The sudden onset of an earthquake, often accompanied by secondary disasters, presents an immense challenge for emergency response. Acquiring accurate and timely information about the disaster zone is paramount for effective rescue operations. While satellite remote sensing offers broad coverage, its spatial resolution is often insufficient for the granular details required in rescue scenarios, such as identifying specific building damage or pinpointing trapped individuals. In contrast, unmanned drones, particularly rotorcraft, offer a powerful solution due to their rapid deployment, high mobility, and ability for close-range, detailed observation. Unlike fixed-wing unmanned drone, rotorcraft possess core advantages like vertical take-off and landing (VTOL), hover capability, low-speed flight at low altitudes, and strong terrain adaptability. They can operate in complex environments like rubble fields without requiring dedicated runways, enabling detailed scanning of structures and ensuring comprehensive data collection.
However, the current paradigm for post-earthquake information gathering using unmanned drones remains underdeveloped. It often relies on manual, large-area scanning coordinated by human operators, leading to information redundancy and inefficiency. A core reason is the inherent limitations of rotorcraft unmanned drone, such as finite flight endurance, payload capacity per sortie, and operational range. Existing methods lack an effective framework for partitioning the disaster area, determining search priorities, and planning flight paths that are specifically adapted to the unique mobility characteristics of unmanned drone fleets. Consequently, the flexible advantages of unmanned drones are underutilized, and poorly planned routes can result in wasted energy and mission interruptions.
The fundamental principle for post-earthquake building inspection is to prioritize structures that are “most likely to contain survivors” and have the “highest potential number of trapped individuals.” This is complemented by prioritizing buildings that are not completely collapsed but contain residual spaces, based on on-site feedback. While substantial research exists on post-earthquake search and rescue, including area zoning and priority determination, a significant gap remains in methodologies centered on the unmanned drone as the primary platform. Existing zoning methods often lack a rapid inventory of target buildings for unmanned drone search and are not designed with unmanned drone performance parameters in mind. Similarly, models for determining regional rescue priorities do not incorporate factors specific to unmanned drone operations, such as the number of drones required per zone or building-specific vulnerability scores relevant to aerial inspection. This oversight hinders the full exploitation of unmanned drone capabilities for efficient search and rescue.
To address these challenges, this research integrates the characteristics of rotorcraft unmanned drone—small-area precision observation and high mobility—with the urgent demands of post-earthquake response. We propose a comprehensive four-stage technical framework: “High-Value POI Screening – Spatial Clustering and Zoning – Rescue Priority Assessment – Ant Colony Algorithm Path Optimization.” This framework provides an end-to-end solution for planning unmanned drone search routes, overcoming the drawbacks of traditional, experience-dependent planning that leads to low data collection efficiency.
The core contributions of this framework are threefold. First, to provide targeted data for unmanned drone searches, we screen six categories of Points of Interest (POIs) characterized by high population density and significant rescue value, such as educational institutions and medical facilities, focusing efforts on areas with a higher probability of survivor presence. Second, to overcome the limitations of existing zoning methods that are not adapted to unmanned drone performance and are susceptible to interference from low-density “blank” areas, we propose a MeanShift-Kmeans two-layer clustering algorithm. This algorithm first eliminates blank area interference using MeanShift and then controls partition size using K-means to match unmanned drone fleet capabilities. The number of unmanned drone required is calculated based on dual “area-point count” constraints, establishing a “zone-capacity” adaptation system. Third, to address the shortcomings of priority models that lack unmanned drone-specific indicators and are subject to subjective bias, we construct a multi-indicator rescue priority determination model that integrates both subjective and objective factors. This model incorporates distinctive factors like building rescue priority (determined using the Entropy Weight Method) and the number of patrol unmanned drone, enabling precise quantitative ranking. Furthermore, the building rescue priority is mapped to the pheromone concentration in the Ant Colony Optimization (ACO) algorithm for path planning, directly linking rescue value with route selection and providing a scientific basis for unmanned drone deployment.
1. Clustering and Zoning of Post-Earthquake Search Areas
1.1 Screening of High-Value POI Data
Points of Interest (POIs) serve as precise descriptors of geographic entities, reflecting urban spatial vitality. Their hierarchical classification system (e.g., commercial facilities, public service facilities) provides a data foundation for locating high-value search areas post-earthquake. Data such as name, category, and geographic coordinates can be acquired from platforms like AMap.
