. Further analysis also revealed that the effect of housing prices was not significant during the morning and evening peaks but showed some significance during the day, night, and overnight periods, reflecting the temporal dependency of housing prices on travel demand.

4.3. Results of Spatial Autocorrelation Test

While the linear regression model highlights the relationships between variables and ride-hailing demand, its limitations must be acknowledged. First, the linear regression model assumes independence between observations without accounting for the spatial aggregation or diffusion properties of the dependent variable (i.e., spatial autocorrelation) [53]. In reality, ride-hailing demand is influenced not only by the characteristics of the current region but also by potential spillover effects from neighboring areas, which might show spatial dependence. Second, the model assumes that the regression coefficients are consistent across the entire study area, yet the influence of variables often varies significantly by spatial location. For instance, the proportion of greenery may promote travel demand in suburban areas, while it may suppress demand in central urban areas. Given these limitations, it is necessary to further analyze the residuals of the model to assess its ability to capture spatial features.

Before examining the spatial autocorrelation of the residuals, this study also measured the global Moran’s I values of the main independent variables to assess their spatial distribution patterns. The results in

Table 3. Results of Moran’s I test for selected variables.

Variable	Moran’s Index	Z-Score
Population	0.665	31.559 ***
Housing Price	0.878	41.208 ***
City Center	0.983	46.047 ***
Road Mileage	0.351	16.533 ***
Bus Stops	0.333	15.723 ***
Finance	0.286	13.593 ***
Auto Repair	0.203	9.679 ***
Buildings	0.573	26.911 ***
Greenery	0.481	22.571 ***
Sky	0.303	14.263 ***
Roads	0.614	28.783 ***
Traffic Control	0.246	11.697 ***
Sidewalks	0.433	20.339 ***

*** Significant at 0.001 level.

indicate that, at the 0.1% significance level (p < 0.001), all the regional characteristic variables exhibit significant positive spatial autocorrelations, as measured by Moran’s I. This suggests that their spatial distributions deviate markedly from randomness and display clear patterns of positive spatial clustering. When the independent variables themselves have strong spatial dependence, a simple OLS regression may fail to fully capture potential neighborhood effects and spatial spillover processes [64], thereby possibly exacerbating the spatial autocorrelation of the residuals.

To test the spatial autocorrelation of the residuals in the linear regression model, this study employed the global Moran’s I statistic. The results in

Table 4. Results of Moran’s I test for models’ residuals.

Time Period	Moran’s Index			Z-Score
	OLS	SEM	SLM	OLS	SEM	SLM
Workday
Morning Peak	0.179	0.007	0.018	8.437 ***	0.425	0.899
Morning Day	0.160	−0.008	−0.020	7.569 ***	−0.318	−0.865
Evening Peak	0.183	−0.005	0.001	8.635 ***	−0.159	0.111
Evening	0.142	−0.001	−0.006	6.737 ***	0.040	−0.183
Overnight	0.254	0.006	0.016	11.972 ***	0.343	0.843
Holiday
Morning Peak	0.144	0.003	0.004	6.821 ***	0.221	0.261
Morning Day	0.118	−0.002	−0.014	5.618 ***	−0.014	−0.600
Evening Peak	0.109	0.000	−0.009	5.158 ***	0.057	−0.338
Evening	0.105	0.001	−0.008	5.033 ***	0.108	−0.286
Overnight	0.284	0.006	0.015	13.388 ***	0.342	0.792

*** Significant at 0.001 level.

show that the Moran’s I values for the OLS residuals are generally high across all ten time periods, and the Z-values are highly significant. This indicates strong spatial autocorrelation in the residuals, meaning that the OLS model fails to fully account for the spatial characteristics of the dependent variable. Particularly during the morning and evening peak hours on workdays, the residual Moran’s I values display a distinct "bimodal" pattern, reflecting a stronger spatial clustering effect. During the overnight periods, whether on workdays or holidays, the Moran’s I values reach their highest point, suggesting that ride-hailing demand in these time periods is highly concentrated in specific areas.

