Data acquisition system for remotely acquiring environmental data based on unmanned aerial vehicle

Abstract

The invention discloses a data acquisition system for remotely acquiring environmental data based on an unmanned aerial vehicle, and the system comprises the following modules: a flight mission modeling module which is used for establishing a flight state diagram; the path optimization module is used for outputting a Pareto track candidate set; the feature selection module is used for initializing a second ant population and generating a Pareto optimal feature subset candidate set; the feature extraction module is deployed on the unmanned aerial vehicle image processing unit, and is used for performing multi-scale feature rapid extraction on the acquired multi-modal vegetation observation data stream, and performing compression according to a Pareto optimal feature subset candidate set to generate a compressed feature data stream; and the classification and inversion module is used for inputting the compressed feature data stream into a kernel Fisher discriminant analysis model, performing online reasoning according to the optimal kernel function type and parameters, and generating a vegetation type classification layer. According to the invention, an unmanned aerial vehicle ecological perception closed loop of flight-calculation-analysis is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data acquisition, and in particular to a data acquisition system based on an unmanned aerial vehicle (UAV) for remotely acquiring environmental data. Background Art

[0002] With the development of remote sensing technology and intelligent unmanned systems, drones have been widely used in highly complex natural environments for forest resource monitoring, ecological environmental assessment, and agricultural information collection. In subtropical to tropical hills, valleys, and monsoon evergreen forests with complex vegetation, drones can carry hyperspectral, visible light, and lidar multimodal sensors at low altitudes to achieve high-resolution observation of vegetation type, structure, and growth status. However, in practical applications, the extraction of environmental information data in highly complex vegetation areas still faces many challenges. Existing technologies have obvious deficiencies in data dimension processing, path planning, and discriminant analysis, which restricts the practicality and intelligence level of drone remote sensing systems.

[0003] First, in terms of multimodal data processing, highly complex vegetation areas often experience "different objects with the same spectrum" and "same objects with different spectra." The spectral curves of vegetation communities highly overlap, making it difficult to distinguish responses in different bands. This makes it difficult for traditional linear dimensionality reduction methods such as principal component analysis (PCA) or linear discriminant analysis (LDA) to effectively separate vegetation classes in high-dimensional feature spaces. Furthermore, the dimensions of hyperspectral and lidar data often exceed 1,000, creating a "curse of dimensionality" problem that increases model computational complexity and reduces the discriminative power of subsequent classification.

[0004] Secondly, in terms of drone path planning, flight missions in complex terrain must balance terrain undulations, obstacle distribution, and flight energy consumption. Traditional methods based on single-target shortest paths or static grid planning cannot dynamically balance mission area coverage efficiency and flight safety. This is especially true in scenarios with limited flight time or frequent wind disturbances, which can easily lead to path blind spots or excessive energy consumption. Furthermore, most path planning algorithms are not deeply coupled with the value of data collection, often resulting in data redundancy ("more photos, less effective results"), which affects overall mission resource utilization.

[0005] Thirdly, in terms of feature selection and classification analysis, existing feature selection algorithms are mostly based on single-objective optimization, failing to take into account the discriminative ability, redundancy rate and computational cost of features at the same time. The lack of a lightweight feature compression mechanism for edge computing platforms makes it difficult to meet the flight end's requirements for model response speed and real-time inference performance. In addition, traditional classification models such as support vector machines or decision trees have low classification accuracy when dealing with highly nonlinear and multi-coupled vegetation structure features, and lack the ability to effectively model complex boundaries between high-dimensional features.

[0006] In summary, there is an urgent need for a multi-objective collaborative method that can realize path-feature-discriminant linkage optimization in complex vegetation environments to improve the intelligence and practicality of UAV information collection. Summary of the Invention

[0007] One purpose of the present invention is to propose a data acquisition system based on a drone to collect environmental data from a long distance. The present invention realizes a drone ecological perception closed loop of flying, computing and analyzing.

[0008] According to an embodiment of the present invention, a data acquisition system for remotely collecting environmental data using a drone includes the following modules:

[0009] The flight mission modeling module is used to construct a three-dimensional flight feasible domain model of the target vegetation area and establish a flight status diagram;

[0010] The path optimization module is used to initialize the first ant population in the flight state graph, perform multi-target path search based on the improved three-dimensional environment perception probability model, and output a set of Pareto trajectory candidates;

[0011] The feature selection module is used to initialize the second ant population, build a feature selection probability model based on the multimodal vegetation observation data sample set, perform feature subset search based on three objectives: discriminant divergence, feature redundancy rate, and computational cost, and generate a Pareto optimal feature subset candidate set;

[0012] The feature extraction module is deployed on the UAV image processing unit and is used to perform multi-scale feature fast extraction on the collected multi-modal vegetation observation data stream and compress it according to the Pareto optimal feature subset candidate set to generate a compressed feature data stream;

[0013] The classification and inversion module is used to input the compressed feature data stream into the kernel Fisher discriminant analysis model, perform online inference based on the optimal kernel function type and parameters, and generate a vegetation type classification layer.

[0014] A data acquisition method for remotely collecting environmental data based on an unmanned aerial vehicle (UAV) is applied to a data acquisition system for remotely collecting environmental data based on an unmanned aerial vehicle (UAV), comprising the following steps:

[0015] S1. Obtain a digital elevation model and obstacle distribution model of the target vegetation area, and generate a 3D flight feasible range model based on the solar altitude angle and terrain type.

[0016] S2. Initialize a first ant population within the three-dimensional flight feasible domain model and use a multi-objective ant colony optimization algorithm to search for flight energy consumption targets, mission area coverage targets, and track safety targets to generate a set of candidate tracks.

[0017] S3. Initialize a second ant population based on a set of multimodal vegetation observation data samples obtained from ground and historical sampling. Use a multi-objective ant colony optimization algorithm to search for discriminant divergence, feature redundancy, and computational cost objectives to generate a candidate set of feature subsets.

