A low-altitude flight management method and system based on multi-dimensional feature map

Abstract

This invention discloses a low-altitude flight management method and system based on multi-dimensional feature maps. The method includes: collecting terrain, meteorological, and static obstacle data, and fusing them to generate a multi-dimensional environmental feature map; establishing a three-dimensional grid airspace network model based on this map and mapping it to real-time traffic flow; calculating the comprehensive travel cost value for each grid by combining the corresponding risk factors and traffic flow in the feature map; generating an initial optimal flight path using a graph search algorithm with the origin and destination as nodes and the comprehensive cost as the weight; dynamically updating the model and re-evaluating the total cost of the current path based on updated meteorological and traffic data during flight; if the total cost exceeds an acceptable range or a better path exists, triggering replanning and outputting a correction command. This invention, through fusion modeling and dynamic cost optimization, can continuously generate flight paths with the minimum comprehensive cost, improving the safety and efficiency of low-altitude flight.

Description

Technical Field

[0001] This invention relates to the field of low-altitude flight management technology, and in particular to a low-altitude flight management method and system based on multi-dimensional feature maps. Background Technology

[0002] With the development of the low-altitude economy and the increase in low-altitude aircraft, low-altitude flight management has become a crucial link in ensuring the safety and orderliness of low-altitude flights. Currently, one of the most fundamental challenges facing low-altitude flight management lies in how to conduct accurate and real-time capacity assessments from massive, heterogeneous, and dynamic environmental and traffic information, and based on this, plan globally optimal flight paths that can adapt to sudden changes.

[0003] Most existing technical solutions follow a hierarchical processing and step-by-step optimization approach. For example, some solutions pre-plan theoretically safe corridors based on static geographic information databases, and then perform local avoidance based on real-time weather or no-fly zone notices. Other solutions borrow algorithms from robotic path planning, such as rapidly expanding random trees or artificial potential fields, to perform reactive avoidance in real time after detecting obstacles. These methods solve problems in specific scenarios to some extent. However, they reveal significant shortcomings when dealing with the overall complexity of low-altitude environments. Static corridors lack adaptability to dynamic risks (such as sudden flocks of birds or temporary drone swarms). Once the pre-planned path is blocked, the lack of global perspective for replanning often leads to lengthy paths or getting stuck in local optima. Pure reactive avoidance, lacking a comprehensive assessment of global costs, can easily lead to short-sighted decisions by the aircraft, such as entering an area with greater overall traffic flow to avoid a brief airflow disturbance, thus increasing the risk of collision.

[0004] A deeper problem lies in the difficulty of unifying and integrating the risks across four dimensions: terrain, weather, static obstacles, and dynamic traffic flow. These are often treated independently, preventing planning systems from answering a crucial question: In the current comprehensive environment, which path from point A to point B has the lowest overall cost? This overall cost is not simply distance or time, but a weighted balance of safety risks, energy consumption, time costs, and even the impact on overall airspace traffic flow. Lacking such a unified, dynamic cost map, existing management methods often fall short in balancing efficiency and safety, either being overly conservative and limiting airspace capacity, or overly aggressive and creating safety hazards. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art, which is that the low-altitude environment information is fragmented, dynamically changing and lacks unified quantitative assessment, making it difficult to achieve a global and adaptive balance between safety and efficiency in flight path planning. The present invention provides a low-altitude flight management method and system based on multi-dimensional feature maps, which can generate and continuously optimize the flight path with the minimum comprehensive traffic cost by integrating and modeling environmental and traffic data and performing dynamic cost calculations, thereby systematically improving flight safety and operational efficiency in complex and ever-changing low-altitude environments.

[0006] To address the aforementioned technical problems, this invention provides a low-altitude flight management method based on multi-dimensional feature maps, comprising the following steps: Collect terrain data, meteorological data, and static obstacle data of the target low-altitude airspace. Through multi-source data fusion processing, generate a multi-dimensional environmental feature map that includes a terrain elevation layer, a meteorological risk layer, and a static obstacle layer. Based on the multi-dimensional environmental feature map, a spatial network model represented by a three-dimensional mesh is established; real-time traffic flow information is obtained and mapped to the corresponding nodes in the spatial network model in real time. For each node in the airspace network model, the comprehensive passage cost value of the node is calculated by combining at least the terrain risk factor, meteorological risk factor, static obstacle risk factor and real-time traffic flow of the corresponding location in the multi-dimensional environmental feature map. Using the flight start and end points as the starting and target nodes of the airspace network model, respectively, and the comprehensive travel cost as the cost weight for path search, a graph search-based path planning algorithm is used to generate the initial optimal flight path. During the flight mission, the system continuously acquires updated meteorological and traffic flow data and dynamically updates the airspace network model accordingly. Based on the updated airspace network model, the system reassesses the estimated total cost of the current flight path. If the estimated total cost is higher than the preset acceptable cost range, or if there is an alternative path with a lower total cost, the system triggers path replanning, generates and outputs the corrected flight path instructions.

[0007] In one embodiment of the present invention, generating a multi-dimensional environmental feature map includes: Acquire raw terrain data, meteorological observation data, and static obstacle vector data of the target low-altitude airspace; perform rasterization processing on the raw terrain data to generate a basic elevation grid with uniform resolution; Based on the elevation values ​​of each unit in the basic elevation grid and the degree of terrain undulation around them, the terrain risk factor of each unit is calculated to form a terrain elevation layer. The wind speed, visibility, and precipitation intensity parameters in the meteorological observation data are compared with preset threat level thresholds to obtain the quantitative risk level of each parameter. The risk levels of each parameter at discrete points are expanded into a continuous risk distribution field covering the low-altitude airspace of the target through spatial interpolation. The risk levels of each parameter at the same location are weighted and fused to form a meteorological risk layer. For each obstacle in the static obstacle vector data, its three-dimensional influence range is determined based on its type and geometric dimensions. The three-dimensional influence ranges of all obstacles are projected onto the basic elevation grid. For each grid cell, it is determined whether it is covered by one or more three-dimensional influence ranges. If it is covered, an obstacle risk factor is assigned to the cell based on the type and distance of the obstacle covering it. The obstacle risk factors of all cells constitute the static obstacle layer. The terrain elevation layer, meteorological risk layer, and static obstacle layer are aligned and superimposed in the same spatial coordinate system to generate a multi-dimensional environmental feature map.

[0008] In one embodiment of the present invention, establishing a spatial network model includes: Based on the horizontal projection of the basic elevation grid, the grid is divided into layers at preset height intervals in the vertical direction. Define the three-dimensional cube space corresponding to each horizontal grid at each height level as a basic spatial unit, and mark the center point of the basic spatial unit as a network node; A node attribute table is established for each network node to store its three-dimensional spatial coordinates and associate it with the terrain risk factor, meteorological risk factor and obstacle risk factor at the corresponding location in the multi-dimensional environmental feature map.

[0009] In one embodiment of the present invention, acquiring real-time traffic flow information and mapping it in real-time to the corresponding node in the spatial network model includes: Receive broadcast information from airspace surveillance data sources, including real-time position reports of multiple moving aerial targets; Find the basic spatial unit where a moving aerial target is located based on its real-time three-dimensional position; The number of different aerial mobile targets appearing in each basic spatial unit during the most recent update cycle is counted and used as the real-time traffic density value of the network node corresponding to that unit, which is then recorded in its node attribute table.

