Spatial Accessibility in Last-Mile Logistics: A New Dimension of Urban–Rural Integration

Abstract

Under the advancing urban–rural integration strategy, last-mile logistics, and their spatial accessibility, have become key indicators for measuring regional coordination. Focusing on Guangzhou as the case study area, this study constructs an urban–rural spatial accessibility assessment model integrating multimodal convolutional neural networks and Graph Neural Networks (GNN) to systematically examine the evolving accessibility patterns in last-mile logistics distribution across urban and rural spaces. The study finds that Guangzhou’s urban space continues to expand while rural space gradually decreases during this period, showing an overall development trend from centralized single-core to multi-polar networked patterns. The spatial accessibility of last-mile logistics in Guangzhou exhibits higher levels in urban core areas and lower levels in peripheral rural areas, but the overall accessibility is progressively expanding and improving in outlying regions. These accessibility changes not only reflect the optimization path of logistics infrastructure but also reveal the practical progress of urban–rural integration development. Through spatial distribution analysis and dynamic simulation of logistics networks, this study establishes a novel explanatory framework for understanding the spatial mechanisms of urban–rural integration. The findings provide decision-making support for optimizing last-mile logistics network layouts while offering both theoretical foundations and practical approaches for promoting co-construction and sharing of urban–rural infrastructure and achieving integrated regional spatial governance.

1. Introduction

Last-mile logistics delivery has become a vital component of urban activities. With the acceleration of urbanization, the demand for logistics distribution continues to rise, particularly in last-mile delivery, which serves as a critical link in modern urban logistics systems [1]. The efficiency of last-mile logistics not only impacts urban economic operations but also directly affects residents’ quality of life [2,3]. In recent years, with the growth of Electronic Commerce (EC) and the diversification of consumer demands, last-mile logistics has gained increasing attention from academia and industry. Research in this field has expanded significantly, covering areas such as logistics distribution models, technological advancements, and delivery efficiency. However, most existing studies focus on technical issues such as delivery efficiency and shortest path analysis, with limited in-depth exploration of the spatial accessibility of last-mile logistics in urban–rural integration and its impact on urban–rural development [4].

Last-mile logistics is not merely a simple delivery issue, but actually reflects disparities in resource allocation, infrastructure development, and socio-economic growth between urban and rural areas [5]. From the perspective of urban–rural spatial accessibility, last-mile logistics provides a unique vantage point to observe the progress of urban–rural integration. Nevertheless, current research primarily emphasizes the efficiency improvement of logistics distribution models, lacking a thorough investigation into the spatial accessibility between urban and rural areas, particularly the logistics connectivity during urban–rural integration [6]. Therefore, this study addresses this research gap by analyzing the role and influence of last-mile logistics in urban–rural integration from the perspective of spatial accessibility for the first time. However, most existing studies fail to address this dimension, lacking systematic analysis that closely links last-mile logistics with the urban–rural integration process [7]. Therefore, this study combines the spatial accessibility of last-mile logistics with the urban–rural integration process, proposes a new research framework, fills this gap, and provides new perspectives and methods for future research in related fields.

Urban–rural integration, as a complex multi–dimensional issue, has become a prominent topic in academic circles both domestically and internationally in recent years [8]. Scholars generally focus on urban–rural development disparities, particularly the imbalances in economic, social, cultural, and resource allocation aspects [9,10]. However, traditional urban–rural integration studies predominantly concentrate on macro-level disparity analyses, such as regional development gaps, income disparities, and population mobility. Research shows that the barriers to urban–rural integration mainly manifest in infrastructure development, public services, and quality-of-life differences between urban and rural residents [11,12]. These studies often emphasize the urban–rural divide during urbanization processes [13]. However, with the advancement of urban–rural integration, scholars are beginning to explore more diverse research perspectives. From a spatial perspective, urban–rural spatial accessibility has gradually become an important indicator reflecting the degree of urban–rural integration. Spatial accessibility no longer merely refers to the construction of transportation networks, as it, more importantly, involves how to connect urban and rural areas through logistics, information flow, and other systems to promote optimal allocation of economic resources [14,15]. As a key link between urban and rural areas, last-mile logistics is gaining increasing attention [16]. However, most existing studies focus on improving delivery efficiency or analyzing urban–rural disparities, lacking in-depth research on the specific role of last-mile logistics in urban–rural integration, particularly how “spatial accessibility of last-mile logistics” can optimize the allocation of urban–rural resources.

Urban–rural spatial identification involves both geographic spatial characteristics and economic activity intensity/distribution, particularly combining land use data and NTL data. Land use data provides the physical distribution characteristics of regions, while NTL data reflects economic activity intensity. Simple weighted average methods, or traditional machine learning models, cannot effectively capture the complex nonlinear relationships between these features or dynamically adjust feature weights. We choose to use multimodal CNN for edge area identification and accessibility analysis, mainly based on the following considerations. First, it should be noted that although traditional machine learning methods (such as random forests, K-nearest neighbors, support vector machines, and maximum likelihood classification) perform well in many tasks, our comprehensive comparison shows that CNN has unique advantages in this study. Specifically, when processing spatial data and image data, CNN demonstrates stronger feature learning capabilities. Compared with traditional methods, CNN demonstrates three key advantages. First, conventional machine learning approaches typically rely on manual feature extraction, which shows clear limitations when processing spatial data, particularly for geographic information and network optimization problems, as they often fail to adequately capture both local and global patterns in the data. In contrast, CNN’s unique multi-layer convolutional architecture automatically extracts important features from data, particularly excelling in recognizing edges, shapes, and spatial dependencies, leading to superior performance when handling complex datasets. Second, in terms of specific method comparisons, approaches like random forests and K-nearest neighbors depend heavily on feature selection and distance metrics. When dealing with high-dimensional, structurally complex spatial data, their performance often becomes constrained. While support vector machines (SVM) can address nonlinear problems through kernel functions, they require substantial training time and memory resources for large-scale datasets. Third, and most importantly, CNN’s weight-sharing mechanism not only enables efficient processing of large datasets but also achieves significantly higher accuracy in edge region identification compared to traditional methods. Moreover, CNN’s exceptional capability in capturing local data dependencies perfectly matches our research needs for edge region recognition. It is precisely these technical advantages that lead us to select CNN as the core methodology for this study. Therefore, we employ a multi-modal CNN to automatically learn nonlinear relationships between land use data and NTL data [17].