Adhering to the core principle of “prioritizing areas with a high survival probability,” we focus on regions that are densely populated, have a high likelihood of occupancy, and possess buildings with critical functions. Six categories of high-value POIs are selected as the data source for unmanned drone search tasks. These include:
- Science, Education & Cultural Services (e.g., middle schools, primary schools): Concentrate vulnerable populations like children and students.
- Healthcare Services (e.g., general hospitals, emergency centers): Function both as potential disaster sites and as critical hubs for subsequent treatment.
- Commercial & Residential Buildings (e.g., office towers, residential compounds): Have high nighttime occupancy rates.
- Accommodation Services (e.g., hotels, hostels): Feature temporary and mobile population gatherings.
- Shopping Services (e.g., supermarkets, shopping malls): Exhibit high daytime population density.
- Corporate Enterprises (e.g., construction companies, pharmaceutical firms): Often concentrate healthy adults.
The specific POI data classification is summarized in Table 1.
| Main Category | Medium Category | Small Category Examples |
|---|---|---|
| Science, Education & Cultural Services | Schools | Middle School, Primary School, Kindergarten, Adult Education, Higher Education, etc. |
| Healthcare Services | General Hospitals, Specialized Hospitals, Emergency Centers | Grade III, Class A Hospital, Gynecological Hospital, Orthopedic Hospital, Neurology Hospital, etc. |
| Commercial & Residential Buildings | Industrial Parks, Buildings, Residential Areas, Related Facilities | Office Building, Dormitory, Residential Community, etc. |
| Accommodation Services | Hotels, Hostels & Guesthouses | Hotel, Star-rated Hotel, Hostel, Guesthouse, etc. |
| Shopping Services | Supermarkets, Shopping Malls, Comprehensive Markets | Supermarket, Shopping Center, Mall, Comprehensive Market, Furniture Mall, etc. |
| Corporate Enterprises | Factories, Companies & Enterprises | Construction Company, Pharmaceutical Company, Machinery & Electronics, etc. |
1.2 Search Area Division Based on a Two-Layer Clustering Algorithm
To achieve the goals of “reducing interference from low-density blank areas” and “ensuring partition size is compatible with the search range of an unmanned drone fleet,” we designed the MeanShift-Kmeans two-layer clustering algorithm. This method performs zoning in two stages: primary density clustering via MeanShift to eliminate blank areas, followed by secondary clustering via a Z-curve optimized K-means to ensure partition scale matches unmanned drone performance.
1.2.1 Primary Density Clustering with Mean-Shift
The Mean-Shift algorithm identifies density peaks through kernel density estimation without predefining the number of clusters. Its core advantages are its data-driven nature (no manual parameter setting), its natural adaptation to varying density distributions via probability density gradient ascent, and its stability due to insensitivity to initial values. These features make it ideal for the critical requirement in earthquake rescue of “not missing any building,” effectively identifying various building distribution patterns.
1.2.2 Secondary Clustering with Z-curve Assisted K-means
While primary clustering addresses blank area interference, the resulting clusters may still be too large for a single unmanned drone sortie, and the number of clusters cannot be manually controlled to match the size of the unmanned drone fleet. Therefore, we employ a K-means clustering algorithm that uses the Z-curve to determine initial cluster centers. The logic is to first calculate a reasonable K value and then optimize cluster center distribution, preliminarily adapting partition size to unmanned drone performance. The K value determines the number of secondary partitions, aligning it with the “total single-sortie search capacity” of the unmanned drone fleet. The Z-curve ensures initial cluster centers are evenly distributed in space, avoiding the cluster center aggregation problem of traditional K-means. The calculation of K is performed as follows:
$$K = \text{round} \left( \frac{S_{\text{total}}}{3 \cdot S_{\text{drone}}} \right)$$
$$S_{\text{drone}} = L \times W = L \times \frac{2 \times \text{GSD} \times f}{d \times \tan\left(\frac{\text{FOV}}{2}\right)}$$
where \(S_{\text{total}}\) is the area of a single primary cluster calculated via convex hull; \(S_{\text{drone}}\) is the predefined search area for a single unmanned drone fleet; \(L\) is the flight range; GSD is the Ground Sample Distance (should be less than 5cm to identify an adult); \(f\) is the camera focal length; \(d\) is the camera pixel size; and FOV is the camera’s field of view.