[0010] In one embodiment of the present invention, the calculation of the comprehensive passage cost value of a node includes: For any network node to be computed in the spatial network model, a cubic spatial region with a predetermined side length is selected with its position in three-dimensional space as the center; the terrain risk factor, meteorological risk factor, obstacle risk factor, and real-time traffic density value obtained from the node attribute table are extracted from the multi-dimensional environmental feature map of all network nodes in the cubic region; these four factors are used as independent feature channels and organized into a three-dimensional multi-channel feature cube as the input of the deep feature learning model. A multi-channel feature cube is input into a pre-trained deep convolutional network structure. The network structure contains multiple cascaded convolutional computation layers and feature map downsampling layers. Each convolutional computation layer performs local scanning and weighted combination operations on the feature map output by the previous layer through a set of learnable feature detectors to extract spatial features at different levels, from simple edges to complex patterns. The feature map downsampling layer is used to expand the effective perception range of subsequent convolutional computations, thereby capturing the coupling relationship and contextual information of environmental factors and traffic flow at different spatial scales within the cube region. The final output layer of a deep convolutional network structure is a fully connected computational layer that maps the extracted and flattened high-dimensional abstract features into a single scalar value, which serves as the comprehensive passage cost value of the network node to be computed.

[0011] In one embodiment of the present invention, the training of the deep feature learning model includes: Collect a large amount of historical flight data and corresponding airspace environment snapshot data; For each historical flight, positive and negative samples are generated based on the actual sequence of airspace network nodes flown over and the corresponding environmental snapshots. Positive samples are nodes that are flown over safely and efficiently and their corresponding feature cubes, and are marked as low overall cost target values. Negative samples are nodes that are dangerous, inefficient, or should be avoided and their corresponding feature cubes, and are marked as high overall cost target values. Using a training dataset consisting of positive and negative samples, the parameters of the deep convolutional network structure are iteratively optimized so that the network's output value for positive samples approaches the low overall cost target value, and the output value for negative samples approaches the high overall cost target value.

[0012] In one embodiment of the present invention, a path planning algorithm based on graph search is used to generate an initial optimal flight path, including: Initialize two node sets: a first set and a second set. The first set is used to store nodes with a determined minimum cumulative cost, and the second set is used to store nodes to be examined. Maintain two attributes for each network node: the current known minimum cumulative cost from the origin to the node, and the identifier of the previous node to reach the node; Place the starting point into the second set and set its current known minimum cumulative cost to 0. Enter the loop, and each time select the node with the smallest known cumulative cost from the second set as the current processing node and move it into the first set; Traverse all network nodes that are adjacent to the current processing node in space. For each adjacent node, calculate the path cost from the starting point to the adjacent node via the current processing node. The path cost is equal to the current known minimum cumulative cost of the current processing node plus the comprehensive travel cost of the adjacent node itself. If the calculated path cost is less than the current known minimum cumulative cost of the current record of the adjacent node, then update the current known minimum cumulative cost of the adjacent node to the path cost, and update the identifier of its predecessor node to the current processing node, while ensuring that the adjacent node is in the second set. Repeat the above loop process until the endpoint is moved into the first set; Starting from the endpoint, the path is traced backwards according to the identifier of the previous node of each node until the starting point is reached. The sequence of nodes traversed constitutes the initial optimal flight path.

[0013] In one embodiment of the present invention, reassessing the estimated total travel cost of the current flight path includes: Based on the updated airspace network model, obtain the latest comprehensive transit cost value of each network node that the current flight path passes through in sequence; The sum of the latest comprehensive transit costs is the estimated total transit cost for the current flight path.

[0014] In one embodiment of the present invention, triggering path replanning and generating and outputting a corrected flight path instruction includes: The network node corresponding to the current real-time location is used as the starting point for the new path search; The network node corresponding to the original flight mission endpoint is used as the unchanging path search endpoint; Based on the latest comprehensive passage cost of all nodes in the updated airspace network model, the path planning process is re-executed to obtain an alternative flight path from the new starting point to the destination. Compare the estimated total cost of alternative flight paths with the estimated total cost of the remaining portion of the current flight path; If the alternative flight path is less costly and the maneuvering required for path switching is within the limits of the aircraft's performance, then the alternative flight path will be output as the corrected flight path command.

[0015] To address the aforementioned technical problems, the present invention also provides a low-altitude flight management system based on multi-dimensional feature maps, used to execute the above method, comprising: The data acquisition and fusion module is configured to: acquire terrain data, meteorological data, and static obstacle data of the target low-altitude airspace, and generate a multi-dimensional environmental feature map containing a terrain elevation layer, a meteorological risk layer, and a static obstacle layer through multi-source data fusion processing; The airspace dynamic modeling module is configured to: establish an airspace network model represented by a three-dimensional mesh based on the multi-dimensional environmental feature map; acquire real-time traffic flow information and map it to the corresponding region or node in the airspace network model in real time; The intelligent cost calculation module is configured to: calculate the comprehensive passage cost value of each region or node in the airspace network model by combining at least the terrain risk factor, meteorological risk factor, static obstacle risk factor and real-time traffic flow at the corresponding location in the multi-dimensional environmental feature map. The path planning module is configured to: use the flight start point and the destination point as the starting node and the target node of the airspace network model, respectively, use the comprehensive passage cost value as the cost weight of the path search, and use a graph search-based path planning algorithm to generate an initial optimal flight path; The dynamic monitoring and replanning module is configured to: continuously acquire updated meteorological and traffic flow data during flight mission execution, and dynamically update the airspace network model accordingly; reassess the estimated total cost of the current flight path based on the updated airspace network model; if the estimated total cost is higher than a preset acceptable cost range, or if there is an alternative path with a lower total cost, trigger path replanning, generate and output the corrected flight path instruction.

[0016] The technical solution of the present invention has the following advantages compared with the prior art: This invention discloses a low-altitude flight management method based on multi-dimensional feature maps. By constructing multi-dimensional environmental feature maps and dynamically updated airspace network models, it achieves unified quantification and fusion of risks from multiple dimensions, including terrain, weather, static obstacles, and real-time traffic flow. This enables accurate and real-time calculation of the comprehensive passage cost of airspace nodes, and based on this, generates a globally optimal initial flight path. During flight mission execution, the system continuously and dynamically monitors environmental changes and triggers path replanning as needed, effectively responding to emergencies, avoiding redundant paths or local optima, and improving the safety, efficiency, and airspace utilization of low-altitude flights. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1This is a flowchart of the steps of the low-altitude flight management method based on multi-dimensional feature maps of the present invention; Figure 2 This is a flowchart of the steps for generating a multi-dimensional environmental feature map according to the present invention; Figure 3 This is a flowchart of the steps for establishing a spatial network model according to the present invention; Figure 4 This is a flowchart of the steps of the present invention to map traffic flow information to the corresponding nodes in the airspace network model in real time; Figure 5 This is a flowchart illustrating the steps involved in calculating the comprehensive passage cost value of a node according to the present invention. Figure 6 This is a flowchart illustrating the steps of generating an initial optimal flight path using a graph search-based path planning algorithm, as described in this invention. Figure 7 This is a structural framework diagram of the low-altitude flight management system based on multi-dimensional feature maps of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Reference Figure 1 As shown, this invention discloses a low-altitude flight management method based on multi-dimensional feature maps. The proposed technical solution constructs a complete methodological system from unified representation of environmental features to dynamic path closed-loop optimization, including: The system collects terrain, meteorological, and static obstacle data from the target low-altitude airspace. Through multi-source data fusion processing, it generates a multi-dimensional environmental feature map containing terrain elevation, meteorological risk, and static obstacle layers. Its principle begins with creating a digital substrate capable of carrying multi-dimensional information. By fusing terrain, meteorological, and static obstacle data, a structured multi-dimensional environmental feature map is generated. This is not merely an overlay of data layers; rather, it uses a unified spatial reference system and quantification standards to assign calculable feature values ​​to each risk factor (such as terrain undulation, wind speed gradient, and building density), thereby integrating previously fragmented environmental information into a framework that can be understood by the algorithm.