Research on last-mile logistics has attracted widespread attention from both academia and industry in recent years. Current studies mainly focus on the spatial layout of last-mile logistics [18]. The first key aspect is the location selection of distribution centers and delivery stations. The placement of distribution centers represents a core issue in last-mile logistics planning [19]. Researchers have proposed various optimization models to address distribution center location problems, particularly in urban environments where proper placement can reduce redundant delivery routes and improve service quality [20,21]. Studies show that rational distribution center and delivery station layouts can significantly decrease delivery time and costs while enhancing delivery efficiency [22]. Some research employs big data and geographic information system-based approaches to analyze urban demand distribution, thereby optimizing distribution center locations and delivery route planning [23,24]. The second key aspect is spatial accessibility and network optimization. Beyond distribution center location selection, the spatial accessibility of last-mile logistics represents another important research direction [25,26]. Scholars have examined how to improve accessibility by optimizing the spatial relationship between delivery stations and urban residents [27]. By developing spatial-based shortest path and optimal distribution network models, researchers have proposed multiple effective network optimization solutions. These solutions can enhance delivery accessibility for both urban and rural residents while significantly reducing logistics costs [28,29]. In large urban areas, the spatial layout of last-mile logistics must integrate with public transportation systems, leveraging urban transit networks to further optimize deliveries and strengthen connectivity between cities and suburbs [30]. However, these studies focus on last-mile delivery efficiency and network optimization, while overlooking the critical role of urban–rural spatial accessibility. Particularly, in the context of urban–rural integration development, last-mile logistics serve as a critical link between urban and rural areas, requiring further research to examine its role in facilitating urban–rural integration.

In this study, we use the term “accessibility” to specifically refer to people’s ability to obtain logistics distribution services (e.g., food delivery, daily necessities, express delivery, etc.). Notably, unlike traditional logistics research, our study not only focuses on path optimization but, more importantly, emphasizes improving residents’ convenience in accessing these services by optimizing transportation networks. Specifically, our core objective is not simply to find the shortest path but to enhance service accessibility for residents in specific areas through efficient logistics planning. Furthermore, our research goes beyond addressing last-mile logistics distribution issues—it centers on accessibility while comprehensively considering multidimensional factors such as geographic distribution, resource allocation, and population density. Building on this foundation, the spatial accessibility of last-mile logistics delivery is also gaining attention. Research on delivery accessibility originated from spatial accessibility analysis in logistics networks, with early studies primarily focusing on quantifying accessibility in delivery areas through mathematical models [31]. With technological progress, researchers have gradually expanded this to more complex multi-dimensional analyses, incorporating factors such as time, cost, traffic conditions, and social elements. Delivery accessibility no longer simply means “whether delivery is possible,” but more importantly emphasizes “how to deliver efficiently and at low cost” [32,33]. Current domestic and international researchers commonly adopt the following methods to measure delivery accessibility. First is shortest path analysis. This method applies graph theory and network analysis to build transportation network models, calculating the shortest delivery routes between consumers and distribution centers to evaluate accessibility. Shortest path analysis typically optimizes by minimizing delivery time or cost. This approach suits urban or regional distribution networks [34,35,36]. However, shortest path analysis typically operates on static network assumptions and relies on fixed route calculations. This approach fails to accurately capture dynamic road network changes in urban–rural fringe areas, particularly in regions with complex urban villages and rapidly developing transportation networks [37]. This occurs because road networks in urban–rural transition zones frequently change within short periods due to construction and demolition activities. These changes significantly impact logistics distribution routes, yet shortest path analysis cannot effectively capture such rapid spatiotemporal variations [38]. Second is coverage area modeling. These studies define service radii around distribution centers to measure accessibility in target areas. Optimizing service radii improves network coverage and reduces service gaps, making this method useful for urban logistics and urban–rural connections [39,40,41]; coverage area modeling typically assesses accessibility by setting a service radius, but this method fails to account for dynamic changes in delivery demand. With the growing penetration of EC and logistics companies expanding into rural markets, delivery demand and service coverage have undergone significant changes. The coverage area modeling fails to consider these changes, leading to inaccurate assessments of logistics service capabilities in urban–rural areas. This limitation becomes particularly evident when service gaps emerge in regions experiencing rapid infrastructure development [42]. More importantly, these two conventional methods are fundamentally based on an assumption of linear or regular relationships between data points. However, in reality, the data we process in urban–rural last-mile logistics systems typically originates from multiple sources, exhibiting distinct multi-source characteristics. Precisely because of this multi-source nature, the data forms complex nonlinear correlations. Consequently, when applying simple linear methods, we encounter a critical issue: the inability to accurately capture and represent the intricate interactive relationships between these data points. This limitation, in turn, leads to research being unable to comprehensively and accurately assess the combined impact of various factors on last-mile logistics accessibility. Third is multi-objective optimization modeling. As delivery systems grow more complex, scholars increasingly employ multi-objective optimization models to assess accessibility. These models integrate multiple factors like time, cost, and service quality, enhancing overall system performance through comprehensive optimization [43,44,45]. It should be noted that evaluating last-mile logistics accessibility involves multiple dimensions of complex factors, including geographical information, traffic conditions, delivery demand, and time constraints. What deserves special attention is that the inter-relationships between these elements often exhibit distinct nonlinear and highly complex characteristics [17]. However, traditional modeling methods show clear limitations when handling these complex multidimensional data. In sharp contrast, neural network technology demonstrates unique advantages in this regard. Specifically, taking Graph Neural Networks (GNN) as an example, this technology not only effectively captures dynamic changes in road networks but, more importantly, it models topological changes in road networks through graph structures, thereby precisely describing the spatiotemporal evolution of delivery routes in urban–rural fringe areas [46]. Building on these technical strengths, such advanced methods successfully overcome the inherent shortcomings of traditional approaches in dealing with complex, dynamic urban–rural logistics networks. As a result, they can more accurately and timely reflect the dynamic evolution characteristics of logistics systems during urban–rural integration. Current research on spatial accessibility mainly uses path optimization methods. However, our study focuses on last-mile logistics delivery differences between urban and rural areas. Last-mile logistics faces significant challenges in urban–rural fringe areas, especially in Guangzhou with its numerous urban villages where delivery routes frequently break and spatial structures change dynamically. Traditional path planning methods struggle to respond quickly to these changes. Conventional path analysis approaches (like weighted shortest path and grid-based accessibility analysis) cannot accurately capture dynamic road network topology changes and spatial-economic coupling relationships. While Graph Neural Networks (GNN) can directly model graph structure features of road networks, learning both local and global path properties through information propagation between nodes, and demonstrating stronger representation capabilities [47]. Therefore, this study adopts GNN to model and analyze last-mile logistics delivery accessibility under urban demolition conditions, comprehensively considering road network changes, land use transitions, and delivery capacity factors, achieving high-precision dynamic modeling of delivery accessibility. The core concept of GNN is as follows: each node updates its own representation through features of neighboring nodes [48]. It is important to clarify that GNN demonstrates unique advantages in spatial accessibility analysis by effectively capturing complex road network topologies and spatial-economic coupling relationships. In comparison, traditional methods like the Huff model, 2SFCA, or gravity-based accessibility models excel at processing continuous geographic features, including distance, travel time, and traffic flow parameters. These models can precisely quantify spatial decay effects, particularly in accurately reflecting how service accessibility decreases with increasing distance [49,50]. However, it must be emphasized that these traditional methods show clear limitations when dealing with dynamic changes and complex network structures. This limitation becomes particularly prominent in urban–rural fringe areas, as these regions often undergo rapid spatial transformations. To precisely address these practical challenges, this study adopts the GNN approach, primarily because GNN not only adapts to dynamic changes in urban–rural transition zones but, more importantly, automatically learns and adjusts to patterns of change in both local and global network topologies.

