A method for unmanned aerial vehicle flight control for meteorological sounding

Abstract

The application discloses a kind of unmanned aerial vehicle flight control methods for meteorological detection, it is related to unmanned aerial vehicle scheduling technical field.The method includes: obtaining meteorological satellite remote sensing data, cloud boundary is identified by deep learning and macroscopic parameter is reversed, and preliminary detection area is screened out;Based on satellite data, the three-dimensional prior potential diagram of each region is constructed, and each grid point in the diagram is assigned with the prior value representing catalytic potential;With prior potential diagram as guide, the customized detection flight path of unmanned aerial vehicle is planned, and local encryption sampling is triggered according to real-time data in the detection process;According to the detection data, a three-dimensional cloud field dynamic grid model is constructed, and the target area is divided into core area and potential area using dynamic adaptive threshold;Finally, accurate broadcast path is planned for the core area, adaptive coverage broadcast path is planned for the potential area, and unmanned aerial vehicle is controlled to execute operation cooperatively, and the application realizes the whole process closed loop from satellite preliminary exploration, intelligent detection to dynamic zoning broadcast.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) scheduling technology, and more specifically, to a UAV flight control method for meteorological observation. Background Technology

[0002] Traditional weather modification operations primarily rely on macroscopic observation data from ground-based radar and meteorological satellites. Operators then determine the target areas based on experience and disseminate the seeds via aircraft or ground-based rockets / anti-aircraft artillery. This method has significant drawbacks: ground-based radar is affected by the Earth's curvature and terrain obstruction, making it difficult to obtain detailed vertical structure information within clouds; satellite remote sensing has low spatiotemporal resolution and cannot reflect the microscopic physical processes within clouds in real time. Furthermore, operational decisions depend on human experience, lacking accurate identification of key locations such as areas with abundant supercooled water and updraft centers within clouds, often leading to "blind operations."

[0003] In recent years, the rapid development of drone technology has provided new technical means for weather modification. Multi-drone collaborative operations offer advantages such as maneuverability, the ability to penetrate cloud layers for detection, and precise seeding. However, current technologies have not yet formed a complete systematic solution encompassing initial satellite remote sensing, detailed drone detection, and dynamic zoned seeding. Summary of the Invention

[0004] The purpose of this invention is to provide a flight control method for unmanned aerial vehicles (UAVs) used for meteorological observation, so as to solve the above-mentioned technical problems.

[0005] To achieve the above objectives, the embodiments of this application provide the following technical solutions:

[0006] This application provides a UAV flight control method for meteorological observation. The method includes: acquiring multi-channel remote sensing data fed back by meteorological satellites and preprocessing it; then identifying cloud boundaries in the processed multi-channel remote sensing data using a deep learning semantic segmentation model to obtain multiple cloud cluster labels; performing cloud cluster parameter inversion on each cloud cluster label based on the multi-channel remote sensing data to obtain macroscopic feature parameters of each cloud cluster; then evaluating the potential index of each cloud cluster label using a weighted algorithm based on the macroscopic feature parameters of each cloud cluster, and selecting cloud cluster labels with a potential index greater than a preset threshold as preliminary detection areas; and constructing each preliminary detection area. The corresponding three-dimensional grid is used, and then the center factor, gradient factor, temperature factor and motion trend factor corresponding to each cell are calculated sequentially based on multi-channel remote sensing data. Then, the prior potential value corresponding to each cell is obtained by weighted summation, and thus the prior potential map corresponding to each preliminary detection area is obtained. The prior potential map is a three-dimensional grid, and each grid point is assigned a potential value representing the prior estimate of catalytic potential. Based on the prior potential map, the detection flight path of the UAV formation is planned, and the UAV is controlled to detect cloud clusters according to the detection flight path, thereby obtaining the corresponding three-dimensional cloud field dynamic grid model, and the corresponding catalytic dissemination path is generated based on the three-dimensional cloud field dynamic grid model.

[0007] Optionally, the multi-channel remote sensing data includes a visible light channel, an infrared channel, a water vapor channel, and a microwave radiometer channel. The preprocessing of the multi-channel remote sensing data includes radiometric calibration and geometric correction, cloud detection and mask generation, and projection transformation. The radiometric calibration and geometric correction are used to eliminate sensor errors and geographic offsets. The cloud detection and mask generation are used to distinguish between cloud pixels and clear sky pixels. The projection transformation is used to unify the satellite image to the geographic coordinate system of the mission area.

