Joint optimization method for vehicle-machine collaborative forestry monitoring

Abstract

The invention provides a joint optimization method for vehicle-aircraft collaborative forestry monitoring, which belongs to the technical field of forestry monitoring and comprises the steps of data preprocessing, initial scheme construction, initial course angle optimization, initial Pareto solution set construction and alternate iterative optimization based on a Fix-and-Optimal framework. Ground vehicles and unmanned aerial vehicles are cooperatively used as an integral monitoring system, sub-region division, stop site selection, unmanned aerial vehicle coverage path and vehicle path scheduling are jointly considered, and strategies such as cluster-level compression, sampling evaluation, variable neighborhood search, isolated block detection and local re-division are adopted. And dual-objective optimization of maximizing the monitoring coverage rate and minimizing the total operation time is realized under the engineering constraint. According to the method, the integrity and reliability of monitoring coverage can be improved, the resource allocation efficiency is improved, the solving quality and the calculation expandability are both considered, and the method is suitable for forestry monitoring tasks in large-range and complex terrains.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forestry monitoring, and in particular to a joint optimization method for vehicle-machine cooperative forestry monitoring. BACKGROUND

[0002] The forest ecosystem is facing many challenges, including pest invasion, ecological degradation, biodiversity reduction, and poor forest resource management. Among them, the harm of forestry pests is more significant. Therefore, it is essential to establish an efficient and accurate forestry monitoring system to timely detect problems, respond quickly and take effective measures.

[0003] Effective forestry monitoring requires both high spatial resolution and short time lag response capability. Although traditional ground patrols have high accuracy, they are limited by manpower, time and accessibility, making it difficult to meet the needs of large-scale, sudden pest control. Satellite remote sensing and high-altitude remote sensing can provide wide coverage, but in areas with dense forest canopy and complex terrain, their spatial resolution and timeliness often cannot meet the engineering needs of tree canopy level lesion detection and rapid response.

[0004] In the prior art, unmanned aerial vehicle-based forestry patrol has become a common engineering practice and has been applied in many practical tasks. This method combines the mobility of ground vehicles with the aerial monitoring capability of unmanned aerial vehicles, with unmanned aerial vehicles as aerial data acquisition units and ground vehicles as mobile platforms for carrying and supplying. The vehicle drives to the selected stop location along the road as the take-off, landing, recovery and battery replacement point for the unmanned aerial vehicle. Then the unmanned aerial vehicle takes off from the stop point to perform aerial photography tasks and returns. After completing the monitoring, the vehicle goes to the next stop point according to the plan, and the patrol task is repeated. In existing engineering practice, vehicle-machine cooperative forestry patrol is usually composed of several relatively independent steps, including candidate stop point selection, monitoring area division, unmanned aerial vehicle route design, and vehicle stop sequence. In actual operation, each of the above steps often relies on the on-site experience of the operator or simple heuristic rules to achieve rapid deployment and immediate adjustment. Although this implementation mode is operable in conventional scenarios, it is often difficult to strike a balance between coverage completeness and total operation time in large areas, sparse roads and complex terrain forest areas. The specific problems are as follows.

[0005] 1) The large scale and high combination dimension of candidate stop points limit the scalability and computational efficiency; 2) Experience-based area division is prone to form isolated fragments and repeated coverage; 3) Road network accessibility is not systematically included in the planning, resulting in feasibility and efficiency problems; 4) The decoupling of decision-making makes it difficult to obtain a global optimum or comparable compromise.

[0006] Therefore, there is an urgent need for a forestry monitoring technical solution that can balance coverage integrity and timeliness, and achieve efficient task planning and scheduling under engineering constraints, to solve the above problems existing in the prior art. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application provides a joint optimization method for vehicle-machine cooperative forestry monitoring to solve the problems of low selection efficiency of candidate stopping points, unreasonable regional division, insufficient consideration of road network accessibility, and poor global optimization effect due to fragmented decision-making in existing vehicle-machine cooperative forestry monitoring technology.