With the acceleration of globalization, urban–rural integration has become a common challenge for many countries and regions, especially in the development of logistics systems and the optimal allocation of resources. By exploring the spatial accessibility of last-mile logistics in urban and rural areas, this study provides a new perspective for global academia, helping to understand how logistics connectivity can promote urban–rural integration and further drive balanced economic, social, and cultural development. Moreover, with the continuous advancement of EC and urbanization, this study holds significant practical implications for local policymakers. The findings offer a theoretical foundation for local governments when formulating policies related to urban–rural integration, particularly in optimizing logistics networks, improving delivery efficiency, and enhancing infrastructure development. These insights help decision-makers better understand and address imbalances in urban–rural development, facilitating resource exchange and economic integration within, and beyond, regions.

Based on the current research gaps, this study examines urban–rural integration through the lens of spatial accessibility in last-mile logistics delivery. The research consists of three main components: first, identifying urban–rural spatial boundaries; second, analyzing the spatial accessibility of last-mile logistics across these areas; and third, evaluating the role of last-mile logistics in urban–rural integration.

2. Materials and Methods

2.1. Study Area

Guangzhou is a major city in southern China, located in the heart of the Pearl River Delta. As the capital of Guangdong Province, it covers approximately 7434 square kilometers with a population exceeding 15 million. Despite rapid urbanization in recent years, significant disparities persist between urban and rural areas in Guangzhou. The differences in infrastructure, logistics services, and economic development levels between Guangzhou’s urban core and surrounding rural regions make it an ideal and representative case for studying logistics accessibility in urban–rural integration [23]. As one of China’s economic hubs, Guangzhou boasts a highly developed logistics industry with comprehensive distribution networks and modern logistics technologies. This context gives Guangzhou high practical value for last-mile logistics research. Particularly, in e-commerce, express delivery, and instant delivery services, Guangzhou leads domestically in logistics innovation and technology application, providing rich field data and research background for this study. Therefore, selecting Guangzhou as the case study helps us to better understand the role of last-mile logistics in urban–rural integration, propose actionable optimization recommendations, and provide reference for other similar cities (Figure 1).

2.2. Data Source

This study requires two key types of data: first, urban–rural spatial identification data, and second, last-mile logistics distribution data. Specifically, for urban–rural spatial identification, we comprehensively utilize land use data and NTL data. Most importantly, we analyze night-time light intensity characteristics of impervious surfaces combined with Landscan data to accurately distinguish between urban and rural spatial boundaries. As for the last-mile logistics distribution data, it primarily comes from Amap POI information. Through spatial correlation analysis with road network data, this ultimately enables effective evaluation of delivery service accessibility levels.

2.2.1. Land Use Data

Land use data represents spatial information describing land resource utilization within a region, covering the distribution and changes in different land types. This data helps researchers analyze development levels, land functions, and spatial distribution characteristics of land use across different areas. Land use data typically includes categories such as agricultural land, construction land, forest land, water bodies, and other specially designated land types. Using this data, researchers can identify distinct characteristics of urban and rural spaces, and analyze allocation differences in land resources between them. The land use data in this study comes from Wuhan University’s CLCD (Figure 2) [51].

2.2.2. Night-Time Light (NTL) Data

NTL data captures surface brightness information at night through satellite remote sensing. This data primarily reflects artificial lighting intensity on Earth’s surface at night and commonly serves to study spatial distribution of human activities, urbanization processes, and economic development. NTL data offers high timeliness and spatial resolution, accurately representing regional economic activity, urban scale, and population density. Particularly in urban–rural spatial studies, it provides unique insights into activity intensity, resource utilization, and living standard differences between urban and rural areas. This study utilizes VIIRS DNB (Day/Night Band) data from NOAA, widely applied in urbanization, economic activity, and population distribution research, with extensive use in geoinformatics, urban planning, and environmental studies. The data is accessible through NASA’s Earth Observing System Data and Information System (EOSDIS) platform (Figure 3) [52].