[0018] S4. Based on the remaining battery level, real-time wind field information, and lighting conditions, the drone selects the current optimal track from the Pareto track candidate set. The drone then flies along this optimal track, continuously collecting hyperspectral image data, visible light image data, lidar point cloud data, and inertial navigation data during flight to form a multimodal vegetation observation data stream.

[0019] S5. Perform multi-scale feature fast extraction on the multimodal vegetation observation data stream on the image processing unit of the UAV, and retain only the corresponding features in the Pareto feature subset candidate set to obtain a compressed feature data stream;

[0020] S6. Input the compressed feature data stream into the kernel Fisher discriminant analysis model. The kernel Fisher discriminant analysis model performs online inference according to the kernel function type and kernel parameters determined by the Pareto feature subset candidate set to generate vegetation type classification results and ecological indicator inversion results.

[0021] Optionally, the S1 includes the following steps:

[0022] S11. The three-dimensional terrain information of the target vegetation area is modeled using remote sensing images, and a digital elevation model (DEM) is constructed to describe the elevation values corresponding to each two-dimensional geographic location within the vegetation area. veg , where the elevation value corresponding to each spatial point position (x, y) is h(x, y), and the spatial point position belongs to the spatial definition domain Ω of the target vegetation area veg ;

[0023] S12. Obtain an obstacle distribution model OBS for representing the positions and heights of all obstacles in the vegetation area veg , each obstacle is identified by its two-dimensional position and vertical height For three-dimensional expression, there are N obs obstacle sample points constitute a complete obstacle set;

[0024] S13. Construct the illumination constraint factor set IL according to the solar altitude angle information of the flight mission period veg , which is used to characterize the imaging illumination conditions of the target vegetation area at different times and different geographical locations. The illumination constraint factor of each spatial point position (x, y) at time t is determined by the solar incident angle θ at that point. s(x, y, t) is jointly determined with the ground object albedo weight factor μ(x, y, t) to reflect the comprehensive weight of the current point imaging illumination quality;

[0025] S14. Based on the digital elevation model DEM veg , obstacle distribution model OBS veg and the lighting constraint factor set IL veg , combined with the mission area terrain type to build a three-dimensional flight feasible domain model FE 3D In the three-dimensional flight feasible domain model, the flight altitude range corresponding to each two-dimensional spatial point position (x, y) in the vegetation area is the ground elevation value h(x, y) at that position plus the minimum flight safety distance δ min To the maximum allowed flight height z affected by lighting constraints max The continuous height interval between (x, y), the three-dimensional flight feasible area space definition domain should exclude the obstacle area Ω defined by the obstacle projection obs ;

[0026] S15. 3D flight feasible domain model FE 3D Perform spatial discretization and divide it into multiple cubic three-dimensional voxel units V i , each 3D voxel unit is defined as a cubic area with fixed spatial resolution Δx, Δy, Δz in the x, y, and z directions. The collection of all 3D voxel units together constitutes the entire 3D flight feasible domain model;

[0027]

[0028] Among them, Δx, Δy, Δz are the spatial resolutions of voxel division, N vox is the total number of voxels in the 3D flight feasible domain model, V i is the i-th three-dimensional voxel unit.

[0029] Optionally, S2 includes the following steps:

[0030] S21. In the three-dimensional flight feasible domain model FE 3D The flight state graph G with dynamic constraints is constructed based on the voxel unit division results. flight ,The nodes of the flight state graph are the voxel unit centers, and the connectivity between adjacent voxel units is dynamically adjusted based on the vegetation height difference, wind field disturbance coefficient, and energy consumption threshold;

[0031] S22. In flight state diagram G flight Initialize the first ant population A route , the first ant population contains N a ant individuals, each ant individual corresponds to an initial drone flight path;

[0032] S23. During each iteration, each ant in the first ant population selects the next pixel unit for extending the drone's flight path based on the improved 3D environment perception probability model. The improved 3D environment perception probability model is defined as:

[0033]

[0034] Among them, τ(V i ,V j ) represents node V i To node V j The path pheromone concentration, τ(V i ,V′ j ) represents node V i To candidate node V′ j The path pheromone concentration, η env (V i ,V j ) represents node V i To node V j The environmental adaptation heuristic factor is combined with the vegetation height difference, terrain complexity and wind field disturbance coefficient between the two nodes. For the task of extracting environmental information in the vegetation area, the flight path of the UAV with small wind disturbance and stable flight is preferred, η env (V i ,V′ j ) represents node V i To candidate node V′ j Environmental adaptation heuristic factor, η cov (V j ) represents node V j The area coverage heuristic factor is the higher the uncovered ratio of the voxel unit where the node is located, the larger its value is, which encourages the UAV to give priority to areas where vegetation information is not collected, η cov (V′ j ) represents the candidate node V′ j The area coverage heuristic factor, N(V i ) represents node V i The set of adjacent nodes that meet the flight constraints, α, β, and γ are the importance weights of pheromone, environmental adaptation heuristic factor, and area coverage heuristic factor, respectively;

[0035] S24. Calculate the multi-objective fitness index of each UAV flight path, including the flight energy consumption index E k , Mission area coverage index C k and track safety index S k ;

[0036] S25. After each round of iteration, the multi-objective fitness index (Ek ,C k ,S k ) Perform non-dominated sorting and congestion evaluation to select the Pareto optimal path set

[0037] S26. Based on the actual performance of each node in the Pareto optimal path set, reversely update the heuristic factors and pheromone concentrations in the improved three-dimensional environment perception probability model:

[0038] For the path edge (V i ,V j ), enhance its environmental adaptation inspiration factor η env (V i ,V j );

[0039] For nodes V that cover previously uncollected areas j , improve the regional coverage heuristic factor η cov (V j );

[0040] Update pheromone concentration according to the new pheromone update rules:

[0041] τ(V i ,V j )←(1-ρ)·τ(V i ,V j )+Δτ i,j ;

[0042] Where ρ is the pheromone volatility coefficient, The flight path of the UAV contains edges (V i ,V j ) when its comprehensive fitness gain;

[0043] S27. Repeat steps S23 to S26 for several rounds of iterations until convergence or the maximum number of iterations is reached, and finally output the Pareto track candidate set with dynamic perception enhancement