Spatial regression models are applied to provide a more comprehensive explanation of the spatial characteristics of ride-hailing demand and to address the spatial autocorrelation present in the residuals of the OLS model [28]. Specifically, SEM and SLM are employed to analyze spatial diffusion effects and spatial error dependencies in demand. The results reveal that both SEM and SLM significantly reduce the Moran’s I values to near zero in most time periods, with the Z-values becoming non-significant, indicating effective control of spatial autocorrelation. However, these models still do not address the second limitation of the OLS model—the assumption of globally consistent regression coefficients. This assumption fails to reflect the significant spatial heterogeneity that may exist in the factors affecting ride-hailing demand. Therefore, we employ local regression models that allow parameters to vary spatially, thereby identifying spatial differences.

4.4. Model Performance Comparison

To thoroughly evaluate the explanatory power of different models and the added value of incorporating street-view features, this study compares five models: three global models (OLS, SEM, and SLM) and two local models (GWR and MGWR). Model performance is assessed using three metrics: goodness of fit (R²), the Akaike Information Criterion (AIC), and log-likelihood. Following [80,87], lower AIC and higher log-likelihood values indicate a better model fit, while higher R² values reflect stronger explanatory power.

To assess the contribution of street-view imagery, we further re-estimated each model excluding all the street-level features and report the corresponding results in parentheses in

Table 5. Comparison of metrics among models.

Metrics	Models	Workday					Holiday
		Morning Peak	Morning Day	Evening Peak	Evening	Overnight	Morning Peak	Morning Day	Evening Peak	Evening	Overnight
R2	OLS	0.631 (0.478)	0.595 (0.516)	0.585 (0.508)	0.550 (0.493)	0.525 (0.430)	0.647 (0.542)	0.598 (0.548)	0.535 (0.484)	0.453 (0.425)	0.493 (0.416)
	SEM	0.621 (0.437)	0.583 (0.483)	0.574 (0.480)	0.543 (0.477)	0.503 (0.391)	0.644 (0.520)	0.593 (0.539)	0.532 (0.475)	0.450 (0.419)	0.464 (0.377)
	SLM	0.700 (0.662)	0.667 (0.646)	0.659 (0.640)	0.609 (0.592)	0.653 (0.632)	0.695 (0.659)	0.644 (0.626)	0.575 (0.560)	0.498 (0.488)	0.645 (0.631)
	GWR	0.684 (0.498)	0.714 (0.644)	0.700 (0.642)	0.692 (0.661)	0.695 (0.646)	0.684 (0.558)	0.711 (0.696)	0.708 (0.645)	0.637 (0.604)	0.672 (0.642)
	MGWR	0.708 (0.670)	0.730 (0.720)	0.743 (0.718)	0.743 (0.636)	0.739 (0.677)	0.753 (0.685)	0.734 (0.707)	0.767 (0.640)	0.699 (0.583)	0.721 (0.695)
AIC	OLS	3249.423 (3451.825)	3659.376 (3760.662)	3895.186 (3991.313)	3733.924 (3798.738)	1810.895 (1916.714)	2886.347 (3040.225)	3570.258 (3635.601)	4100.249 (4156.637)	4007.649 (4032.723)	2098.582 (2179.379)
	SEM	3176.105 (3272.187)	3598.441 (3642.488)	3823.250 (3867.748)	3688.120 (3716.736)	1677.416 (1725.665)	2841.113 (2929.767)	3536.945 (3566.437)	4073.787 (4098.941)	3982.385 (3990.883)	1942.843 (1974.013)
	SLM	3150.248 (3238.794)	3564.797 (3607.279)	3801.391 (3837.524)	3669.210 (3694.423)	1660.456 (1706.378)	2817.123 (2897.291)	3513.789 (3545.224)	4060.329 (4081.269)	3970.760 (3980.266)	1928.569 (1959.647)
	GWR	3227.244 (3453.368)	3601.558 (3718.567)	3831.804 (3933.904)	3653.714 (3719.302)	1694.277 (1815.345)	2877.098 (3040.669)	3517.803 (3581.691)	3991.538 (4074.240)	3914.372 (3950.603)	1987.015 (2070.072)
	MGWR	3194.350 (3252.818)	3538.801 (3547.899)	3757.113 (3775.209)	3575.447 (3667.173)	1616.246 (1746.544)	2824.211 (2894.910)	3446.273 (3476.958)	3893.158 (4026.761)	3829.960 (3921.047)	1903.229 (1905.268)
Log-likelihood	OLS	−1615.712 (−1721.912)	−1819.688 (−1874.330)	−1938.593 (−1990.657)	−1856.962 (−1893.369)	−895.448 (−951.357)	−1436.174 (−1516.113)	−1775.129 (−1810.800)	−2041.125 (−2073.318)	−1995.825 (−2011.361)	−1039.291 (−1082.690)
	SEM	−1579.052 (−1632.094)	−1789.220 (−1815.244)	−1902.625 (−1928.874)	−1834.060 (−1852.368)	−828.708 (−855.832)	−1413.556 (−1460.884)	−1758.472 (−1776.218)	−2027.893 (−2044.471)	−1983.192 (−1990.442)	−961.422 (−980.006)
	SLM	−1565.124 (−1614.397)	−1771.398 (−1796.639)	−1890.695 (−1912.762)	−1823.605 (−1840.212)	−819.228 (−845.189)	−1400.562 (−1443.645)	−1745.895 (−1764.612)	−2020.165 (−2034.634)	−1976.380 (−1984.133)	−953.284 (−971.824)
	GWR	−1568.097 (−1709.922)	−1713.791 (−1780.320)	−1839.837 (−1893.544)	−1740.765 (−1770.456)	−760.038 (−805.903)	−1402.475 (−1505.392)	−1674.759 (−1689.950)	−1898.747 (−1958.875)	−1870.498 (−1897.056)	−906.407 (−933.267)
	MGWR	−1544.608 (−1581.570)	−1695.489 (−1707.230)	−1792.590 (−1820.405)	−1685.285 (−1792.032)	−713.168 (−777.449)	−1327.596 (−1401.283)	−1649.221 (−1678.633)	−1829.961 (−1963.374)	−1813.594 (−1912.874)	−856.773 (−884.729)