[0008] The technical solution of the present application is: a joint optimization method for vehicle-machine cooperative forestry monitoring, comprising the following steps: S1), preprocessing the original space and stopping point data to obtain a final candidate stopping point set that can be reached by vehicles and used for unmanned aerial vehicle take-off and landing; S2), introducing a clustering-based partition strategy on the final candidate stopping point set to divide the sub-regional monitoring units and allocate initial stopping points; S3), based on the pre-computed unit shortest distances from the garage to each stopping point and the pairwise road network distance matrix between stopping points, a greedy nearest neighbor heuristic algorithm is used to construct the vehicle visit order; S4), a genetic algorithm is used to optimize the heading angle of each sub-regional unit; S5), an initial Pareto solution set is constructed; S6), based on the Fix-and-Optimize framework, an alternating iterative optimization is performed to obtain the current Pareto solution set, which contains a set of non-dominated solutions that balance coverage rate and total operation time.

[0009] As a preferred, step S1) specifically comprises the following steps: S11), based on remote sensing images and field survey information, label the non-monitoring units of residential areas, farmland and river channels in the monitoring area as internal holes, remove them from the monitoring area, and form a set of hole polygons after correction ; and define the monitoring area as the difference set of the outer ring and the hole geometry ; S12), regular gridding is performed in the monitoring area to generate discrete task units; S13), road accessibility and safety screening is performed on the original candidate stopping point set to obtain a reasonable candidate set that can be reached by vehicles and used for unmanned aerial vehicle take-off and landing.

[0010] As preferred, step S2) specifically comprises the following steps: S21), for each given UAV dispatch scale , the grid points are clustered into clusters by using a clustering method based on least square error to construct initial sub-area units; S22), after obtaining connected and effective sub-area monitoring units, an initial parking point is selected for each sub-area monitoring unit as a UAV take-off and landing and vehicle parking point.

[0011] As preferred, in step S3), the greedy nearest neighbor heuristic algorithm aims to minimize the total travel time of the vehicle , wherein the total travel time of the vehicle is: In the formula, represents the shortest distance from the parking point to the parking point . .

[0012] The vehicle departs from the garage, and at each step, the current node is selected to the remaining unvisited parking point with the minimum road network cost as the next target for visiting, until all parking points are visited and returned to the garage.

[0013] The beneficial effects of the present application are: 1. The present application ensures that the generated sub-area and selected parking point are reachable and executable in engineering by gridding the monitoring area, removing non-monitoring units and incorporating road accessibility constraints; meanwhile, the point-by-point energy consumption recursion and segment-by-segment return detection are introduced in the route planning, which avoids the interruption of coverage due to insufficient energy, thereby significantly reducing the coverage blind area and missed detection risk, and improving the reliability and safety of the monitoring task; 2. The present application can output a group of comparable and engineering executable scheduling schemes by constructing a dual-objective model of maximum coverage rate and minimum total system operation time, and using a Pareto solution set maintenance mechanism, for the manager to make decisions between resource investment and time efficiency requirements; 3. The present application reduces the combination evaluation of invalid or redundant parking points by cluster-level compression and candidate parking point screening based on location similarity, combined with parking point combination sampling and local search, thereby reducing unnecessary detours of the vehicle and repeated coverage of the UAV; in the case of limited resources, a high-coverage and low-investment compromise solution can be obtained by reasonably regulating the UAV sortie, thereby improving resource utilization; 4. In the case of fixed docking points, this invention provides isolated block detection and two types of local re-partitioning strategies, which can correct local infeasible or inefficient partitioning without changing the global framework, avoid isolated cells being missed or large-scale duplicate coverage, and thus improve the overall coverage efficiency and the engineering adaptability of the solution. 5. This invention adopts a Fix-and-Optimize alternating iterative framework, combined with cluster-level compression, sampling evaluation, variable neighborhood search and parallel evaluation strategies, so that even in engineering instances with a large number of candidate docking points and fine grid granularity, it can still steadily expand the non-dominated solution set within a limited computational budget and maintain the diversity and quality of solutions, thus possessing scalability and practicality on an engineering scale. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a flowchart of the alternating iterative optimization process based on the FO framework of this invention; Figure 3 This is a flowchart illustrating the optimization of stop point combinations using a fixed partitioning scheme according to the present invention. Figure 4 This is a flowchart illustrating the sub-region re-division under a fixed stopping point according to the present invention. Detailed Implementation