2.2.3. Last-Mile Logistics Facility Distribution Data

Last-mile logistics facility distribution data records the spatial distribution of last-mile logistics facilities (such as distribution centers, warehouses, express delivery stations, and logistics hubs). These facilities play a crucial role in logistics systems and are key components for last-mile delivery. The rational layout of last-mile logistics facilities directly affects delivery efficiency and spatial accessibility. Studying their spatial distribution reveals differences between urban and rural logistics systems and helps analyze resource allocation and spatial planning of logistics facilities during urban–rural integration. We obtained location information (including longitude, latitude, facility type, and operational status) for all relevant logistics facilities in Guangzhou from 2015 to 2024 through AutoNavi Map’s (Amap’s) API interface. To ensure data accuracy and validity, we cleaned and preprocessed the last-mile logistics facility distribution data obtained from Amap (Figure 4) [53].

2.3. Methods

To more accurately identify urban and rural spaces, we define them based on the urban–rural classification standards from China’s National Bureau of Statistics. Specifically, urban areas mainly refer to nationally designated urban built-up areas and their surrounding regions, including prefecture-level cities, county-level cities, and urban districts. These areas typically exhibit three key characteristics: first, they have high economic and population density; second, they possess relatively well-developed infrastructure; and third, their economic activities are primarily focused on secondary and tertiary industries, with comprehensive public services and social security systems. From a spatial perspective, urban areas include not only core cities and sub-central cities but also surrounding metropolitan regions, which concentrate the abundant economic, cultural, and social resources. In contrast, rural areas refer to non-urban built-up zones primarily engaged in agriculture and related industries for residence and production. These regions exhibit two main characteristics: on one hand, they feature lower population density and relatively limited infrastructure; on the other hand, their economic activities have long been dominated by primary industries. However, with recent progress in urban–rural integration, rural areas have gradually developed diversified industrial structures. It should be particularly noted that rural areas encompass not only traditional agricultural regions but also urban–rural transitional zones and some outer suburban counties. Although these areas are administratively classified as rural or township territories, they maintain close economic and logistics connections with urban centers. Finally, it is important to emphasize that urban–rural classification criteria may vary across different countries and regions due to differences in socioeconomic development levels. Therefore, both the research scope and conclusions of this study are specifically based on China’s unique socioeconomic context.

2.3.1. Multi-Modal Convolutional Neural Network (Multi-Modal CNN)

Compared to linear regression or shallow machine learning models, multi-modal CNN extracts features from each modality through convolutional layers and effectively fuses features from both modalities using nonlinear activation functions, improving urban–rural spatial classification accuracy. The multi-modal CNN structure contains two parallel convolutional paths, with each path processing one data modality. Land use data and NTL data serve as input feature vectors, undergoing convolution and pooling operations to extract their respective feature information. The features from both modalities then merge through a feature fusion layer, with final classification performed by fully connected layers. Compared to traditional methods, multi-modal CNN dynamically adjusts the influence of each modality and automatically optimizes feature weights through end-to-end training [54,55]:

$$\Gamma =-\sum _{i=1}^N{y}_{i}log\left({\widehat{y}}_i\right)+(1-y_i)log(1-{\widehat{y}}_i)$$

(1)

where ${\mathrm{y}}_{\mathrm{i}}$ represents the true label of the $i$-th sample and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ is the predicted probability from the network output. By optimizing this loss function, the neural network learns the nonlinear mapping relationship between input features and urban–rural spatial classification. In this study, our multi-modal CNN structure contains one input layer, two convolutional paths, one feature fusion layer, multiple hidden layers, and one output layer. Each convolutional path includes several convolutional layers with ReLU activation function and max-pooling layers. After convolution, features concatenate through the fusion layer, pass through a fully connected layer (128 neurons), and, finally, output classification results via Softmax activation function.

The multi-modal CNN model primarily uses two key data sources as input: NTL data and Landscan data. Specifically, NTL data directly reflects regional light intensity characteristics; urban areas typically show higher light intensity values, while rural areas exhibit relatively lower levels [56]. However, it should be noted that NTL data has a spatial resolution of only 500 m, and this relatively coarse resolution may lead to loss of detail in urban–rural boundary areas, thereby affecting accurate classification of transition zones. In comparison, land use data demonstrates superior 30 m resolution, providing more detailed regional distribution information that can more precisely capture micro-features of urban–rural spaces. Meanwhile, as another important data source, Landscan data provides annual population density information for global geographic grids. Although its 1000 m resolution suits macro-regional analysis, its performance at micro-scales proves less precise than the other two datasets [57]. Precisely because of these significant resolution differences among the three data types, each plays a distinct role in urban–rural spatial identification. By feeding these data sources into different convolutional layers for feature fusion and processing, the model effectively learns complex relational patterns between various features, ultimately achieving high-precision urban–rural spatial classification tasks [58].

During network training, first, it should be noted that NTL data essentially contains light intensity information from various regions worldwide. To adapt to CNN model input requirements, we standardize the raw data; specifically, we normalize each pixel value to the [0,1] range, a critical step that effectively eliminates scale differences in light intensity across regions. Similarly, for Landscan data, we perform specialized preprocessing: on one hand, we convert it into a format suitable for CNN model input, and on the other hand, we discretize the population density data by dividing it into several intervals, enabling the model to more effectively interpret this information. After completing data preprocessing, we jointly input the standardized NTL data and discretized Landscan data into the CNN model. It is particularly important to note that, given this study’s fundamental nature as a typical urban–rural binary classification problem, we select Cross-Entropy Loss as the optimization objective for model training [59]. To ensure training quality, we use manually interpreted results based on remote sensing images and urban planning data as ground truth, with these labeled data containing 5481 tag points possessing clear urban–rural characteristics. Regarding the specific network training process, we use the Adam optimizer with an initial learning rate of 0.001, batch size of 32, and 100 training epochs. To prevent overfitting, we add dropout layers after each hidden layer with a dropout rate of 0.5. We also implement early stopping to terminate training when the validation loss stops decreasing, avoiding overfitting from excessive training. For data stability and training efficiency, we standardize both the land use data and NTL data. To evaluate model performance, we employ Accuracy and Confusion Matrix as primary validation metrics—Accuracy measures overall classification performance while Confusion Matrix analyzes urban/rural classification accuracy, particularly false positives (urban misclassified as rural) and false negatives (rural misclassified as urban). The final results show the network achieves 92.35% accuracy, 91.08% precision, 93.17% recall, and an F1-score of 0.8953, demonstrating high training accuracy. We also conduct 5-fold cross-validation to verify model generalization capability and ensure robustness across different data subsets.