Note: Values in parentheses indicate model results without street-view features.

. This side-by-side comparison demonstrates the incremental improvement brought by the inclusion of street-level visual attributes in modeling spatial variations in ride-hailing demand.

4.4.1. Global Model Results

The results in Table 5 show that the OLS model has the highest AIC and log-likelihood values in all the time periods, indicating that it is the poorest-performing model. Additionally, its R² value is significantly lower than that of all the other models except SEM, suggesting that OLS fails to effectively capture the spatial characteristics of ride-hailing demand. Among the global spatial models, SLM outperforms SEM by significantly improving the R² value, with an increase of up to 39% during the overnight period on holidays, indicating that the lag model is more appropriate for addressing spatial autocorrelation issues in this case. This conclusion is also supported by the results of the Lagrange multiplier (LM) test in

Table 6. Results of Lagrange multiplier (LM) diagnostics.

Time Period	LM-Error		LM-Lag		RLM-Error		RLM-Lag
	Value	p	Value	p	Value	p	Value	p
Workday
Morning Peak	69.079	0.000	110.439	0.000	0.842	0.3590	42.201	0.000
Morning Day	55.470	0.000	113.420	0.000	0.981	0.3221	58.931	0.000
Evening Peak	72.385	0.000	112.630	0.000	0.363	0.5468	40.608	0.000
Evening	43.843	0.000	77.088	0.000	0.457	0.4991	33.702	0.000
Overnight	139.838	0.000	182.903	0.000	3.042	0.0811	46.107	0.000
Holiday
Morning Peak	44.950	0.000	79.194	0.000	0.056	0.8126	34.300	0.000
Morning Day	30.343	0.000	66.847	0.000	1.365	0.2426	37.869	0.000
Evening Peak	25.515	0.000	47.373	0.000	0.874	0.3498	22.733	0.000
Evening	24.279	0.000	43.471	0.000	1.324	0.2498	20.515	0.000
Overnight	165.068	0.000	209.493	0.000	3.276	0.0703	47.700	0.000