[0044] Optionally, the flight energy consumption index E kIt represents the comprehensive energy consumption required by the UAV to complete the entire flight path. The comprehensive energy consumption index consists of two parts: the first is the basic flight energy consumption caused by the horizontal projection distance between adjacent flight points in the path, which is weighted by the horizontal flight energy consumption coefficient; the second is the climbing or descending energy consumption caused by the vertical height difference between adjacent flight points in the path. On the basis of the weighted vertical flight energy consumption coefficient, a canopy density adjustment factor is introduced to reflect the energy consumption differences caused by flight altitude changes under different vegetation canopy densities. The flight energy consumption index comprehensively evaluates the energy consumption changes caused by the path slope and vegetation resistance in a highly complex vegetation environment:

[0045]

[0046] in, It represents the horizontal projection distance between the i-th node and the i+1-th node on the path, reflecting the basic energy consumption caused by horizontal displacement, η hor and η ver are the energy consumption coefficients for horizontal flight and vertical climb flight, κ(z i ,z i+1 ) represents the vegetation canopy density difference adjustment factor between the two node heights, which is used to reflect the additional flight energy consumption caused by the height difference of the vegetation area;

[0047] The mission area coverage index C k Indicates the degree of coverage of effective information of the vegetation area that has not been collected by the current flight path. The mission area coverage index is calculated by accumulating the effective vegetation information area of all voxel units in the path that has not been covered by the historical path and comparing it with the total area of the remaining uncollected vegetation information. It measures the actual contribution of the current path to improving the completion of the mission. The mission area coverage index focuses on reflecting the coverage effectiveness of the UAV in collecting environmental information in complex terrain:

[0048]

[0049] in, It represents the effective vegetation information area within the i-th voxel unit in the path that is not covered by the previous path. It is the sum of all areas within the vegetation area of the mission zone for which valid information has not been collected;

[0050] The track safety index S kIndicates the overall safety performance of the flight path. The track safety index is composed of the flight safety factors of all voxel units in the path. The flight safety factor of each voxel unit is the safety distance score between it and the obstacle set, obtained after normalization by the minimum flight safety distance and evaluated in combination with the obstacle distribution model. The track safety index reflects whether the path can continuously maintain a safe flight channel, whether it is away from high-risk areas of obstacles, and whether it ensures the safety of the UAV during flight in complex vegetation areas:

[0051]

[0052] Among them, λ(V k,i ) represents node V k,i Voxel unit and obstacle distribution model OBS veg The safety distance between the normalized score and the minimum flight safety distance δ min Conduct an assessment.

[0053] Optionally, S3 includes the following steps:

[0054] S31. Multimodal vegetation observation data sample set D obtained based on ground-truth measurements and historical sampling veg Initialize the second ant population A feature , the second ant population contains N f ant individuals, each ant individual corresponds to a feature subset candidate solution;

[0055] S32. Each ant individual is based on the multimodal vegetation observation data sample set D veg The feature subset selection is performed in the multi-scale feature space of the feature subset, and the feature subset selection probability model is defined as:

[0056]

[0057] Among them, f i The i-th candidate feature corresponds to a specific spectral, texture, structural or spatiotemporal feature component in the multimodal vegetation observation data sample set, P(f i ) represents the selection of candidate features f i The probability of being adopted by the current ant individual, τ f (f i ) represents the candidate feature f i The corresponding characteristic pheromone concentration is used to reflect the candidate feature f i The contribution degree to the high discriminant performance subset in the first few iterations. The higher the concentration, the better the performance of the feature in the historical collection. η div (f i ) represents the candidate feature f iThe discriminant divergence heuristic factor is used to measure the degree of improvement in the ability to distinguish between vegetation categories after the introduction of this feature. The larger the value, the more effective the feature is in amplifying the differences between different categories. red (f i ) represents the candidate feature f i The redundancy suppression heuristic factor is used to measure the redundancy between the feature and the selected feature set. The larger the value, the less information overlap between the feature and the existing features, and the stronger the complementarity. f represents the characteristic pheromone concentration τ f (f i ) is used to control the influence of pheromone concentration on feature selection probability, β f Denotes the discriminant divergence heuristic factor η div (f i ), which is used to regulate the weight of the contribution of the feature to the class separability in the selection probability, γ f represents the redundancy suppression heuristic factor η red (f i ) is used to control the role of complementarity between features in feature selection probability;

[0058] S33. Calculate the multi-objective fitness index for each feature subset candidate solution, including the discriminant divergence index J m , feature redundancy index R m and the computational cost index T m ;

[0059] S34. After each round of iteration, the multi-objective fitness index (J m ,R m ,T m ) to perform non-dominated sorting and crowding evaluation, and select the Pareto optimal feature subset candidate set of multimodal vegetation observation data

[0060] S35. Based on the actual performance of each feature in the Pareto optimal feature subset candidate set, reversely update the heuristic factors and feature pheromone concentrations in the feature selection probability model:

[0061] Enhance the discriminant divergence heuristic factor η for the feature that improves the discriminant divergence index div (f i );

[0062] Improve the redundancy suppression heuristic factor η for features that reduce feature redundancy rate indicators red (f i );

[0063] Update the characteristic pheromone concentration according to the characteristic pheromone update rule:

[0064] τ f (f i )←(1-ρ f )·τ f (f i )+Δτ f (f i );

[0065] Among them, ρ f is the characteristic pheromone volatility coefficient, The candidate solution for the feature subset includes feature f i The comprehensive fitness gain when

[0066] S36. Repeat steps S32 to S35 for several rounds of iterations until convergence or the maximum number of iterations is reached, and finally output the Pareto optimal feature subset candidate set with dynamic perception enhancement