[0020] Based on this, a spatial network model represented by a three-dimensional mesh is established according to the multi-dimensional environmental feature map; real-time traffic flow information is acquired and mapped to the corresponding nodes in the spatial network model in real time; by establishing a spatial network model represented by a three-dimensional mesh, the continuous spatial domain is discretized into computable units. The complex spatial navigation problem is transformed into a classic node and edge problem in graph theory. The mapping of real-time traffic flow information injects dynamic elements into this static network, making the cost of each mesh unit no longer a fixed value.

[0021] Next, for each node in the airspace network model, the comprehensive passage cost value of the node is calculated by combining at least the terrain risk factor, meteorological risk factor, static obstacle risk factor and real-time traffic flow of the corresponding location in the multi-dimensional environmental feature map. It is not a simple addition, but based on the risk factors in the feature map and the real-time traffic flow, a dynamic and scalar passage price is output for each grid cell through a model that embeds the multi-objective trade-off relationship of safety, energy efficiency and efficiency. This price intuitively reflects the instantaneous difficulty and risk level of flying between nodes.

[0022] Using the flight start and end points as the starting and target nodes of the airspace network model, respectively, and the comprehensive travel cost as the cost weight for path search, a graph search-based path planning algorithm is used to generate the initial optimal flight path. A physical path constructed with the flight start and end points corresponds to a node sequence in the graph model, and its total cost is the sum of the costs of all nodes along the way. Finding the optimal path is equivalent to finding the path with the lowest total cost in the graph. The graph search algorithm can efficiently find this globally optimal initial optimal flight path in a large network.

[0023] During flight mission execution, continuously updated meteorological and traffic flow data are acquired and dynamically updated based on this data to refine the airspace network model. Based on the updated airspace network model, the estimated total cost of the current flight path is reassessed. If the estimated total cost exceeds a preset acceptable cost range, or if an alternative path with a lower total cost exists, path replanning is triggered, generating and outputting a revised flight path instruction. Ultimately, the method of this invention constructs a real-time "perception-assessment-decision-execution" loop. By continuously acquiring the latest meteorological and traffic data, the system can update the cost values ​​of affected nodes in the airspace network model in real time, ensuring that its internal world model is synchronized with the external real environment. The assessment phase no longer checks the safety of a single point in isolation, but rather reassesses the estimated total cost of the current flight path from a mission-wide perspective, defining the safety and efficiency baseline through a preset acceptable cost range. Simultaneously, it actively seeks alternative paths with lower total costs. This design enables bidirectional optimization: it can proactively avoid risks when the environment deteriorates (e.g., the total cost of the path exceeds the limit), and proactively improve efficiency when the environment improves or a better option emerges (e.g., a lower-cost alternative path is discovered). Once the triggering conditions are met, path replanning is quickly initiated based on the latest cost map, generating revised flight path instructions. This process elevates dynamic risk response from passive, local obstacle avoidance to proactive, global path reconstruction.

[0024] In summary, the synergistic effect of a series of technical features ultimately brings significant benefits, including: First, by creating a multi-dimensional environmental feature map and a dynamic airspace network model, a comprehensive, structured, and computable representation of the complex low-altitude environment and traffic conditions is achieved, fundamentally solving the problem of information fragmentation and providing a reliable data foundation for intelligent decision-making. Second, by introducing comprehensive transit cost as a unified quantitative indicator and performing global optimization through a graph search algorithm, the generated flight path is no longer the optimal solution for a single objective, but rather the optimal solution after balancing multiple constraints such as safety, energy efficiency, and time, thereby fundamentally improving the scientific and economical nature of the path. Finally, based on a real-time data-driven dynamic evaluation and replanning mechanism, the system possesses resilience to cope with emergencies and the potential for continuous performance optimization. It can not only avoid risks in a timely manner to ensure flight safety, but also flexibly adjust when opportunities arise to improve operational efficiency. This allows it to always fly along the trajectory with the optimal comprehensive cost in complex and ever-changing environments, thereby improving the overall utilization efficiency, safety level, and operational predictability of low-altitude airspace.

[0025] In low-altitude flight management methods, it is necessary to generate multi-dimensional environmental feature maps that include terrain elevation layers, meteorological risk layers, and static obstacle layers. However, the key to building a reliable flight management system lies in how to refine and effectively integrate raw, heterogeneous terrain, meteorological, and static obstacle data to accurately and quantitatively characterize the risks of the airspace environment. If the generation process of the environmental feature map is not refined enough or lacks a unified standard, it will directly affect the accuracy of subsequent airspace network models and the effectiveness of path planning.

[0026] In this regard, refer to Figure 2 As shown, this application further proposes a specific method for generating multi-dimensional environmental feature maps, including the following steps: First, acquire raw topographic data, meteorological observation data, and static obstacle vector data for the target low-altitude airspace. Then, rasterize the raw topographic data to generate a base elevation grid with uniform resolution. The raw topographic data can originate from digital elevation models, lidar scan data, or aerial photogrammetry data, providing surface elevation information. Meteorological observation data can come from ground weather stations, weather radar, satellite remote sensing, or UAV-borne sensors, containing real-time or forecast meteorological parameters such as wind speed, wind direction, temperature, humidity, visibility, and precipitation intensity. Static obstacle vector data typically originates from geographic information system databases, building information models, or manual surveying, including the precise location, height, shape, and type information of fixed structures such as high-rise buildings, transmission towers, chimneys, and bridges. Rasterizing the raw topographic data transforms irregular or continuous raw topographic data into a regular grid structure. This typically involves defining a uniform grid cell size and then sampling or averaging the topographic elevation within each grid cell to generate a base elevation grid with uniform resolution, facilitating subsequent spatial analysis and data overlay.

[0027] Secondly, based on the elevation values ​​of each cell in the basic elevation grid and the degree of terrain undulation in the surrounding area, a terrain risk factor is calculated for each cell, forming a terrain elevation layer. The terrain risk factor aims to quantify the potential impact of terrain on low-altitude flight. The elevation value itself may affect flight altitude limits, while the degree of terrain undulation, such as slope, aspect, roughness, and terrain obstruction, directly relates to the aircraft's maneuverability, collision risk, and airflow disturbance. For example, slope can be assessed by calculating the elevation difference between a grid cell and its adjacent cells; the greater the slope, the higher the risk factor. These calculation results are normalized and mapped to preset risk levels, thus forming the terrain elevation layer.