2.3.2. Graph Neural Networks (GNN)

Before conducting last-mile logistics calculations, we first perform systematic data cleaning, which mainly includes four key steps. First, in the data validation stage, we focus on identifying and removing invalid data entries, especially those with null location information, missing values, or duplicate records. For geographic location data, we strictly apply filtering criteria to eliminate all entries that cannot be matched to valid geographic coordinates. Second, during the outlier processing phase, we conduct comprehensive verification of all geographic coordinate data. Specifically, by establishing a spatial index of Guangzhou’s administrative boundaries, we ensure each logistics facility’s latitude and longitude coordinates fall within reasonable ranges. When coordinates exceed municipal boundaries or show obvious positional anomalies, we promptly correct these data entries. Third, for categorical data validation, we carefully classify the data by facility type, including distribution centers, delivery stations, and logistics hubs. During this process, we not only ensure accurate labeling for each record but also manually review special cases that cannot be clearly categorized, removing them when necessary. Finally, given the extended temporal scope of our research data (from 2015 to 2024), we implement an additional time-series verification step. Specifically, we conduct annual checks on the completeness of facility distribution data. Whenever we identify missing annual data or detect abnormal fluctuations, we immediately trace the data sources and perform necessary supplementation or corrections [60].

This capability proves especially valuable for handling complex scenarios like demolition and road network updates [61]. Specifically, GNN’s key advantage lies in its unique message-passing mechanism between adjacent nodes. Through this mechanism, GNN captures real-time dynamic changes in road networks—a capability that traditional distance-based or travel-time-based spatial accessibility models lack. Particularly in urban–rural fringe areas where road networks experience frequent disruptions and significant traffic flow variations, GNN’s adaptive strengths become even more pronounced. It effectively models these dynamic changes, thereby enabling more scientific route planning in complex urban–rural logistics environments [62].

Given a graph G = (V, E), the embedding of each node updates at layer l as follows:

$$h_i^{\left(l\right)}=UPDAT{E}^{\left(l\right)}(h_i^{(l-1)}),AGGREGAT{E}^{\left(l\right)}(\{h_j^{(l-1)}:j\in N(i)\})$$

(2)

where $\mathrm{N}\left(\mathrm{i}\right)$ is the set of neighboring nodes for node $\mathrm{i}$. $\mathrm{A}\mathrm{G}\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{G}\mathrm{A}\mathrm{T}{\mathrm{E}}^{\left(\mathrm{l}\right)}$ (·) serves as the aggregation function that collects information from neighboring nodes, while $\mathrm{U}\mathrm{P}\mathrm{D}\mathrm{A}\mathrm{T}{\mathrm{E}}^{\left(\mathrm{l}\right)}$ (·) acts as the update function that combines neighbor information with self-information. For this study, our GNN architecture consists of an input layer, multiple graph convolution modules, hidden layers, and an output layer. The model takes graph-structured data as input, with its core components including nodes, edges, and their associated attribute features. Specifically, first, regarding node representation, each node in the graph corresponds to a spatial unit. In the actual last-mile logistics network, these nodes represent different types of geographical areas. On one hand, nodes can denote physical spaces like urban areas, rural settlements, or logistics centers. On the other hand, they may also represent key locations such as transportation hubs or critical logistics nodes. For node selection criteria, we primarily consider two key factors: first, the node’s influence within the logistics network, and second, its connectivity characteristics. These specifically include important spatial elements like urban street networks, rural residential distributions, and major transportation hubs. Second, regarding edge definition, edges represent transportation connections between nodes. These connections include both urban road networks and interurban infrastructure like highways and railways. Importantly, our edge definition incorporates not just physical connectivity but also dynamic traffic attributes. Specifically, each edge integrates multiple features: travel time, spatial distance, transportation cost, and road capacity. These attributes critically influence the GNN’s learning process. Finally, for feature engineering, we equip each node with comprehensive feature vectors. These specifically include the following: regional geographic information, traffic flow dynamics, population density distribution, and road classification levels. These features provide GNN with complete regional profiles, enabling precise capture of last-mile logistics accessibility characteristics. Simultaneously, edge attributes are carefully designed with core parameters like path distance, transportation time efficiency, traffic load, and road classification; these edges are directional, representing the flow direction of delivery routes. Through these well-constructed edge attributes, the model accurately evaluates path quality, thereby providing scientific basis for final spatial accessibility calculations. For each edge (path), its weight is calculated through a weighted sum of these factors, with the specific formula as follows:

$$w_{ij}={\alpha ·D}_{ij}+\beta {·T}_{ij}+\gamma {·F}_{ij}+\delta ·C_{ij}$$

where $w_{ij}$ is the edge weight between nodes $i$ and $j$, $D_{ij}$ is the physical distance between two nodes, $T_{ij}$ is the delivery time, $F_{ij}$ is traffic flow, $C_{ij}$ is road class and capacity, and $\alpha ,\beta ,\gamma ,\delta$ are corresponding weighting coefficients. This weighting calculation method dynamically reflects changes in the transportation network, making it particularly suitable for addressing the complex logistics environment in urban–rural fringe areas.

The calculation of spatial accessibility metrics primarily relies on GNN’s unique message-passing mechanism, which involves three key implementation steps: First, during feature initialization, each node’s initial features are constructed based on multidimensional regional information, including critical data indicators such as transportation infrastructure status and population density distribution. These initial features establish an essential foundation for subsequent computations. Second, in the graph convolution process, nodes interact with their neighboring nodes through specific information aggregation mechanisms. Specifically, each node receives and integrates feature information from surrounding nodes, then updates its own feature representation through nonlinear transformations. This step enables nodes to fully assimilate topological relationship information within the network. Finally, through iterative optimization across multiple convolutional layers, the model progressively refines each node’s feature representation and ultimately outputs quantifiable accessibility scores for all nodes. Based on these scoring results, we can not only evaluate relative accessibility between any two nodes but also generate continuous accessibility distribution maps through spatial interpolation methods, thereby comprehensively reflecting the last-mile logistics convenience patterns across the study area. Higher scores indicate stronger last-mile logistics accessibility between corresponding regions, and vice versa.