[0067] Optionally, the discriminant divergence index J m It is used to measure the ability of feature subsets to distinguish different vegetation categories in a multimodal vegetation observation data sample set, and the discriminant divergence index J m It is defined as the ratio of the trace of the inter-class scatter matrix between vegetation categories to the trace of the intra-class scatter matrix within the category in the kernel Fisher discriminant analysis mapping space. The trace of the inter-class scatter matrix is used to measure the degree of dispersion of the center distribution between different categories, and the trace of the intra-class scatter matrix is used to measure the degree of aggregation of sample points within the same category. Therefore, when the selected feature subset can enlarge the inter-class distance and compress the intra-class difference in the kernel mapping space, its discriminant divergence index J is m The higher it is, the more beneficial the feature subset is for class discrimination in highly complex vegetation areas;

[0068] The feature redundancy index R m It is used to measure the degree of information duplication between all features within the feature subset, and the feature redundancy rate index R m It is defined as the average value of the correlation coefficients between all pairs of features in the feature subset, where the correlation coefficient of each pair of features is used to represent the linear similarity between the two in a statistical sense. The number of features contained in the current feature subset is recorded as M, and there are M(M-1) / 2 pairs of feature combinations whose correlations need to be calculated. The average of all correlation values is taken as the redundancy value. If the average value is low, it means that the features in the feature subset are highly complementary and the information redundancy is low, which is conducive to improving the efficiency of extracting environmental information in vegetation areas.

[0069] The calculation cost index T m It is used to measure the computational overhead required for inference of feature subsets in the subsequent kernel Fisher discriminant analysis model, and calculate the cost index Tm It consists of two parts: the first part is the fixed basic inference time overhead t base , which represents the normal time required for model initialization and infrastructure calculation regardless of the feature subset selected; the second part is the linear growth term related to the number of features, indicating that each new feature will introduce a unit of computational overhead t unit , the number of features contained in the current feature subset is M, then the total computational cost index T m The product of the basic inference time plus the unit cost and the number of features, that is, t base With t unit M jointly decides and calculates the cost index T m The lower the value, the more suitable the feature subset is for deployment on the UAV edge computing platform, meeting the requirements of real-time and resource constraints in highly complex vegetation areas.

[0070] Optionally, the S5 includes the following steps:

[0071] S51. Receive, on the image processing unit of the UAV, a multimodal vegetation observation data stream collected in real time by the UAV, the data stream including hyperspectral image data, visible light image data, lidar point cloud data, and inertial navigation data;

[0072] S52. Build a multi-scale fast feature extraction operator set for multimodal vegetation observation data streams, including a hyperspectral feature extraction operator, an image texture feature extraction operator, a point cloud structure feature extraction operator, and a navigation spatiotemporal feature extraction operator. This operator set performs parallel fast feature extraction for vegetation data of different modalities.

[0073] S53. Use the hyperspectral feature extraction operator to extract vegetation band features from the hyperspectral image data and obtain the hyperspectral feature vector F used to characterize the spectral reflectance characteristics of vegetation. spec ,The different components in each hyperspectral feature vector represent the spectral reflectance intensity of different bands, reflecting the species information of the target vegetation area;

[0074] S54. Use the image texture feature extraction operator to extract vegetation texture features from the visible light image data to obtain a texture feature vector F for describing the texture pattern of the vegetation area. tex ,The components of each texture feature vector represent the spatial texture uniformity and heterogeneity characteristics of the vegetation area, reflecting the differences in the community structure of the vegetation area;

[0075] S55. Use the point cloud structure feature extraction operator to extract the three-dimensional structure features of the lidar point cloud data and obtain the three-dimensional structure feature vector F used to express the vegetation canopy structure. lidar,The components of each three-dimensional structural characteristic vector represent the vegetation canopy height, canopy density and vertical spatial distribution characteristics, reflecting the three-dimensional structural differences of vegetation;

[0076] S56. Use the navigation spatiotemporal feature extraction operator to extract the flight attitude and position information features of the inertial navigation data to obtain the navigation feature vector F used to characterize the relationship between the UAV's flight state and spatial position. nav ,The components of each navigation eigenvector represent the UAV’s attitude stability, flight trajectory and ,spatiotemporal information at the observation moment, reflecting the spatial and ,temporal context of vegetation data collection;

[0077] S57. Obtain the multi-scale feature set F multi ={F spec ,F tex ,F lidar ,F nav All features in} are based on the Pareto optimal feature subset candidate set The feature retention rules determined in the above are used for rapid screening, and only the feature components belonging to the Pareto optimal feature subset candidate set are retained, and the unselected redundant feature components are removed to obtain the compressed feature data stream F comp .

[0078] Optionally, the S6 includes the following steps:

[0079] S61. compress the feature data stream F comp Input Kernel Fisher Discriminant Analysis Model KFDA veg ,The kernel Fisher discriminant analysis model is a classification and inversion model constructed based on the Pareto optimal feature subset candidate set. ,The kernel Fisher discriminant analysis model uses a nonlinear mapping method to map the original feature space to a high-dimensional ,feature space;

[0080] S62. Automatically match the kernel function type and kernel function parameters used in the kernel Fisher discriminant analysis model based on the feature combination structure in the Pareto optimal feature subset candidate set;

[0081] S63. Perform kernel mapping on the input compressed feature data stream to generate a high-dimensional kernel mapping feature expression, perform multi-class classification based on the established discriminant hyperplane, and output the vegetation type classification result of the current drone observation area;

[0082] S64. The vegetation type classification results are spatially projected based on the UAV's geolocation information and the image's spatial distribution information to generate a continuous rasterized classification layer. Each pixel is assigned a vegetation type label to describe the spatial distribution pattern of complex vegetation areas.

[0083] S65. Based on the output vegetation type classification layer, based on the nonlinear mapping relationship between the weight structure between category features implicit in the kernel Fisher discriminant analysis model and the input features, combined with the known ecological indicator label information in the ground samples, perform simultaneous inversion processing of ecological indicators in highly complex vegetation areas and output the ecological indicator inversion results;

[0084] S66. The ecological index inversion results include the following ecological indicators:

[0085] Vegetation coverage index: represents the vertical projection ratio of vegetation within the pixel range at each spatial position;

[0086] Canopy density index: Estimates the canopy volume distribution density per unit area based on the structural characteristics of the lidar point cloud;

[0087] Vegetation health index: The normalized vegetation index is calculated based on the ratio of red and near-infrared band reflectance in the hyperspectral characteristics;

[0088] Community vertical structure index: The three-dimensional structure complexity of the community is derived by integrating the point cloud height distribution and spectral variation amplitude;

[0089] Spatial heterogeneity index: reflects the characteristics of vegetation landscape patches based on the spatial continuity and texture variation of the same type of vegetation in the classification layer.