[0028] Furthermore, the wind speed, visibility, and precipitation intensity parameters from meteorological observation data are compared with preset threat level thresholds to obtain the quantified risk level of each parameter. Using spatial interpolation methods, the risk levels of each parameter at discrete points are expanded into a continuous risk distribution field covering the target low-altitude airspace. The risk levels of parameters at the same location are then weighted and fused to form a meteorological risk layer. Meteorological parameters such as wind speed, visibility, and precipitation intensity are crucial to the safety of low-altitude flight. The preset threat level thresholds are determined based on aircraft performance, flight mission type, and safety standards. Comparing the observed parameter values ​​with these thresholds quantifies them into risk levels. Since meteorological observation data is typically discrete station data, spatial interpolation methods, such as Kriging interpolation, inverse distance weighted interpolation, and spline interpolation, are used to extrapolate the continuous meteorological risk distribution throughout the entire target low-altitude airspace based on the risk levels of these discrete points, ensuring a reasonable meteorological risk assessment even in areas without direct observation points. At the same location, different meteorological parameters may pose risks simultaneously. Weighted fusion combines the risk levels of these independent parameters to form a unified meteorological risk value. The weights can be set according to the degree of impact of different meteorological parameters on flight safety.

[0029] Furthermore, for each obstacle in the static obstacle vector data, its three-dimensional influence range is determined based on its type and geometric dimensions. The three-dimensional influence ranges of all obstacles are projected onto a base elevation grid. For each grid cell, it is determined whether it is covered by one or more three-dimensional influence ranges. If covered, an obstacle risk factor is assigned to the cell based on the type and distance of the covering obstacle. The obstacle risk factors of all cells constitute the static obstacle layer. Static obstacles not only occupy physical space but also affect the surrounding airspace. For example, tall towers may generate turbulence or require a certain safety margin for avoidance. Determining the three-dimensional influence range requires considering the physical dimensions and type of the obstacle and setting a safety buffer zone according to flight safety standards. If a grid cell is covered by an obstacle influence range, an obstacle risk factor is assigned to the cell based on the type of obstacle and the distance from the cell to the obstacle, reflecting the potential collision or avoidance risk in that area due to static obstacles.

[0030] Finally, the terrain elevation layer, meteorological risk layer, and static obstacle layer are aligned and overlaid in the same spatial coordinate system to generate a multi-dimensional environmental feature map. Since all risk layers are generated based on a uniform resolution base elevation grid, they naturally share the same spatial reference and resolution. Alignment and overlay refers to logically or physically combining these independent risk layers to form a multi-channel or multi-attribute raster data structure. Each grid cell now contains a set of attribute values ​​representing the terrain risk, meteorological risk, and obstacle risk at that location.

[0031] Furthermore, after generating a multi-dimensional environmental feature map containing a terrain elevation layer, a meteorological risk layer, and a static obstacle layer, this application needs to effectively transform this two-dimensional risk information into a spatial model that can be used for three-dimensional path planning. To this end, referring to... Figure 3 As shown, this application further proposes a method for establishing a spatial network model, including: The horizontal projection of the baseline elevation grid with uniform resolution obtained during the generation of multi-dimensional environmental feature maps is used as a two-dimensional plane reference. Above this reference, in the vertical direction, i.e., the height dimension, the space is divided layer by layer according to preset height intervals. These preset height intervals can be set according to the needs of the actual application scenario. For example, they can be set to fixed height values ​​(such as 10 meters or 20 meters) to ensure uniform airspace coverage; or non-uniform height intervals can be used depending on the complexity of the airspace, the performance limitations of the aircraft, or the specific flight mission requirements. For example, smaller height intervals can be used in areas with large terrain undulations or dense obstacles, while larger height intervals can be used in open airspace to balance model accuracy and computational efficiency. Through this layered processing, the continuous vertical airspace is discretized into a series of discrete height layers.

[0032] After vertical layering, each horizontal grid cell intersects with each height layer, forming a cubic space with specific three-dimensional dimensions. This cubic space is defined as a basic spatial unit, representing a minimal identifiable region in the spatial domain. To facilitate subsequent graph search path planning, the geometric center point of each basic spatial unit is abstracted as a network node. This network node, as a fundamental element in the spatial network model, carries all relevant information about the basic spatial unit and serves as a traversable point in the path planning algorithm. This method of discretizing three-dimensional space into basic spatial units and using their center points as network nodes effectively transforms a continuous spatial domain into a discrete, computable graph structure.

[0033] An independent node attribute table is created for each network node, storing its precise coordinates in three-dimensional space (e.g., longitude, latitude, and altitude). More importantly, this attribute table also associates the risk factors of the basic spatial unit represented by the network node with the corresponding location in the multi-dimensional environmental feature map, including terrain risk factors, meteorological risk factors, and obstacle risk factors. This association can be achieved by querying the corresponding layer of the multi-dimensional environmental feature map, extracting the corresponding risk value based on the network node's three-dimensional coordinates or the range of its basic spatial unit. For example, the risk value can be obtained through interpolation or direct sampling. In this way, each network node not only represents a spatial location but also contains comprehensive environmental risk information for that location, providing a direct data foundation for subsequent comprehensive access cost calculations.

[0034] Furthermore, after establishing the airspace network model, it is also necessary to map traffic flow to the corresponding nodes in the airspace network model in real time, referring to... Figure 4 As shown, the specific steps include: First, receiving broadcast information from airspace surveillance data sources. This information typically contains real-time position reports of multiple moving aerial targets. Airspace surveillance data sources can be ADS-B systems, radar systems, or other cooperative / non-cooperative surveillance systems, which continuously transmit standardized data packets containing parameters such as aircraft identity, three-dimensional position (e.g., longitude, latitude, altitude), and speed. The receiving module (e.g., an ADS-B receiver or data link port) continuously listens for and parses these broadcast messages, extracting the real-time three-dimensional position information of each detected moving aerial target.

[0035] Subsequently, based on the real-time 3D position of an aerial moving target, the system needs to locate its basic spatial unit. A basic spatial unit is a predefined 3D grid cell with a specific volume within the spatial network model. By converting the geographic coordinates of the aerial moving target to the local coordinate system used by the spatial network model, and combining this with preset horizontal grid resolution and vertical height layering, the target's current horizontal grid cell and vertical height layer can be accurately determined, thus uniquely identifying its basic spatial unit. This lookup process can be achieved using spatial indexing structures (such as kd-trees or octrees) or through direct calculation based on the grid origin and resolution to ensure efficient positioning.

[0036] Based on this, the system counts the number of different aerial moving targets appearing in each basic spatial unit within the most recent update cycle and uses this number as the real-time traffic density value of the corresponding network node for that unit, recording it in its node attribute table. To ensure the timeliness of traffic information, the system sets a time window as the most recent update cycle (e.g., 5 seconds or 10 seconds). Within each update cycle, the system tracks and records the unique aerial moving targets entering or passing through each basic spatial unit. After the cycle ends, the system counts the number of different aerial moving targets within each basic spatial unit and uses this number as the real-time traffic density value of the corresponding network node for that unit, updating the node's attribute table. This process continues to cycle, ensuring that the traffic density value remains up-to-date.

[0037] In practical applications, simply linearly combining or weighting these independent risk factors and traffic information may fail to fully capture the complex and nonlinear interactions between various factors in the low-altitude airspace environment. For example, the combination of specific terrain conditions with specific wind direction and speed, or the superposition of high traffic flow and low visibility, may generate complex risks far exceeding the sum of a single factor. These deep-seated coupling relationships and contextual information are crucial for accurately assessing the true cost of flight paths.