, which includes LM-Error (spatial error test), LM-Lag (spatial lag test), and the corresponding Robust LM values, along with their significance (p-values). The LM-Error test is used to detect spatial correlation in the error terms [90]. If significant, it suggests that unobserved factors in the model cause neighboring regions’ error terms to influence each other, indicating that SEM should be used for adjustment. On the other hand, the LM-Lag test is used to determine whether the dependent variable exhibits spatial lag effects. If significant, it indicates that ride-hailing demand in one area is influenced not only by its own factors but also by the demand in neighboring areas, making SLM the more appropriate model.

In Table 6, the LM-Error and LM-Lag tests are significant (p < 0.001) across all the time periods, suggesting that ride-hailing demand exhibits spatial dependence and requires spatial regression models. However, the Robust LM-Error test is not significant (p > 0.05) in any of the time periods, while the Robust LM-Lag test is significant (p < 0.001) across all the periods. This indicates that the spatial correlation in ride-hailing demand primarily arises from the spatial lag of the dependent variable rather than the spatial correlation of the error terms. In other words, ride-hailing demand is influenced not only by local factors but also by spillover effects from neighboring areas, and this spatial effect cannot be captured by OLS. Therefore, SLM is more appropriate for fitting this demand. Additionally, spatial effects are generally stronger on workdays than on holidays, particularly during the overnight period, where the LM-Error and LM-Lag values are the highest (workdays: 139.838 and 182.903; holidays: 165.068 and 209.493), indicating that ride-hailing demand is highly concentrated during this time, with significant interaction between regions. This may be due to the imbalance of ride-hailing supply and demand during late-night hours, making the demand highly dependent on neighboring areas’ spillover effects. In contrast, during daytime and evening peak hours, spatial effects are still significant, but the spatial correlation is weaker on holidays, indicating that holiday travel patterns are more dispersed, with less impact between regions. Therefore, during workday peak hours and overnight periods, the focus should be on the spillover effects between regions, while, on holidays, due to more random demand distribution, a more nuanced analysis considering other spatial factors is needed.

4.4.2. Local Model Results

Although SLM effectively addresses the spatial autocorrelation issue in ride-hailing demand and significantly improves model fit, it assumes consistent regression coefficients across the entire study area, which fails to capture the spatial heterogeneity of variable effects between regions. In reality, ride-hailing demand is influenced by several factors, including population density, commercial activities, and public transportation accessibility, which may have significantly different effects in different spatial locations. This spatial heterogeneity cannot be accurately captured by global models such as SLM, so local models are necessary to analyze the spatial variation in the influence of variables, thereby improving the model’s explanatory power and prediction accuracy.

Compared to the global models, the GWR and MGWR models outperform in terms of goodness of fit and spatial adaptability. In all the time periods, the R² values for MGWR are higher than for both GWR and SLM. For instance, during the holiday evening peak period, the R² for MGWR reaches 0.767, significantly higher than GWR (0.708) and SLM (0.575), indicating that MGWR is better at capturing spatial heterogeneity in variable effects. Additionally, the R² improvement for MGWR during the overnight period is especially significant workdays: 0.739; holidays: 0.721), further indicating that this model provides stronger explanatory power during periods of complex demand distribution. Regarding model selection metrics, the results for the AIC and log-likelihood further confirm these findings. The AIC for the OLS model is the highest in all the periods, indicating poor performance in balancing goodness of fit and model complexity. For instance, in the workday overnight period, the AIC for OLS is 1810.895, while, for SLM and MGWR, it drops to 1660.456 and 1616.246, respectively, with the latter being the best among all the models.