[0090] Optionally, the kernel function type includes a radial basis function kernel, a polynomial kernel or a hybrid kernel, and the kernel function parameters include kernel width, polynomial order and hybrid kernel weight coefficient, which are specifically determined by the model fitness index corresponding to the feature subset and the online sample distribution pattern.

[0091] The beneficial effects of the present invention are:

[0092] (1) The present invention introduces an improved three-dimensional environmental perception probability model in flight path planning. The model integrates pheromone concentration, local terrain disturbance index and density of uncovered vegetation areas to form a composite path selection strategy based on three factors: flight cost, mission value and path safety. This strategy not only avoids the limitation of decoupling path planning from information acquisition in traditional algorithms, but also uses real-time flight results for dynamic adjustment of the flight status diagram through a structural feedback loop, effectively solving the problems of conventional trajectory schemes in vegetation areas, such as path blind spots, excessive energy consumption or risk aggregation.

[0093] (2) The present invention introduces multi-objective ant colony optimization into the multimodal feature compression task in UAV vegetation information collection, and simultaneously optimizes the three objectives of discriminant divergence, feature redundancy rate and model calculation cost in the feature selection process. The selected subset can significantly compress the model input dimension and maintain a high category discrimination ability. The compressed feature subset further drives the kernel Fisher discriminant analysis model to perform dynamic adaptive matching of kernel function type and kernel parameters, so that the model can maintain stable discrimination performance when processing different objects with the same spectrum or the same objects with different spectra.

[0094] (3) The present invention integrates a lightweight kernel Fisher discriminant analysis model into the UAV image processing unit and constructs a compressed feature data stream based on the selected Pareto optimal feature subset, significantly reducing the model input dimension and computational complexity. Simultaneously, through the inter-class divergence projection relationship in the model structure, multi-dimensional ecological indicators such as vegetation coverage, canopy density, and health are inferred during classification and reasoning, realizing a closed loop of UAV ecological perception that allows for simultaneous flight, computation, and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0096] Figure 1 This is a flow chart of a data acquisition system based on a drone to collect environmental data from a distance, proposed by the present invention;

[0097] Figure 2 This is a flow chart of a multi-target feature subset search performed by a second ant population in a data acquisition system based on a drone for remotely acquiring environmental data, as proposed by the present invention;

[0098] Figure 3 This is a schematic diagram of the reasoning path for vegetation type classification and ecological indicator inversion in the kernel Fisher discriminant analysis model using compressed feature data streams in a data acquisition system based on long-distance environmental data acquisition by drones proposed in the present invention. DETAILED DESCRIPTION

[0099] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0100] refer to Figure 1-Figure 3 , a data acquisition system based on long-distance collection of environmental data by drones, including the following modules:

[0101] The flight mission modeling module is used to construct a three-dimensional flight feasible domain model of the target vegetation area and establish a flight status diagram;

[0102] The path optimization module is used to initialize the first ant population in the flight state graph, perform multi-target path search based on the improved three-dimensional environment perception probability model, and output a set of Pareto trajectory candidates;

[0103] The feature selection module is used to initialize the second ant population, build a feature selection probability model based on the multimodal vegetation observation data sample set, perform feature subset search based on three objectives: discriminant divergence, feature redundancy rate, and computational cost, and generate a Pareto optimal feature subset candidate set;

[0104] The feature extraction module is deployed on the UAV image processing unit and is used to perform multi-scale feature fast extraction on the collected multi-modal vegetation observation data stream and compress it according to the Pareto optimal feature subset candidate set to generate a compressed feature data stream;

[0105] The classification and inversion module is used to input the compressed feature data stream into the kernel Fisher discriminant analysis model, perform online inference based on the optimal kernel function type and parameters, and generate a vegetation type classification layer.

[0106] A data acquisition method for remotely collecting environmental data based on an unmanned aerial vehicle (UAV) is applied to a data acquisition system for remotely collecting environmental data based on an unmanned aerial vehicle (UAV), comprising the following steps:

[0107] S1. Obtain a digital elevation model and obstacle distribution model of the target vegetation area, and generate a 3D flight feasible range model based on the solar altitude angle and terrain type.

[0108] S2. Initialize a first ant population within the three-dimensional flight feasible domain model and use a multi-objective ant colony optimization algorithm to search for flight energy consumption targets, mission area coverage targets, and track safety targets to generate a set of candidate tracks.

[0109] S3. Initialize a second ant population based on a set of multimodal vegetation observation data samples obtained from ground and historical sampling. Use a multi-objective ant colony optimization algorithm to search for discriminant divergence, feature redundancy, and computational cost objectives to generate a candidate set of feature subsets.

[0110] S4. Based on the remaining battery level, real-time wind field information, and lighting conditions, the drone selects the current optimal track from the Pareto track candidate set. The drone then flies along this optimal track, continuously collecting hyperspectral image data, visible light image data, lidar point cloud data, and inertial navigation data during flight to form a multimodal vegetation observation data stream.

[0111] S5. Perform multi-scale feature fast extraction on the multimodal vegetation observation data stream on the image processing unit of the UAV, and retain only the corresponding features in the Pareto feature subset candidate set to obtain a compressed feature data stream;

[0112] S6. Input the compressed feature data stream into the kernel Fisher discriminant analysis model. The kernel Fisher discriminant analysis model performs online inference according to the kernel function type and kernel parameters determined by the Pareto feature subset candidate set to generate vegetation type classification results and ecological indicator inversion results.