[0008] Optionally, cloud cluster parameter inversion is performed on each cloud cluster label based on the multi-channel remote sensing data to obtain the macroscopic feature parameters of each cloud cluster. Then, based on the macroscopic feature parameters of each cloud cluster, a weighted algorithm is used to evaluate the potential index of each cloud cluster label, including:

[0009] The cloud top height is derived by inverting the infrared split window channel, and then the cloud top temperature is obtained based on the cloud top height and atmospheric temperature profile.

[0010] Optical thickness is derived from visible light reflectance, and cloud phase categories are determined based on the ratio of infrared brightness temperature to visible light reflectance. The cloud phase categories include water cloud phase, ice cloud phase, and mixed phase.

[0011] Based on the cloud top height, a cloud top height score is calculated; based on the cloud top temperature, a cloud top temperature score is calculated; based on the optical thickness, an optical thickness score is calculated; and based on the cloud phase category, a mixed-phase cloud score is calculated. Then, based on the cloud top height score, cloud top temperature score, optical thickness score, and mixed-phase cloud score, a potential index corresponding to the cloud cluster is calculated using a weighted algorithm. Cloud clusters with a potential index greater than a preset threshold are designated as preliminary detection areas.

[0012] Optionally, the construction of a three-dimensional grid corresponding to each preliminary detection area, followed by sequentially calculating the center factor, gradient factor, temperature factor, and motion trend factor for each cell based on multi-channel remote sensing data, and then obtaining the prior potential value for each cell through weighted summation, includes:

[0013] A three-dimensional grid covering each preliminary detection area is constructed, with a horizontal resolution of 500-800 meters and a vertical resolution of 200-300 meters. The grid extends from the cloud base to the cloud top and then upwards by 100-200 meters.

[0014] Local maxima of the cloud top height field are identified as strong center candidates. Multiple strong center candidates are used as centers, and Gaussian diffusion is performed in the horizontal direction and in the vertical direction with the height of the strong center candidate as the center, so as to obtain the center factor corresponding to each cell.

[0015] Based on the cloud top temperature and atmospheric temperature profiles, if the temperature of the grid point corresponding to the cell is within the catalytic window, it is assigned a high value of 1, otherwise it is assigned a low value of 0, thus obtaining the temperature factor corresponding to the cell.

[0016] Based on the gradient magnitude of the cloud top height, the gradient factor corresponding to the cell is calculated;

[0017] Based on the cloud cluster motion vector and the continuous time-phase cloud cluster area change rate, the motion trend factor corresponding to the cell is calculated.

[0018] Optionally, planning the reconnaissance flight path of the UAV formation based on the prior potential map includes:

[0019] Based on the fast travel method, a skeleton detection path is planned with the goal of minimizing the path cost function. The cost function includes a priori potential attraction term and dynamic constraint term, and a local encrypted detection mode is triggered based on real-time detection data during flight.

[0020] Optionally, a corresponding three-dimensional cloud field dynamic mesh model is obtained, and a corresponding catalytic dissemination path is generated based on the three-dimensional cloud field dynamic mesh model, including:

[0021] Control the UAV to collect meteorological observation data along the detection flight path, and construct a preliminary three-dimensional cloud field dynamic grid model of the detection area based on the meteorological observation data;

[0022] Based on the three-dimensional cloud field dynamic grid model, the target operation area is divided into a core area and a potential area using a dynamic adaptive threshold division method. The core area is the continuous spatial area with the highest comprehensive quantitative value of the catalytic potential index, and the potential area is the area with catalytic potential but whose index is lower than that of the core area.

[0023] Based on the spatial distribution and physical characteristics of the core area and the potential area, catalyst dispersal paths for the core area and the potential area are planned for the UAV formation, and the UAVs are controlled to perform dispersal operations.

[0024] Optionally, based on the aforementioned three-dimensional cloud field dynamic mesh model, a dynamic adaptive threshold partitioning method is used to divide the target operation area into a core area and a potential area, including:

[0025] The system calculates the statistical distribution characteristics of meteorological parameters within the target area in real time and adaptively adjusts the threshold coefficient according to the cloud development stage. It also generates the core area threshold by weighted fusion of global and local statistics, extracts the core area and potential area through three-dimensional connected domain analysis, and predicts the evolution of the zoning boundary based on the three-dimensional wind field.

[0026] Optionally, catalyst dissemination paths are planned for the core area and the potential area of ​​the drone formation, respectively, including:

[0027] The core area seeding path adopts a spiral seeding, vertical shuttle or multi-core collaborative path, and the seeding corridor is corrected based on the three-dimensional wind field;

[0028] The seeding path in the potential area employs either an adaptive density-adjusted grid scan or a spiral expansion path.