Given the typically limited endurance and flight speed of rotorcraft unmanned drone, a medium-sized emergency response team配置 with 2-4 unmanned drone is assumed to ensure mission timeliness. Since the K-means algorithm requires a predefined K and cannot directly control final partition area, we use the maximum search area of a 3-unmanned drone fleet as the benchmark for determining K. This ensures the basic area of each post-clustering partition falls within the search range of a small-to-medium unmanned drone fleet, preventing失控 partition sizes from affecting subsequent information collection efficiency. This stage establishes a preliminary theoretical framework for adapting “partition scale to unmanned drone performance,” providing a foundation for subsequent unmanned drone allocation but not yet fully incorporating post-earthquake time constraints and data load limits, which require further optimization based on actual needs.
1.3 Calculating the Required Number of Patrol Unmanned Drone
After the secondary K-means clustering, a preliminary theoretical adaptation between partition scale and unmanned drone performance is achieved. However, due to algorithmic characteristics, the actual area of the final generated partitions may deviate from the predefined \(S_{\text{drone}}\). Furthermore, the stringent time requirements of post-earthquake rescue and the data load limit per unmanned drone sortie necessitate precise calculation of the required number of unmanned drone for each partition.
Therefore, we propose a dual-dimensional calculation method for unmanned drone quantity based on “area constraint” and “point count constraint,” using the following combined formula:
$$N = \max\left(\left\lceil\frac{S_{\text{zone}}}{S_{\text{drone}}}\right\rceil, \left\lceil\frac{P_{\text{zone}}}{P_{\text{max}}}\right\rceil\right)$$
where \(N\) is the number of unmanned drone required for a single secondary cluster partition; \(S_{\text{zone}}\) is the total area of the secondary cluster partition; \(P_{\text{zone}}\) is the total number of POI points within the partition; \(P_{\text{max}}\) is the maximum number of POI points a single unmanned drone can search per sortie; and \(\lceil x \rceil\) is the ceiling function.
By combining this with the actual constraints of post-earthquake rescue, the final number of unmanned drone to be deployed for each secondary partition is determined, ensuring that the partition search task is compatible with unmanned drone performance while meeting rescue timeliness and data completeness requirements.
2. Rescue Priority Determination Model
2.1 Screening and Determination of Influencing Factors
Building upon prior research for traditional ground rescue, we adapt the evaluation framework for unmanned drone operations through targeted optimizations in three areas: indicator replacement, indicator addition, and calculation logic refinement, resulting in a model with 10 core factors. The model maintains the original “subjective factors” and “objective factors” classification and its three-tier structure.
Indicator Replacement: The original “secondary geological hazard risk level” is replaced with “Topography,” as it directly impacts unmanned drone flight efficiency and safety. “Number of deaths” and “number of injuries” are merged into “Casualties” to simplify quantification given the difficulty of obtaining precise early post-earthquake data.
Indicator Addition: “Building Rescue Priority (K)” is added to quantify the importance of different building types based on POI functionality (population density, vulnerability, social function urgency, secondary disaster risk). “Number of Patrol Unmanned Drone (N)” is also added to reflect regional task load.
Calculation Logic Optimization: The calculation scope for factors like intensity, epicenter location, and population density is adjusted. Instead of using administrative regions, the basic statistical unit is now the secondary partition from the two-layer clustering, avoiding assessment bias due to mismatched boundaries.
2.1.1 Subjective Factors
These are factors determined by人为设定的评估标准 with clear goal orientation, primarily including the following five items.
(1) Seismic Intensity (I): Describes the degree of ground shaking and damage. Calculated for the partitioned area as the basic unit using standard attenuation formulas.
(2) Number of Deaths (D) and Injuries (C): Estimated based on the partitioned area’s total population (derived from POI type counts and typical population ranges per building type, as shown in Table 2) multiplied by casualty rates corresponding to the estimated intensity for urban, town, and rural settings. Data is normalized to a 3-7 score range for model stability.