[0113] In this embodiment, S1 includes the following steps:

[0114] S11. The three-dimensional terrain information of the target vegetation area is modeled using remote sensing images, and a digital elevation model (DEM) is constructed to describe the elevation values corresponding to each two-dimensional geographic location within the vegetation area. veg , where the elevation value corresponding to each spatial point position (x, y) is h(x, y), and the spatial point position belongs to the spatial definition domain Ω of the target vegetation area veg ;

[0115] S12. Obtain an obstacle distribution model OBS for representing the positions and heights of all obstacles in the vegetation area veg , each obstacle is identified by its two-dimensional position and vertical height For three-dimensional expression, there are N obs obstacle sample points constitute a complete obstacle set;

[0116] S13. Construct the illumination constraint factor set IL according to the solar altitude angle information of the flight mission period veg , which is used to characterize the imaging illumination conditions of the target vegetation area at different times and different geographical locations. The illumination constraint factor of each spatial point position (x, y) at time t is determined by the solar incident angle θ at that point. s (x, y, t) is jointly determined with the ground object albedo weight factor μ(x, y, t) to reflect the comprehensive weight of the current point imaging illumination quality;

[0117] S14. Based on the digital elevation model DEM veg , obstacle distribution model OBS veg and the lighting constraint factor set IL veg , combined with the mission area terrain type to build a three-dimensional flight feasible domain model FE 3D In the three-dimensional flight feasible domain model, the flight altitude range corresponding to each two-dimensional spatial point position (x, y) in the vegetation area is the ground elevation value h(x, y) at that position plus the minimum flight safety distance δ min To the maximum allowed flight height z affected by lighting constraintsmax The continuous height interval between (x, y), the three-dimensional flight feasible area space definition domain should exclude the obstacle area Ω defined by the obstacle projection obs ;

[0118] S15. 3D flight feasible domain model FE 3D Perform spatial discretization and divide it into multiple cubic three-dimensional voxel units V i , each 3D voxel unit is defined as a cubic area with fixed spatial resolution Δx, Δy, Δz in the x, y, and z directions. The collection of all 3D voxel units together constitutes the entire 3D flight feasible domain model;

[0119]

[0120] Among them, Δx, Δy, Δz are the spatial resolutions of voxel division, N vox is the total number of voxels in the 3D flight feasible domain model, V i is the i-th three-dimensional voxel unit.

[0121] In this embodiment, S2 includes the following steps:

[0122] S21. In the three-dimensional flight feasible domain model FE 3D The flight state graph G with dynamic constraints is constructed based on the voxel unit division results. flight ,The nodes of the flight state graph are the voxel unit centers, and the connectivity between adjacent voxel units is dynamically adjusted based on the vegetation height difference, wind field disturbance coefficient, and energy consumption threshold;

[0123] S22. In flight state diagram G flight Initialize the first ant population A route , the first ant population contains N a ant individuals, each ant individual corresponds to an initial drone flight path;

[0124] S23. During each iteration, each ant in the first ant population selects the next pixel unit for extending the drone's flight path based on the improved 3D environment perception probability model. The improved 3D environment perception probability model is defined as:

[0125] Among them, τ(V i ,V j ) represents node V i To node V j The path pheromone concentration, τ(V i ,V′ j ) represents node V i To candidate node V′ jThe path pheromone concentration, η env (V i ,V j ) represents node V i To node V j The environmental adaptation heuristic factor is combined with the vegetation height difference, terrain complexity and wind field disturbance coefficient between the two nodes. For the task of extracting environmental information in the vegetation area, the flight path of the UAV with small wind disturbance and stable flight is preferred, η env (V i ,V′ j ) represents node V i To candidate node V′ j Environmental adaptation heuristic factor, η cov (V j ) represents node V j The area coverage heuristic factor is the higher the uncovered ratio of the voxel unit where the node is located, the larger its value is, which encourages the UAV to give priority to areas where vegetation information is not collected, η cov (V′ j ) represents the candidate node V′ j The area coverage heuristic factor, N(V i ) represents node V i The set of adjacent nodes that meet the flight constraints, α, β, and γ are the importance weights of pheromone, environmental adaptation heuristic factor, and area coverage heuristic factor, respectively;

[0126] The significance of the S23 formula:

[0127] The multi-objective coordinated path selection mechanism embodies the principle of "local intelligent decision-making" during path construction. Each ant's path selection at each step considers both the existing pheromone concentration (historical experience), the environmental adaptability of the current path (energy consumption risk), and a dynamic assessment of the "information value" (data collection priority) of the target area. This constitutes a real-time, multi-factor, task-oriented local optimization strategy.

[0128] Dynamic mapping path search guided by environmental information, by modeling environmental factors, wind speed, slope, and height difference into η env ,path selection not only depends on static terrain, but also responds to dynamic changes in the real-time flight environment.,Compared to traditional ant colonies that only rely on pheromones, this mechanism has stronger,environmental perception capabilities and is suitable for tasks with high requirements on path stability and,safety in highly complex terrain.

[0129] Collect value-driven autonomous path scheduling, η covBy guiding ants to prioritize areas that have not yet been covered, a positive feedback loop is formed between path planning and information gain. The system does not blindly pursue the shortest path, but rather prefers the path segment with the largest amount of information, significantly improving the data utilization rate of a single flight.

[0130] Beneficial effects after implementation:

[0131] The mission adaptability and stability of trajectory planning are improved. By simultaneously considering the three dimensions of energy consumption, safety, and sampling value, the method of the present invention can dynamically avoid areas with high energy consumption, strong wind disturbance, and repeated information, making the trajectory generated in complex mountainous and forested areas more executable and effective.

[0132] The convergence efficiency and solution quality of the multi-objective search algorithm are enhanced. The collaborative participation of the heuristic factor guides the search space to the task-related area, effectively reducing the probability of the algorithm falling into the local optimum in the early stage. env ,η cov The collaborative update mechanism makes the distribution of solutions more concentrated in the efficient area, ultimately improving the quality and distribution balance of the Pareto solution set.

[0133] To achieve the linkage optimization of flight path and information collection value, compared with the "path is target" scheduling of traditional ant colony algorithm, this paper introduces the regional coverage factor η cov , achieving joint optimization of path planning and maximizing data value, so that the drone can focus on the "high information area" in every flight.