[0029] The beneficial effects of this invention are as follows:

[0030] This invention generates a priori potential maps using satellite remote sensing data, guiding UAVs to prioritize the exploration of key areas within cloud clusters (such as strong centers and supercooled water layers), avoiding wasted flight time in areas with no potential. Customized exploration paths combined with a dynamic encryption triggering mechanism ensure high-density sampling at key locations such as strong centers and supercooled water layers. The resulting 3D cloud field dynamic mesh model, reconstructed through interpolation, accurately reflects the complex structure within the cloud cluster, providing a high-quality data foundation for subsequent decision-making.

[0031] Secondly, a dynamic adaptive threshold method is used to delineate the core area and potential area. The threshold is automatically adjusted based on real-time cloud field statistical characteristics and cloud development stage, completely eliminating the limitations of fixed thresholds. By introducing a weighted fusion of local and global statistics, multiple core areas and potential areas within the cloud cluster can be accurately identified. Even if the cloud splits, merges, or moves, the partitioning results can be updated in real time to adapt to the dynamic evolution of the cloud cluster.

[0032] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic flowchart of a UAV flight control method for meteorological detection as described in an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0036] Example 1:

[0037] like Figure 1 As shown, this embodiment provides a flight control method for a UAV used for weather detection, the method including steps S100, S200, S300 and S400.

[0038] Step S100: Obtain multi-channel remote sensing data fed back by meteorological satellites, preprocess it, and then use a deep learning semantic segmentation model to identify cloud boundaries in the processed multi-channel remote sensing data to obtain multiple cloud labels.

[0039] The multi-channel remote sensing data includes a visible light channel (solar radiation reflected by cloud droplets, reflectivity is related to the effective radius and optical thickness of cloud droplets, used to detect cloud top reflectivity and cloud optical thickness), an infrared window channel (thermal radiation emitted from cloud tops, temperature is inverted according to Planck's law, used to detect cloud top brightness and cloud top temperature), an infrared split window channel (difference in ice / water absorption between two channels, the difference reflects cloud phase, used to detect cloud phase and cloud top height correction), a water vapor channel (detecting water vapor absorption in the middle atmosphere, indirectly reflecting the vertical development of cloud bodies, and thus obtaining the approximate middle-layer water vapor content and cloud body vertical structure), and a microwave radiometer channel (microwaves can penetrate non-precipitating clouds, are absorbed and scattered by liquid water, used to detect liquid water paths and liquid water content in clouds).

[0040] The preprocessing of the multi-channel remote sensing data includes radiometric calibration and geometric correction, cloud detection and mask generation, and projection transformation. The radiometric calibration and geometric correction are used to eliminate sensor errors and geographic offsets. The cloud detection and mask generation are used to distinguish between cloud pixels and clear sky pixels. The projection transformation is used to unify the satellite image to the geographic coordinate system of the mission area.

[0041] The specific implementation of identifying cloud boundaries in processed multi-channel remote sensing data using a deep learning semantic segmentation model is as follows: A deep learning semantic segmentation model is used to fuse visible light, infrared, and water vapor three-channel data to extract the horizontal projection contour of each independent cloud.

[0042] Step S200: Perform cloud cluster parameter inversion on each cloud cluster label based on the multi-channel remote sensing data to obtain the macroscopic feature parameters of each cloud cluster. Then, evaluate the potential index of each cloud cluster label based on the macroscopic feature parameters of each cloud cluster using a weighted algorithm, and select cloud cluster labels with potential index greater than a preset threshold as the initial detection area.

[0043] Its specific implementation method is as follows:

[0044] Step S210: Invert cloud top height based on infrared split window channel; Invert cloud top temperature based on infrared window channel;

[0045] Step S220: Based on the reflectivity of the visible light channel, the optical thickness is inverted, and then the cloud phase category is determined based on the ratio of infrared split window temperature difference and visible light reflectivity. The cloud phase categories include water cloud phase, ice cloud phase, and mixed phase. The cloud motion vector is calculated based on the optical flow method of continuous temporal infrared images.

[0046] Step S230: Calculate the cloud top height score based on the cloud top height, calculate the cloud top temperature score based on the cloud top temperature, calculate the optical thickness score based on the optical thickness, and calculate the mixed phase cloud score based on the cloud phase category. Then, calculate the potential index corresponding to the cloud cluster using a weighted algorithm based on the cloud top height score, cloud top temperature score, optical thickness score, and mixed phase cloud score, and take the cloud cluster with the potential index greater than a preset threshold as the preliminary detection area.