3. Results

3.1. Urban–Rural Spatial Identification

Analysis of spatial distribution maps from 2015, 2018, 2021, and 2024 reveals Guangzhou’s urban space has expanded continuously while rural space has shrunk, showing distinct phased changes and spatial restructuring. Based on urban–rural spatial identification results from 2015 to 2024, we obtain the following key findings: First, regarding spatial scale changes, urban areas show significant expansion, with four-period measurements of 1520.08, 2243.92, 2605.85, and 3040.15 km², respectively. In sharp contrast, rural areas demonstrate continuous shrinkage during the same periods, measuring 5818.38, 4994.54, 4560.23, and 4198.31 km². Second, in terms of change rates, statistical analysis reveals the following: urban areas achieve an average annual growth rate of 26.80%, peaking at 47.62%, while rural areas show negative growth with −10.26% average annual rate and −7.94% peak rate (

Table 1. Urban–Rural Spatial Identification Results.

	2015	2018	2021	2024	Average Growth Rate (%)
Urban Area (km2)	1520.08	2243.92	2605.85	3040.15	26.80
Rural Area (km2)	5818.38	4994.54	4560.23	4198.31	−7.9

). Combining these patterns, we draw two major conclusions: On one hand, Guangzhou’s urban expansion contrasts sharply with rural contraction, clearly reflecting ongoing urbanization; on the other hand, the urban growth rate initially accelerates then gradually decelerates.

The four-period comparison clearly demonstrates urban space progressively extending outward with expanding boundaries. In 2015, urban areas concentrated mainly in central and southern Guangzhou. By 2018, urban space expanded significantly with broader distribution and more complex boundaries. Starting in 2021, urban growth showed both area expansion and merging between fringe zones. By 2024, urban space became more continuous and networked, with clearer core-subcenter-periphery structures. The urban spatial expansion shows nonlinear characteristics, demonstrating a “rapid-slow-stable” phased pattern. From 2015 to 2018, urban expansion reached its peak with the largest quantity and area of new urban spaces. Multiple urban fringe areas developed rapidly, including Nansha New District, Panyu, and Huangpu Technology City. From 2018 to 2021, the expansion pace slowed significantly, with existing urban patches beginning to connect and form more complete structural units. Urban development shifted from outward expansion to internal optimization. From 2021 to 2024, urban boundary expansion further decelerated, mainly featuring local edge adjustments, while the spatial pattern evolved from “fragmented-multicore” to “contiguous-network” (Figure 5).

In summary, from 2015 to 2024, Guangzhou’s urban–rural space underwent significant transformation from concentrated to expanded patterns, with clear boundaries becoming increasingly blurred. The urban space continuously expanded outward and growing in scale, particularly from 2015 to 2018 when expansion reached its fastest pace. Meanwhile, rural space gradually decreased and became more marginalized. These changes collectively reflect accelerated spatial restructuring dominated by urban expansion and the progressing trend of urban–rural integration.

3.2. Spatial Accessibility Analysis of Last-Mile Logistics Delivery

The results show significant spatial differentiation in last-mile logistics delivery accessibility across Guangzhou. Core urban areas demonstrate high accessibility, while rural fringe areas show low accessibility, reflecting substantial differences in transportation network density, infrastructure development, and delivery network layout between urban and rural regions. By integrating last-mile logistics accessibility data across different periods, we obtain the following key findings: First, in terms of numerical performance, urban areas show a steady increasing trend in average delivery accessibility, with four-period measurements of 68.74, 73.65, 79.01, and 85.09, respectively. Meanwhile, rural areas also maintain growth momentum, corresponding to values of 33.09, 38.31, 45.64, and 51.08. Second, regarding growth rates, statistical analysis reveals the following: urban areas achieve an average annual growth rate of 7.37%, while rural areas reach a higher rate of 15.61% (

Table 2. Average Accessibility of Last-Mile Logistics Delivery in Urban–Rural Areas.

	2015	2018	2021	2024	Average Growth Rate (%)
Urban Area	68.74	73.65	79.01	85.09	7.37
Rural Area	33.09	38.31	45.64	51.08	15.61

). This disparity warrants special attention. Combining these data patterns, we draw three major conclusions: First, Guangzhou’s urban areas generally maintain relatively high logistics accessibility levels. Second, the accessibility gap between urban and rural areas gradually narrows over time. Third, and most notably, rural areas demonstrate continuous improvement in accessibility scores—a phenomenon likely closely related to two key factors: ongoing infrastructure upgrades and expanding transportation networks.

From a spatial distribution perspective, the highest accessibility areas are concentrated in Guangzhou’s main urban districts, including Yuexiu, Liwan, Tianhe, and Haizhu districts. These traditional core areas feature developed transportation networks, high road density, dense delivery stations, and excellent transport connectivity with robust infrastructure support. Additionally, their high commercial density and large order volumes further motivate logistics companies to establish comprehensive last-mile delivery systems, creating highly concentrated logistics accessibility core zones. In contrast, peripheral areas—particularly urban–rural fringe zones and rural spaces like Conghua, Zengcheng, southern Panyu, and northern Huadu—show relatively lower accessibility. These regions commonly face issues including sparse road networks, discontinuous arterial roads, and rugged terrain. Simultaneously, last-mile logistics facilities distribute unevenly, with many villages still relying on traditional postal or manual delivery systems that cannot efficiently connect with urban logistics networks. Consequently, a clear accessibility gradient emerges in spatial distribution.

Further analysis of accessibility trends over time reveals that core urban areas maintain high accessibility overall but show limited growth. This mainly occurs because these areas reached relative saturation early, leaving little room for marginal improvement. Against the backdrop of Guangzhou’s expanding urban–rural space and continuous transportation infrastructure upgrades, rural peripheral areas demonstrate particularly noticeable accessibility gains. Especially after 2018, as infrastructure extended outward and smart delivery points reached community and village levels, coupled with EC platforms and logistics companies accelerating rural network expansion, previously underserved areas gradually entered last-mile logistics coverage. This structural accessibility shift indicates Guangzhou’s last-mile logistics system is evolving from “single-center intensive delivery” to “multi-tier network coverage,” narrowing but not eliminating urban–rural logistics service disparities. Particularly in blurred urban–rural boundary zones, delivery accessibility remains vulnerable to transportation bottlenecks and infrastructure gaps (Figure 6).