[0134] S24. Calculate the multi-objective fitness index of each UAV flight path, including the flight energy consumption index E k , Mission area coverage index C k and track safety index S k ;

[0135] S25. After each round of iteration, the multi-objective fitness index (E k ,C k ,S k ) Perform non-dominated sorting and congestion evaluation to select the Pareto optimal path set

[0136] S26. Based on the actual performance of each node in the Pareto optimal path set, reversely update the heuristic factors and pheromone concentrations in the improved three-dimensional environment perception probability model:

[0137] For the path edge (V i ,V j ), enhance its environmental adaptation inspiration factor η env (V i ,Vj );

[0138] For nodes V that cover previously uncollected areas j , improve the regional coverage heuristic factor η cov (V j );

[0139] Update pheromone concentration according to the new pheromone update rules:

[0140] τ(V i ,V j )←(1-ρ)·τ(V i ,V j )+Δτ i,j ;

[0141] Where ρ is the pheromone volatility coefficient, The flight path of the UAV contains edges (V i ,V j ) when its comprehensive fitness gain;

[0142] S27. Repeat steps S23 to S26 for several rounds of iterations until convergence or the maximum number of iterations is reached, and finally output the Pareto track candidate set with dynamic perception enhancement

[0143] In this embodiment, the flight energy consumption index E k It represents the comprehensive energy consumption required by the UAV to complete the entire flight path. The comprehensive energy consumption index consists of two parts: the first is the basic flight energy consumption caused by the horizontal projection distance between adjacent flight points in the path, which is weighted by the horizontal flight energy consumption coefficient; the second is the climbing or descending energy consumption caused by the vertical height difference between adjacent flight points in the path. On the basis of the weighted vertical flight energy consumption coefficient, a canopy density adjustment factor is introduced to reflect the energy consumption differences caused by flight altitude changes under different vegetation canopy densities. The flight energy consumption index comprehensively evaluates the energy consumption changes caused by the path slope and vegetation resistance in a highly complex vegetation environment:

[0144]

[0145] in, It represents the horizontal projection distance between the i-th node and the i+1-th node on the path, reflecting the basic energy consumption caused by horizontal displacement, η hor and η ver are the energy consumption coefficients for horizontal flight and vertical climb flight, κ(z i ,z i+1) represents the vegetation canopy density difference adjustment factor between the two node heights, which is used to reflect the additional flight energy consumption caused by the height difference of the vegetation area;

[0146] Mission area coverage indicator C k Indicates the degree of coverage of effective information of the vegetation area that has not been collected by the current flight path. The mission area coverage index is calculated by accumulating the effective vegetation information area of all voxel units in the path that has not been covered by the historical path and comparing it with the total area of the remaining uncollected vegetation information. It measures the actual contribution of the current path to improving the completion of the mission. The mission area coverage index focuses on reflecting the coverage effectiveness of the UAV in collecting environmental information in complex terrain:

[0147]

[0148] in, It represents the effective vegetation information area within the i-th voxel unit in the path that is not covered by the previous path. It is the sum of all areas within the vegetation area of the mission zone for which valid information has not been collected;

[0149] Track safety index S k Indicates the overall safety performance of the flight path. The track safety index is composed of the flight safety factors of all voxel units in the path. The flight safety factor of each voxel unit is the safety distance score between it and the obstacle set, obtained after normalization by the minimum flight safety distance and evaluated in combination with the obstacle distribution model. The track safety index reflects whether the path can continuously maintain a safe flight channel, whether it is away from high-risk areas of obstacles, and whether it ensures the safety of the UAV during flight in complex vegetation areas:

[0150]

[0151] Among them, λ(V k,i ) represents node V k,i Voxel unit and obstacle distribution model OBS veg The safety distance between the normalized score and the minimum flight safety distance δ min Conduct an assessment.

[0152] In this embodiment, S3 includes the following steps:

[0153] S31. Multimodal vegetation observation data sample set D obtained based on ground-truth measurements and historical sampling veg Initialize the second ant population A feature , the second ant population contains N f ant individuals, each ant individual corresponds to a feature subset candidate solution;

[0154] S32. Each ant individual is based on the multimodal vegetation observation data sample set D veg The feature subset selection is performed in the multi-scale feature space of the feature subset, and the feature subset selection probability model is defined as:

[0155]

[0156] Among them, f i The i-th candidate feature corresponds to a specific spectral, texture, structural or spatiotemporal feature component in the multimodal vegetation observation data sample set, P(f i ) represents the selection of candidate features f i The probability of being adopted by the current ant individual, τ f (f i ) represents the candidate feature f i The corresponding characteristic pheromone concentration is used to reflect the candidate feature f i The contribution degree to the high discriminant performance subset in the first few iterations. The higher the concentration, the better the performance of the feature in the historical collection. η div (f i ) represents the candidate feature f i The discriminant divergence heuristic factor is used to measure the degree of improvement in the ability to distinguish between vegetation categories after the introduction of this feature. The larger the value, the more effective the feature is in amplifying the differences between different categories. red (f i ) represents the candidate feature f i The redundancy suppression heuristic factor is used to measure the redundancy between the feature and the selected feature set. The larger the value, the less information overlap between the feature and the existing features, and the stronger the complementarity. f represents the characteristic pheromone concentration τ f (f i ) is used to control the influence of pheromone concentration on feature selection probability, β f Denotes the discriminant divergence heuristic factor η div (f i ), which is used to regulate the weight of the contribution of the feature to the class separability in the selection probability, γ f represents the redundancy suppression heuristic factor η red (f i ) is used to control the role of complementarity between features in feature selection probability;

[0157] S33. Calculate the multi-objective fitness index for each feature subset candidate solution, including the discriminant divergence index J m , feature redundancy index R m and the computational cost index T m ;

[0158] S34. After each round of iteration, the multi-objective fitness index (J m ,R m ,T m ) to perform non-dominated sorting and crowding evaluation, and select the Pareto optimal feature subset candidate set of multimodal vegetation observation data