[0047] Step S300: Construct a three-dimensional grid corresponding to each preliminary detection area, and then calculate the center factor, gradient factor, temperature factor and motion trend factor corresponding to each cell based on multi-channel remote sensing data. Then, calculate the prior potential value corresponding to each cell by weighted summation, and then obtain the prior potential map corresponding to each preliminary detection area. The prior potential map is a three-dimensional grid, and each grid point is assigned a potential value that represents the prior estimate of catalytic potential.

[0048] Step S400: Based on the prior potential map, plan the detection flight path of the UAV formation, and control the UAV to detect cloud clusters according to the detection flight path, thereby obtaining the corresponding three-dimensional cloud field dynamic mesh model, and generating the corresponding catalytic dissemination path based on the three-dimensional cloud field dynamic mesh model.

[0049] Secondly, as described in step S300, a three-dimensional grid is constructed corresponding to each preliminary detection area. Then, based on multi-channel remote sensing data, the center factor, gradient factor, temperature factor, and motion trend factor corresponding to each cell are calculated sequentially. Finally, the prior potential value corresponding to each cell is obtained by weighted summation, including:

[0050] Step S310: Construct a three-dimensional grid covering the preliminary detection area, with the spatial range of each preliminary detection area as the boundary. The grid has a horizontal resolution of 500-800 meters and a vertical resolution of 200-300 meters. The grid range extends from the cloud base to the cloud top and then upwards by 100-200 meters.

[0051] Step S320: Identify local maxima in the cloud top height field as strong center candidates. Using these points as centers, perform Gaussian diffusion horizontally and vertically around these heights to obtain the center factor for each cell (numerator: spatial distance from the grid point to the nearest strong center candidate, after Gaussian function transformation; denominator: preset maximum influence radius). In cloud physics, the strong center of a cloud often corresponds to the strongest updraft. Updrafts are the "power engine" that sustains cloud development and continuously transports lower-level water vapor to higher levels. By observing the cloud top bulge from satellites, the position of the downdraft column can be deduced, and the center factor tells the drone where the "engine" is.

[0052] Step S330: Based on cloud top temperature and atmospheric temperature profiles (weather forecast data), if the temperature of the grid point corresponding to the cell is within the catalytic window (-15℃ to 0℃), a high value of 1 is assigned; otherwise, a low value of 0 is assigned. This yields the temperature factor corresponding to the cell. Temperature directly determines the physical mechanism and efficiency of catalysis. -15℃ to 0℃ is the optimal temperature range for cold cloud catalysts such as silver iodide (where ice crystal nucleation efficiency is highest). Above 0℃, it belongs to warm cloud catalysis, requiring hygroscopic flame agents, and the mechanism is completely different. A "suitable for operation" temperature constraint is added to the potential assessment to ensure that the detected area has at least the possibility of operation, avoiding invalid detection.

[0053] Step S340: Based on the cloud top height gradient magnitude (calculated from the cloud top height field), calculate the gradient factor corresponding to each cell (numerator: cloud top height gradient magnitude corresponding to the projection of this grid point, denominator: preset maximum reference gradient magnitude). Regions with large gradients (steep edges of the cloud cluster) are projected onto lower height layers, and the potential value decreases with depth. Regions with large gradients (i.e., the steepest changes in cloud top height) are usually areas where the main body of the cloud cluster is violently entangling with the environment. Here, dry and cold air meets warm and humid clouds, and the gradient factor tells the drone where the "front line" of the cloud cluster is, serving as the basis for predicting cloud cluster dynamics and judging newly formed areas.

[0054] Step S350: Based on the cloud cluster motion vector (optical flow method result) and the continuous temporal cloud cluster area change rate, calculate the motion trend factor corresponding to each cell (numerator: the probability that the grid point will still be within the potential zone in the future after advection extrapolation, determined by both the cloud cluster motion vector and the area change rate; denominator: the preset maximum retention probability). Clouds are in motion. Satellite images represent the state of the past few minutes; it takes time for the drone to reach the target location. The motion trend factor transforms the static map into a dynamic prediction map, ensuring that the drone flies towards the future potential zone, not the past potential zone.