| Main Category | Medium Category | Population Range (Persons) |
|---|---|---|
| Science, Education & Cultural Services | Schools | 1000-4000 |
| Healthcare Services | General Hospitals | 300-800 |
| Healthcare Services | Specialized Hospitals | 100-300 |
| Healthcare Services | Emergency Centers | 50-150 |
| Commercial & Residential Buildings | Industrial Parks | 1500-3000 |
| Commercial & Residential Buildings | Buildings | 300-800 |
| Commercial & Residential Buildings | Residential Areas | 2000-4000 |
| Accommodation Services | Hotels | 100-300 |
| Shopping Services | Supermarkets | 500-1500 |
| Shopping Services | Shopping Malls | 1000-2000 |
| Shopping Services | Comprehensive Markets | 300-1000 |
| Corporate Enterprises | Companies & Enterprises | 100-400 |
(3) Number of Patrol Unmanned Drone (N): The number of unmanned drone required for the partition, determined by the dual constraints in Section 1.3. The final score is assigned based on this number, up to 10.
(4) Building Rescue Priority (K): Determined using the Entropy Weight Method. An objective scoring system is constructed based on four dimensions: Population Density, Personnel Vulnerability, Social Function Urgency of the Building, and Secondary Disaster Risk. Each dimension is scored 1-10 for the six POI categories (higher score = higher priority). The entropy weight method calculates the weight for each dimension (see Table 3), and a weighted sum yields the comprehensive rescue priority score for each POI type (see Table 4). The regional Building Rescue Priority score is the arithmetic mean of all building priority scores within the region.
| Indicator | Information Entropy (e) | Information Utility (d) | Weight (%) |
|---|---|---|---|
| Population Density | 0.831 | 0.169 | 21.49 |
| Personnel Vulnerability | 0.806 | 0.194 | 24.72 |
| Social Function Urgency | 0.747 | 0.253 | 32.30 |
| Secondary Disaster Risk | 0.831 | 0.169 | 21.49 |
| POI Type | Total Score |
|---|---|
| Healthcare Services | 8.71 |
| Science, Education & Cultural Services | 5.85 |
| Accommodation Services | 2.68 |
| Commercial & Residential Buildings | 4.31 |
| Corporate Enterprises | 2.46 |
| Shopping Services | 2.61 |
2.1.2 Objective Factors
These are inherent regional attributes or directly measurable physical parameters, including five items.
(1) Epicenter Location (L): If the partition contains the epicenter, score=10; otherwise, score=0.
(2) Population Density (PD): Calculated for the partition by averaging the population density of all buildings within it, using standard values per building function, rather than total population/total area.
(3) Distance from Epicenter to Partition (S): The distance from the epicenter to the partition’s center. Normalized to a 3-7 score.
(4) Population Aggregation (PA): The ratio of the partition’s population density to the population density of its parent primary cluster. Normalized to a 3-7 score.
(5) Topography (T): Classified as plain, hill, mountain, or plateau, directly affecting unmanned drone flight difficulty.
2.2 Construction of the Post-Earthquake Rescue Priority Model
After refining the factors, weights are assigned and calibrated based on the research context and practical needs to complete the new model, as shown in Table 5.
| Level-1 Indicator | Weight (W1) | Level-2 Indicator | Weight (W2) | Level-3 Indicator & Scoring (S) |
|---|---|---|---|---|
| Subjective Factors | 0.4 | Intensity (I) | 0.35 | VI/VII:0, VIII:1, IX:2, X:3, XI:4 |
| Number of Deaths (D) | 0.20 | Normalized score 3-7 | ||
| Number of Injuries (C) | 0.08 | Normalized score 3-7 | ||
| Number of Patrol Unmanned Drone (N) | 0.05 | 1-10 based on actual count | ||
| Building Rescue Priority (K) | 0.09 | Regional average score (1-10) | ||
| Objective Factors | 0.6 | Epicenter Location (L) | 0.05 | In zone:10, Not in zone:0 |
| Population Density (PD) | 0.03 | Greater than avg:7, Less than avg:3 | ||
| Distance from Epicenter (S) | 0.04 | Normalized score 3-7 | ||
| Population Aggregation (PA) | 0.06 | Normalized score 3-7 | ||
| Topography (T) | 0.05 | Plain:1, Hill:2, Mountain:3, Plateau:4 |
The rescue priority score for a region \(i\) is calculated as:
$$M_i = \sum_{k=1}^{10} (W_{1,k} \times W_{2,k} \times S_k)$$
where \(W_{1,k}\) is the Level-1 weight, \(W_{2,k}\) is the Level-2 weight, and \(S_k\) is the Level-3 score for indicator \(k\). A higher \(M_i\) indicates a higher rescue priority for the region, providing a quantitative basis for unmanned drone rescue force deployment and prioritization.
