Urban Medical Emergency Logistics Drone Base Station Location Selection

Highlights

What are the main findings?

• The proposed “dynamic-static” collaborative drone base station model achieves 96.18% coverage and 673 s average response time in Guangzhou, significantly enhancing emergency logistics efficiency and resilience.
• The integration of multi-source data and multi-objective optimization effectively balances coverage, response time, and cost, providing a Pareto-optimal solution set for flexible decision making.

What are the implications of the main findings?

• The model offers a scalable and data-driven framework for urban medical emergency logistics planning, enhancing system resilience and adaptability to spatiotemporal demand fluctuations.
• It supports policy-making and infrastructure investment by quantifying trade-offs between service efficiency and cost, facilitating the design of robust drone network in complex urban environments.

Abstract

In densely populated and traffic-congested major cities, medical emergency rescue incidents occur frequently, making the use of drones for emergency medical supplies delivery a new emergency distribution method. However, establishing drone transportation networks in urban areas requires balancing spatiotemporal fluctuations in emergency needs, meeting hospitals’ mandatory constraints on response time, and addressing factors like airspace restrictions and weather impacts. By analyzing the spatiotemporal distribution characteristics of medical emergency logistics in large cities, this study constructs a drone base station location optimization model integrating dynamic and static factors. The model combines multi-source data including emergency needs, geographic information, and airspace limitations. It employs kernel density estimation to identify hotspot areas, uses DBSCAN clustering to detect long-term stable demand hotspots, and applies LSTM methods to predict short-term and sudden demand fluctuations. The model optimizes coverage rate, response time, and cost budget control for drone transportation networks through a multi-objective genetic algorithm. Using Guangzhou as a case study, the results demonstrate that through “dynamic-static” collaborative deployment and multi-model drone coordination, the network achieves 96.18% demand coverage with an average response time of 673.38 s, significantly outperforming traditional vehicle transportation. Sensitivity analysis and robustness testing further validate the model’s effectiveness in handling demand fluctuations, weather changes, and airspace restrictions. This research provides theoretical support and decision-making basis for scientific planning of urban medical emergency drone transportation networks, offering practical significance for enhancing urban emergency rescue capabilities.

1. Introduction

In densely populated metropolitan areas, frequent medical emergencies such as traffic accidents, sudden illnesses, and public safety incidents create substantial demands for medical rescue and emergency logistics. For example, in London (UK), the London Ambulance Service receives around 2000 emergency responses per day [1]. In New York City (USA), the number of hospital emergency room (ER) visits is 452 (per 1000 population) [2]. In 2023, the 11 districts of Guangzhou (CHN) collectively handled 1.6639 million emergency calls to the 120 hotline, with pre-hospital medical emergency vehicles dispatched 339,000 times [3]. Statistics indicate that major cities may receive hundreds to thousands of emergency calls daily [4], posing significant challenges to rescue response efficiency and supply chain operations. Both medical emergency logistics and rescue efforts operate under strict time constraints. Critical conditions like cardiac arrest and severe trauma require immediate treatment within the “golden 4 min” or “golden 1 h” window [5], as survival rates plummet otherwise [6]. Medical emergency logistics must reach designated locations within the treatment window specified by healthcare institutions to form a complete medical assistance cycle. Urban medical emergency logistics involves multi-stakeholder collaboration in sporadic medical supply processes [7], After completing storage operations, medical emergency supplies face extremely limited transportation timeframes (see Figure 1). In the final delivery phase, road transport is often constrained by traffic congestion, vehicles, and human resources. This extends the entire service cycle time and increases distribution costs.

With advancements in logistics drone technology, the high flexibility and rapid response capabilities of drones have made drone delivery an increasingly vital supplementary approach for urban medical emergency logistics. While drones show great potential in urban medical emergency logistics, large-scale deployment of emergency drone network still faces multiple challenges [8]. The nature of medical emergency logistics demands being sporadic and subject to spatiotemporal fluctuations makes fixed stations inadequate for meeting dynamic needs. With constraints from government budgets and commercial project funding, drone base station construction must simultaneously optimize coverage, response times, and control system-wide operational costs, presenting complex multi-objective optimization challenges. Drone performance metrics like endurance and payload capacity directly impact logistics efficiency and accessibility. Urban airspace restrictions including no-fly zones and restricted flight areas require strict compliance with aviation and public safety regulations during deployment. Additionally, drones are vulnerable to environmental and weather disruptions, with extreme conditions like strong winds and heavy rain potentially forcing operational suspensions and base station failures.

Current studies focus on handling the spatiotemporal heterogeneity of urban medical emergency demand, oversimplified constraints, and lack of dynamic resource coordination. Most studies focus on static, long-term average demand for site selection, failing to effectively integrate the significant spatiotemporal fluctuations of emergency needs [9]. They lack dynamic response mechanisms for sudden events and seasonal variations, leading to a disconnect between static planning and dynamic demand. Current research models exhibit insufficient realism in site selection constraints [10]. Drones are significantly affected by wind and precipitation in meteorological conditions [11], yet many models oversimplify urban environmental complexity. They fail to systematically incorporate multidimensional dynamic constraints such as airspace control and weather changes into optimization frameworks, compromising the practical feasibility of solutions. Additionally, there is insufficient research on dynamic allocation of drone transportation resources. Existing studies predominantly rely on single-model drones [12], failing to fully leverage the potential of multi-model drone collaborative delivery in addressing diverse delivery distances, payload capacities, and time-sensitive requirements. This limitation restricts the overall efficiency and robustness of the system. Furthermore, the singular optimization objectives primarily focusing on cost or coverage rate fail to adequately reveal the inherent trade-offs between multiple goals such as coverage, response time, and construction costs, making it challenging to support complex real-world decision making.

Therefore, the current medical drone transportation network development shows a significant gap in creating an integrated site selection framework. This framework must simultaneously (i) account for the dynamic spatiotemporal heterogeneity of emergency needs [13], (ii) incorporate real-world multidimensional constraints (e.g., weather and airspace), (iii) achieve coordinated multimodal resource allocation, and (iv) strike a balance between coverage, response time, and cost [14]. To address these challenges, this study aims to develop a decision making framework integrating multi-source data and multi-objective optimization for the scientific deployment of urban medical emergency logistics drone base stations, thereby enhancing the efficiency and network resilience of urban medical emergency logistics. The main contributions of this study include the following aspects.

Methodologically, this study overcomes the limitations of traditional static location models by developing a dual-layer planning framework that integrates static and dynamic approaches. The static layer utilizes historical data to establish long-term stable demand hotspots for infrastructure coverage, while the dynamic layer activates temporary takeoff and landing points through demand forecasting to address short-term fluctuations and emergencies. This integrated framework effectively combines infrastructure planning with real-time operational scheduling, significantly enhancing the system’s adaptability to the spatiotemporal heterogeneity of urban emergency logistics.

In terms of modeling, this study has refined the multidimensional real-world constraints that were previously oversimplified. The model not only addresses conventional limitations such as drone endurance and budget, but also innovatively incorporates dynamic interference factors based on meteorological warning levels and airspace control intensity. These factors are embedded as key parameters in both the objective function and constraint conditions. This approach ensures the optimization results demonstrate high feasibility and safety in real-world operational environments.

At the application level, this study has developed a comprehensive data-driven technical framework. By integrating multi-source data including medical orders, geographic information, airspace, and meteorological data, the system employs kernel density estimation and DBSCAN clustering to precisely identify demand hotspots. Building on this foundation, an enhanced NSGA-III multi-objective genetic algorithm is utilized to generate Pareto-optimal solutions for coverage, response time, and cost, providing decision-makers with multiple non-dominated alternatives. The model also supports coordinated delivery between multi-vehicle drones and ground transportation, enabling intelligent resource allocation and flexible capacity expansion based on cargo characteristics and mission urgency. This innovation marks a leap from single-tool selection to intelligent scheduling systems, achieving a paradigm shift in logistics coordination.

The structure of this study is organized as follows. Section 2 reviews literature on medical emergency logistics requirements, drone application scenarios, and base station siting. Section 3 introduces a modeling framework for medical emergency logistics drone base stations, detailing a dual-layer siting method that combines dynamic and static approaches. This section covers multi-source data fusion, multi-objective optimization techniques, and the development of an optimization model with algorithm selection. Section 4 provides background information on the study area, urban data sources, data processing workflows, and key parameter settings. Section 5 presents experimental design and results, analyzing spatiotemporal characteristics of hospital emergency logistics demands. It evaluates the sensitivity and robustness of siting schemes and examines how changes in critical parameters affect their performance. Conclusions, limitations, and future research directions are presented in Section 6.

2. Literature Review

Identifying the unique characteristics of medical emergency logistics demands within urban environments is fundamental to establishing an efficient drone response system. Unlike large-scale disaster relief operations, frequent medical emergencies in cities require more refined specifications for the timeliness of emergency supply distribution, spatial accessibility, and dynamic resource allocation. A thorough analysis of core features in urban medical emergency delivery serves as the foundation for evaluating existing drone logistics models, identifying critical bottlenecks in drone transportation of emergency medications, and exploring innovative solutions. This section will outline the distinctive features of urban medical emergency logistics that differentiate it from conventional logistics and large-scale disaster response. It will also analyze existing limitations and research gaps in drone-assisted medical delivery within urban areas, thereby establishing an analytical framework for exploring the applicability and operational models of drones in urban emergency logistics.

2.1. Characteristics of Emergency Distribution of Urban Medical Rescue

Existing research has extensively explored medical emergency response to large-scale disasters such as earthquakes, terrorist attacks, and pandemics. However, urban areas frequently experience more frequent occurrences of diseases, traffic accidents, construction site incidents, fires, and chemical poisoning, making it crucial to examine the characteristics of medical emergency response within urban spaces.

Emergency medical service (EMS) systems across the world have established clear response time standards for different types of incidents, which serve as core KPIs for evaluating service efficiency and quality. Cardiac arrest is the most time-sensitive scenario. The 2025 AHA guidelines emphasize immediately starting cardiopulmonary resuscitation (CPR) and performing it with a rate of 100–120 compressions per minute [15] and the European Resuscitation Council (ERC) emphasize the “Chain of Survival,” [16] where early defibrillation is a critical link. The goal is to deliver defibrillation within 3–5 min of an out-of-hospital cardiac arrest (OHCA), as survival rates decrease by 7–10% per minute of delay. This underpins the widespread concept of the “golden 4 min.” For patients with major trauma, the “golden hour” [17] concept exists, meaning the time from injury to receiving definitive care (e.g., surgery) should be within 60 min to minimize mortality. Especially for ischemic stroke, guidelines recommend that time from symptom onset to IV alteplase should be as short as possible and never more than 4.5 h, with faster treatment directly linked to better functional outcomes [18]. For conditions like surgical preparation (requiring artificial eyeball, joint), anaphylaxis (requiring epinephrine), severe infection (requiring antibiotics), or opioid overdose (requiring naloxone) [19], rapid medication delivery is also lifesaving. While there is no universal absolute minute standard, “as fast as possible” is the fundamental principle, and clinical studies often use the “call-to-administration” interval as an evaluation metric.