[0055] Choosing these four factors for weighting is essentially a multi-view remote sensing inversion of the invisible physical processes within a cloud. It utilizes limited information observable by satellites (shape, temperature, brightness) to inversely deduce the most probable distribution of internal potential through the inherent correlations in cloud physics. This design ensures that, in the preliminary stages, limited detection resources (drones) can be guided to areas most likely to have operational value, avoiding blind spatial searches.

[0056] The prior potential map is essentially a three-dimensional weighted grid, with each grid cell storing a value. The value represents the probability that "this location may be a key structure of the cloud". 0 represents an almost irrelevant area (such as the outside of the cloud), and 1 represents the most likely key area (such as a strong center). The potential value is calculated by weighting indirect features observable by the satellite, such as the center factor, gradient factor, and temperature factor.

[0057] This embodiment generates a priori potential map using satellite remote sensing data, guiding UAVs to prioritize the detection of key areas within cloud clusters (such as strong centers and supercooled water layers), avoiding wasting flight time in areas with no potential. Customized detection paths combined with a dynamic encryption triggering mechanism ensure high-density sampling at key locations such as strong centers and supercooled water layers. The resulting 3D cloud field dynamic mesh model, reconstructed through interpolation, accurately reflects the complex structure within the cloud cluster, providing a high-quality data foundation for subsequent decision-making.

[0058] Secondly, a dynamic adaptive threshold method is used to delineate the core area and potential area. The threshold is automatically adjusted based on real-time cloud field statistical characteristics and cloud development stage, completely eliminating the limitations of fixed thresholds. By introducing a weighted fusion of local and global statistics, multiple core areas and potential areas within the cloud cluster can be accurately identified. Even if the cloud splits, merges, or moves, the partitioning results can be updated in real time to adapt to the dynamic evolution of the cloud cluster.

[0059] Example 2:

[0060] This embodiment is based on Embodiment 1 and is used to further illustrate the detailed implementation principle of step S400:

[0061] The specific implementation method for planning the detection flight path of the UAV formation based on the aforementioned prior potential map is as follows:

[0062] Step S410: Based on the fast travel method, a skeleton detection path is planned with the goal of minimizing the path cost function. The cost function includes a priori potential attraction term and dynamic constraint term, and a local encrypted detection mode is triggered based on real-time detection data during flight.

[0063] Specifically, the detection skeleton path planning:

[0064] Extract the morphological centerline (central axis transformation) of the cloud cluster's horizontal projection as a horizontal skeleton. Plan vertical profiles at key nodes of the horizontal skeleton (such as endpoints and branch points), including vertical spirals and zigzag shuttles. Plan a tracking path along the cloud cluster boundary while performing small vertical oscillations to obtain the boundary's vertical structure.

[0065] The path cost function guided by prior potential can be:

[0066] ;in, The algorithm aims to find a path from the starting point to the ending point that accumulates the total cost along the entire path, where the total cost is the sum of the integrals. The shortest path.

[0067] In the formula As a priori potential guide, it encourages paths to pass through high-potential areas. These are dynamic constraints that ensure path safety and feasibility.

[0068] As an integral variable, it represents the infinitesimal element of the path, i.e., a very small segment of the path. It expresses the path cost as an accumulation along the path, meaning that the longer the path, the higher the cost. This naturally penalizes excessively long paths and encourages paths to be as short as possible while still satisfying the objective.

[0069] For position The prior potential value at a location is obtained through the prior potential map. The larger the value, the more likely that the location is to be a key structure of the cloud (strong center, supercooled water layer, etc.), and is worth focusing on for detection. This is the potential attraction coefficient, used to adjust the weighting of prior potential on the path attraction strength.

[0070] As a dynamic constraint term, it is used to characterize the dynamic constraint cost that needs to be considered at a certain point on the path. It is a dimensionless comprehensive index, and its components can be: kinematic feasibility cost (the cost increases at positions that do not meet the maneuverability performance of the UAV, such as excessive path curvature or steep climb angle) and energy cost (factors that increase energy consumption, such as flying against the wind or high-altitude low-oxygen environment, can be converted into cost). It changes in real time, dynamically adjusting as environmental perception and drone status are updated. The dynamic constraint weighting coefficient is used to balance the importance of prior potential guidance and dynamic constraints, adjusting the stringency of safety / feasibility constraints relative to the target being detected.

[0071] use The integral form reflects the path length penalty mechanism on the one hand – even if the path passes through low-cost areas, the total cost will still increase if the path is too long. This naturally balances "probing key areas" and "saving energy / time". On the other hand, the path needs to be continuous and smooth, because the integral is accumulated along the continuous path and the transition point cannot be covered by the integral. The algorithm considers the cumulative cost of the entire path, rather than being greedy point by point. Therefore, the globally optimal solution of "going a little detour to explore a high-potential area and then returning" may occur.