Furthermore, the log-likelihood for MGWR is the highest in all the periods, such as −1327.596 during the holiday morning peak, higher than GWR (−1402.475) and SLM (−1400.562), suggesting that MGWR not only significantly improves model fit but also demonstrates greater robustness in accounting for model complexity. Additionally, the impact of time periods on model performance shows significant differences. During peak hours, the spatial autocorrelation in ride-hailing demand is stronger, and models like SLM, GWR, and MGWR perform better, while OLS is unable to capture these features effectively. Particularly during the overnight period, whether on workdays or holidays, the spatial concentration of demand is most significant, further highlighting the advantages of the MGWR model, which shows optimal performance in the AIC and log-likelihood. This suggests that MGWR has a significant advantage in capturing the spatial heterogeneity of variable effects across different regions, especially during periods of complex and fluctuating demand distribution.

For local models like GWR and MGWR, bandwidth is an important indicator of the spatial effect range of variables [73]. GWR uses a single bandwidth, meaning all the variables share the same spatial scale, while MGWR allows different variables to have different bandwidths, making it more flexible in capturing spatial heterogeneity. In this study, by comparing the bandwidths of MGWR and GWR, we further explore the spatial variation in the impact of variables and analyze the characteristics of ride-hailing demand in different time periods.

Table 7. Comparison of variable bandwidths between GWR and MGWR.

Variable	Workday										Holiday
	Morning Peak		Morning Day		Evening Peak		Evening		Overnight		Morning Peak		Morning Day		Evening Peak		Evening		Overnight
	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR	MGWR	GWR
Population	—	—	—	—	—	—	—	—	—	—	61	305	418	187	—	—	—	—	—	—
Housing Price	—	—	602	182	—	—	276	184	609	181	—	—	267	187	—	—	—	—	604	181
City Center	601	305	604	182	496	189	281	184	610	181	601	305	267	187	195	151	195	151	610	181
Road Mileage	610	305	—	—	—	—	—	—	125	181	—	—	—	—	—	—	—	—	125	181
Bus Stops	610	305	54	182	54	189	54	184	49	181	54	305	54	187	54	151	54	151	49	181
Finance	—	—	102	182	76	189	76	184	115	181	—	—	151	187	76	151	76	151	114	181
Auto Repair	—	—	610	182	610	189	610	184	610	181	—	—	610	187	421	151	610	151	610	181
Buildings	597	305	256	182	—	—	—	—	125	181	213	305	247	187	159	151	182	151	132	181
Greenery	418	305	353	182	343	189	594	184	—	—	310	305	—	—	—	—	—	—	—
Sky	511	305	421	182	506	189	506	184	420	181	540	305	599	187	588	151	574	151	416	181
Roads	—	—	—	—	604	189	311	184	—	—	—	—	—	—	579	151	—	—	—
Traffic Control	238	305	189	182	182	189	189	184	353	181	—	—	193	187	76	151	193	151	353	181
Sidewalks	93	305	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—

shows that the bandwidth for GWR is mostly between 181 and 305, indicating that a larger smoothing window is applied to all the variables, averaging their spatial variation and potentially ignoring some spatial heterogeneity. In contrast, the bandwidth distribution for MGWR is more diverse, ranging from 49 to 610, indicating significant differences in the spatial effects of variables at different scales. This is the key reason why MGWR outperforms GWR in capturing spatial heterogeneity. Examining the variables themselves, there are significant differences in their spatial effects. For example, the bandwidth for bus stop density is consistently small (MGWR: 49–54), indicating that its impact is limited to a smaller range. Ride-hailing primarily serves as a short-distance connector, complementing surrounding public transportation facilities. The bandwidth for distance from the city center reaches 610 during overnight periods, indicating that late-night travel has a broader spatial impact and is more influenced by the overall urban structure than by local facilities. Additionally, the bandwidth for financial POIs is smaller during workdays (76–115) but significantly expands during holidays, suggesting that financial districts may have a greater impact on a wider range of travel due to commercial and leisure activities.