Official System Targets: Many national and urban EMS systems have statutory response time targets. For example, the UK National Health Service (NHS) sets a target average response time of 7 min for the most critical calls (Category 1) [20]. In China, major cities typically require the 120 emergency centers to achieve an average urban response time within 10–15 min of receiving a call [21].

How are these response time standards used to evaluate services?

Direct measurement of “call-to-scene arrival time,” “scene-to-hospital time,” etc. Correlation of response time with patient survival rates, neurological outcomes, and other health endpoints. Regular public reporting of metrics like average response time and compliance rates (e.g., 90% of calls responded to within X minutes) for inter-system comparison, internal quality improvement, and public accountability.

Studies indicate that among factors affecting medical treatment efficacy, emergency logistics proves to be the most critical factor compared to medical treatment and environmental conditions [22]. Urban medical emergency logistics exhibits the following features. (1) Significant spatiotemporal heterogeneity and suddenness [23]. Medical rescue events show temporal fluctuations, with winter and spring being peak periods for disease outbreaks and mortality [24], while the demand locations for emergency medical supplies remain unpredictable. (2) Strict time constraints for rescue operations. Different medical emergencies require varying response timelines. For instance, life-saving interventions for heart attacks [12] and cerebral hemorrhage demand extremely short response times, imposing high requirements on medical supply delivery schedules [5]. (3) Relatively fixed delivery destinations. Unlike specific disaster scenarios like earthquake epicenters [22,25], fire rescue sites [26], or high-altitude, forest, and snowy environments [27] where transportation vehicles and personnel may be inaccessible, urban medical emergency facilities are primarily fixed at hospitals and emergency stations. (4) Social value outweighs material value. The significance of medical emergency supplies extends beyond drug pricing, as their true worth lies in their critical role in saving lives.

2.2. Application Status of Drones in Medical Emergency

Drones have become indispensable in medical emergency response and search-and-rescue operations. Their rapid deployment capabilities and adaptability to diverse geographical environments enable efficient transportation of vaccines, diagnostic reagents, medical devices, automated external defibrillators (AEDs), pharmaceuticals, and blood products. These aerial vehicles also assist in locating victims and aiding medical professionals in assessing injuries, while extending healthcare coverage to remote areas with limited medical resources and safeguarding public health [28]. The integration of drones into medical services not only enhances treatment quality and accessibility but also reduces operational costs, establishing them as a vital supplementary solution for emergency medical distribution [29]. Current research on drone applications in medical rescue primarily focuses on the following key areas.

(1) Personnel Search and Rescue: The deployment of drones in search and rescue operations can significantly reduce response time [30]. Through coordinated missions involving multiple drones, rescue efficiency can be substantially improved [31,32]. Drones also demonstrate strong adaptability to diverse scenarios, including mountainous areas and snowy environments [33]. When integrated with artificial intelligence, drones can effectively detect casualties in various settings [34,35,36]. Furthermore, combining drones with other transportation modes such as vehicles, boats, and motorcycles in joint rescue operations has dramatically shortened search and rescue timelines [37].

(2) Material Delivery: The transportation of medical supplies stands as a primary application scenario for drones in emergency medical rescue. After rescue operations commence, deploying drones significantly reduces response time, with particularly notable time savings observed in congested traffic zones [38]. A Montreal-based medical drone pilot study demonstrated that drone-assisted blood delivery achieved an average response time of 17 min, significantly outperforming ground transportation, which exhibited longer durations and greater variability (25–38 min) [39]. A real-world study on opioid overdose delivery networks revealed that Virginia Beach drones reduced response times by 82% while increasing patient survival rates by over 273% [40]. A Swedish prospective observational study comparing drone delivery of automated external defibrillators (AEDs) with ambulance arrival for suspected out-of-hospital cardiac arrest patients showed that drones outperformed ambulances in 67% of successful deliveries, with a median time advantage of 3 min and 14 s [41]. Consequently, drones are extensively utilized for delivering time-sensitive medical supplies such as emergency medications [42], defibrillators [43], blood products [44,45], vaccines [46], and diagnostic reagents [47].

While the medical supplies listed have varying payload requirements for drones, many time-critical life-saving initial interventions demand relatively low weight and volume. A standard AED weighs approximately 1–2 kg [48,49], a dose of naloxone weighs only a few hundred grams [50], and a unit of blood (plasma/platelets: about 250–300 milliliters) typically weighs less than 0.5 kg [51]. Many vaccines and reagents are also lightweight, often transported in refrigerated containers weighing 5–10 kg [39]. These transportation weights fall entirely within the payload capacity range of many professional-grade drones, as demonstrated by the drone models selected in this study (10–60 kg).

In terms of cost and benefit, while the construction and operation of drone delivery networks require time to reach break-even [52], a comprehensive economic evaluation must consider the substantial financial benefits of accelerated patient care, typically manifested as reduced long-term healthcare burdens and improved economic outcomes. Health economists commonly use metrics such as quality-adjusted life years (QALYs) or statistical life value to monetize health benefits [53]. Although precise application to drone logistics presents challenges, from a societal perspective, investments in faster emergency response may demonstrate exceptionally high cost-effectiveness, even achieving cost savings. Regarding pure logistics costs, there are significant differences between hospital emergency supplies and routine medical supplies. Since emergency deliveries fall outside the daily order processing workflows of logistics centers, commercial pharmaceutical distributors must independently coordinate transport vehicles and delivery personnel. Each shipment requires professionally trained drivers and delivery staff, resulting in persistently high ground transportation costs. In contrast, drone delivery offers advantages in reducing transport vehicle and labor costs [54], as evidenced by cost analysis cases in specific scenarios [46]. This conclusively demonstrates the technical and economic feasibility of drone applications for such tasks. The technology not only optimizes logistics processes but also effectively reduces operational costs through scientifically planned delivery schedules and cost-effective operational models.

(3) Information and Communication: Timely acquisition of post-disaster personnel and environmental conditions is crucial after emergency incidents. In medical rescue operations, drones are employed to transmit critical information and establish medical communication links [55]. Furthermore, medical drones can assist in restoring communication networks in affected areas while transporting medical supplies [56,57,58].

The application of drones in medical scenarios has become increasingly widespread, with numerous case studies demonstrating their use in emergency medical supply distribution [59]. While existing research has explored integrating drones with other transportation modes to compensate for their limitations in flight range and payload capacity, most studies have focused on single-model drone applications. There remains limited research on utilizing multiple drone variants for emergency delivery [60]. Additionally, urban drone delivery operations face challenges such as population density, building distribution patterns, environmental constraints, and numerous restricted flight zones mandated by regulatory policies. How to effectively model these multifaceted urban factors and integrate them with medical drone deployment strategies requires further investigation.

2.3. Site Selection of Medical Emergency Delivery Drone Base Station

The primary objective of medical emergency logistics is to minimize response time and ensure prompt fulfillment of medical rescue needs. Research indicates a strong correlation between the accessibility of medical facilities and mortality rates in emergency medical interventions [61], making the strategic location of medical rescue facilities crucial.

The selection of locations for medical emergency distribution facilities involves multiple influencing factors. Current research often separates considerations such as cost, demand, response time [62,63], and workload due to research limitations, yet these factors are interconnected and require comprehensive evaluation. Additionally, different scenarios in medical emergency logistics necessitate varying decision-making factors. Implementing multi-objective collaborative optimization better aligns with practical emergency drone deployment needs [64]. For instance, humanitarian aid operations consider personnel allocation, energy consumption, cost efficiency, and delivery timelines [65,66]. Furthermore, developing hierarchical planning frameworks [67] to address seasonal demand fluctuations in medical emergency distribution remains an area requiring further exploration [68].

Unlike humanitarian relief operations, urban hospital emergency logistics require balancing multiple factors, while prioritizing rescue success rates, cost-effectiveness and operational efficiency are equally crucial for sustainable drone base station management. Effective deployment involves seasonal adjustments. During peak winter and spring seasons when disease outbreaks increase emergency delivery demands, resources are strategically allocated to ensure order fulfillment. Conversely, during low-risk periods, deployment scales back accordingly. Notably, current research remains limited in integrating hospital emergency orders’ temporal patterns into dynamic drone base station deployment strategies.

3. Methodology

3.1. Problem Description

3.1.1. System Context and Stakeholders

This study addresses a critical logistics gap within the urban medical emergency ecosystem: the rapid and reliable transportation of time-sensitive medical commodities from supply points to healthcare facilities under strict time constraints and amidst urban traffic congestion. The core problem is the strategic and tactical planning of drone logistics network to fulfill this role. Specifically, we focus on determining the optimal locations for a hybrid network of static (permanent) and dynamic (temporary) drone base stations to serve the fluctuating emergency supply demands of urban hospitals.

The system involves multiple stakeholders (see Figure 1). (i) Demand Points (Hospitals): Fixed locations that generate urgent requests for medical supplies (e.g., drugs, blood products, reagents). (ii) Supply and Logistics Providers: Entities (e.g., central pharmacies, blood banks, medical distribution companies) that store and dispatch emergency supplies. (iii) Drone Network Operator: Manages the fleet of drones and the network of base stations. (iv) Regulators: Aviation and urban authorities governing airspace use, safety, and noise.

3.1.2. Core Decision Problem

The problem is a multi-period, multi-objective facility location-allocation problem under uncertainty. Key decisions include (i) where to establish permanent base stations to cover long-term, stable demand hotspots, (ii) when and where to activate temporary take-off/landing points (e.g., on hospital helipads, designated urban areas) in response to predicted short-term demand surges or unexpected events, (iii) which drone type (varying in speed, range, payload) to assign from which station to serve a specific demand, considering real-time constraints.