[0038] In this regard, refer to Figure 5 As shown, this application further proposes a method for calculating the comprehensive passage cost value of a node. Specifically, in calculating the comprehensive passage cost value of any network node, firstly, a cubic spatial region with a predetermined side length is delineated, centered on the precise location of the network node in three-dimensional space. This predetermined side length can be flexibly configured according to the scale of the actual application scenario, aircraft performance, or environmental characteristics; for example, it can be set to tens to hundreds of meters to ensure that the selected region can cover the local environmental range that significantly affects flight decisions. This cubic region aims to provide the target node with sufficient local contextual information, avoiding isolated judgments based solely on the node's own attributes.

[0039] Subsequently, the system traverses all network nodes within the cubic spatial region and, for each node, extracts its corresponding terrain risk factor, meteorological risk factor, and obstacle risk factor from the multi-dimensional environmental feature map. Simultaneously, it retrieves the node's latest real-time traffic density value from its attribute table. These four factors—terrain risk factor, meteorological risk factor, obstacle risk factor, and real-time traffic density value—are treated as independent feature dimensions or channels. They are organized into a three-dimensional data structure, namely a multi-channel feature cube. For example, if the cubic region contains N x N x N network nodes, then this feature cube will be an N x N x N x 4 tensor, where 4 represents the four feature channels. This structured representation fully preserves the distribution of each factor in three-dimensional space and their relative positions, providing rich spatial context information for subsequent deep learning models.

[0040] Next, this multi-channel feature cube is fed into a pre-trained deep convolutional network structure. This network structure is specifically designed to process spatially correlated data, and its core capability lies in its ability to automatically learn and extract complex patterns from the data. The network typically consists of multiple cascaded convolutional computation layers and feature map downsampling layers. Each convolutional computation layer contains a set of learnable feature detectors (i.e., convolutional kernels), which perform local sliding window-like scanning and weighted combination operations on the feature map output from the previous layer. Through this operation, the network can progressively extract spatial features at different levels of abstraction from the raw risk factor and traffic data, from simple local edges and gradient changes to more complex, semantically meaningful patterns, such as the shape of a danger zone formed by the superposition of specific terrain and meteorological conditions, or the combination pattern of high traffic volume and densely obstructed areas. Feature map downsampling layers, such as max pooling layers or average pooling layers, play a role between the convolutional layers. They effectively expand the effective perception range of subsequent convolutional computations by reducing the spatial resolution of the feature maps. This means that as data is passed layer by layer through the network, deep convolutional kernels can see a larger area within the original input cube, thus capturing the coupling relationship and contextual information between environmental factors and traffic flow at different spatial scales within the cube. For example, shallow layers may identify strong wind shear in a small area, while deep layers can link this strong wind shear to complex terrain and high-density traffic flow in a wider area, forming a more comprehensive assessment of the overall risk.

[0041] Finally, after multiple convolutions and downsampling processes, the high-dimensional abstract features extracted by the network are flattened and input into the final output layer of the deep convolutional network structure, which is a fully connected computation layer. This fully connected layer synthesizes these abstract features and maps them to a single scalar value. This scalar value is the comprehensive passage cost of the network node to be computed, which quantifies the comprehensive difficulty, risk, or efficiency loss encountered when passing through that node.

[0042] Specifically, the training method for deep feature learning models includes: First, collecting a large amount of historical flight data and corresponding airspace environment snapshot data. Historical flight data can include information such as the actual flight trajectory, speed, altitude, and attitude, as well as records of events that occurred during the flight, such as successful passage, encounters with turbulence, avoidance maneuvers, and even accidents or emergencies. Airspace environment snapshot data consists of environmental information such as terrain data, meteorological data, static obstacle data, and real-time traffic flow in the target low-altitude airspace at the time these historical flights occurred. This data needs to be time-synchronized and spatially aligned to ensure that each segment of historical flight data corresponds to the specific environmental conditions at the time of its occurrence.

[0043] Secondly, for each historical flight, positive and negative samples are generated based on the actual sequence of airspace network nodes flown over and the corresponding environmental snapshots. Specifically, for airspace network nodes flown safely and efficiently in historical flights, a three-dimensional multi-channel feature cube composed of terrain risk factors, meteorological risk factors, static obstacle risk factors, and real-time traffic flow at the corresponding location in the multi-dimensional environmental feature map is marked as a positive sample and assigned a low comprehensive cost target value, such as close to zero or a preset minimum safe cost value. These nodes represent areas where aircraft can pass smoothly under specific environmental conditions. Conversely, for airspace network nodes that were dangerous, inefficient (e.g., prolonged circling, significant detours), or should be avoided (e.g., encountering strong winds, sudden drops in visibility, approaching obstacles) in historical flights, their corresponding feature cubes are marked as negative samples and assigned a high comprehensive cost target value, such as a preset maximum risk cost value. These nodes represent areas where aircraft should avoid or proceed with caution under specific environmental conditions. In this way, actual flight experience is transformed into training data with clear cost labels.

[0044] Finally, the parameters of the deep convolutional network structure are iteratively optimized using a training dataset consisting of these positive and negative samples. This optimization process aims to adjust the weights and biases within the network so that when the network receives a feature cube input of a positive sample, its output comprehensive passage cost value approaches a preset low comprehensive cost target value; while when it receives a feature cube input of a negative sample, its output comprehensive passage cost value approaches a preset high comprehensive cost target value. This is typically achieved by minimizing a loss function that measures the difference between the network's prediction and the target value. For example, mean squared error (MSE) or cross-entropy loss functions can be used, combined with optimization algorithms such as gradient descent for parameter updates. Through repeated iterative training, the deep convolutional network structure can learn to extract complex patterns and nonlinear relationships related to flight risk and efficiency from multi-dimensional environmental features and real-time traffic flow, thereby accurately predicting the comprehensive passage cost value of nodes.

[0045] In actual low-altitude flight management, the airspace environment is complex and ever-changing, with a massive number of nodes. Efficiently and accurately searching for a truly optimal flight path that meets real-time requirements from this vast number of nodes is crucial for ensuring flight safety and efficiency. Traditional path search methods may suffer from low computational efficiency and an inability to effectively handle dynamic cost changes, leading to less than ideal path planning results.

[0046] In this regard, refer to Figure 6 As shown, this application further proposes a path planning algorithm based on graph search to generate an initial optimal flight path, specifically including the following steps: Initialize two sets of nodes: a first set and a second set. The first set stores nodes whose minimum cumulative cost has been determined, and the second set stores nodes to be examined. These two sets are the core data structures used by the graph search algorithm to manage the search state. The first set typically stores nodes whose shortest paths from the starting point have been determined; once a node is moved to this set, its shortest path is considered finalized. The second set stores all nodes whose shortest paths have not yet been determined. These nodes are candidate nodes that the algorithm is currently exploring or waiting to be processed, and are usually implemented in the form of a priority queue to efficiently select the minimum cost node.

[0047] Each network node maintains two attributes: the current known minimum cumulative cost from the origin to the node, and the identifier of the node preceding the destination node. The current known minimum cumulative cost records the sum of the total travel costs of all nodes along the path from the origin to the current node. This value is continuously updated during algorithm execution to ensure that it always records the cost of the currently discovered shortest path. Initially, the cumulative cost of the origin is 0, and the cumulative cost of other nodes is infinity. The identifier of the node preceding the destination node records which predecessor node the node was reached from when the current known minimum cumulative cost was achieved. This identifier is crucial for path backtracking, allowing the complete flight path to be traced backward from the destination node after the algorithm ends.