At the same time, the spatial effect range of ride-hailing demand is dynamic. During peak periods, spatial dependency is more localized, with smaller MGWR bandwidths for variables such as road and building density, reflecting the direct influence of local conditions on commuting demand. In the overnight period, spatial diffusion effects become more prominent, with larger MGWR bandwidths, such as for distance from the city center and buildings, indicating that late-night demand is more influenced by the overall urban space structure rather than local factors. Additionally, holiday demand is more dispersed, and the spatial effect range of variables is larger than during workdays. For instance, the bandwidth for greenery rate and building density is larger during holidays, suggesting that holiday travel patterns are influenced more by leisure and living environment factors than by commuting needs. MGWR dynamically adjusts the spatial effect range of different variables, providing a more accurate depiction of the variations in ride-hailing demand across regions and times.

4.5. Local Estimates in the MGWR Model

In this section, we analyze the significant explanatory variables of ride-hailing demand across different time periods using the MGWR model. The results are presented in

Table 8. Descriptive statistics of variable coefficients during peak hours.

Variable	Workday										Holiday
	Morning Peak					Evening Peak					Morning Peak					Evening Peak
	Mean	Std	Min	Median	Max	Mean	Std	Min	Median	Max	Mean	Std	Min	Median	Max	Mean	Std	Min	Median	Max
Population	–	–	–	–	–	–	–	–	–	–	0.304	0.111	0.119	0.322	0.549	–	–	–	–	–
City Center	−1.091	0.017	−1.106	−1.101	−1.052	−1.852	0.032	−1.898	−1.857	−1.793	−0.684	0.018	−0.712	−0.686	−0.652	−2.745	0.665	−4.803	−2.597	−1.895
Road Mileage	0.382	0.029	0.310	0.388	0.439	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–
Bus Stops	1.645	0.031	1.595	1.644	1.707	2.426	1.202	0.711	2.303	5.261	1.563	0.164	1.282	1.542	1.832	2.546	2.292	-1.124	1.986	9.252
Finance	–	–	–	–	–	0.810	0.702	−0.259	0.646	2.450	–	–	–	–	–	0.798	0.421	0.198	0.814	1.606
Auto Repair	–	–	–	–	–	−0.258	0.071	−0.410	−0.246	−0.110	–	–	–	–	–	−0.396	0.396	−2.531	−0.095	0.338
Building	1.391	0.236	0.923	1.379	1.869	–	–	–	–	–	1.318	0.421	0.455	1.384	2.155	0.569	0.075	0.440	0.583	0.691
Greenery	0.409	1.086	−1.991	0.124	4.504	−0.447	1.541	−3.985	−0.650	7.374	0.413	0.739	−1.323	0.287	3.002	–	–	–	–	–
Sky	−0.972	0.425	−2.040	−0.895	−0.426	−2.775	0.062	−2.864	−2.785	−2.659	−0.518	0.019	−0.563	−0.518	−0.474	−2.524	0.070	−2.652	−2.518	−2.412
Roads	–	–	–	–	–	3.838	0.036	3.760	3.839	3.911	–	–	–	–	–	2.801	0.065	2.686	2.798	2.911
Traffic Control	0.391	0.224	0.068	0.353	1.039	0.641	1.285	−1.251	0.222	5.784	–	–	–	–	–	0.768	0.768	0.674	0.772	0.861
Sidewalks	0.393	1.002	−1.301	0.277	3.591	–	–	–	–	–	–	–	–	–	–	–	–	–	–	–

, summarizing the distribution and variability of the parameter estimates across the time periods. We then visualize the spatial distribution of these coefficients, focusing on the morning and evening peak hours of workdays and holidays, to further illustrate the spatiotemporal dynamics of the explanatory factors.