[0015] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1 like Figure 1 As shown, this embodiment provides a joint optimization method for vehicle-machine collaborative forestry monitoring, including the following steps: S1) Preprocess the original space and docking point data to obtain the final candidate docking point set that vehicles can reach and that can be used for drone take-off and landing. This includes the following steps: S11) Based on remote sensing imagery and field survey information, non-monitoring units such as residential areas, farmland, and waterways within the monitoring area are marked and removed from the monitoring area as internal voids. After correction, a set of void polygons is formed. ,in, This represents the number of non-monitoring units; Indicates non-monitoring units within the area; monitoring area Defined as the outer ring of the region With void geometry The difference set, i.e.: ; By eliminating non-monitored units, the number of grid units involved in partitioning and path generation is significantly reduced, thereby improving the efficiency and feasible solution ratio of subsequent algorithms. S12) Regularly grid the monitoring area after removing non-monitoring units, retaining those that meet the requirements. The center of the grid is used as the monitoring unit; units that are internal cavities are shielded in subsequent division and coverage assessment. S13) Perform road accessibility and safety screening on the original candidate docking point set to obtain the final candidate docking point set that is accessible to vehicles and can be used for drone take-off and landing. ; This embodiment defines candidate points through deduplication and coordinate cleaning, mapping to road network nodes, and calculation based on road network connectivity and shortest path. Accessibility indication for: In the formula, The shortest travel distance based on the road network; The maximum threshold is determined by the acceptable delay time and vehicle speed. Decide; Represents the set of grids in the monitoring area. Represents a set of subregions; only The points were included in the final candidate docking point set. To facilitate subsequent partitioning schemes and docking point matching, each candidate docking point is pre-mapped according to the grid to obtain the corresponding cell representative, which is used for fast nearest neighbor matching and cluster compression. S2) Introduce a clustering-based partitioning strategy on the final candidate stop point set to divide sub-region monitoring units and allocate initial stop points, specifically including the following steps: S21), for each given drone dispatch size The grid points are clustered using a least-squares error-based clustering method. The initial sub-regions are constructed using clusters, and an initial docking point is assigned to each sub-region. In this embodiment, a clustering-based partitioning strategy is introduced on the discretized set of grid centers, such that the position of each grid point in the projected coordinate system is... The cluster label function is obtained using a partitioning method based on least squares error. ,satisfy: In the formula, For the first The centroid of a cluster; The output of clustering is the set of clusters: ; in, The sub-region monitoring unit is monitored by a sub-region monitoring unit; each cluster represents an initial sub-region unit set; Based on the sub-region unit set of each cluster, the discrete point set is converted into a polygon representation using a buffered convex hull for subsequent visualization and route boundary judgment; when the number of units in the cluster is extremely small, a small-scale buffer is formed around the centroid to avoid invalid areas, and if there are units not covered by any cluster, the unit is assigned to the centroid of the nearest cluster.

[0016] S22), get connected and effective sub-region monitoring units Then, an initial parking point is selected for each sub-region monitoring unit As a UAV landing and vehicle parking point; in order to consider the proximity of the set and the accessibility of the road network, this embodiment adopts a two-stage matching criterion for selection: First, take the sub-region centroid as the representative point, and sort the parking point candidate set According to the projection Euclidean distance, take several nearest candidate points as the candidate parking point set of the sub-region monitoring unit .

[0017] For each candidate point , calculate the shortest driving time from the previous parking point to the candidate point ; if it exceeds the set threshold, it is removed from the candidate set. The initialization scheme and parking point assignment result under each sub-region constitute an initialization scheme, which is used as input for subsequent heading angle optimization and Pareto screening.

[0018] S3), based on the pre-computed unit shortest distance from the garage to each parking point and the pairwise road network distance matrix between parking points, a greedy nearest neighbor heuristic algorithm is used to construct the vehicle access order, as follows: In this embodiment, the ground vehicle is responsible for connecting multiple parking points to provide landing for the UAV, and the path order directly affects the total driving time of the vehicle , and in turn affects the overall operation time, in order to quickly obtain a feasible vehicle route in the initialization phase, this embodiment constructs the vehicle access order based on the shortest driving cost of the road network, and uses the road network graph as the basis for vehicle accessibility and distance measurement. If two points are not reachable on the road network, the road network distance between the two points is given a large penalty , and under the premise of a given average driving speed of the vehicle, the total driving time of the vehicle is as follows, and the goal is to minimize the total driving time of the vehicle: where represents the parking point to the stop point with the shortest distance; .