3. Unmanned Drone Route Planning
3.1 Model Formulation
For planning post-earthquake unmanned drone search paths, the optimization model focuses on two objectives under multiple constraints.
Objectives:
- Minimize Total Path Distance for the Unmanned Drone Fleet: Optimize flight paths to reduce total里程, lowering energy consumption and improving operational efficiency.
$$\min \sum_{i=1}^{n} \sum_{j=0}^{m} \sum_{k=1}^{m} d_{jk} x_{ijk}$$
- Maximize Priority-Weighted Coverage: Prioritize covering high-priority POIs to enhance rescue value, quantified by the weighted sum (higher priority \(P_j\) receives greater weight).
$$\max \sum_{i=1}^{n} \sum_{j=1}^{m} y_{ij} P_j$$
Constraints: To ensure feasibility, multiple constraints are introduced based on unmanned drone physical performance, mission requirements, and environmental factors.
- Maximum Endurance Constraint: The total distance flown by unmanned drone \(i\) cannot exceed its maximum range \(L_i\).
$$\sum_{j=1}^{m} \sum_{k=1}^{m} x_{ijk} \cdot d_{jk} \leq L_i \quad \forall i$$
- Path Closedness: Each unmanned drone’s path must start and end at the launch/landing point (indexed as ‘start’/’end’).
$$\sum_{j=1}^{m} x_{i,\text{start},j} = \sum_{j=1}^{m} x_{i,j,\text{end}} = 1 \quad \forall i$$
- Complete and Non-Repeating Coverage: Every POI \(j\) must be visited by exactly one unmanned drone.
$$\sum_{i=1}^{n} y_{ij} = 1 \quad \forall j$$
- Task Load Constraint: The number of POIs assigned to unmanned drone \(i\) is limited to \(M\) to avoid overloading and ensure timely data transmission.
$$\sum_{j=1}^{m} y_{ij} \leq M \quad \forall i$$
- Task Balancing (Multi-TSP Capacity Limit): Ensure the number of POIs searched by each unmanned drone is approximately equal to achieve balanced workload.
$$\left| \sum_{j=1}^{m} y_{ij} – \frac{m}{n} \right| \leq \text{round}\left(0.15 \cdot \frac{m}{n}\right) \quad \forall i$$
- Priority Heuristic Constraint: During path construction, the next target point selection considers both POI priority \(P_j\) and distance heuristic factor \(\eta_{jk}\).
$$\arg \max_{k \in \text{allowed}_i} (P_k \cdot \eta_{jk})$$
- Communication Radius Constraint: The farthest point on unmanned drone \(i\)’s path must be within its communication radius \(R_{\text{comm}}\) from the launch point.
$$\max_{j \in \text{Path}_i} (d_{\text{start},j}) \leq R_{\text{comm}} \quad \forall i$$
- Mission Time Window Constraint: The total time (flight + search) for unmanned drone \(i\) must not exceed the maximum allowed time \(T_{\text{max}}\).
$$\sum_{j=1}^{m} \sum_{k=1}^{m} x_{ijk} \cdot \frac{d_{jk}}{v_i} + \sum_{j=1}^{m} y_{ij} \cdot t_j \leq T_{\text{max}} \quad \forall i$$
The search time \(t_j\) at each POI is estimated based on its representative area and the unmanned drone’s camera parameters and flight speed, accounting for straight flight and turning time.
3.2 Model Solution
To solve this constrained Multi-Traveling Salesman Problem (MTSP) for post-earthquake unmanned drone operations, the Ant Colony Optimization (ACO) algorithm is selected as the core optimizer. Its suitability for this scenario surpasses other intelligent algorithms like Genetic Algorithms (GA).
Compared to GA, which relies on fixed-weight fitness functions and may generate invalid paths requiring complex repair mechanisms, ACO offers superior adaptability to complex constraints. Its pheromone-guided search with evaporation naturally encourages exploration to escape local optima and inherently generates closed, feasible paths from the launch point, perfectly matching the complex post-earthquake environment with scattered POIs and multiple constraints like endurance and complete coverage. A key innovation of this study is mapping the quantitatively derived Building Rescue Priority (\(K\)) scores directly to the pheromone concentration in the ACO algorithm. This ensures high-priority POIs (e.g., healthcare or educational buildings) have higher pheromone values and are therefore prioritized for selection during path construction, directly binding “rescue value” to “path planning” and aligning with the core rescue principle of “prioritizing high-survival-probability areas.”