Upon receiving an urgent order from a hospital, the system decides the fulfillment strategy. If feasible and optimal, a drone is dispatched from the most suitable base station to transport the item directly to the hospital, bypassing ground traffic. For long-term stable demand hotspots, static base stations are strategically deployed as strategic medical logistics hubs. They would store a pre-positioned inventory of high-priority, time-sensitive emergency supplies. To address sudden emergency demand fluctuations, dynamic base station (temporary takeoff/landing zones (TZLs)) is designed, enabling real-time adjustment of drone base station resource allocation. As temporary operational nodes, they often leverage pre-negotiated and pre-approved temporary TZLs at or near hospitals or other secure urban locations (e.g., rooftops, parking lots). These TZLs would not hold significant inventory but would serve as forward launch/retrieval points for drones carrying supplies dispatched from the main hubs or other sources. If drone delivery is infeasible due to weather, airspace restrictions, or capacity limits, the order is routed via traditional ground vehicles. The base stations act as forward-deployed operational hubs, not necessarily holding bulk inventory but enabling rapid last-mile (or last-several-miles) aerial delivery from larger warehouses or between facilities. In modelled system, drones are primarily allocated for inter-facility logistics, transporting urgent medical supplies from strategic supply points (the drone base stations) to receiving hospitals or major medical facilities. Considering the seasonal variations in the quantity, weight, and urgency of medical supplies delivered each time, this paper introduces relevant parameters to investigate their impact on drone operational efficiency.

These decisions must simultaneously optimize for maximum service coverage, minimized average response time, and controlled total system cost, while adhering to constraints such as drone flight range, hospital time windows, airspace regulations, and weather conditions.

3.1.3. Positioning Within Advanced Air Mobility for Emergency Logistics

This work on strategic network design (station location) is a foundational component of the broader Advanced Air Mobility (AAM) framework for emergency medical services [69,70]. It complements related research streams focusing on vehicle scheduling and routing [71], air traffic management at vertiports [72], and the integration of aerial and ground operations. Our model provides the essential “where to place infrastructure” analysis that precedes and informs detailed operational planning studied in those works. The proposed “dynamic-static” hierarchy specifically addresses the spatial and temporal variability of urban medical demand, a challenge that must be resolved for effective AAM deployment in emergency logistics.

3.2. Methodology Framework

This study proposes a multi-objective optimization method based on multi-source data fusion for urban medical emergency logistics drone base station site selection planning (see Figure 2). The methodology framework is as follows:

(1) The system integrates multi-source data through fusion processing, incorporating the following. (1) Emergency logistics data from historical medical distribution incidents targeting hospitals (including time, location, item type, and demand volume). These are urgent deliveries of medications and medical supplies needed to maintain hospital pharmacy stocks for emergency and routine care, especially for time-sensitive treatments. It also includes ad hoc requests triggered by unexpected surges in demand (e.g., due to a multi-casualty incident or stock depletion of a critical drug). (2) Urban geographic information such as road networks, population density, and building distribution. (3) Drone airspace constraints (including policy-designated no-fly zones and altitude restrictions in the study area). (4) Weather data like rainfall intensity and wind speed. Through data cleansing, spatiotemporal features and environmental constraints were extracted as model inputs.

(2) Analyze the spatiotemporal characteristics of emergency needs by employing kernel density estimation methods based on Geographic Information Systems (GIS) and the DBSCAN spatiotemporal statistical model to identify spatiotemporal clustering patterns and hotspot areas in rescue operations.

(3) Developing a Multi-Objective Optimization Model. This study establishes a “dynamic-static” two-tier site selection framework. For long-term stable demand hotspots such as large hospitals, a fixed-base station layout optimization model is developed to ensure coverage in high-frequency rescue zones. For short-term fluctuating demands, a temporary takeoff and landing point scheduling strategy is designed using kernel density estimation methods, with real-time data updates to dynamically adjust resource allocation. Subsequently, a multi-objective optimization model is constructed, with objective functions including minimizing response time, maximizing coverage, and minimizing cost budget. The model also incorporates constraints such as drone flight radius, airspace limitations, budget upper/lower bounds, and hospital service time requirements.

(4) The proposed solution employs an enhanced multi-objective genetic algorithm, utilizing non-dominated sorting and reference point guidance to generate Pareto-optimal solutions that balance trade-offs among objectives. A dynamic weighting mechanism is introduced, combined with Long Short-Term Memory (LSTM) algorithm predictions for emergency logistics demand, which are embedded into the optimization process to enhance the model’s adaptability to dynamic demands.

(5) The model undergoes validation and sensitivity/robustness analysis. Through simulation experiments, its effectiveness is verified. Using large cities as case studies, the research compares rescue efficiency of different site selection strategies through various drone deployment configurations. Evaluation metrics include coverage rate, average response time, and cost-effectiveness ratio. Sensitivity analysis assesses the impact of key parameters on site selection outcomes. Additionally, the study examines how weather conditions, airspace restrictions, and demand fluctuations affect the model’s robustness.

3.3. Identification of Hot Spots in Medical Emergency Logistics Demand Space

To identify spatial hotspots for medical emergency logistics needs, this study employs Geographic Information Systems (GIS) and kernel density estimation to analyze distribution patterns. As a non-parametric method that operates directly on raw data without assumptions, kernel density estimation demonstrates strong adaptability for spatial distance or connectivity-based analyses, making it a widely adopted approach across various fields.

Two-dimensional kernel density estimation methods are widely used in spatial feature analysis. In spatial clustering analysis, the dataset for kernel density estimation is two-dimensional. The principle involves selecting an appropriate kernel function and calculating the kernel density estimate for each pixel within the raster. The formula is as follows:

$$f\left(w_{ci}\right)={\displaystyle \frac{1}{n{h}^2}}{\sum }_{j=1}^{n}K\left({\displaystyle \frac{D(w_{ci},w_{cj})}{h}}\right)$$

(1)

where $f\left(w_{ci}\right)$ denotes the kernel density estimate at position $w_{ci}$, $\mathrm{h}$ represents the bandwidth, $D(w_{ci},w_{cj})$ indicates the distance between positions $w_{ci}$ and $w_{cj}$, $n$ is the number of observed points within bandwidth, and $K$ is the kernel function as defined in Equation (2).

$$K\left(u\right)={\displaystyle \frac{1}{\sqrt{2\pi }}}{e}^{-{\displaystyle \frac{1}{2}}{u}^2}$$

(2)

Through analysis, the spatial distribution hotspots of emergency distribution events were identified, and the spatial units were divided into cold spots, warm spots and hot spots according to the density value, and the kernel density estimation value was taken as an important input for the subsequent emergency drone base station site selection.

3.4. Drone Base Station Site Selection Optimization Model

3.4.1. Model Assumptions

(1) The spatial and temporal distribution of historical emergency logistics demand is quantified by the prediction model, and the weight $W_i$ is determined by the kernel density estimation, demand quantity and shipment interval weighting.

(2) Static base station candidate points $B_{static}$ are selected by integrating DBSCAN spatial–temporal clustering centers, while avoiding no-fly zones and height restriction areas.

(3) Dynamic layer base stations $B_{dynamic}$ can be activated or deactivated based on real-time requirements.

(4) The total construction budget $C_{max}$ is proportionally allocated to the static layer $\lambda C_{max}$ and dynamic layer $(1-\lambda )C_{max}$.

(5) The single flight radius of the drones and the response time must $R\le {\displaystyle \frac{V\ast T_{maxfly}}{2}}{,t}_{ij}^m\le T_{max}$. The model simplifies the energy consumption of drones by not distinguishing between different modes of energy consumption with and without payload.

3.4.2. Parameter and Symbol Definitions

Table 1. Parameter and symbol definitions.

Symbol	Definition
$i$	The medical emergency logistics demand point
${lon}_i$	The longitude of the medical demand point
${lat}_i$	The latitude of the medical demand point
$j$	Indicates a drone base station
${lon}_j$	Indicates the longitude of the drone base station
${lat}_j$	Indicates the latitude of the drone base station
$R^m$	Indicates the radius of the type $\mathrm{m}$ drone
$V^m$	Indicates the average speed of the type $\mathrm{m}$ drone
$E_{max}^m$	Indicates the maximum energy capacity of the type $\mathrm{m}$ drone
${EP}^m$	Indicates the energy consumption per kilometer of drone transportation.
$D$	Emergency demand points collection total number is $\mathrm{N}$
$M$	The set of transportation options includes a variety of types of drones, $\mathrm{m}\in \mathrm{M}$.
$y_i$	Indicates whether the drone covers the demand point $\mathrm{i}$
$x_j^{nominal}$	Normal base station deployment plan
$P_i$	The priority factor of the demand point $\mathrm{i}$ actually represents the urgency of the demand
${\eta }^m$	The unit time energy consumption unit price of the type $\mathrm{m}$ drone
${\xi }_j$	The social cost of base station $\mathrm{j}$
$Q_t$	Real-time demand forecast for the time window $\mathrm{t}$
$B_{static}$	Static layer candidate base station set
$B_{dynamic}$	Dynamic layer candidate base station set
$t_{ij}^m$	Flight time of type $\mathrm{m}$ drone from the base station $\mathrm{j}$ to the demand point $\mathrm{i}$
$W_i$	The weight of demand points $\mathrm{i}$ is determined by a combination of kernel density values, required quantities, and shipping time intervals.
$T_{maxfly}$	Maximum flight time of the drone
$T_{max}$	The maximum allowed response time threshold is the required arrival time for the hospital
$c_j$	Base station $\mathrm{j}$ construction and one year operation costs
$x_j\in \left\{\mathrm{0,1}\right\}$	Decision variables indicate ${\mathrm{x}}_{\mathrm{j}}=1$ the construction of base stations at locations $\mathrm{j}$
$\lambda$	Static layer budget allocation ratio
$y_{i_{static}}\in \left\{\mathrm{0,1}\right\}$	Indicates the variable, which ${\mathrm{y}}_{{\mathrm{i}}_{\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{c}}}=1$ is the demand point is covered by the static base station and 0 otherwise.
$proximity$	The spatial distance function between the base station $\mathrm{j}$ and the restricted airspace.

summarizes the definitions, symbols, and units of the key parameters and decision variables used in the proposed two-stage optimization model. The contents are categorized into three main groups: Decision Variables, which denote the unknowns to be solved by the model; Input Parameters, including static facility parameters, dynamic demand parameters, drone performance parameters, and environmental constraint parameters, which serve as the known conditions for the model; and Key Performance Indicators, which are used to evaluate the solution’s performance.