[0048] The starting node is placed in the second set, and its current known minimum cumulative cost is set to 0. This is the algorithm's initialization step; the starting node is used as the starting point of the search, and its cumulative cost to itself is naturally 0. Placing it in the second set to be examined indicates that the algorithm will start exploring its neighboring nodes from this node.

[0049] Entering the loop, each time the node with the smallest known cumulative cost is selected from the second set as the current processing node and moved to the first set. In each iteration, the algorithm always selects the node with the smallest cumulative cost from the starting point to itself among all the nodes to be examined. This node is considered the current processing node because its shortest path has been determined. Moving it to the first set indicates that the node has been processed and does not need to be examined again.

[0050] The algorithm iterates through all network nodes spatially adjacent to the current processing node. For each adjacent node, it calculates the path cost from the starting point to the adjacent node via the current processing node. The path cost equals the current known minimum cumulative cost of the current processing node plus the adjacent node's own comprehensive travel cost. When a node is selected as the current processing node, the algorithm checks all its directly connected neighbors. For each neighbor, it calculates a new path cost: the known minimum cumulative cost from the starting point to the current processing node, plus the cost of the "edge" from the current processing node to that neighbor, which is the neighbor node's own comprehensive travel cost, since the node cost in the spatial network model represents the cost of passing through that node.

[0051] If the calculated path cost is less than the current known minimum cumulative cost of a neighboring node, then update the current known minimum cumulative cost of that neighboring node to the path cost, and update the identifier of its predecessor node to the currently processed node. Simultaneously, ensure that the neighboring node is in the second set. If the path cost to a neighboring node via the currently processed node is less than any previously known path cost, then update the minimum cumulative cost and the identifier of its predecessor node. This means a better path has been found. At the same time, ensure that the neighboring node remains in the second set so that it can be selected and further explored in subsequent iterations.

[0052] Repeat the above loop until the destination node is moved into the first set. The algorithm continues to execute the above loop until the target destination node is selected and moved into the first set. At this point, the shortest path from the starting point to the destination has been determined.

[0053] Starting from the endpoint, the path is traced backwards according to the identifier of the preceding node of each node, back to the starting point. The sequence of nodes traversed constitutes the initial optimal flight path. Once the endpoint is processed, it can be traced back to its predecessor node through its preceding node identifier, and then back to its predecessor node, repeating this process until the starting node is reached. This reverse-traced sequence of nodes, after being reversed, is the initial optimal flight path from the starting point to the endpoint.

[0054] During low-altitude flight missions, the airspace environment is not static; weather conditions, traffic flow, and other factors continuously change dynamically. To ensure flight safety and efficiency, the airspace network model needs to be updated in a timely manner based on these changes. To this end, this application further proposes a method for reassessing the estimated total passage cost of the current flight path, which includes: obtaining the latest comprehensive passage cost value for each network node traversed sequentially by the current flight path based on the updated airspace network model; and summing the latest comprehensive passage cost values ​​to obtain the estimated total passage cost of the current flight path.

[0055] Specifically, during flight mission execution, the system continuously acquires updated meteorological and traffic flow data and dynamically updates the airspace network model accordingly. When the airspace network model is updated, the previously calculated comprehensive passage cost values ​​for nodes may no longer be accurate. Therefore, it is necessary to query or recalculate the latest comprehensive passage cost value for each network node included in the current flight path from the updated airspace network model. For example, if the meteorological conditions in a node area suddenly deteriorate, its meteorological risk factor will increase, leading to an increase in the comprehensive passage cost value of that node. This step ensures that subsequent path assessments are based on the most real-time and accurate environmental conditions. Subsequently, by simply summing the latest comprehensive passage cost values ​​of all network nodes on the current flight path, the overall estimated total passage cost of the path under the current environment is obtained. This cumulative approach intuitively reflects the total cost an aircraft needs to travel along the path from its current position to its destination. This cost incorporates various risk factors such as terrain, weather, obstacles, and traffic flow. By accumulating these costs, the overall feasibility, safety, and efficiency of the path can be quantitatively assessed, providing a unified metric for comparison with a pre-defined acceptable cost range or with the costs of alternative paths.

[0056] Furthermore, regarding the updating of the airspace network model, this application proposes a method for triggering path replanning and generating and outputting revised flight path instructions, specifically including: First, the network node corresponding to the current real-time location is used as the starting point for the new path search. During mission execution, the aircraft's position is constantly changing. When path replanning is required, the new path search should not start from the original starting point, but rather from the aircraft's current actual position to ensure that the planned path is an effective guide for the current state. Specifically, the system obtains the aircraft's real-time 3D coordinates through its navigation system (such as GPS, inertial navigation system, etc.). Then, these real-time 3D coordinates are matched with nodes in the spatial network model to find the network node corresponding to the nearest basic spatial unit that contains the real-time location, and this node is determined as the new starting point for the path search. This ensures that the replanned path seamlessly integrates with the aircraft's current state.

[0057] Meanwhile, the network node corresponding to the original flight mission endpoint is used as the unchanging endpoint for path search. Although the flight path needs to be dynamically adjusted, the ultimate goal of the flight mission is usually fixed. Therefore, during path replanning, the endpoint remains unchanged to ensure that the aircraft can eventually reach the predetermined destination regardless of the adjustments. During mission initialization, the flight mission endpoint is determined and mapped to a specific network node in the airspace network model. During the replanning process, this endpoint node identifier is preserved as the target of the path search algorithm.

[0058] Based on this, the path planning process is re-executed using the latest comprehensive travel cost values ​​of all nodes in the updated airspace network model, resulting in a new alternative flight path from the starting point to the destination. The core of path replanning lies in utilizing the latest environmental information and traffic conditions. Therefore, it must be based on a dynamically updated airspace network model, which includes the latest terrain risk factors, weather risk factors, static obstacle risk factors, and comprehensive travel cost values ​​calculated from real-time traffic flow. Specifically, the system employs the same graph search-based path planning algorithm as the initial path planning, but the input data is the updated airspace network model and the latest comprehensive travel cost values ​​of its nodes. The algorithm starts from the new path search starting point, using these latest cost values ​​as cost weights, to search for the optimal path to the unchanged path search destination, thereby generating an alternative flight path reflecting the current airspace conditions.

[0059] Subsequently, the estimated total cost of the alternative flight path is compared with the estimated total cost of the remaining portion of the current flight path. Deciding whether to switch to the alternative path requires a quantifiable comparison standard. This standard is the estimated total cost of each path. Specifically, for the alternative flight path, its estimated total cost is the sum of the combined costs of all its nodes. For the current flight path, the estimated total cost from the aircraft's current position to the original destination needs to be calculated, similarly obtained by summing the latest combined costs of all nodes on this remaining path.

[0060] If the alternative flight path is less costly, and the maneuvers required for the path switch are within the aircraft's performance limits, the alternative flight path will be output as the corrected flight path command. Path switching considers not only whether the new path is less costly, but also the feasibility and safety of the switch itself. If the switching maneuver is too drastic or exceeds the aircraft's performance limits, even if the new path is theoretically superior, it may be unfeasible or introduce new risks. Specifically, the system first compares the estimated total cost of the two paths. If the alternative path is indeed less costly, the maneuvers required to switch from the current flight path to the alternative path are further evaluated. This includes, but is not limited to, turning radius, climb / descent rate, and speed changes. These parameters are compared with the aircraft's preset performance envelope. Only when the alternative path is less costly and the switching maneuvers are within the aircraft's performance limits will the system confirm the alternative path as a feasible corrected path and convert it into flight commands (such as waypoint sequences, speed commands, etc.) and output it to the flight control system.