In summary, Guangzhou’s last-mile logistics delivery shows high spatial accessibility in urban core areas and low accessibility in rural peripheral areas, but the overall accessibility level is gradually expanding and improving in peripheral regions. These results not only reveal the evolution path of urban–rural logistics service capabilities but also provide spatial guidance for future optimization of delivery systems and resource allocation.

3.3. Analysis of Last-Mile Logistics Spatial Accessibility and Urban–Rural Integration Development

Recent studies on China’s urban–rural integration consistently indicate that as the new urbanization and rural revitalization strategies deepen, infrastructure levels, public service capacity, and spatial connectivity between urban and rural areas continue to improve. The urban–rural integration process shows a sustained positive trend, with gaps gradually narrowing. This macro-level observation is particularly evident in megacities like Guangzhou, especially in the expansion and restructuring of transportation networks, digital infrastructure, and logistics distribution systems. The spatial accessibility analysis of last-mile logistics in this study provides clear quantitative evidence of this urban–rural integration trend. During the three key phases from 2015 to 2024, rural areas show continuous and significant improvement in logistics accessibility, while core urban areas maintain high accessibility but with slower or stabilized growth. This spatial pattern strongly indicates a profound transformation in urban–rural logistics relationships. From 2015 to 2018, driven by rapid EC penetration and the “express delivery to rural areas” policy, logistics resources expand into rural areas for the first time, establishing township-level logistics infrastructure and enabling the first major leap in accessibility. From 2018 to 2021, the comprehensive implementation of county-level logistics hubs and village-level delivery points further enhances rural spatial accessibility.

Particularly noteworthy is that from 2021 to 2024, rural areas maintain steady growth in accessibility while urban areas show almost no significant changes. This phenomenon reflects two key aspects: first, urban last-mile logistics systems approach saturation or optimization limits, with their distribution network density and service efficiency reaching maturity, leaving little room for further improvement; second, it indicates that policies and corporate resources increasingly focus on rural areas, aiming to bridge the long-standing urban–rural logistics gap.

Therefore, we can consider that the spatial accessibility of last-mile logistics systems not only quantifies the infrastructure gap between urban and rural areas but also serves as a dynamic indicator of urban–rural integration progress. Urban–rural integration manifests not only in physical spatial structures, but more concretely in service accessibility and daily convenience, providing solid support for achieving substantive urban–rural integration.

4. Discussion and Conclusions

As urban–rural integration advances and logistics infrastructure expands rapidly, last-mile logistics—a crucial link between urban and rural daily life and production—has seen its spatial distribution and accessibility become key metrics for assessing balanced regional development. This study takes Guangzhou as a representative case and adopts an innovative perspective of last-mile logistics accessibility to systematically reveal the urban–rural spatial evolution process and accessibility change characteristics from 2015 to 2024, while deeply analyzing their spatial response mechanisms under urban–rural integration development. By constructing a spatial accessibility evaluation model based on Graph Neural Networks (GNN) and combining urban–rural spatial classification methods, we obtain the following key conclusions: First, from the perspective of spatial pattern evolution, Guangzhou demonstrates a distinct “urban expansion-rural contraction” characteristic over the past decade. Specifically, continuous urban boundary expansion leads to gradual blurring of traditional urban–rural transition zones, while the urban spatial structure transforms from monocentric agglomeration to polycentric networking. Simultaneously, rural areas undergo a process of fragmentation and nested marginal retreat. Second, regarding logistics accessibility, the study area shows significant spatial differentiation. On one hand, the core urban areas consistently maintain high-level logistics accessibility. On the other hand, rural areas achieve continuous improvement in accessibility from 2015 to 2024, particularly after 2018 when transportation infrastructure expansion and logistics network penetration substantially enhance rural coverage capacity. Third, examining temporal evolution patterns, changes in logistics accessibility clearly reflect the dynamic development trajectory of urban–rural connectivity. Specifically, urban areas reach near-saturation accessibility levels, while steadily improving rural accessibility marks a substantive transition in urban–rural integration, shifting from institutional design to tangible progress in spatial service capacity. It should be noted that although Guangzhou’s rural logistics accessibility shows significant improvement, a noticeable gap remains compared to urban areas. Cross-city comparisons reveal that Guangzhou’s progress in urban–rural logistics accessibility integration lags behind cities like Shanghai and Beijing, primarily due to substantial disparities in transportation infrastructure between urban and rural areas [63].

The extensive literature indicates that China’s urban–rural integration level continues to improve in recent years, with urban resources gradually spreading to rural areas and the urban–rural gap narrowing. This convergence manifests in infrastructure, living conditions, income levels, and social security systems [64,65,66]. This trend demonstrates distinctive characteristics when compared with urban–rural integration processes in other countries. For instance, studies in Europe and North America focus more on the integration of economic activities and job markets [67,68]. However, as China’s urban–rural integration progresses, our quantitative analysis of logistics accessibility reveals that integration manifests not only in institutional and service equalization, but more significantly in the evolution of spatial accessibility and logistics network structures [69]. This finding differs from traditional macro-level perspectives and highlights last-mile logistics accessibility as a crucial spatial indicator of urban–rural integration [70]. At the same time, both urban and rural delivery accessibility scores improved during this period, with rural areas showing much higher growth rates than urban areas. This indicates that rural logistics accessibility has significantly improved with infrastructure development and transportation network upgrades [71,72]. At the same time, our analysis shows that improved spatial accessibility in last-mile logistics not only indicates enhanced urban-to-rural logistics supply capacity but also suggests an evolution from “one-way radiation” to “two-way interaction” in urban–rural spatial relationships. This two-way interaction reflects the “deep structural” transformation of urban–rural integration and provides a complementary spatial empirical perspective to existing urban–rural integration studies [73].