3.4.3. Objective Function

Static layer target:

(1) Maximize the coverage of static layer. The weight $W_i$ is determined by the historical data kernel density value, the demand quantity and the time interval, $P_i\in \left\{\mathrm{1,2}\right\}$, which represents emergency orders categorized into two urgent types: “Emergency” and “Urgent”, where both order types must be within $T_{max}$.

$$max{\sum }_{i\in D}{p_i\ast w}_i\ast y_{i_{static}}$$

(3)

$y_{i_{static}}\in \left\{\mathrm{0,1}\right\}$, an indicator variable that when $y_{i_{static}}=1$ represents the demand point $\mathrm{i}$ is covered by a static base station and 0 otherwise. The coverage condition is:

$$y_{i_{static}}=\{\begin{array}{c}1,\exists j\in , t_{ij}^m\le T_{max} and x_j=1\\ 0\end{array}$$

(4)

The comprehensive weight $w_i$ calculation method is:

$$w_i=w_{kenel}\ast w_{demand}\ast w_{time}$$

(5)

where $w_{kenel}$ represents the kernel density weight, the demand weight $w_{demand}$, and the time interval weight $w_{time}$.

$$w_{kenel}=Demand point kernel density value/max(kernel density values)$$

(6)

$$w_{demand}=quantity demanded/max(quantity demandeds)$$

(7)

$$w_{time}=1-(shiping interval-min(shiping interval\left)\right)/\left(max\right(shiping interval)-min(shiping interval)$$

(8)

(2) Minimization Static layer cost. The cost involves capital and operational expenditure. $c_j$ represent one-time costs for drone procurement and base station construction/setup (including landing pad installation, basic shelter, charging infrastructure, and initial site preparation). Recurring costs include routine maintenance for drones and stations, energy consumption for drone flights (incorporated via ${\eta }^m$), insurance, and personnel costs for network monitoring and basic management. Here, it is simplified to the price per kilometer. A proxy cost ${\xi }_j$ representing expenses related to securing landing permissions, community engagement, noise mitigation, and environmental impact assessments at base station locations guides the model towards selecting base station locations with better overall socio-community acceptance and composite benefits, beyond mere financial costs.

$$min{\sum }_{j\in B_{static}}{(c}_j{x}_j+{\xi }_j)+{\sum }_{i\in D}{\sum }_{j\in B_{static}}{\sum }_{m\in M}{{\eta }^{m}t}_{ij}^m$$

(9)

Dynamic layer target:

(1) Minimize dynamic response time. The dynamic base station, as a temporary take-off and landing point, selects a suitable drone for delivery based on the urgency of orders represented by $w_i$, aiming to achieve the shortest response time.

$$min{\sum }_{i\in \mathrm{D}}{\sum }_{j\in B_{dynamic}}{\sum }_{m\in M}{t}_{ij}^m{w}_i{x}_j$$

(10)

(2) Maximize dynamic coverage supplement rate. ${\mathrm{y}}_{{\mathrm{i}}_{\mathrm{d}\mathrm{y}\mathrm{n}\mathrm{a}\mathrm{m}\mathrm{i}\mathrm{c}}}\in \left\{\mathrm{0,1}\right\}$, an indicator variable that when ${\mathrm{y}}_{{\mathrm{i}}_{\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{c}}}=1$ represents the demand point $\mathrm{i}$ covered by the dynamic base station and 0 otherwise. The coverage condition is $1-{\mathrm{y}}_{{\mathrm{i}}_{\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{c}}}$, meaning the dynamic layer only contributes to the objective function when the demand point $\mathrm{i}$ is not covered by the static layer, demonstrating its targeted supplementation of the static layer’s blind spots.

$$max{\sum }_{i\in \mathrm{D}}\left(1-y_{i_{static}}\right)\ast y_{i_{dynamic}}$$

(11)

$$y_{i_{dynamic}}=\{\begin{array}{c}1,\exists j\in B_{dynamic},t_{ij}^m\le T_{max} and x_j=1\\ 0\end{array}$$

(12)

3.4.4. Constraint Condition

(1) Coverage constraints: Due to the limited response time of emergency distribution demand, only one drone model is selected for each demand point to cover, without transport relay and transfer.

$$\sum _{m\in M}{y}_i^m\le 1,\forall i\in D$$

(13)

(2) Dynamic Demand Adaptation: Real-time demand forecasting drives base station activation, where ${\mathrm{Q}}_{\mathrm{j}}$ represents the theoretical service capacity of the base station $\mathrm{j}$, ${\tau }_{\mathrm{j}}\left(\mathrm{t}\right)$ represents the base station $\mathrm{j}$ service efficiency under weather warnings during the time period $\mathrm{t}$, ${\delta }_{\mathrm{i}}$ represents the demand type weight (divided into emergency medical and super emergency medical delivery categories), $\alpha$ represents the seasonal fluctuation factor of demand, and ${\mathrm{Q}}_{\mathrm{t}}$ represents the predicted demand volume for the time period $\mathrm{t}$.

$$\sum _{j\in B}{x}_j\ast Q_j\ast {\tau }_j(t)\ge \sum _{i\in D}{\delta }_i\ast \alpha \ast Q_t,\forall t\in T$$

(14)

(3) Multiple drone type selection: Automatically selects the drone model according to the characteristics of drug delivery, and chooses to use car transportation when the drone cannot be used within the time limit of transportation.

$$t_{ij}^m\le T_{max},\forall i\in D,j\in B,m\in M$$

(15)

(4) Energy efficiency: The endurance constraint of the selected model of drone should be less than the maximum energy capacity of drone.

$$2{\ast {EP}^m\ast t}_{ij}^m\le E_{max}^m,\forall i\in D,j\in B,m\in M$$

(16)

(5) Robustness constraints: Redundant base station deployment under extreme events.

$$\sum _{j\in B}{x}_j\ge \beta \ast \sum _{j\in B}{x}_j^{nominal},\forall i\in D,j\in B,m\in M$$

(17)

(6) Social–environmental constraints limit the expansion of base stations in specific regions, increasing social penalties and transportation noise costs. ξj represents expenses related to securing landing permissions, community engagement, noise mitigation, and environmental impact assessments at base station locations; the total estimated social cost $Proximity(j,Restricted areas)$ generated by all base stations is summed and subjected to a budget limit in the objective function.

$${\xi }_j=\gamma \ast Proximity(j,Restricted areas)$$

(18)

Dynamic weights and priority mechanism

(1) Weight update rules

$$\alpha ,\beta ,\gamma =f(demand side fluctuation,weather warning,policy changes)$$

(19)

The time fluctuation of emergency logistics $\Delta {\mathrm{Q}}_{\mathrm{t}}$ demand is:

$$\Delta Q_t={\displaystyle \frac{\left|Q_t-\overline{Q_h}\right|}{{\sigma }_h}}$$

(20)

where ${\mathrm{Q}}_{\mathrm{t}}$ represents the forecasted demand for the current period, $\overline{{\mathrm{Q}}_{\mathrm{h}}}$ represents the historical average for the same period, and ${\sigma }_{\mathrm{h}}$ represents the historical demand standard deviation.

Environmental and airspace control interference ${{\rm E}A}_t$ factors are:

$${{\rm E}A}_t={\omega }_{weather}·{\tau }_t+{\omega }_{air}·A_t$$

(21)

${\tau }_{\mathrm{t}}$ denotes the weather warning level ${\tau }_{\mathrm{t}}\in \left[\mathrm{0,1}\right]$. ${\mathrm{A}}_{\mathrm{t}}$ indicates the airspace control intensity ${\mathrm{A}}_{\mathrm{t}}\in \left[\mathrm{0,1}\right]$, ${\mathrm{A}}_{\mathrm{t}}=0.9$ for areas with severe restrictions (e.g., near airports) and ${\mathrm{A}}_{\mathrm{t}}=0.$1 for mild restrictions, based on buffer distances to no-fly zone centroids. ${\omega }_{\mathrm{w}\mathrm{e}\mathrm{a}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{r}}$ and ${\omega }_{\mathrm{a}\mathrm{i}\mathrm{r}}$ represents the weighting coefficients for weather and airspace restrictions $\varrho \in \left[\mathrm{0,1}\right]$, ${\omega }_{weather}=0.7$, ${\omega }_{air}=0.3$, assuming weather has a stronger immediate operational impact.

$${\tau }_t(WF,RF)=\varrho \ast f\left(WF\right)+(1-\varrho )g\left(RF\right)$$

(22)

$f\left(WF\right)$ represents the wind effect function:

$$f\left(WF\right)={\displaystyle \frac{1}{1+e^{-k_1(wf-w_0)}}}$$

(23)

$wf$ indicates the wind speed required for demand, and the risk threshold for wind speed $w_0$. A Beaufort Force 5 wind (29–38 km/h, 8–10.7 m/s) is set as the nominal risk threshold in the wind effect function, beyond which the risk will increase sharply. $k_1$ indicates the slope parameter, which controls the rate of risk growth. $k_1=0.5$ (slope for wind risk), $w_0=8.5$ m/s (Beaufort Force 5 threshold).

The rainfall effect function is:

$$g\left(RF\right)=min(1,c\ast R^d)$$

(24)

$R$ represents the rainfall grade; $c,d$ is a regulating parameter, for which we set $\mathrm{c}=0.3$, $\mathrm{d}=1.5$, and use the power function to simulate the accelerated growth of meteorological risk caused by the change of rainfall grade.

$$A_t=\{\begin{array}{c}1,Airspace restrictions\\ 0,Airspace prohibited\end{array}$$

(25)

$${\omega }_{weather}+{\omega }_{air}=1$$

(26)

Because the intensity of tasks varies at different times of the day, a time-sensitive coefficient $S_t$ is introduced.

$$S_t=\{\begin{array}{c}2,Emergency peak hours\\ 1.5,Regular work hours\\ 0.8,Night and holidays\end{array}$$

(27)

The dynamic weight $\alpha ,\beta ,\gamma$ update function is:

$${\alpha }_t={\alpha }_0·(1+\ln (1+\Delta {\mathrm{Q}}_{\mathrm{t}}))·S_t$$

(28)

$${\beta }_t={\beta }_0·(1+\Delta {\mathrm{Q}}_{\mathrm{t}}·S_t)/(1+{EA}_t)$$

(29)

$${\gamma }_t={\gamma }_0·(1+{EA}_t)·exp(-Q_t)$$

(30)

The initial weight is set to ${\alpha }_0,{\beta }_0,{\gamma }_0\left[\mathrm{0.4,0.4,0.2}\right]$, and the normalization $\alpha ,\beta ,\gamma$ is performed as follows:

$$\{\begin{array}{c}\alpha ={\alpha }_t/({\alpha }_t+{\beta }_t+{\gamma }_t)\\ \beta ={\beta }_t/({\alpha }_t+{\beta }_t+{\gamma }_t)\\ \gamma ={\gamma }_t/({\alpha }_t+{\beta }_t+{\gamma }_t)\end{array}$$

(31)

(2) Priority layering coverage

$$y_i^m=\{\begin{array}{c}1,if{p}_i\ge 2 and t_{ij}^m\le T_{max}\\ 0, otherwise\end{array}$$

(32)

3.4.5. Algorithm Selection and Implementation

(1) Algorithm selection

To address the multi-objective optimization problem of drone base station site selection, this study employs the improved Non-Dominating Sorting Genetic Algorithm (NSGA-III) [66]. By introducing a reference point mechanism, NSGA-III effectively tackles high-dimensional multi-objective optimization problems while outperforming traditional algorithms in maintaining solution diversity and convergence. The specific selection rationale is as follows. (1) The model requires simultaneous optimization of three objectives: coverage, response time, and cost. NSGA-III’s population evolution guided by reference points effectively balances multi-objective trade-offs. (2) The dynamic weighting mechanism allows adjusting objective function priorities based on real-time demand predictions, enhancing the model’s responsiveness to emergencies. (3) The improved NSGA algorithm supports integrated constraints such as budget limitations and airspace regulations, filtering feasible solutions through constraint dominance principles. Furthermore, this paper uses NSGA-II as a baseline model to compare key performance metrics between NSGA-III and NSGA-II in computational efficiency.