[0061] Reference Figure 7As shown, in order to implement the above method, this application further proposes a low-altitude flight management system based on multi-dimensional feature maps, including a data acquisition and fusion module, an airspace dynamic modeling module, an intelligent cost calculation module, a path planning module, and a dynamic monitoring and replanning module.

[0062] The data acquisition and fusion module is primarily responsible for obtaining raw environmental information from multiple data sources and performing preprocessing and integration. Specifically, it can be configured to receive topographic elevation data from Geographic Information Systems (GIS), meteorological observation data from meteorological stations or satellites, and static obstacle databases from airspace management departments via interfaces. After receiving this raw data, the module performs data cleaning, format conversion, and spatial registration to ensure that data from different sources can be fused within a unified spatial coordinate system. For example, topographic data may be rasterized to generate a basic elevation grid, meteorological data may be spatially interpolated to form a continuous risk distribution field, and static obstacle data may be converted into a three-dimensional impact range. Finally, the module overlays these processed data to generate a multi-dimensional environmental feature map containing a topographic elevation layer, a meteorological risk layer, and a static obstacle layer, providing comprehensive environmental background information for subsequent airspace modeling.

[0063] The core function of the airspace dynamic modeling module is to construct and maintain an airspace network model that reflects the three-dimensional structure and dynamic changes of the airspace based on the generated multi-dimensional environmental feature map. Specifically, this module can be configured to use the horizontal projection of the environmental feature map as a reference, and then divide the entire airspace into a series of three-dimensional cubic spatial units according to preset height intervals in the vertical direction. The center point of each spatial unit is defined as a network node, and a node attribute table is established for it to store its three-dimensional spatial coordinates and the terrain risk factors, meteorological risk factors, and obstacle risk factors associated with it from the multi-dimensional environmental feature map. In addition, this module is responsible for continuously acquiring real-time traffic flow information from airspace monitoring systems (such as ADS-B and radar) and mapping it to the corresponding areas or nodes in the airspace network model. For example, by counting the number of aerial moving targets passing through a spatial unit within a specific time window, this number is used as the real-time traffic density value of the network node corresponding to that unit and updated in the node attribute table, thus giving the airspace network model real-time dynamics.

[0064] The intelligent cost calculation module aims to comprehensively assess the passage cost of each region or node in a spatial network model, quantifying the difficulty or risk of passing through that region or node. Specifically, this module can be configured to select a cubic spatial region of predetermined side length centered on the location of any region or node to be calculated in three-dimensional space. It then extracts terrain risk factors, meteorological risk factors, obstacle risk factors, and real-time traffic density values ​​obtained from the node attribute table for all network nodes within this region, based on their corresponding multi-dimensional environmental feature maps. These factors are organized into a multi-channel feature cube, serving as input to a pre-trained deep feature learning model. This deep feature learning model, such as a deep convolutional network structure, extracts different levels of spatial features and contextual information from the local environment and traffic flow through multi-layer convolutional computation and feature map downsampling. Finally, the network's output layer maps these abstract features into a single scalar value—the comprehensive passage cost of the region or node. This intelligent calculation method can capture complex nonlinear relationships that are difficult to detect using traditional methods, thus providing a more accurate cost assessment.

[0065] The path planning module is responsible for generating the optimal flight path from the flight origin to the destination based on the airspace network model and the comprehensive travel cost values ​​of regions or nodes. Specifically, this module can be configured to employ a graph search-based path planning algorithm. During algorithm execution, the flight origin and destination are designated as the starting and target nodes of the airspace network model, respectively, while the comprehensive travel cost value of each region or node is used as the cost weight for path search. This module maintains the cumulative cost of regions or nodes and information on preceding regions or nodes, and iteratively expands the set of explored regions or nodes until the path with the minimum cumulative cost from the origin to the destination is found. Finally, the module outputs the optimal region or node sequence obtained through reverse tracing as the initial optimal flight path, providing preliminary route guidance for the flight mission.

[0066] The dynamic monitoring and replanning module is crucial for ensuring the safe and efficient execution of flight missions in dynamic environments. Specifically, this module is configurable to continuously receive and process updated meteorological and traffic flow data during flight mission execution. Based on this real-time data, the module dynamically updates the airspace network model, including risk factors and traffic density values ​​for regions or nodes. Subsequently, the module reassesses the latest comprehensive passage cost values ​​for all regions or nodes on the current flight path and sums them to obtain the estimated total passage cost for the current path. The module continuously monitors this estimated total passage cost and compares it with a preset acceptable cost range. If the estimated total passage cost exceeds the acceptable range, or if analysis reveals an alternative path with a lower total passage cost, the module immediately triggers the path replanning process. During replanning, the module uses the network node corresponding to the aircraft's current real-time position as the new starting point, the original flight mission endpoint as the target node, and re-executes the path planning algorithm using the updated airspace network model to generate and output a revised flight path command, thereby achieving real-time adjustment and optimization of the flight path.