4.5.1. Workdays

Figure 4 illustrates the distribution of correlation coefficients for eight explanatory variables in the workday morning peak. These variables include distance to the city center, road mileage, bus stops (from the socioeconomic and transportation accessibility categories), and street-view features such as buildings, greenery, sky, traffic control facilities, and sidewalks. It is observed that the distance to the city center generally shows a negative correlation with ride-hailing usage, with minimal spatial variation (−1.106 to −1.052). In contrast, road mileage and bus stops both have positive coefficients, indicating that better transportation accessibility attracts more demand for ride-hailing services. These variables exhibit similar distribution patterns, centered around Jinyu and Fucheng Streets, extending towards the two main roads, Ye Hai Avenue and Yingbin Avenue, supported by the well-developed transport networks around Haikou East Station and Haikou Passenger Station, which significantly enhance ride-hailing demand in the morning peak. The traffic control facilities exhibit a distribution pattern similar to that of bus stops, while the proportion of sky elements shows a negative correlation with ride-hailing demand, particularly pronounced in the central urban districts (Binhai, Jinmao, Haifu, and Datong Streets). At the same time, the coefficient distribution of the buildings in these regions shows positive peaks, suggesting that users in developed urban areas are more likely to use ride-hailing, supported by well-maintained greenery and sidewalk designs. The presence of both large positive and negative coefficients of greenery further supports the hypothesis that greenery interacts with land-use intensity, positively associated with ride-hailing in central business areas but negatively in more walkable residential districts.

In the evening peak, the significant variables include distance to the city center, bus stops, financial and auto repair POIs from the land-use category, and street-view features such as greenery, sky, roads, and traffic control facilities (Figure 5). The coefficient distribution for distance to the city center is similar to that of the morning peak, further amplifying its negative effect on ride-hailing demand. The impact of bus stops remains positive throughout the study area, emphasizing that bus stop density is closely linked to ride-hailing usage. The two land-use variables show distinct coefficient patterns: financial services have a strong positive influence in the coastal areas of Jinmao, Binhai, and Renmin Road, which may be related to local industrial patterns and increased opportunities for ride-hailing use. Conversely, auto repair services exhibit a negative correlation across the entire area (−4.105 to −1.110), suggesting that areas dominated by auto repair services reduce reliance on ride-hailing. Among the street-view features, the effect of greenery is similar to that in the morning peak, with a positive correlation centered on the central urban districts and spreading outward but with more pronounced negative coefficients in several residential sectors, particularly the southern and southeastern peripheries, indicating that, unlike morning commuting, which is often time-sensitive and destination-specific, return trips may not generate the same level of urgency, especially in quieter, green neighborhoods. The proportion of sky elements shows a negative correlation globally, with minimal variation (−2.864 to −2.659), while road elements show a positive correlation with little variation (3.760 to 3.911). The coefficient distribution for traffic control facilities differs notably from the morning peak, now primarily concentrated in the central urban areas, with negative correlations in the suburban areas, likely due to more open spaces and less traffic congestion in these areas.

4.5.2. Holidays

During holidays, as shown in Figure 6, significant variables for the morning peak include population, distance to the city center, bus stops, and street-view features such as buildings, greenery, and sky. The population variable, bus stops, and building elements are positively correlated across the entire area, while distance to the city center and sky elements are negatively correlated. Notably, the population variable shows a peak concentration in Meilan District, a residential area in Haikou located farther from the city center, which increases the reliance on ride-hailing. The peak area for bus stops shifts to the southeast, away from the central urban area, indicating that, on holidays, users are more likely to combine public transportation with ride-hailing for their travel needs. The coefficient distribution for buildings shows two centers: the western Jinmao and Binhai Streets, and the eastern Lantian and Guoxing Streets, which are Haikou’s financial and cultural districts, respectively, generating different travel demands. The distribution of greenery is similar to that of workdays but more dispersed, with an expanded positive impact, reflecting the more diverse nature of holiday travel purposes, such as leisure, shopping, or family visits. In some areas, higher greenery coverage enhances travel comfort and the attractiveness of destinations, thus encouraging ride-hailing use, while, in others, the same factor may coincide with low activity intensity and high self-sufficiency, suppressing mobility demand.