[0019] In this embodiment, two types of pre-computations are performed on the set of stop points : Compute and cache the cell shortest distances from the garage to each stop point ; ; Compute and cache the pairwise road network distance matrix between stop points; The pre-computations transform the repeated degree of the road network path problem from online queries to one batch computation and caching, significantly improving the speed of subsequent path construction and evaluation. For any inadmissible pair, let automatically generate a huge disadvantage in sorting or optimization, facilitating subsequent screening.

[0020] To obtain a feasible and reasonably good initial solution in a short time, this embodiment uses a greedy nearest neighbor heuristic algorithm based on the pre-computed road network distance matrix in the initialization stage. The idea of the greedy nearest neighbor heuristic algorithm is that the vehicle starts from the garage, and at each step, the node with the smallest road network cost to the remaining unvisited stop point is selected as the next target for visiting, until all stop points are visited and the vehicle returns to the garage. If the current node is , then , where is the set of remaining unvisited nodes; denotes the distance from the current stop point to the next nearest stop point .

[0021] For the starting step, the starting stop point is selected to minimize . This method is simple and convenient to implement and can give a relatively reasonable initial solution under most measured road networks.

[0022] S4), a genetic algorithm is used to optimize the heading angle of each sub-regional cell, specifically including the following steps: In this embodiment, the heading angle is defined as the angle between the flight direction of the UAV and the geographic north, with a value range of ; In this embodiment, the coding method of the genetic algorithm is binary coding. Let the chromosome length be , then the angle corresponding to the chromosome is mapped as: ; In this embodiment The angular resolution is approximately This approach can meet the angular granularity requirements of flight path missions. In this embodiment, the genetic algorithm uses roulette wheel selection for individuals with high fitness, and employs a binary crossover operator with a crossover probability of 0.7 to exchange gene fragments between parents to generate new individuals. The mutation probability is set to 0.12 to ensure sufficient exploration capability even in a small population and to avoid premature convergence.

[0023] In this embodiment, given the docking points and operational parameters, finding the heading angle that maximizes sub-region coverage and minimizes total operational time becomes a crucial local decision problem. To construct the scan sequence, the sub-regions are first... Based on heading angle The division is performed in the projected coordinate system, that is... The set of points is projected onto a one-dimensional coordinate system along the heading angle, and then... The system is divided into rows, and the starting direction of each row is determined by the docking point location to determine the operation sequence of that row, thus forming a complete UAV operation sequence. This grid division process is mathematically equivalent to segmenting a polygon by vector directions.

[0024] In this embodiment, fitness evaluation functions are constructed for coverage priority and job time priority, one of which is coverage priority fitness. Firstly, through based on The simulation of ox-plowing method was used to obtain the flight scheme of the sub-region at this angle and calculate the coverage. With the assignment time; if If the acceptable threshold is not reached, the fitness is determined by the coverage rate; if full coverage is achieved or nearly achieved, a penalty term equal to the inverse of the time is applied to the coverage rate to encourage schemes with shorter operation times to achieve higher fitness. Another is the job time priority fitness, which uses the negative total job time directly as the fitness.

[0025] In this embodiment, the parameters of the ox-plowing method include the heading angle. , heading spacing Lateral spacing In this embodiment, the heading angle is... The angle between the drone's flight direction and due north; the heading distance. The axial distance between adjacent shooting points in the heading direction; the lateral spacing. This represents the cross-band interval between adjacent scan bands. In projected coordinates, the sub-region... Projected onto the heading angle Rotated coordinate system, and spaced laterally The sub-region is divided into several rows in the lateral direction, forming several rows of scanning bands, and then each row is divided into several scanning bands in the heading direction with a heading interval A series of shooting points are generated. Among them: ; Among them, are the effective coverage distances of the camera in the lateral and heading directions respectively at a certain height; , are the heading overlap rate and the lateral overlap rate respectively; wherein the heading overlap rate and the lateral overlap rate are determined by the flight height, the camera field of view and the image overlap rate requirement, and the effective coverage distance , is calculated according to the flight height and the camera inherent width.