[0067] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for low altitude flight management based on multi-dimensional feature map, characterized in that, Includes the following steps: Collect terrain data, meteorological data, and static obstacle data of the target low-altitude airspace. Through multi-source data fusion processing, generate a multi-dimensional environmental feature map that includes a terrain elevation layer, a meteorological risk layer, and a static obstacle layer. Based on the multi-dimensional environmental feature map, a spatial network model represented by a three-dimensional mesh is established; real-time traffic flow information is obtained and mapped to the corresponding nodes in the spatial network model in real time. For each node in the airspace network model, the comprehensive passage cost value of the node is calculated by combining at least the terrain risk factor, meteorological risk factor, static obstacle risk factor and real-time traffic flow of the corresponding location in the multi-dimensional environmental feature map. Using the flight start and end points as the starting and target nodes of the airspace network model, respectively, and the comprehensive travel cost as the cost weight for path search, a graph search-based path planning algorithm is used to generate the initial optimal flight path. During the flight mission, continuously acquire updated meteorological and traffic flow data, and dynamically update the airspace network model accordingly; based on the updated airspace network model, reassess the estimated total traffic cost of the current flight path; If the estimated total cost of travel is higher than the preset acceptable cost range, or if there is an alternative route with a lower total cost of travel, route replanning is triggered, and a revised flight path instruction is generated and output. 2.The method of claim 1, wherein: Generate multi-dimensional environmental feature maps, including: Acquire raw terrain data, meteorological observation data, and static obstacle vector data of the target low-altitude airspace; perform rasterization processing on the raw terrain data to generate a basic elevation grid with uniform resolution; Based on the elevation values ​​of each unit in the basic elevation grid and the degree of terrain undulation around them, the terrain risk factor of each unit is calculated to form a terrain elevation layer. The wind speed, visibility, and precipitation intensity parameters in the meteorological observation data are compared with preset threat level thresholds to obtain the quantitative risk level of each parameter. The risk levels of each parameter at discrete points are expanded into a continuous risk distribution field covering the low-altitude airspace of the target through spatial interpolation. The risk levels of each parameter at the same location are weighted and fused to form a meteorological risk layer. For each obstacle in the static obstacle vector data, its three-dimensional influence range is determined based on its type and geometric dimensions. The three-dimensional influence ranges of all obstacles are projected onto the basic elevation grid. For each grid cell, it is determined whether it is covered by one or more three-dimensional influence ranges. If it is covered, an obstacle risk factor is assigned to the cell based on the type and distance of the obstacle covering it. The obstacle risk factors of all cells constitute the static obstacle layer. The terrain elevation layer, meteorological risk layer, and static obstacle layer are aligned and superimposed in the same spatial coordinate system to generate a multi-dimensional environmental feature map. 3.The method of claim 2, wherein: Establishing a spatial network model includes: Based on the horizontal projection of the basic elevation grid, the grid is divided into layers at preset height intervals in the vertical direction. Define the three-dimensional cube space corresponding to each horizontal grid at each height level as a basic spatial unit, and mark the center point of the basic spatial unit as a network node; A node attribute table is established for each network node to store its three-dimensional spatial coordinates and associate it with the terrain risk factor, meteorological risk factor and obstacle risk factor at the corresponding location in the multi-dimensional environmental feature map. 4.The method of claim 3, wherein: Acquire real-time traffic flow information and map it to the corresponding nodes in the spatial network model in real time, including: Receive broadcast information from airspace surveillance data sources, including real-time position reports of multiple moving aerial targets; Find the basic spatial unit where a moving aerial target is located based on its real-time three-dimensional position; The number of different aerial mobile targets appearing in each basic spatial unit during the most recent update cycle is counted and used as the real-time traffic density value of the network node corresponding to that unit, which is then recorded in its node attribute table. 5.The method of claim 4, wherein: The calculated comprehensive passage cost of a node includes: For any network node to be computed in the spatial network model, a cubic spatial region with a predetermined side length is selected with its position in three-dimensional space as the center; the terrain risk factor, meteorological risk factor, obstacle risk factor, and real-time traffic density value obtained from the node attribute table are extracted from the multi-dimensional environmental feature map of all network nodes in the cubic region; these four factors are used as independent feature channels and organized into a three-dimensional multi-channel feature cube as the input of the deep feature learning model. A multi-channel feature cube is input into a pre-trained deep convolutional network structure. The network structure contains multiple cascaded convolutional computation layers and feature map downsampling layers. Each convolutional computation layer performs local scanning and weighted combination operations on the feature map output by the previous layer through a set of learnable feature detectors to extract spatial features at different levels, from simple edges to complex patterns. The feature map downsampling layer is used to expand the effective perception range of subsequent convolutional computations, thereby capturing the coupling relationship and contextual information of environmental factors and traffic flow at different spatial scales within the cube region. The final output layer of a deep convolutional network structure is a fully connected computational layer that maps the extracted and flattened high-dimensional abstract features into a single scalar value, which serves as the comprehensive passage cost value of the network node to be computed. 6.The method of claim 5, wherein: Training a deep feature learning model includes: Collect a large amount of historical flight data and corresponding airspace environment snapshot data; For each historical flight, positive and negative samples are generated based on the actual sequence of airspace network nodes flown over and the corresponding environmental snapshots. Positive samples are nodes that are flown over safely and efficiently and their corresponding feature cubes, and are marked as low overall cost target values. Negative samples are nodes that are dangerous, inefficient, or should be avoided and their corresponding feature cubes, and are marked as high overall cost target values. Using a training dataset consisting of positive and negative samples, the parameters of the deep convolutional network structure are iteratively optimized so that the network's output value for positive samples approaches the low overall cost target value, and the output value for negative samples approaches the high overall cost target value. 7.The method of claim 1, wherein: A graph search-based path planning algorithm is used to generate an initial optimal flight path, including: Initialize two node sets: a first set and a second set. The first set is used to store nodes with a determined minimum cumulative cost, and the second set is used to store nodes to be examined. Maintain two attributes for each network node: the current known minimum cumulative cost from the origin to the node, and the identifier of the previous node to reach the node; Place the starting point into the second set and set its current known minimum cumulative cost to 0. Enter the loop, and each time select the node with the smallest known cumulative cost from the second set as the current processing node and move it into the first set; Traverse all network nodes that are adjacent to the current processing node in space. For each adjacent node, calculate the path cost from the starting point to the adjacent node via the current processing node. The path cost is equal to the current known minimum cumulative cost of the current processing node plus the comprehensive travel cost of the adjacent node itself. If the calculated path cost is less than the current known minimum cumulative cost of the current record of the adjacent node, then update the current known minimum cumulative cost of the adjacent node to the path cost, and update the identifier of its predecessor node to the current processing node, while ensuring that the adjacent node is in the second set. Repeat the above loop process until the endpoint is moved into the first set; Starting from the endpoint, the path is traced backwards according to the identifier of the previous node of each node until the starting point is reached. The sequence of nodes traversed constitutes the initial optimal flight path.

8. The low-altitude flight management method based on multi-dimensional feature maps according to claim 1, characterized in that: Reassess the estimated total cost of the current flight path, including: Based on the updated airspace network model, obtain the latest comprehensive transit cost value of each network node that the current flight path passes through in sequence; The sum of the latest comprehensive transit costs is the estimated total transit cost for the current flight path.

9. The low-altitude flight management method based on multi-dimensional feature maps according to claim 8, characterized in that: Trigger path replanning, generate and output the corrected flight path instructions, including: The network node corresponding to the current real-time location is used as the starting point for the new path search; The network node corresponding to the original flight mission endpoint is used as the unchanging path search endpoint; Based on the latest comprehensive passage cost of all nodes in the updated airspace network model, the path planning process is re-executed to obtain an alternative flight path from the new starting point to the destination. Compare the estimated total cost of alternative flight paths with the estimated total cost of the remaining portion of the current flight path; If the alternative flight path is less costly and the maneuvering required for path switching is within the limits of the aircraft's performance, then the alternative flight path will be output as the corrected flight path command.

10. A low-altitude flight management system based on multi-dimensional feature maps, used to execute the method described in any one of claims 1 to 9, characterized in that: include: The data acquisition and fusion module is configured to: acquire terrain data, meteorological data, and static obstacle data of the target low-altitude airspace, and generate a multi-dimensional environmental feature map containing a terrain elevation layer, a meteorological risk layer, and a static obstacle layer through multi-source data fusion processing; The airspace dynamic modeling module is configured to: establish an airspace network model represented by a three-dimensional mesh based on the multi-dimensional environmental feature map; acquire real-time traffic flow information and map it to the corresponding region or node in the airspace network model in real time; The intelligent cost calculation module is configured to: calculate the comprehensive passage cost value of each region or node in the airspace network model by combining at least the terrain risk factor, meteorological risk factor, static obstacle risk factor and real-time traffic flow at the corresponding location in the multi-dimensional environmental feature map. The path planning module is configured to: use the flight start point and the destination point as the starting node and the target node of the airspace network model, respectively, use the comprehensive passage cost value as the cost weight of the path search, and use a graph search-based path planning algorithm to generate an initial optimal flight path; The dynamic monitoring and replanning module is configured to: continuously acquire updated meteorological and traffic flow data during flight mission execution, and dynamically update the airspace network model accordingly; and reassess the estimated total traffic cost of the current flight path based on the updated airspace network model. If the estimated total travel cost is higher than the preset acceptable cost range, or if there is an alternative path with a lower total travel cost, then path replanning is triggered, and a corrected flight path instruction is generated and output.

Images