[0072] Within completely unknown cloud formations, drones can autonomously plan a safe and efficient detection path based on limited prior clues, maximizing the acquisition of real data from key locations and laying the foundation for subsequent 3D modeling and seeding decisions.

[0073] Secondly, the specific implementation method for triggering the local encrypted detection mode based on real-time detection data during flight, as described in step S400, can be as follows:

[0074] As the drone flies along the skeleton path, it processes onboard sensor data in real time. A local encrypted detection mode is automatically triggered when the following conditions are met:

[0075] Step S411: When the measured liquid water content exceeds the preset liquid water content threshold and the gradient change of liquid water content is large, the local encrypted detection mode is triggered. At this time, the UAV automatically deviates from the original skeleton path to perform encrypted detection. After completion, it returns to the skeleton path. The encrypted detection method is to perform spiral conical surface flight detection according to the preset flight path.

[0076] All UAV detection data are collected, and spatiotemporal alignment (wind field advection correction), outlier removal, and confidence level assignment are performed. Then, based on the detection data density and cloud morphology, an adaptive resolution grid is generated, for example: a dense detection zone of 50m×50m×25m; a transition zone of 100m×100m×50m; and an outer zone of 200m×200m×100m. The detection data includes cloud droplet spectrum distribution (detected by a cloud particle probe CDP, using the number concentration within the 2-50μm particle size range to determine the cloud droplet spectrum width and phase), liquid water content (detected by a hot-wire liquid water content meter or a combined probe), ice crystal concentration (detected by a cloud particle imager CIP or a polarization probe to identify the ice phase process and determine whether supercooled water has frozen naturally), and three-dimensional wind speed (measured by a three-dimensional ultrasonic anemometer).

[0077] An anisotropic Kriging interpolation algorithm is used to reconstruct the meteorological parameter fields by considering the correlation length differences in the horizontal, vertical and wind directions, thereby generating a three-dimensional cloud field dynamic grid model. The three-dimensional cloud field dynamic grid model discretizes the entire mission airspace into a regular three-dimensional grid. Each grid point stores a set of parameters describing the cloud physical state at that location and its statistical characteristics. Specifically, each grid point stores a set of parameters, including location coordinates and estimated meteorological parameters (including ice crystal concentration, liquid water content and cloud droplet effective radius, etc.).

[0078] Example 3:

[0079] This embodiment is based on Embodiment 2 and is used to further illustrate the specific implementation of generating the corresponding catalytic dissemination path based on the three-dimensional cloud field dynamic mesh model in step S400:

[0080] Step S420: Based on the three-dimensional cloud field dynamic grid model, the target operation area is divided into a core area and a potential area using a dynamic adaptive threshold division method. The core area is the continuous spatial area with the highest comprehensive quantitative value of the catalytic potential index, and the potential area is the area with catalytic potential but whose index is lower than that of the core area.

[0081] Step S430: Based on the spatial distribution and physical characteristics of the core area and potential area, plan the catalyst dispersal paths for the core area and potential area for the UAV formation, and control the UAVs to perform the dispersal operation. The path planning, after clarifying the boundaries of the core area and potential area, assigns the access order and path (i.e., catalyst dispersal path) of each UAV using a conventional multi-traveling salesman algorithm. Specifically, the core area dispersal path uses a hovering dispersal, vertical shuttle, or multi-core collaborative path; the potential area dispersal path uses an adaptive density-adjusted grid scanning or spiral expansion path. It should be noted that the specific planning method for the catalyst dispersal path is a built-in preset flight mode, which is a conventional existing technology. That is, how to generate a flight path traversing a fixed spatial range is a conventional technical means in the UAV field, and this embodiment will not elaborate further, only providing some conventional path planning algorithms, such as coverage path planning (applicable to the potential area) and dynamic target tracking algorithms (applicable to the core area).

[0082] The core region is the critical area within a cloud cluster, characterized by the highest liquid water content, strongest updrafts, and greatest catalytic potential. It is typically small in size but has high value per unit volume. The seeding path planning for the core region follows the principle of "precise targeting and full coverage," employing existing mature UAV trajectory planning technologies. For example, dynamic target tracking algorithms enable the UAV to track the movement trajectory of the core region's centroid in real time. Once the UAV enters the core region, it needs to fully seed the catalyst within this limited space, using methods such as hovering seeding patterns, figure-eight patterns, or elliptical trajectories.