To forecast real-time demand, this study employs Long Short-Term Memory (LSTM) networks to analyze emergency logistics orders delivered to hospitals by a pharmaceutical distribution company over the past year, aiming to predict hospital order demands across different time periods. Compared to traditional backpropagation (BP) neural networks, LSTM demonstrates significant advantages in time series prediction. By introducing gating mechanisms, it effectively resolves gradient vanishing and exploding issues, enabling more accurate capture of long-term dependencies in temporal data.

(2) Algorithm Implementation

(i) Genetic algorithm

Step 1. Initialize static layer candidate solutions by generating an initial base station location set based on spatiotemporal cluster centers, ensuring candidate points are located in demand hotspots while avoiding no-fly zones. For dynamic layer candidate solutions, randomly generate initial mobile base stations by incorporating real-time hotspot density values. The encoding method employs binary encoding, where each gene bit indicates whether to deploy a base station at the candidate location.

Step 2. Normalize coverage, response time, and cost functions to eliminate dimensional differences. The weighted fitness calculation is as follows:

$$Fitness=\alpha Coverage+\beta {\displaystyle \frac{1}{responsetime}}+\gamma {\displaystyle \frac{1}{cost}}$$

(33)

The weight coefficients ($\alpha ,\beta ,\gamma$) are adjusted through a dynamic weighting mechanism to respond to real-time demand fluctuations.

Step 3. Perform genetic operations using tournament selection, prioritizing non-dominated solutions and high-diversity individuals. Apply simulated binary crossover (SBX) with a crossover probability of 0.9 to ensure global search capability. Implement polynomial variation with a mutation probability of 0.1 to enhance local search capability [73].

Step 4. The Das and Dennis method was used for generating reference points in a 3-objective space. Generate reference points uniformly distributed across the target space to ensure solution diversity. Individuals are associated with the nearest reference point, balancing convergence and distribution through niche technology.

Step 5. Dynamically adjust weights by incorporating parameters such as kernel density estimation, time, and demand volume to update weight coefficients.

Step 6. Process model constraints by applying penalty functions to eliminate solutions exceeding the budget, and introduce a flight prohibition zone distance penalty in the fitness function to ensure path legality.

Step 7. Set the termination criteria: Either limit the model to 100 iterations, or halt when the Pareto front shows no significant changes for 50 consecutive generations, to balance diversity and computational cost for our problem scale.

Step8. Select a trade-off solution with high coverage, short response time, and moderate cost from the Pareto frontier, and finalize the decision based on the decision-maker’s preferences.

(ii) LSTM algorithm

Enter the current state information of the memory gate, calculate ${\mathrm{i}}_{\mathrm{t}}$ the value of the input gate and the candidate state value ${\tilde{C}}_t$ of the input cell at the moment $\mathrm{t}$.

$$i_t=\sigma ({\omega }_i·[h_{t-1},x_t]+b_i)$$

(34)

$${\tilde{C}}_t=tanh({\omega }_c·[h_{t-1},x_t]+b_c)$$

(35)

where ${\omega }_i{,\omega }_c$ represents the corresponding weight and $b_i, b_c$ represents the corresponding bias.

Forget the information of the forget gate, and calculate the activation value $f_t$ of the forget gate at the moment $t$.

$$f_t=\sigma ({\omega }_f·[h_{t-1},x_t]+b_f)$$

(36)

where ${\omega }_f,b_f$ represents the weight and bias of the forget gate, and $\sigma$ denotes the sigmoid function.

Update cell state: Update the cell state based on the calculated results of the input gate and forget gate, and obtain the cell state update value $C_t$ at the current time $t$:

$$C_t=i_t\odot {\tilde{C}}_t+f_t\odot C_{t-1}$$

(37)

The output gate outputs the information and updates the calculated cell state value $C_t$, resulting in the calculation formula of the output gate $o_t$:

$$o_t=\sigma \left(W_O·\left(h_{t-1}·X_t\right)+b_O\right)$$

(38)

$$h_t=o_t\odot tanh\left(C_t\right)$$

(39)

In the formula, $W_O{,b}_O$ represents the weight and bias of the output gate; ${\mathrm{h}}_{\mathrm{t}}$ denotes the output value of the current unit. A two-layer stacked LSTM with 128 units in each layer was employed for demand forecasting, followed by a Dropout layer (dropout rate = 0.2) and a Dense output layer. The model was trained on 70% of the chronological data, validated on 15%, and tested on 15%. We utilized the Adam optimizer with a learning rate of 0.001, Mean Squared Error (MSE) as the loss function. The model was trained for 200 epochs, with early stopping triggered by a plateau in the validation loss. The input sequence length was 7 days, and the model was designed to predict the demand for the next day. The features incorporated lagged demand values and day-of-week indicators.

By integrating the LSTM method, we can efficiently process large volumes of historical demand data. Using historical demand as input, this approach enables prediction of future emergency logistics needs in hospitals, identifies high-demand facilities, and provides references for selecting optimal locations for emergency logistics drone base stations, determining appropriate drone models, and allocating related resources.

4. Case Study

4.1. Study Area

This study selects Guangzhou, a mega city in China, as the research area. The selection is mainly based on its representativeness as a typical ultra-high-density city, the uneven spatial distribution of medical resources, the complexity of low-altitude airspace management, and the diversity of meteorological conditions.

As the capital of Guangdong Province and a national central city, Guangzhou is densely populated and economically active, serving as an important transportation hub and logistics center in southern China. The city administers 11 districts with a total area of 7434.4 square kilometers. According to data from the Guangzhou Bureau of Statistics, the permanent population of Guangzhou reached 18.978 million by the end of 2024, with an urbanization rate of 87.24%. The high urbanization level and significant population mobility exhibit typical characteristics of a megacity [74]. While Guangzhou boasts abundant medical resources, the spatial distribution of these resources within the city is uneven. According to the 2023 “Guangzhou Health Statistics Yearbook” (see Appendix A 

Table A1. The number of medical institutions in Guangzhou.

Organization Classification	Total	Lai Wan	Yue Xiu	Hai Zhu	Tian He	Bai Yun	Huang Pu	Pan Yu	Hua Du	Nan Sha	Cong Hua	Zeng Cheng
Total	6677	291	398	552	1075	1021	493	649	698	321	410	769
1.hospital	331	28	34	22	69	67	24	28	17	16	9	17
general hospital	162	13	11	12	29	31	14	17	7	12	6	10
hospital of traditional Chinese medicine	44	5	7	4	6	12	4	3	0	1	1	1
hospitals of traditional Chinese and Western medicine	4	0	0	1	0	0	0	1	2	0	0	0
special hospital	103	8	15	5	30	17	5	7	7	2	2	5
nursing home	18	2	1	0	4	7	1	0	1	1	0	1
2. Primary health care institutions	6132	257	334	521	994	946	385	602	672	294	387	740
Community Health Service Center	163	20	18	18	26	19	16	16	11	9	4	6
Community health service station	160	12	0	21	16	24	43	28	3	8	1	4
health clinics in towns and townships	31	0	0	0	0	4	1	0	8	0	8	10
Village clinic	942	0	0	0	0	116	28	0	195	110	222	271
ambulant clinic	1928	95	134	183	451	311	102	255	131	62	36	168
Clinic, health post, clinic, nursing station, etc	2908	130	182	299	501	472	195	303	324	105	116	281
3. Professional public health institutions	67	3	12	5	7	5	3	10	5	4	6	7
Centers for Disease Control and Prevention	16	1	2	1	3	2	1	2	1	1	1	1
Specialized disease prevention and treatment institutions	7	0	2	1	1	0	0	2	0	0	0	1
Health Education Institutions	4	0	1	0	1	0	0	1	0	0	1	0
maternity and child care institution	12	1	2	1	1	1	1	1	1	1	1	1
Emergency Center (Station)	8	0	1	0	0	1	0	1	1	1	1	2
Blood collection and supply institutions	5	0	1	0	0	0	0	1	1	0	1	1
Health surveillance agencies	14	1	3	2	1	1	1	1	1	1	1	1
Family planning technical service institutions	1	0	0	0	0	0	0	1	0	0	0	0
4. Other health institutions	147	3	18	4	5	3	81	9	4	7	8	5
Rehabilitation medical institutions	10	1	1	2	1	0	2	0	1	0	1	1
Medical research institutions	6	0	6	0	0	0	0	0	0	0	0	0
Medical on-the-job training institution	0	0	0	0	0	0	0	0	0	0	0	0
Clinical laboratory	37	0	2	0	1	1	28	3	0	2	0	0
Statistics Center	2	0	1	0	0	0	0	0	0	0	1	0
others	92	2	8	2	3	2	51	6	3	5	6	4

) released by the Guangzhou Health Commission, the city had 6677 medical institutions by the end of 2023, including 331 hospitals, with over 40% concentrated in core areas such as Tianhe, Yuexiu, and Baiyun districts, while suburban and emerging development zones have significantly lower medical resource availability. This disparity between core and peripheral areas provides an ideal scenario for testing whether drone systems can effectively bridge resource gaps and improve the equity of urban services. The airspace environment in Guangzhou is complex, with strict airspace control. It includes multiple no-fly zones such as large international airports, military management areas, the Central Business District (CBDs), and scenic spots, resulting in highly fragmented airspace. Under these stringent constraints, the model validation demonstrates its feasibility in real-world policy environments. Additionally, Guangzhou’s subtropical monsoon climate features frequent rainfall, summer typhoons, and occasional winter cold waves, creating complex meteorological conditions that provide a testing ground for further validating the robustness of the drone network in real-world environments.