[0026] Let the sequence points on each scanning band be , then the flight time of the scanning band is approximately , wherein represents the flight time of adjacent shooting points on the scanning band, represents the additional time cost of the UAV changing rows; this changing action is described as first moving laterally to the starting position of the adjacent scanning band at the end of the band, and then returning to the scanning in reverse order. This process is not an arc turn, which does not produce significant turning energy consumption or speed loss, and the time is determined by the lateral displacement distance and the speed of the UAV. The flight time of the entire sub-region is the sum of the flight times of each scanning band plus the time for the UAV to go from the take-off point to the first sequence point and return to the take-off point from the last sequence point. The corresponding coverage rate can be obtained by the ratio of the number of cells accessed by the UAV to the total number of cells in the sub-region. When increases, the number of scanning bands decreases, thereby reducing the total flight distance and the number of row changes, but the gap covered between each band increases, which may cause the problem of insufficient overlap of the shooting frames in subsequent image stitching and registration; on the contrary, reducing will improve the image overlap and stitching quality, but at the cost of longer flight routes and longer operation time.

[0027] In this embodiment, the sub-region is divided into several rows in the heading angle direction in the coordinate system, and each row is divided into several scanning bands with an interval to generate a sequence of shooting points. This wavy sequence follows a reciprocating flight path, forming a candidate flight path for a single flight task, wherein each flight point corresponds to a shooting position. Each sub-region is completed by a flight task, and the flight time cost can be decomposed into flight time, shooting time and UAV take-off and landing preparation time; that is: ; wherein, is the UAV flight time; is the shooting time; is the UAV take-off and landing preparation time; is a binary variable, if = 1 indicates that there is a flight leg from cell to cell in region ; is the flight distance of the UAV from grid to ; is the average flight speed of the UAV; is the single shooting time of the UAV; sub-region and the corresponding parking point and the local optimization objective of the heading angle is to maximize the coverage rate of this flight under the given energy and time constraints and to minimize the flight time of this flight under the premise of ensuring the coverage requirement as a double objective function, that is: Objective function one: ; Objective function two: ; wherein, is the global region coverage rate; is the sub-region set; is a binary variable, if = 1 indicates that there is a flight leg from cell to cell in region ; is the total operation time; is the flight time cost; is the vehicle path arc, if the vehicle drives from parking point to , then , otherwise ; is the shortest driving distance of the vehicle from parking point to ; is the average driving speed of the vehicle.

[0028] The constraint conditions are: UAV flight constraints: , , lower and upper limits of the number of UAV flights; Sub-region assignment constraint: each region grid can be assigned to at most 1 sub-region; Stop point assignment constraint: each sub-region must be assigned a unique reachable stop point; Vehicle path constraint: if a sub-region chooses a stop point, the vehicle path must contain the point; the vehicle departs from the garage and finally returns to the garage, ensuring vehicle flow conservation and avoiding sub-loops; Heading angle constraint: the heading angle ; UAV path constraint: the UAV path is continuous and has no branches, starting and ending at the selected stop point, avoiding repeated coverage between two sub-regions; Single-pass coverage constraint: limit the number of grid cells covered in a single pass; Stop point reuse constraint: each stop point can be selected by multiple sub-regions.

[0029] S5), construct an initial Pareto solution set, as follows: For each sub-region, perform initial heading angle optimization in parallel and generate the corresponding plough flight scheme, calculate the coverage rate and operation time of each scheme, and use a hybrid seed sampling strategy (70% random sampling; 15% extreme coverage rate priority, 15% extreme operation time priority) to extract several seeds from the initial scheme to construct an initial Pareto solution set.