4. Scenario Verification and Analysis
4.1 Parameter and Scenario Setup
To validate the model’s feasibility and quantitative assessment capability, a simulated earthquake scenario is established. The epicenter is set at a hypothetical location. The study area is a representative city district known for its seismic activity, dense population, and diverse economy, providing an ideal sample for model verification. An industrial-grade unmanned drone (model: DJI Matrice 300 RTK) equipped with a zoom camera (DJI H20T) is selected as the experimental platform. Key parameters are listed below.
| Parameter Category | Value |
|---|---|
| Unmanned Drone Max Speed (S-mode) | 23 m/s |
| Unmanned Drone Max Flight Time | 55 min |
| Camera Focal Length (f) | 60 mm |
| Camera Pixel Size (d) | 1.43 μm |
| Camera Field of View (FOV) | 84° |
| Target Ground Sample Distance (GSD) | <5 cm |
| Calculated Single Unmanned Drone Search Area (\(S_{\text{drone}}\)) | ~17.08 km² |
4.2 Post-Earthquake Search Area Clustering
Based on recent POI data from the study area, 1,443 high-value POIs were screened. The total area covered by these points was 1774.97 km². Applying the two-layer clustering algorithm: the primary MeanShift clustering resulted in 6 partitions with a total area of 1079.48 km². Subsequent K-means clustering produced 25 final secondary partitions with a total area of 681.21 km². This process effectively reduced blank area by 61.6%, making the partitions more compact and suitable for unmanned drone fleet operations.
4.3 Rescue Priority Determination for Search Areas
Taking one primary cluster (Cluster 5) and its four secondary partitions (5-0, 5-1, 5-2, 5-3) as an example, the rescue priority was calculated. Table 7 shows the final scores derived from the model based on the original factor data.
| Partition | Final Score (M_i) | Key Contributing Factors |
|---|---|---|
| 5-3 | 0.90 (Highest) | Highest scores in Population Density, Distance from Epicenter, Population Aggregation, and Building Rescue Priority. High weight of objective factors elevated total score. |
| 5-1 | 0.74 | Highest scores in estimated Deaths and Injuries. |
| 5-2 | 0.72 | Benefited from a higher Number of Patrol Unmanned Drone (4). |
| 5-0 | 0.60 (Lowest) | Most indicators at relatively lower levels. |
The results demonstrate the model’s ability to precisely differentiate rescue priorities by integrating both subjective and objective factors.
4.4 Unmanned Drone Patrol Route Planning
4.4.1 Parameter Settings for ACO
Pheromone importance (α) = 2, Heuristic information importance (β) = 3. Constraints: Mission time window \(T_{\text{max}} = 30\) min, Maximum flight distance per sortie = 40 km, Communication radius \(R_{\text{comm}} = 15\) km. Search times \(t_j\) per POI type were calculated based on typical area and flight speed in P-mode (17 m/s).
4.4.2 Case Comparison and Experimental Area
To直观体现 the advantages of the planning scheme, the unmanned drone operations during the 2023 earthquake in Jishishan, Gansu, are used as a comparative real-world case. For the experimental test, “Partition 0-4” from the study area is selected.
Real Case: During the Jishishan earthquake, a team deployed测绘 unmanned drone, flying a total of 17 sorties to cover approximately 32.05 km² of key disaster areas.
Experimental Area (Partition 0-4): Contains 99 POI points. According to the dual-constraint formula, 4 unmanned drone are required. Its area is 32.19 km², which is highly comparable to the real case area (32.05 km²), providing a basis for横向对比.
4.4.3 Route Planning Results for the Experimental Area
Using the ACO algorithm, flight paths for 4 unmanned drone were generated for Partition 0-4. Key results are summarized in Table 8.
| Path Parameter | Path 1 | Path 2 | Path 3 | Path 4 |
|---|---|---|---|---|
| Path Length (km) | 14.97 | 21.97 | 22.82 | 12.86 |
| Max Comm. Distance (km) | 4.02 | 4.01 | 3.49 | 3.74 |
| POIs Covered | 26 | 25 | 23 | 25 |
| Total Time (Flight+Search) | 17.23 min | 27.64 min | 24.84 min | 15.59 min |
Analysis:
- Performance Constraint Adherence: All path lengths are under 40 km, communication distances are far below 15 km, total times are under 30 min, and POI counts per path are under 30. This confirms that each unmanned drone can safely and completely execute its mission.