Guangzhou, a city characterized by high demand pressure, uneven resource distribution, strict airspace restrictions, and complex meteorological conditions, serves as a prime example for evaluating the effectiveness, robustness, and practical potential of the “dynamic-static” collaborative site selection model proposed in this study. By focusing on Guangzhou as a case study, this research not only validates the efficacy of the drone base station location model but also provides valuable references for other major cities developing similar medical emergency logistics systems.

4.2. Data Sets and Data Processing

(1) Historical medical emergency logistics data: This study collaborated with a large pharmaceutical distribution company, selecting 11,973 emergency logistics orders placed by the company in Guangzhou, China throughout 2023 (1 January to 31 December) for hospitals. During each trip, only one order can be delivered, by 1–2 dedicated drivers (sometimes, more people with a pharmaceutical background are needed) and vehicle. The data includes parameters such as customer name, delivery and pickup locations, order time, dispatch time, time requirements, transportation method, quantity, product name, category, unit price, total amount, quantity, and weight. Due to the cross-city air transport nature of medical emergency logistics orders and the impact of urgent demand, some spatial ranges of the data do not align with this study. Additionally, discrepancies in the timeliness of order content maintenance and information completeness have led to non-standardized data in emergency logistics. Therefore, this paper has cleaned and processed the obtained emergency logistics data, while also anonymizing personal medical privacy information such as customer names to ensure the non-use of personal data. The data cleaning pipeline involved the following. (i) Temporal Filtering: Retained only orders marked as “emergency” or “urgent” by the logistics provider. (ii) Spatial Filtering: Excluded orders with origin or destination outside the administrative boundary of Guangzhou City. (iii) Completeness Check: Removed records with missing critical fields (order timestamp, hospital location coordinates, product weight/quantity). (iv) Anonymization: All personally identifiable information such as patient names and specific hospital codes were irreversibly anonymized to preserve privacy. For the few records (<0.1%) with missing product weight, we applied a median imputation strategy based on the weight distribution of other orders for the same product category.

After categorizing by transportation timeliness, spatial range, and method, and after data standardization and anonymization, the total number of emergency logistics orders is 9415 (see Figure 3a). (1)Among them, injections accounted for the largest share at 42.3% of total usage, followed by tablets/capsules (25.42%), other medications (including antiviral, immune repair, cancer treatment, and cosmetic therapies) (19.25%), implantable medical devices (11.61%), and biologics (1.27%) (see Figure 3b). Products weighing 0–5 kg represented approximately 70.24% of the total volume, while those under 20 kg accounted for 92.74% (see Figure 3c). The value of these goods was 7–8 times higher than that of general pharmaceutical shipments. (2) Urban geographic information such as road networks, population density, and building distribution. (3) Drone airspace constraints (including policy-designated no-fly zones and altitude restrictions in the study area). The simulated emergency movements primarily involve the urgent transportation of pharmaceuticals and medical supplies between distribution hubs/central pharmacies and these hospitals (i.e., inter-facility logistics), as well as urgent replenishment from central warehouses to hospital pharmacies.

(2) Urban geographic information data: The urban geographic information data mainly originates from open-source GIS platforms, covering information such as the urban road network, population density, and building distribution of Guangzhou City. This study cleans the acquired urban road network, population density, and other data by removing duplicate, erroneous, or missing records, standardizing the data format to the standard GIS format, and then converting it into a unified WGS-84 spatial coordinate system to ensure compatibility and accuracy. The spatial data is also processed through grid division, and a city spatial grid model is constructed by integrating road network, population density, and other data, with each grid cell containing attributes such as road network density. Then, the KDE method is used to quantify the spatial distribution density of emergency delivery events and generate a hotspot map, while the DBSCAN clustering algorithm is employed to identify the spatiotemporal clustering characteristics and hotspot areas of rescue events. For spatiotemporal DBSCAN clustering, we treated latitude, longitude, and time as the three dimensions; the Eps tested in a range from 500 m to 2000 m (spatial) and 2 to 8 h (temporal). The final values of Eps_spatial = 1000 m and Eps_temporal = 4 h were chosen. The MinPts was tested between 5 and 20, and a final MinPts = 10 was selected to ensure clusters were meaningful.

(3) Meteorological data: The meteorological data is the historical meteorological data of Guangzhou City in 2023, including temperature, wind speed, wind Beaufort force and direction, precipitation, etc., obtained through the public dataset of China’s National Meteorological Administration (see Figure 4).

(4) Drone airspace restriction data: The drone airspace restriction data originates from airspace control rules issued by aviation authorities. According to China’s “Regulations on the Management of Drone Piloted Aircraft Flight”, the drone-controlled airspace includes areas above 120 m true altitude, air forbidden and restriction zones and their surrounding airspace, military ultra-low altitude flight airspace. Also, airspace above specific areas such as airports shall be designated as controlled airspace (see

Table 2. The drone-controlled airspace [75].

Number	Spatial Scale
(1)	The airport and the surrounding area;
(2)	The area within a certain range on our side of the boundary line, the line of actual control and the border line;
(3)	Military restricted zones, military management zones, places of supervision and other classified units and the surrounding areas;
(4)	Key military facilities protection areas, nuclear facilities control areas, areas for the production and storage of inflammable and explosive dangerous goods, and large storage areas for inflammable important materials;
(5)	Power plants, substations, filling (gas) stations, water supply plants, public transport hubs, aviation and electricity hubs, major water conservancy facilities, ports, expressways, railway electrified lines and other public infrastructure, as well as the surrounding area and drinking water source protection areas;
(6)	Radio observatories, satellite measurement and control (navigation) stations, aeronautical radio navigation stations, radar stations and other facilities requiring special protection of electromagnetic environment, as well as the surrounding area;
(7)	Important revolutionary sites, important immovable cultural relics and the surrounding area;
(8)	Other areas specified by the national air traffic management authority.

).

By combining relevant restrictions, the geographical boundary data of no-fly zones, restricted flight zones, and altitude restriction areas within Guangzhou is delineated. The geographical boundary data of no-fly zones and strengthened warning area is converted into a GIS-recognizable format. Through GIS technology, the airspace restriction data is overlaid with urban geographic information data to generate airspace restriction boundary data (see Figure 5).

4.3. Main Parameter Settings of the Model

This study examines three types of logistics drones and one automobile transportation method, as shown in Figure 6 and

Table 3. Logistics drone brand and parameters.

Drone Model Parameters	DJI FlyCart100	Fengzhou 90	Ark 40
velocity	20 m/s	30 m/s	14 m/s
air-range	12 km	70 km	20 km
load	30 kg	20 kg	10 kg
Number of packages per shipment	6	2	1
Suitable transport distance	6 km	35 km	10 km
Wind resistance level	6	6	7
Water resistance level	IP55	IP55	IP54

. Considering the energy consumption of the drone’s return trip, the suitable transport distance for drones is half of their air-range. In compliance with Guangzhou’s airspace management regulations, we delineated no-fly zones and conducted site selection research for drone base stations. We used Python 3.8 for coding and computational analysis, NumPy, pandas, scikit-learn for data processing; GeoPandas, scipy for spatial analysis; TensorFlow for LSTM; pymoo for NSGA-III implementation, and the operating system is Windows 11. The study evaluates the coverage of urban emergency medical needs, operational costs, and response times under different $T_{max}$ base station locations. Key experimental parameters include a maximum coverage radius of 20 km for drone base stations, a city-wide budget of USD 10 million, and a maximum response time of 3600 s (1 h) for hospitals.

5. Results

5.1. Spatiotemporal Characteristics of Hospital Emergency Logistics

Hospital emergency logistics demand exhibits distinct spatiotemporal heterogeneity. Spatially, high-demand areas cluster along transportation corridors and at urban–rural transition zones like Guangzhou’s Tianhe District CBD, reflecting the spatial coupling between population density and economic activity (see Figure 7a). The temporal pattern follows a “winter-spring double peak, monthly peak, midweek trough, and daytime double peak” cycle, closely aligned with disease prevalence seasons, healthcare institution work rhythms, and patient visitation behaviors (see Figure 7b–e).While suburban areas and emerging development zones currently show lower demand density, their underdeveloped infrastructure may position them as key targets for future drone coverage expansion.

5.2. Drone Base Station Site Selection Results

By integrating “dynamic-static” drone base station coverage with complementary vehicle transportation, the model effectively met emergency logistics demands in Guangzhou. The medical emergency coverage achieved 100% fulfillment, with 11 static drone base stations and 1 dynamic base station selected (see Figure 8). This significantly outperformed the 1.68 h average transportation time of vehicles. The optimized network achieves a drone coverage rate of 96.18% with an average response time of 673.38 s (≈11.2 min), reducing the average delivery time by approximately two-thirds compared to ground vehicles. This demonstrates the potential for drones to dramatically compress the logistical delay within the critical ‘golden hour’ for emergency care.

The optimal solution set reveals a clear trade-off among budget cost, drone coverage rate, and response time. Solutions with higher coverage rates incur higher costs but achieve shorter response times, while those with lower coverage rates have lower costs but longer response times (see Figure 9). When comparing the coverage ranges of solutions from NSGA-III and NSGA-II algorithms, both algorithms successfully identified optimal solutions, though neither demonstrated a significant relative advantage. NSGA-II exhibited a broader solution distribution. In contrast, NSGA-III produced a more concentrated solution set.

5.3. Sensitivity Analysis

To evaluate the sensitivity of key model parameters on drone base station site selection and operational performance, this study systematically analyzed how variations in hospital maximum response time, drone flight speed, static layer budget allocation ratio, total budget, and “dynamic-static” drone base station coverage radius affect coverage rate, response time, and total cost. The results demonstrate that different parameters exhibit significantly distinct impacts on emergency drone base station performance.

Research demonstrates that establishing appropriate maximum response time thresholds can enhance drone delivery efficiency (see Figure 10). We set hospitals’ maximum response time within the 2700–4500 s range, and the drone coverage rate stabilized at 96.18%, indicating that the maximum response time allowed by medical institutions has limited impact on drone base station coverage capabilities. However, the response time exhibited fluctuating patterns; when $T_{max}=3420$ it increased to 521.32 s and decreased to 329.54 s when $T_{max}=2880$. Total costs also fluctuated between USD 2.8 million and USD 4.8 million, suggesting that adjustments in response time require recalibrating the system’s time–cost trade-offs. While extending the maximum response time permitted by hospitals may improve delivery efficiency, its effect on coverage remains constrained.