In the evening peak on holidays, the significant variables include distance to the city center, bus stops, financial and auto repair services, and street-view features such as buildings, sky, roads, and traffic control facilities (Figure 7). The coefficient distribution for distance to the city center changes from the hierarchical pattern of the morning peak to the lowest point in the central urban area, rising gradually outward, with more distinct spatial variation. Globally, the negative correlation remains, indicating that users are more sensitive to distance from the city center at this time, with a more negative attitude. The distribution for bus stops also shifts, with the peak areas now concentrated in Haifu, Heping, Lantian, Guoxing Streets, and Haikan Street on the west side, showing positive correlations in the central area and negative correlations in the peripheral areas, which may be related to holiday travel patterns. The effect of financial services diminishes compared to workdays, while the negative impact of auto repair services on ride-hailing usage intensifies. The positive impact of buildings is weaker in the evening peak, and the distribution becomes smoother (0.440 to 0.693), suggesting that users are less likely to rely on ride-hailing in developed urban areas during this period, with the remaining variables showing similar coefficient distributions to those of workdays, with little spatial variation.

5. Discussion and Conclusions

As an emerging mode of transportation, the relationship between ride-hailing demand and the built environment has attracted increasing attention from researchers, policymakers, and platform operators [17,19,24]. However, due to data limitations and methodological constraints, accurately identifying the key determinants of ride-hailing demand at the urban scale remains challenging. This study addresses this gap by integrating multi-source heterogeneous data, including POIs, socioeconomic indicators, and street-view imagery, to examine the spatiotemporal impacts of the built environment on ride-hailing demand in Haikou, China. Using a one-month dataset at a 0.5 km × 0.5 km grid level, we extract 26 variables across three categories and apply global and local spatial models to analyze their effects.

Our findings reveal substantial spatial and temporal heterogeneity in ride-hailing demand, with distinct influencing factors emerging across different time periods and urban zones. Among the models tested, the multi-scale geographically weighted regression (MGWR) model demonstrates superior explanatory power, effectively capturing local variations and resolving scale differences among variables. For instance, proximity to the city center and density of bus stops are consistently influential, although their effects vary by time and location. Land-use types such as financial services and auto repair areas show divergent patterns, and street-view features, especially the proportions of greenery, buildings, and sky, further illuminate the role of micro-scale environmental aesthetics in shaping mobility preferences. These results offer important implications for urban transportation planning and policy. First, they suggest that ride-hailing infrastructure, such as designated pick-up and drop-off zones, should be strategically located based on spatial clusters of demand and the local built-environment characteristics. Second, the findings on the time-sensitive effects of specific land uses and street features can inform adaptive traffic management strategies, particularly during holidays or peak periods. Finally, the integration of street-level visual indicators provides a novel and scalable approach for assessing mobility equity and service coverage, especially in underserved peripheral regions. By demonstrating the applicability of the MGWR model to high-resolution multi-source data, this study contributes both methodological insight and practical guidance for improving demand forecasting, optimizing dispatch operations, and promoting the sustainable development of ride-hailing systems.

While this study provides valuable insights into the spatial determinants of ride-hailing demand, it has several limitations that warrant consideration: (1) The analysis focuses on built-environment factors at the community level, which may overlook individual or group-level behavioral heterogeneity and personal preferences. (2) The street-view features were extracted using the DeepLabV3+ model, which was pre-trained on public datasets. This limits the types of recognizable environmental elements and may not fully align with subjective human perceptions of streetscapes. (3) Due to data availability constraints, the ride-hailing data used in this study were collected in 2017. While the analysis aims to uncover structural spatial relationships that are relatively stable over time, we acknowledge that changes in mobility patterns or platform operations since then may affect the generalizability of some findings. Further validation using updated or longitudinal data is encouraged. These limitations also suggest several avenues for future research. First, future research could incorporate emotional responses to different street views to bridge the gap between objective visual features and subjective perception. For example, combining panoramic imagery with human sentiment ratings or emotion-aware visual models may yield a deeper understanding of how environmental aesthetics influence ride-hailing behavior. Moreover, with increasing access to fine-grained and multimodal mobility data, it would be valuable to examine the interactions between ride-hailing and other transportation services. For instance, bike-sharing systems, commonly used for short non-motorized trips in dense urban cores, are often spatially and temporally intertwined with ride-hailing. Finally, future studies should investigate the complementary or competitive dynamics between these two services. Understanding their spatial and temporal interplay is essential for optimizing multi-modal transport systems and informing integrated urban mobility strategies.