- Task Load Balance: The number of POIs per path varies by only 3 (23-26), demonstrating good load balancing and avoiding resource waste.
- Priority Coverage Effectiveness: High-priority POIs (e.g., healthcare, education) consistently appear early in the planned paths, validating the integration of the priority model into the ACO algorithm.
4.4.4 Route Planning Results for the Real Case Area
Applying the proposed framework (POI screening, blank area reduction) to the Jishishan data, the search area was reduced from 32.05 km² to 22.84 km² (a 28.7% reduction). The remaining 49 high-value POIs were then planned for, requiring only 2 unmanned drone sorties instead of the original 17.
4.4.5 Comparative Analysis of Results
- Within-Case Comparison (Jishishan): Applying our framework to the real case itself yielded significant optimization: search area reduced by 28.7% and flight sorties reduced from 17 to 2. This demonstrates a dramatic improvement in efficiency, saving critical time and resources during the golden rescue period.
- Comparison of Similarly-Sized Areas: Partition 0-4 (32.19 km²) and the original Jishishan task area (32.05 km²) are virtually identical in size. Under a similar unmanned drone configuration (4 unmanned drone), the proposed method for Partition 0-4 requires only 4 sorties to cover its high-value targets, compared to the 17 sorties initially used in Jishishan. This represents a 76% reduction in required sorties for an area of equivalent scale, highlighting the profound efficiency gains of the integrated planning approach.
5. Conclusion
This research addresses the challenges of post-earthquake information collection by unmanned drone by proposing an integrated four-stage technical framework: “High-Value POI Screening – MeanShift-Kmeans Two-Layer Clustering – Multi-Factor Rescue Priority Assessment – Ant Colony Algorithm Route Optimization.” This forms a complete technical chain from data to flight paths. The screening of POIs focuses on six categories of densely populated and functionally critical buildings, providing a precise “target list” for the unmanned drone fleet. The two-layer clustering first eliminates blank areas with MeanShift and then adapts partition size to unmanned drone performance using Z-curve K-means, compressing the initial data coverage from 1775 km² to 681 km²—a 61.6% reduction in ineffective area. The “area-point count” dual-constraint formula allows for calculating the required unmanned drone fleet size, establishing a closed-loop “zone-capacity” adaptation system.
The rescue priority model builds upon an established three-level indicator system by introducing unmanned drone-specific factors—Building Rescue Priority (K) and Number of Patrol Unmanned Drone (N)—enabling priority scores to directly inform subsequent path planning. The Ant Colony Algorithm maps the K values to pheromone concentration, quantitatively translating “where to rescue” into “where to fly first.” Under the dual constraints of rescue task requirements and unmanned drone performance, the model successfully generated 4 closed flight paths for 4 unmanned drone in Partition 0-4, covering 99 POIs. Compared to a real-world earthquake case of similar area, this approach reduced the required number of sorties by 13, representing a 76% efficiency improvement, while high-priority buildings were consistently positioned early in the flight paths.
Nevertheless, the study has areas for future improvement. Regarding terrain adaptation, the impact of extreme post-earthquake topography (e.g., large-scale collapses, landslide dams) on clustering and route planning was not fully considered. Future work could integrate real-time terrain monitoring data to further optimize zoning logic and path avoidance strategies. The parameter settings, such as population ranges and building area assumptions for different POI types, as well as the scoring criteria for buildings, require further calibration with empirical data from diverse cities of varying scales to enhance generalizability. Algorithmically, exploring hybrid approaches that combine ACO with other intelligent algorithms could improve the efficiency and robustness of path planning for more complex and dynamic post-disaster scenarios.
Overall, this research addresses the practical needs of post-earthquake rescue by integrating and innovating multiple technologies to construct a scientific, efficient, and feasible method for rapid unmanned drone search route planning. It not only solves key pain points of traditional search modes but also enriches the technical system for post-disaster information collection. This work provides significant practical reference for the application of the low-altitude economy in the field of emergency response and holds important practical value for improving post-earthquake rescue efficiency and reducing casualties.