Enhancing drone flight speed proves to be an effective strategy for improving $\mathrm{V}$ overall efficiency (see Figure 11). When the speed increases from 16 m/s to 24 m/s, the coverage rate remains stable at 96.18%, indicating minimal impact of flight speed on coverage area. However, the response time shows significant improvement; it decreases from 495.90 s at V = 16 m/s to 328.66 s at V = 23 m/s, representing a 33.72% reduction. While faster drone speeds substantially improve system responsiveness, they do not significantly increase total costs, demonstrating that increasing drone speed is an effective approach to optimize response time.

Moderately increasing the budget allocation ratio $\lambda$ for the static layer enhances coverage rates and improves response efficiency. Research findings indicate that drone coverage increases with higher static layer base station budget allocations, while the budget ratio significantly impacts both response time and operational costs (see Figure 12). When the static layer budget ratio $\lambda$ rises from 0.3 to 0.8, drone coverage gradually increases from 95.84% to 96.18%, stabilizing after reaching ≥0.4. Response time shows significant improvement, decreasing from 749.31 s ($\lambda =0.3)$ to 385.96 s ($\lambda =0.7)$—a 48.5% reduction. Total costs also exhibit an upward trend with increased budget allocation. The results demonstrate that moderately raising the static layer budget ratio can optimize coverage and response time while controlling costs.

When the total budget is sufficient, the system can effectively meet and respond to urban demands (see Figure 13). With the budget fluctuating between USD 6.8 million and USD 9.2 million, the coverage rate remains stable at 96.18%, while the response time varies between 385 and 424 s. The actual total expenditure stays within USD 2.8 million to USD 4.3 million. This indicates that under adequate budget conditions, the system can further reduce response time through resource reallocation, though there is limited room for improvement in coverage rate.

Expanding the coverage radius of static base stations effectively enhances both coverage efficiency and response time, while the coverage radius of dynamic base stations has limited impact on overall performance (see Figure 14 and Figure 15). With dynamic base stations maintaining stable coverage within a 12–18 km radius and fluctuating response times between 367 and 424 s, their minimal influence on system performance demonstrates the dynamic layer’s effectiveness as a “supplementary coverage” solution. When the static base station coverage radius increased from 16 km to 24 km, coverage efficiency rose from 94.13% to 97.26%, response time improved from 421.34 s to 330.03 s, and total costs increased from USD 4.05 million to USD 4.55 million. This indicates that expanding static base station coverage significantly boosts coverage capacity and response speed, though at a higher cost.

Sensitivity analysis demonstrates that the coverage capability of the drone system remains resilient to most parameter variations, particularly maintaining stability in coverage rate under changes in $T_{max}$, flight speed V and total budget. However, response time shows sensitivity to flight speed, static budget allocation ratio, and static base station radius. For practical deployment, we recommend enhancing flight speed, optimizing static layer budget distribution, and moderately extending the endurance of static base stations through appropriate drone model selection to improve coverage radius. This approach achieves optimal balance between coverage and response efficiency within limited budgets. Meanwhile, dynamic base station coverage radius having minimal impact on the emergency logistics system shows that the model has robust performance in addressing sudden demand fluctuations.

5.4. Robustness Testing

To evaluate the adaptability and stability of the proposed drone base station site selection model in uncertain environments, this study conducted systematic robustness tests under demand fluctuations, weather variations, airspace regulations, and base station failures. The tests compared the performance of fixed schemes and robust optimization schemes in terms of coverage, response time, and total cost.

Research indicates that demand fluctuations have limited impact on base station coverage, while significantly affecting response time and cost. Through ±30% variations in total demand analysis, the study demonstrated that the drone system maintained 96.17% coverage, highlighting the model’s strong adaptability to demand fluctuations. However, demand volatility caused dynamic changes in response time and cost. When demand decreased by 30%, cost reduction was achieved through optimized strategies, effectively lowering expenses from USD 4.8 million to USD 3.3 million. Conversely, a 30% increase in demand resulted in a slight 357.41-s increase in response time to 399.83 s, with coverage remaining stable throughout the process.

Weather conditions, particularly severe weather, significantly impact the operational efficiency of drone systems. To analyze weather effects, the robustness test simulated five weather interference levels: clear, mild, moderate, severe, and extremely severe. Under clear skies, the coverage rate remained at 96.18%. Mild weather interference reduced coverage to 94.13% with a response time of 455.45 s. Moderate interference further lowered coverage to 90.16%, while extreme conditions like strong winds or heavy rain caused coverage to plummet to 70.57% and response time to 1382.37 s. Although the optimized solution slightly improved coverage in extreme weather (elevating it from 70.57% to 75.52% in severe conditions) and reduced response time, numerous service failures occurred. In terms of cost, weather-induced flight cancellations significantly increase transportation costs. To maintain service continuity during such periods, the system must rely on backup ground transportation, as ground transport typically incurs higher costs for emergency, small-batch shipments compared to drones., around USD 70–150 per trip (including vehicle lease, fuel, maintenance, and driver salary). This substitution method directly increases the variable operating costs of specific transportation. Frequent flight bans also lead to underutilization of drone base stations, reducing ROI. Meanwhile, to meet service standards, planners must overconfigure the drone network to compensate for weather-induced downtime. In this study weather interference factor ${\tau }_t$ directly reduces the effective service capacity $Q_{j\ast }{\tau }_t$ in the demand adaptation constraint.

This study also simulated airspace control with varying intensities ranging from 0.1 to 0.9. Results showed that as the control area expanded to 0.9, coverage decreased from 96.18% to 92.77%, while response time increased from 357.41 s to 468.04 s. The re-optimization proved effective in adjusting resource allocation under moderate-intensity controls, but performance gains became limited when control intensity reached or exceeded 0.5 under high-intensity conditions. This indicates that the model’s dynamic airspace adaptation mechanism requires further refinement.

This study also simulated scenarios of base station failures, analyzing the impact of random failure rates ranging from 5% to 40%. Results indicate that when failure rates remain below 20%, system coverage can still be maintained above 95.96% with response times reaching 369.82 s. However, when failure rates reach 40%, coverage drops to 95.27% and drone response times increase to 638.64 s. While the re-optimized solution partially compensates for failure impacts through resource reallocation, it still faces significant operational challenges under high failure rates.

This study validated the base station site selection model’s ability to maintain robust service capabilities under uncertainties through multidimensional disturbance testing. While demonstrating strong adaptability to demand fluctuations and spatial noise, the model’s performance significantly deteriorates in extreme scenarios such as severe weather, high-intensity airspace restrictions, and large-scale base station failures. The findings indicate that practical emergency logistics drone deployment requires integrating real-time meteorological alerts, dynamic airspace management, redundant base station deployment, and establishing coordinated emergency mechanisms with ground transportation systems to enhance overall system resilience and service continuity.

5.5. Discussion

This study develops a drone base station location planning and evaluation framework that integrates medical emergency logistics spatiotemporal characteristics with multi-objective optimization. Through a case study of Guangzhou and proof of conceptual simulation, it demonstrates its effectiveness. The experimental results not only validate the model’s reliability under ideal conditions but also highlight multiple practical challenges and complex cost–efficiency trade-offs in real-world deployment. These findings provide valuable insights for urban medical emergency logistics management practices and policy formulation.

The core innovation of this study lies in effectively integrating the dynamic spatiotemporal characteristics of medical emergency logistics with infrastructure planning frameworks. Medical emergency demands exhibit a “peak at month’s start, trough mid-week, and daytime peaks” temporal pattern with spatial clustering, which significantly differs from traditional models’ assumptions of uniform demand distribution and reliance on empirical and macro-demographic data. The proposed “dynamic-static” two-tier optimization model maintains long-term stable demand hotspots through static base station coverage while responding to sudden and seasonal fluctuations via dynamic layers. This approach identifies dynamic demand hotspots often overlooked in conventional planning, enhancing the system’s adaptability to heterogeneous spatiotemporal demands. This design addresses existing research gaps in considering temporal characteristics of emergency demands and dynamic collaborative scheduling.

The Pareto solution set generated by the NSGA-III algorithm shows trade-offs between coverage, cost, and response time, offering decision makers diversified non-dominated configuration options to improve the model’s practical applicability. This study emphasizes the need for a holistic approach to evaluate drone systems’ economic viability. Using Guangzhou case data, we calculated typical costs of local medical delivery services under standard traffic. Ground vehicles’ average cost per emergency inter-hospital delivery is USD 70–150, which serves as a baseline variable cost. Our optimization model simulated deployment with an average system cost of USD 2.05 million–2.24 million, achieving 96.18% coverage. The total system cost includes capital and operational expenses, validating the need for a break-even period for drones. The main advantage of drone transportation is significantly reducing response times (from over an hour to about 11 min), bringing healthcare cost savings far outweighing incremental logistics costs. There is a systemic trade-off: investing in a more resilient but costly and weather-vulnerable drone network for faster emergency logistics or relying on a cheaper but slower and traffic vulnerable ground fleet while accepting delay risks. Our model helps quantify this trade-off for holistic decision making.

The integration of multi-drone systems with coordinated emergency delivery through vehicle transportation has proven to be crucial for enhancing system efficiency and operational resilience. Experimental data demonstrates that drone flight speed and coverage radius significantly influence key performance metrics such as service coverage and response time. By strategically combining different drone types, operators can intelligently select transport vehicles based on delivery distance and cargo weight, effectively expanding service coverage while maintaining cost-effectiveness. This approach effectively addresses inherent limitations of single-drone models in payload capacity and endurance, highlighting the importance of “multi-drone coordination” as a key optimization strategy for emergency logistics systems. Furthermore, continuous technological advancements in drone performance will remain essential for improving the overall efficiency of drone transportation networks.