[0083] The potential zone is a transitional area surrounding the core zone or a large area within a cloud cluster that possesses certain catalytic potential but with slightly lower indicators; it is typically large in volume and wide in area. Seeding path planning for the potential zone follows the principle of "wide coverage and efficiency first," employing existing mature coverage path planning technologies, such as coverage path planning commonly used in agricultural drone operations and ground robot cleaning, employing bow-shaped reciprocating scanning or spiral expanding scanning.

[0084] Specifically, based on the aforementioned three-dimensional cloud field dynamic mesh model, the specific implementation method for dividing the target operation area into a core area and a potential area using a dynamic adaptive threshold partitioning approach is as follows:

[0085] Step S421: Calculate the statistical distribution characteristics of meteorological parameters within the target area in real time, and adaptively adjust the threshold coefficient according to the cloud development stage. At the same time, generate the core area threshold by weighted fusion of global statistics (calculating the mean and standard deviation of parameters such as liquid water content and effective cloud droplet radius of each grid in the three-dimensional cloud field dynamic grid model) and local statistics (constructing a three-dimensional sliding window (e.g., with a radius of 500 meters) centered on each grid point in the three-dimensional cloud field dynamic grid model, and calculating the local mean and local standard deviation within the window). Then, extract the core area and potential area through three-dimensional connected domain analysis, and predict the evolution of the partition boundary based on the three-dimensional wind field.

[0086] Mesh that meet the preset core area threshold conditions are marked as core candidate points, and meshes that meet the preset potential area threshold conditions but do not meet the core criteria are marked as potential candidate points.

[0087] Then, using the 3D 26-neighborhood connectivity algorithm, adjacent core candidate points are aggregated into one or more core area connectivity domains; adjacent potential candidate points are aggregated into potential area connectivity domains, thus obtaining the core area and potential area. Combining 3D wind field data and cloud physics model, the optical flow method is used to predict the centroid movement trajectory of each core area / potential area within a future time window (30-120 seconds).

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A flight control method for an unmanned aerial vehicle (UAV) used for meteorological observation, characterized in that, The method includes: Multi-channel remote sensing data fed back by meteorological satellites is acquired and preprocessed. Then, a deep learning semantic segmentation model is used to identify cloud boundaries in the processed multi-channel remote sensing data, thereby obtaining multiple cloud labels. Based on the multi-channel remote sensing data, cloud cluster parameters are inverted for each cloud cluster label to obtain the macroscopic feature parameters of each cloud cluster. Then, based on the macroscopic feature parameters of each cloud cluster, the potential index of each cloud cluster label is evaluated by a weighted algorithm, and cloud cluster labels with potential indices greater than a preset threshold are selected as the initial detection area. A three-dimensional grid is constructed for each preliminary exploration area. Then, based on multi-channel remote sensing data, the center factor, gradient factor, temperature factor, and motion trend factor corresponding to each cell are calculated sequentially. The prior potential value corresponding to each cell is then obtained by weighted summation, thereby obtaining the prior potential map corresponding to each preliminary exploration area. The prior potential map is a three-dimensional grid, and each grid point is assigned a potential value that characterizes the prior estimate of catalytic potential. Based on the prior potential map, the detection flight path of the UAV formation is planned, and the UAV is controlled to detect cloud clusters according to the detection flight path, thereby obtaining the corresponding three-dimensional cloud field dynamic mesh model, and generating the corresponding catalytic dissemination path based on the three-dimensional cloud field dynamic mesh model. This includes obtaining a corresponding three-dimensional cloud field dynamic mesh model, and generating a corresponding catalytic dissemination path based on the three-dimensional cloud field dynamic mesh model, including: Control the UAV to collect meteorological observation data along the detection flight path, and construct a preliminary three-dimensional cloud field dynamic grid model of the detection area based on the meteorological observation data; Based on the three-dimensional cloud field dynamic grid model, the target operation area is divided into a core area and a potential area using a dynamic adaptive threshold division method. The core area is the continuous spatial area with the highest comprehensive quantitative value of the catalytic potential index, and the potential area is the area with catalytic potential but whose index is lower than that of the core area. Based on the spatial distribution and physical characteristics of the core area and the potential area, catalyst dispersal paths for the core area and the potential area are planned for the UAV formation, and the UAVs are controlled to perform dispersal operations.