In the study of spatial accessibility, many existing works use geographic information systems, shortest path analysis, or raster network analysis methods to evaluate the distribution of accessibility for transportation facilities, medical services, and public resources in urban spaces [74,75,76]. These studies often focus on optimizing facility layouts or balancing resource allocation within cities, emphasizing the impact of spatial accessibility on urban residents’ travel behavior and service equity [77]. However, most of these studies limit their analysis units to urban areas, rarely conducting structured comparisons of accessibility differences between urban and rural regions. In particular, there is a lack of systematic analysis on the evolution of accessibility across urban–rural divisions. Building upon the measurement logic of spatial accessibility, this study clearly separates urban and rural spaces and quantitatively analyzes their respective changes in last-mile logistics accessibility, thereby revealing temporal differences and structural imbalances in accessibility improvements across different spatial types. The results show that urban core areas maintain high accessibility throughout with limited marginal improvement, while rural areas demonstrate continuous and significant growth trends. Especially after 2018, rural areas experience substantial enhancements in logistics accessibility, benefiting from the extension of transportation infrastructure and the expansion of logistics networks into these regions [78]. Overall, existing studies primarily focus on institutional foundations of urban–rural integration, such as infrastructure equalization and public service balancing [9]. However, these studies overlook disparities in logistics infrastructure and service accessibility between urban and rural areas, resulting in limited understanding of spatial mechanisms in integration.

This study’s core contribution lies in analyzing urban–rural integration through the lens of last-mile logistics spatial accessibility. It proposes an analytical framework that combines spatial continuity with structural differentiation capability, offering new tools to understand urban–rural integration evolution paths. By explicitly dividing urban and rural spaces into two independent subsystems for temporal comparative analysis, this approach reveals structural differences previously obscured in mixed measurements and enables more detailed characterization of their accessibility features.

Based on the research findings, we propose several development recommendations. First, regarding infrastructure development, we should prioritize investment in and improvement of rural logistics systems. Specifically, rural logistics currently faces three major challenges: (1) overall accessibility remains relatively low; (2) spatial coverage is highly uneven; and (3) service shortages are particularly acute in urban–rural transition zones and transportation dead-end areas. Second, regarding evaluation system development, we should incorporate logistics accessibility as a core indicator for assessing urban–rural integration. This recommendation stems from three key considerations: First, logistics accessibility objectively reflects regional infrastructure levels and spatial connectivity; second, it effectively measures physical connection strength and economic vitality between urban and rural areas; and third, our quantitative analysis confirms that improved logistics accessibility significantly facilitates urban–rural factor flows. Therefore, we propose establishing a three-dimensional evaluation framework incorporating logistics accessibility: (1) making it a key dimension in coordinated urban–rural development assessments; (2) assigning it greater weight in urban–rural planning; and (3) strengthening its guiding role in public service allocation. This institutional design yields dual benefits: on one hand promoting more equitable resource distribution, and on the other ensuring balanced infrastructure and public service development. By systematically evaluating logistics accessibility, we can effectively enhance urban–rural spatial connectivity efficiency and provide institutional safeguards for high-quality urban–rural integration.

Although this study systematically analyzes urban–rural integration and spatial evolution in Guangzhou from the perspective of last-mile logistics accessibility and obtains relatively clear conclusions, it still has some notable limitations that require further optimization and supplementation in future research. This study uses GNN to simulate and predict spatial accessibility, which depends to some extent on the completeness and accuracy of input data. Although we select key spatial features including land use data and NTL data, the model still shows some abstraction in reflecting actual delivery chains due to lacking high-resolution real-time traffic data, logistics path trajectories, and station distribution information, making it difficult to fully reproduce the dynamic operation mechanisms of real logistics networks. Although GNNs perform well in spatial relationship modeling, they still have several key limitations. First, GNNs may rely too heavily on graph structure information, potentially overlooking the independence between certain geographic regions. Additionally, GNNs require substantial computational resources when processing large-scale datasets, which becomes particularly problematic in real-time accessibility evaluation and may significantly reduce processing efficiency [79]. More importantly, the GNN model in this study primarily applies to static spatial accessibility assessment, while its suitability for dynamic distribution networks—such as those affected by real-time traffic congestion or weather changes—still requires further validation [80]. In terms of model optimization, hyperparameter selection significantly impacts the results. However, our methodology section does not fully discuss the criteria and process for hyperparameter selection, leading to a lack of transparency in model tuning. Although we reduce overfitting risks through cross-validation, this process still involves some subjectivity.

Considering the limitations of this study, our future research will focus on improvements in three key aspects: First, regarding model interpretability, we will introduce explainable artificial intelligence methods. Through visualization techniques and feature importance analysis, we aim to enhance the interpretability of GNN. This approach helps to clearly demonstrate how the model makes decisions based on different node and edge features, thereby improving the credibility of our research findings. Second, regarding hyperparameter optimization, we will adopt more systematic and transparent approaches. Specifically, we will establish clear selection criteria for hyperparameters and implement structured search methods like grid search and Bayesian optimization. This makes the entire model optimization process more controllable and reproducible, thereby effectively reducing potential subjective influences on the results. Finally, regarding research depth expansion, although this study thoroughly analyzes temporal changes in urban–rural logistics accessibility in Guangzhou, it does not systematically examine the relationship mechanisms between logistics accessibility and socioeconomic variables (such as urban–rural population flow, consumption behavior, and land use). Therefore, future research will focus on constructing an integrated impact model for urban–rural integration, incorporating socioeconomic indicators to deeply investigate the interaction mechanisms between logistics accessibility and factors like population flow and resource allocation. Furthermore, with the development of smart logistics and intelligent transportation systems, optimizing urban–rural logistics networks using real-time data and developing dynamic accessibility evaluation models will also become key research directions. This study innovatively adopts last-mile logistics accessibility as a core evaluation metric with dual functions: it effectively assesses the balance of infrastructure development while precisely quantifying the spatial characteristics of urban–rural integration. Future research can deepen this approach within broader geographical regions, across multiple scales, and through diverse dimensions to provide theoretical support and decision-making references for coordinated urban–rural development and spatial governance optimization.