[0030] S6), alternating iterative optimization based on the Fix-and-Optimize (FO) framework, as shown in Figure 2 , as follows: S61), extract several seed solutions from the current Pareto solution set according to the hybrid strategy as the input of the outer iteration; S62), fix the division scheme and optimize the stop point combination, as shown in Figure 3 , as follows: S621), spatial clustering of local candidate stop points for each sub-region based on location similarity, clustering into several clusters, recording the member set and cluster center position of each cluster, and the number of clusters is adaptively set according to the number of sub-region candidate points; S622), maintain the historical number of attempts , the mean value of performance samples , the baseline score of the cluster is obtained by logit transformation , and the upper confidence bound UCB sampling score is introduced, that is: wherein, is the adjustment of search intensity; is the total number of attempts; S623), based on Extracting several clusters with SoftMax and top-m strategy, SoftMax sampling is sampling according to probability Sampling clusters, where temperature Controlling the flatness of the probability distribution; op-m sampling is sampling according to Sorting the top m clusters with the highest scores, and then randomly selecting specific stops within each cluster, and calculating the resource budget potential cluster in the resource-limited or low-flight interval set; S624), using variable neighborhood search to perturb and locally tune the constructed stop combination candidate, gradually amplifying the perturbation intensity according to the preset neighborhood sequence; S625), eliminating obviously infeasible or inferior solutions through rapid feasibility checking, which includes checking the reachability of each pair of adjacent stops and the road network from the garage to the stops, and estimating the coverage approximation; performing complete simulation evaluation on the candidate solutions that pass the rapid checking, based on the plowing path simulation with selected heading angle, point-by-point energy consumption recursion and return verification, to obtain accurate coverage and total operation time; S626), for each candidate solution that has passed complete simulation and is written into the Pareto solution set, updating the performance sample mean of the cluster with learning rate, that is: = ; wherein, is the learning rate; S63), sub-regional redivision under fixed stops, as shown in Figure 4 , specifically as follows: S631), establishing an adjacency structure and connectivity representation at the grid cell level, extracting the cell set of each sub-region k, performing connectivity component decomposition based on the adjacency mapping to obtain a block set , which is sorted by size, and the largest component is the main block , and the remaining components are identified as isolated blocks if they meet the minimum size threshold; for sub-region , define the total number of isolated cells as , and measure with an encouragement ratio; and calculate the geometric center and the shortest adjacency distance from the nearest neighbor cell of a different region for each isolated block , and estimate the potential marginal contribution of the isolated block to coverage or operation time if it is migrated; S632), for isolated blocks, based on their size, geometric relationship with neighboring areas, and rough target improvement estimates, generate two types of migration candidates: block-priority bulk migration and cell backtracking. Block-priority bulk migration is suitable for isolated blocks with large size and bulk migration can significantly improve coverage or reduce flight time. A candidate target area set is found by nearest neighbor query and sorted by rough estimated benefit to take the top-m targets for trial. Cell backtracking strategy selects several cells from the isolated block in border priority, finds the closest acceptable area for each cell, and tries to migrate; S633), prohibit any migration operation that results in the number of area cells below a preset threshold. After migration is completed, the cell sets of the source and target areas need to be checked for adjacency to ensure spatial connectivity within each area. S634), for each migration candidate, only generate heading angle candidates for the affected sub-areas and perform ploughing simulation. The heading angle candidate set is composed of the excellent heading of the original solution and the fast heuristic generated by aligning the long side. Replace the original values of coverage and operation time for the affected sub-areas with the values obtained by local re-evaluation to obtain global performance indicators. If the variant brings positive gain in global coverage, adopt the migration and update the global partition and performance record. S64), update the Pareto solution set as follows: Determine whether the newly generated candidate solution is dominated by any solution in the current Pareto solution set. If it is dominated, discard it. Otherwise, remove the solutions in the solution set that are dominated by the candidate and add it to the solution set. Use a two-stage refinement mechanism to control the size of the solution set. In the first stage, merge similar solutions using fixed cells in the target space. In the second stage, calculate the representative score of each solution using the normalized coordinates of the target space and retain the solution with the highest score. S65), repeat steps S62)-S64) until there are no new non-dominated solutions in several consecutive internal alternating rounds, or the number of newly added solutions is below the saturation threshold, or the total computation budget reaches the upper limit, completing the internal alternation. Repeat steps S61)-S65) until the maximum number of iterations or a continuous number of rounds without new non-dominated solutions, completing the outer iteration. S66), after the iteration is terminated, output the current Pareto solution set, which contains a set of non-dominated solutions that balance coverage and total operation time, for decision makers to choose.

[0031] The above embodiments and descriptions only illustrate the principles and best embodiments of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements fall within the scope of the claimed application.