2. The UAV flight control method for meteorological observation according to claim 1, characterized in that, The multi-channel remote sensing data includes a visible light channel, an infrared channel, a water vapor channel, and a microwave radiometer channel. The preprocessing of the multi-channel remote sensing data includes radiometric calibration and geometric correction, cloud detection and mask generation, and projection transformation. The radiometric calibration and geometric correction are used to eliminate sensor errors and geographic offsets. The cloud detection and mask generation are used to distinguish between cloud pixels and clear sky pixels. The projection transformation is used to unify the satellite image to the geographic coordinate system of the mission area.

3. The UAV flight control method for meteorological observation according to claim 1, characterized in that, Based on the multi-channel remote sensing data, cloud cluster parameters are inverted for each cloud cluster label to obtain the macroscopic feature parameters of each cloud cluster. Then, based on the macroscopic feature parameters of each cloud cluster, a weighted algorithm is used to evaluate the potential index of each cloud cluster label, including: The cloud top height is derived by inverting the infrared split window channel, and then the cloud top temperature is obtained based on the cloud top height and atmospheric temperature profile. Optical thickness is derived from visible light reflectance, and cloud phase categories are determined based on the ratio of infrared brightness temperature to visible light reflectance. The cloud phase categories include water cloud phase, ice cloud phase, and mixed phase. Based on the cloud top height, a cloud top height score is calculated; based on the cloud top temperature, a cloud top temperature score is calculated; based on the optical thickness, an optical thickness score is calculated; and based on the cloud phase category, a mixed-phase cloud score is calculated. Then, based on the cloud top height score, cloud top temperature score, optical thickness score, and mixed-phase cloud score, a potential index corresponding to the cloud cluster is calculated using a weighted algorithm. Cloud clusters with a potential index greater than a preset threshold are designated as preliminary detection areas.

4. The UAV flight control method for meteorological observation according to claim 1, characterized in that, The process involves constructing a 3D grid corresponding to each initial detection area, then sequentially calculating the center factor, gradient factor, temperature factor, and motion trend factor for each cell based on multi-channel remote sensing data. Finally, a weighted summation is used to calculate the prior potential value for each cell, including: A three-dimensional grid covering each preliminary detection area is constructed, with a horizontal resolution of 500-800 meters and a vertical resolution of 200-300 meters. The grid extends from the cloud base to the cloud top and then upwards by 100-200 meters. Local maxima of the cloud top height field are identified as strong center candidates. Multiple strong center candidates are used as centers, and Gaussian diffusion is performed in the horizontal direction and in the vertical direction with the height of the strong center candidate as the center, so as to obtain the center factor corresponding to each cell. Based on the cloud top temperature and atmospheric temperature profiles, if the temperature of the grid point corresponding to the cell is within the catalytic window, it is assigned a high value of 1, otherwise it is assigned a low value of 0, thus obtaining the temperature factor corresponding to the cell. Based on the gradient magnitude of the cloud top height, the gradient factor corresponding to the cell is calculated; Based on the cloud cluster motion vector and the continuous time-phase cloud cluster area change rate, the motion trend factor corresponding to the cell is calculated.

5. The UAV flight control method for meteorological observation according to claim 1, characterized in that, Based on the aforementioned prior potential map, the detection flight path of the UAV formation is planned, including: Based on the fast travel method, a skeleton detection path is planned with the goal of minimizing the path cost function. The cost function includes a priori potential attraction term and dynamic constraint term, and a local encrypted detection mode is triggered based on real-time detection data during flight.

6. The UAV flight control method for meteorological observation according to claim 1, characterized in that, Based on the aforementioned three-dimensional cloud field dynamic mesh model, a dynamic adaptive threshold partitioning method is used to divide the target operation area into a core area and a potential area, including: The system calculates the statistical distribution characteristics of meteorological parameters within the target area in real time and adaptively adjusts the threshold coefficient according to the cloud development stage. It also generates the core area threshold by weighted fusion of global and local statistics, extracts the core area and potential area through three-dimensional connected domain analysis, and predicts the evolution of the zoning boundary based on the three-dimensional wind field.

7. The UAV flight control method for meteorological observation according to claim 1, characterized in that, The catalyst dissemination paths for the core area and the potential area of ​​the drone formation are planned respectively, including: The core area seeding path adopts a spiral seeding, vertical shuttle or multi-core collaborative path, and the seeding corridor is corrected based on the three-dimensional wind field; The seeding path in the potential area employs either an adaptive density-adjusted grid scan or a spiral expansion path.

Images