While the model demonstrates strong performance in simulation experiments, its operational effectiveness remains highly susceptible to external environmental factors, with airspace regulations and meteorological conditions being the primary sources of uncertainty. As shown in the robustness tests, although the model exhibits strong adaptability to demand fluctuations and routine noise interference, drone coverage rates plummet under extreme weather conditions. The current model relies on static airspace control data, making it ill-equipped to handle the dynamic adjustments required in real-world flight restriction zones. This highlights the necessity for urban airspace management to move beyond simplistic flight bans, emphasizing the critical need for establishing a smart airspace system with tiered and categorized dynamic adjustments. Policy makers should leverage these findings to establish permanent drone corridors for urban medical emergency logistics, while reserving flexible temporary takeoff and landing spaces in core hotspot areas. By creating expedited airspace approval channels, priority access for rescue operations during peak demand periods can be ensured. Furthermore, future systems must integrate real-time weather forecasting with dynamic airspace management capabilities, while fully utilizing ground transportation as backup solutions to maintain the continuity of emergency logistics networks.

This study also highlights the “last mile” challenge between model optimization and practical implementation. While kernel density analysis can accurately identify demand hotspots, the actual site selection of base stations is constrained by micro-level factors such as building ownership, community acceptance, power access, and physical security elements that are difficult to fully quantify in macro models. For instance, a theoretically optimal site might be located on a hospital rooftop, but implementation is hindered by regulatory policies. Conversely, suboptimal sites may prove more feasible due to existing infrastructure. Therefore, the findings of this study should serve as decision support and supplementary tools for location planning, helping identify priority development areas rather than directly specifying geographic coordinates.

While the current study takes Guangzhou as a case, the proposed “dynamic-static” optimization framework is inherently adaptable to other urban contexts. Its generalizability stems from the modular incorporation of city-specific data. For application, key adaptations would include the following. Spatial Data Layer Replacement: (i) Inputting the target city’s road network, population density, building footprints, and hospital locations to recalculate demand hotspots via kernel density estimation and DBSCAN. (ii) Constraint Parameter Recalibration: Adjusting the airspace restriction map, weather risk functions, and budget level to reflect local regulations, climate, and economic conditions. (iii) Demand Profile Integration: Utilizing the target city’s historical emergency logistics order data or proxy indicators to model its unique spatiotemporal demand patterns. For example, applying the model to a smaller or less dense city might result in a higher proportion of static stations due to more stable demand clusters and a different optimal drone mix. Conversely, in a city with a highly dispersed or polycentric layout, the dynamic layer’s role in bridging coverage gaps between static hubs might become even more critical. The model’s sensitivity analyses (Section 5.3) provide a guide for understanding how key parameters might need to be tuned in different scenarios. Therefore, the framework offers not a rigid solution but a configurable planning tool whose outputs are shaped by the unique data and constraints of the target environment.

In conclusion, the methodology proposed in this study provides a systematic, data-driven decision support framework for the scientific planning of an urban medical emergency drone network. While demonstrating significant potential in enhancing resource allocation efficiency, establishing a truly robust and reliable emergency logistics system still requires deeper coordination among technological advancement, policy frameworks, infrastructure development, and operational management. Future research should focus on promoting cross-disciplinary integration to fully leverage the critical role of drone technology in life-saving operations.

6. Conclusions

6.1. Summary of Research Findings

This study focuses on the site selection problem of drone base stations in urban medical emergency logistics. A “dynamic-static” collaborative site selection model integrating multi-source data and multi-objective optimization is constructed. The research systematically advances from multiple aspects such as spatial–temporal feature analysis, demand prediction, resource allocation and constraint handling, and obtains the following main conclusions.

The demand for medical emergency logistics exhibits significant spatiotemporal heterogeneity. Through kernel density estimation and DBSCAN clustering analysis, this study precisely identifies the fluctuation patterns of medical emergencies in Guangzhou: “dual peaks in winter and spring, sudden surges at month’s start, and daytime peaks.” Spatially, these incidents are concentrated in densely populated areas and transportation hubs. This discovery provides solid data support for the hierarchical deployment of drone base stations, highlighting the critical role of spatiotemporal analysis in emergency logistics planning.

The “dynamic-static” collaborative optimization model significantly enhances overall system performance. The static layer primarily addresses long-term, stable demand hotspots (such as hospitals in core urban areas), while the dynamic layer effectively compensates for static infrastructure limitations by responding to real-time demand fluctuations and emergencies. Case results demonstrate that this model achieves 96.18% drone coverage and 673.38 s average response time within budget constraints, proving its dual advantages of efficiency and robustness.

The integration of multi-drone coordination and multi-objective optimization serves as an effective approach to achieve efficient resource allocation. By applying the NSGA-III algorithm to solve Pareto solutions, an optimal balance is achieved between coverage, response time, and cost. Furthermore, the combined delivery strategy of multi-drone systems significantly extends service radius and enhances cargo adaptability, providing a viable solution for emergency logistics in complex urban environments.

In summary, this study not only proposes a data-driven and optimization-oriented methodology for drone base station site selection, but also offers a scalable approach and provides a transferable methodological framework and a set of evaluative criteria that can inform planning in other large cities, contingent on local data integration and stakeholder priorities. In addition, simultaneously balancing response time and cost-effectiveness, the model realizes synergistic optimization of coverage scope, response time, and cost-efficiency. This provides urban managers with quantifiable decision support tools to avoid resource waste caused by excessive pursuit of single objectives.

6.2. Limitations and Directions for Improvement

While the study has achieved certain results, its direct application requires calibration with city-specific data on demand, geography, regulations, and resource availability. The following limitations remain:

(1) The model’s performance is highly dependent on the quality and representativeness of input data. The medical emergency logistics demand order data in this research primarily originates from large pharmaceutical logistics enterprises. Although the studied companies have service agreements covering all hospitals in Guangzhou and are highly representative, they cannot represent all emergency logistics demands, potentially underestimating the needs of community medical institutions during public health emergencies. The reliance on one year of demand data, while sufficient to capture seasonal patterns and demonstrate the model’s functionality, limits the analysis of long-term trends and rare extreme events (e.g., city-wide public health emergencies or large-scale disasters). Such events could drastically alter demand patterns and stress-test the network’s resilience in ways not observable within a single typical year. Consequently, the robustness claims are bounded within the range of fluctuations present in the 2023 data. Future work could integrate longer-cycle, multi-source data from government emergency centers, hospital emergency departments, and other leading commercial pharmaceutical logistics enterprises to enhance the model’s universality.

(2) The model does not fully incorporate real-time adjustments of airspace control and complex weather scenarios within urban areas, which may overestimate the feasibility of certain urban spaces and environments. There is a need to move from static constraints to real-time, data-driven dynamic risk assessments for operational robustness. Future development could involve integrating dynamic airspace management systems with high-precision meteorological warning modules into drone transportation management systems.

(3) The current model focuses on macro-level analysis of equipment parameters, hospital emergency hard time windows, drone resource allocation, and meteorological impacts on coverage, response time, and costs, but lacks sufficient consideration of micro-level factors such as building load-bearing capacity, payload energy consumption, power access, noise and privacy acceptance, and feasibility of property rights negotiations; these should be addressed in subsequent feasibility studies and site-specific engineering phases. Additionally, the study does not adequately account for the social benefits of drone network, including life-saving value and long-term operational maintenance costs, resulting in insufficient cost accounting accuracy. Future research should adopt a cost–benefit comprehensive evaluation framework to support more thorough economic decision-making analysis.

(4) This study assumes three types of drone models and one type of automobile transportation mode, with relatively fixed endurance and payload capacity of drones, without considering the potential impact of emerging drone technologies. Subsequent research could explore dynamic optimization pathways for location strategies under technological evolution.

(5) While Guangzhou serves as an ideal, high-complexity testbed due to its megacity characteristics, uneven resource distribution, and strict airspace controls, the findings are inherently contextual. The optimal network configuration, specific parameter sensitivities (e.g., the static budget allocation ratio $\lambda$), and even the relative importance of certain constraints (e.g., weather vs. airspace) may differ in cities with distinct urban forms, traffic patterns, medical infrastructure layouts, and regulatory environments. Therefore, the model’s outputs for Guangzhou should not be directly extrapolated as universal blueprints but rather as a validated methodological framework that requires city-specific data and calibration for application elsewhere.

(6) The sensitivity and the robustness conclusions are derived from computational simulations and modeled scenarios. While the simulation incorporates realistic constraints and historical demand patterns, it lacks external validation against actual response times from a deployed medical drone network. This is because such a large-scale, urban medical drone network for emergency logistics, as proposed here, is planned but not yet fully operational. Consequently, the absolute performance figures should be interpreted as well-informed projections rather than guaranteed outcomes. Future real-world drone deployments are essential to calibrate the models and validate their predictive accuracy. Subsequent studies could conduct drone deployments in selected districts to collect real-world operational data (actual flight times, weather-induced delays, ground coordination times) for calibrating and validating the simulation models, thereby transitioning the framework from a planning tool to a validated operational decision-support system.

6.3. Policy Recommendations and Future Research Prospects

Based on the research insight, the potential directions, policy recommendations and future research directions are proposed. (1) Promote tiered and categorized opening of urban low-altitude airspace, establish dedicated routine channels for medical drones, and streamline approval processes for temporary takeoff/landing points in medical emergencies. (2) Develop standardized urban medical drone network planning guidelines specifying base station density, coverage radius, and hospital response time thresholds to ensure equitable service coverage across urban areas. (3) Build an integrated “medical-transport-air traffic control” emergency platform to achieve real-time coordination of demand forecasting, route planning, and resource allocation.

Future researchers can further explore the following aspects. (1) Study real-time dynamic optimization of drones by developing edge computing and IoT-based real-time demand sensing systems to enhance model responsiveness to emergencies. (2) Explore collaborative delivery networks between drones and intelligent terminals such as autonomous vehicles and robots to enhance rescue reliability in complex environments. (3) Investigate public acceptance of drone delivery services. Assess privacy concerns and noise sensitivity related to medical drone deliveries. Evaluate the environmental changes and costs of drone traffic noise exposure based on population noise levels, and design user-friendly operational plans to facilitate technology adoption. (4) Expand data sources and model generalization by integrating multi-source data from government emergency centers and community medical institutions, incorporating extreme scenarios like urban epidemics and natural disasters to enhance model versatility and robustness. (5) Develop detailed modeling of complex drone site selection constraints by refining airspace regulations, meteorological risks, and social factors to establish a multi-dimensional constraint framework. (6) The core “dynamic-static” optimization logic of this framework is generalizable across cities, but the optimal parameter sets (e.g., budget allocation ratio $\lambda$, coverage radius) will vary significantly with urban form and demand patterns. Future research should test this hypothesis through comparative case studies. (7) Promote policy coordination and standardization for pharmaceutical delivery, establish medical drone network standards including takeoff/landing density and drug safety protocols, and explore collaborative operation models involving “government-enterprise-community” partnerships to advance technology implementation and sustainable development.