Claims

1. A joint optimization method for vehicle-machine collaborative forestry monitoring, characterized in that, Includes the following steps: S1) Preprocess the original space and docking point data to obtain the final candidate docking point set that vehicles can reach and that can be used for drone take-off and landing; S2) Introduce a clustering-based partitioning strategy on the final candidate stop point set to divide the monitoring units into sub-regions and allocate initial stop points; S3) Based on the pre-calculated shortest distance between the garage and each parking point and the pairwise road network distance matrix between parking points, a greedy nearest neighbor heuristic algorithm is used to construct the vehicle access order; S4) A genetic algorithm is used to determine the heading angle of each sub-region unit. Optimize; S5) Construct the initial Pareto solution set; S6) Based on the Fix-and-Optimize framework, alternating iterative optimization is used to obtain the current Pareto solution set, which contains a set of non-dominated solutions that balance coverage and total job time.

2. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 1, characterized in that: Step S1) specifically includes the following steps: S11) Based on remote sensing imagery and field survey information, non-monitoring units such as residential areas, farmland, and waterways within the monitoring area are marked and removed from the monitoring area as internal voids. After correction, a set of void polygons is formed. ; and the monitoring area Defined as the outer ring of the region With void geometry The difference set; S12) Regularly grid the monitoring area to generate discretized task units; S13) Perform road accessibility and safety screening on the original candidate stop point set to obtain a reasonable candidate set that can be reached by vehicles and used for drone take-off and landing. .

3. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 2, characterized in that: Step S2) specifically includes the following steps: S21), for each given drone dispatch size The grid points are clustered using a least-squares error-based clustering method. Use clusters to construct the initial sub-region units; S22) In obtaining connected and effective sub-region monitoring units, select an initial docking point for each sub-region monitoring unit as the take-off and landing point for UAVs and the docking point for vehicles. Preferably, in step S3), the greedy nearest neighbor heuristic algorithm minimizes the total vehicle travel time. The objective is the total driving time of the vehicle. for: In the formula, Indicates the stop point Arrive at the stop The shortest distance; . The vehicle starts from the garage and at each step selects the point with the minimum road network cost among the remaining unvisited stops from the current node as the next visit target, until all stops are visited and the vehicle returns to the garage.

4. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 3, characterized in that: In step S4), the heading angle Defined as the angle between the drone's flight direction and geographic north, with a value range of [value range missing]. The genetic algorithm described above uses binary encoding.

5. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 4, characterized in that: In step S4), sub-region and corresponding stop and heading angle The local optimization objective is to maximize the coverage of this sortie under given energy and time constraints. Minimize the flight time of this sortie while ensuring coverage requirements are met. As a biobjective function, that is: Objective function 1: ; Objective function two: ; In the formula, This refers to the global area coverage rate. For a set of sub-regions; Representing a binary variable, if = 1 indicates that in the region There exists a unit To unit The flight segment; Total work time; For flight time cost; This represents the vehicle path arc, if the vehicle starts from the stop point. driving towards ,but Otherwise ; Indicates that the vehicle departed from the stop. arrive The shortest driving distance; This indicates the average speed of the vehicle.

6. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 5, characterized in that: In step S4), the constraints of the biobjective function are: Drone sortie constraints: , , Minimum and maximum number of drone sorties; Sub-region allocation constraint: Each region grid can be assigned to at most one sub-region; Dock assignment constraint: Each sub-region must be assigned a unique reachable dock; Vehicle path constraints: If a stop point is selected in a sub-area, the vehicle path must include that point; vehicles start from the garage and eventually return to the garage to ensure vehicle flow conservation and avoid sub-loops; Heading angle constraint: Heading angle ; Drone path constraints: The drone path is continuous and without branches, starting and ending at the selected stop point, to avoid overlapping coverage between two sub-regions; Single-flight coverage constraint: Limits the number of grid cells captured per single flight; Stop reuse constraint: Each stop can be selected by multiple sub-regions.

7. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 6, characterized in that: In step S5), the initial heading angle is optimized in parallel for each sub-region and the corresponding ox-plowing flight scheme is generated. The coverage and operation time of each ox-plowing scheme are calculated. A mixed seed sampling strategy is used to extract several seeds from the initial scheme to construct the initial Pareto solution set.

8. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 7, characterized in that: Step S6) specifically includes the following steps: S61) Extract several seed solutions from the current Pareto solution set according to the hybrid strategy, and use them as inputs for the outer iteration; S62) Optimize the combination of fixed division schemes for stopping points; S63) Re-division of sub-regions under fixed stopping points; S64), Pareto solution set update, as follows: Determine whether the newly generated candidate solution is dominated by any solution in the current Pareto solution set. If it is dominated, discard it directly; otherwise, remove the solution dominated by the candidate from the solution set and add it to the solution set. A two-stage simplification mechanism is used to control the size of the solution set. In the first stage, similar solutions are merged by using fixed cells in the target space. In the second stage, the representative score of each solution is calculated using the normalized coordinates of the target space and the solution with the highest score is retained. S65), repeat steps S62)-S64) until no new non-dominated solutions are added in several consecutive internal alternation rounds, or the number of new global solutions is lower than the saturation threshold, or the total computational budget reaches the upper limit, and the internal alternation is completed; repeat steps S61)-S65) until the maximum number of iterations or no new non-dominated solutions are added in several consecutive rounds, and the outer iteration is completed; (S66) After the iteration terminates, the current Pareto solution set is output, which contains a set of non-dominated solutions that balance coverage and total job time, for the decision-maker to choose from.

9. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 8, characterized in that: Step S62) specifically includes the following steps: S621) Spatial clustering of local candidate docking points in each sub-region is performed based on location similarity, forming several clusters. The member set and cluster center position of each cluster are recorded. The number of clusters is adaptively set according to the number of candidate points in the sub-region. S622) To maintain the historical number of attempts and the mean of performance samples for each cluster, the historical mean of the cluster is transformed by logit to obtain the baseline score, and the upper confidence bound UCB sampling score is introduced; S623) Several clusters are extracted based on the UCB sampling scores at the upper confidence bound using the SoftMax and top-m strategies; S624) The system disturbance and local deep adjustment are carried out by variable neighborhood search for the constructed docking point combination candidates, and the disturbance intensity is amplified step by step according to the preset neighborhood sequence. S625) Eliminate obviously infeasible or inferior solutions through rapid feasibility testing. Rapid feasibility testing includes checking the accessibility of each pair of adjacent stops and the road network from the garage to the stops in the candidate stop combination, as well as estimating the approximate coverage value; perform a complete simulation evaluation on the candidate solutions that pass the rapid test, and obtain the accurate coverage rate and total operation time by simulating the cattle plowing path based on the selected heading angle, recursively extrapolating the energy consumption point by point, and verifying the return flight. (S626) For each candidate solution that has been fully simulated and written into the Pareto solution set, the performance sample mean of the cluster is updated using an exponentially weighted method with a learning rate.

10. The joint optimization method for vehicle-machine collaborative forestry monitoring according to claim 9, characterized in that: Step S63) specifically includes the following steps: (S631) Establish adjacency structure and connectivity representation at the grid cell level, and extract the cell set for each sub-region k. Connectivity component decomposition based on adjacency mapping yields a block set. Among them, sorted by size, the largest component is the main block. The remaining components are identified as isolated blocks if they meet the minimum scale threshold; for sub-regions Define the total number of isolated cells as And measured by the proportion of encouragement At the same time for each isolated block Calculate the geometric center and its shortest adjacency distance to the nearest neighboring cell, and estimate the potential marginal contribution of the isolated block to coverage or job time if it is migrated. S632) For isolated blocks, based on their size, geometric relationship with neighboring cells and rough target improvement prediction, two types of migration candidates are generated: block-priority overall migration and cell rollback. Block-first overall migration is suitable for large-scale isolated blocks where overall migration can significantly improve coverage or reduce flight time. It finds a set of candidate target areas through nearest neighbor query and sorts them by estimated benefit, then tries the top-m targets. The cell rollback strategy selects several cells from the isolated block according to boundary priority, finds the closest acceptable area for each cell and tries to migrate. S633) Any migration operation that causes the number of regional units to fall below a preset threshold is prohibited. After the migration is completed, the unit sets of the source region and the target region must pass the adjacency test to ensure that the spatial connectivity within each region is maintained. (S634) For each migration candidate, generate heading angle candidates only for the affected sub-regions and perform ox-plowing simulation. The heading angle candidate set consists of a fast heuristic generation that reuses the good headings of the original solution and aligns with the long side. Replace the original values ​​of the corresponding sub-regions with the coverage and operation time obtained from the local re-evaluation to obtain the global performance index. If the variant brings a positive gain in global coverage, then the migration is adopted and the global partition and performance record are updated.

Images