[0159] S35. Based on the actual performance of each feature in the Pareto optimal feature subset candidate set, reversely update the heuristic factors and feature pheromone concentrations in the feature selection probability model:

[0160] Enhance the discriminant divergence heuristic factor η for the feature that improves the discriminant divergence index div (f i );

[0161] Improve the redundancy suppression heuristic factor η for features that reduce feature redundancy rate indicators red (f i );

[0162] Update the characteristic pheromone concentration according to the characteristic pheromone update rule:

[0163] τ f (f i )←(1-ρ f )·τ f (f i )+Δτ f (f i );

[0164] Among them, ρ f is the characteristic pheromone volatility coefficient, The candidate solution for the feature subset includes feature f i The comprehensive fitness gain when

[0165] S36. Repeat steps S32 to S35 for several rounds of iterations until convergence or the maximum number of iterations is reached, and finally output the Pareto optimal feature subset candidate set with dynamic perception enhancement

[0166] In this embodiment, the discriminant divergence index J m It is used to measure the ability of feature subsets to distinguish different vegetation categories in a multimodal vegetation observation data sample set, and the discriminant divergence index J mIt is defined as the ratio of the trace of the inter-class scatter matrix between vegetation categories to the trace of the intra-class scatter matrix within the category in the kernel Fisher discriminant analysis mapping space. The trace of the inter-class scatter matrix is used to measure the degree of dispersion of the center distribution between different categories, and the trace of the intra-class scatter matrix is used to measure the degree of aggregation of sample points within the same category. Therefore, when the selected feature subset can enlarge the inter-class distance and compress the intra-class difference in the kernel mapping space, its discriminant divergence index J is m The higher it is, the more beneficial the feature subset is for class discrimination in highly complex vegetation areas;

[0167] Feature redundancy index R m It is used to measure the degree of information duplication between all features within the feature subset, and the feature redundancy rate index R m It is defined as the average value of the correlation coefficients between all pairs of features in the feature subset, where the correlation coefficient of each pair of features is used to represent the linear similarity between the two in a statistical sense. The number of features contained in the current feature subset is recorded as M, and there are M(M-1) / 2 pairs of feature combinations whose correlations need to be calculated. The average of all correlation values is taken as the redundancy value. If the average value is low, it means that the features in the feature subset are highly complementary and the information redundancy is low, which is conducive to improving the efficiency of extracting environmental information in vegetation areas.

[0168] Calculate the cost index T m It is used to measure the computational overhead required for inference of feature subsets in the subsequent kernel Fisher discriminant analysis model, and calculate the cost index T m It consists of two parts: the first part is the fixed basic inference time overhead t base , which represents the normal time required for model initialization and infrastructure calculation regardless of the feature subset selected; the second part is the linear growth term related to the number of features, indicating that each new feature will introduce a unit of computational overhead t unit , the number of features contained in the current feature subset is M, then the total computational cost index T m The product of the basic inference time plus the unit cost and the number of features, that is, t base With t unit M jointly decides and calculates the cost index T m The lower the value, the more suitable the feature subset is for deployment on the UAV edge computing platform, meeting the requirements of real-time and resource constraints in highly complex vegetation areas.

[0169] In this embodiment, S5 includes the following steps:

[0170] S51. Receive, on the image processing unit of the UAV, a multimodal vegetation observation data stream collected in real time by the UAV, the data stream including hyperspectral image data, visible light image data, lidar point cloud data, and inertial navigation data;

[0171] S52. Build a multi-scale fast feature extraction operator set for multimodal vegetation observation data streams, including a hyperspectral feature extraction operator, an image texture feature extraction operator, a point cloud structure feature extraction operator, and a navigation spatiotemporal feature extraction operator. This operator set performs parallel fast feature extraction for vegetation data of different modalities.

[0172] S53. Use the hyperspectral feature extraction operator to extract vegetation band features from the hyperspectral image data and obtain the hyperspectral feature vector F used to characterize the spectral reflectance characteristics of vegetation. spec ,The different components in each hyperspectral feature vector represent the spectral reflectance intensity of different bands, reflecting the species information of the target vegetation area;

[0173] S54. Use the image texture feature extraction operator to extract vegetation texture features from the visible light image data to obtain a texture feature vector F for describing the texture pattern of the vegetation area. tex ,The components of each texture feature vector represent the spatial texture uniformity and heterogeneity characteristics of the vegetation area, reflecting the differences in the community structure of the vegetation area;

[0174] S55. Use the point cloud structure feature extraction operator to extract the three-dimensional structure features of the lidar point cloud data and obtain the three-dimensional structure feature vector F used to express the vegetation canopy structure. lidar ,The components of each three-dimensional structural characteristic vector represent the vegetation canopy height, canopy density and vertical spatial distribution characteristics, reflecting the three-dimensional structural differences of vegetation;

[0175] S56. Use the navigation spatiotemporal feature extraction operator to extract the flight attitude and position information features of the inertial navigation data to obtain the navigation feature vector F used to characterize the relationship between the UAV's flight state and spatial position. nav ,The components of each navigation eigenvector represent the UAV’s attitude stability, flight trajectory and ,spatiotemporal information at the observation moment, reflecting the spatial and ,temporal context of vegetation data collection;

[0176] S57. Obtain the multi-scale feature set F multi ={F spec ,F tex ,F lidar ,F nav All features in} are based on the Pareto optimal feature subset candidate set The feature retention rules determined in the above are used for rapid screening, and only the feature components belonging to the Pareto optimal feature subset candidate set are retained, and the unselected redundant feature components are removed to obtain the compressed feature data stream F comp .

[0177] In this embodiment, S6 includes the following steps:

[0178] S61. compress the feature data stream F comp Input Kernel Fisher Discriminant Analysis Model KFDA veg ,The kernel Fisher discriminant analysis model is a classification and inversion model constructed based on the Pareto optimal feature subset candidate set. ,The kernel Fisher discriminant analysis model uses a nonlinear mapping method to map the original feature space to a high-dimensional ,feature space;

[0179] S62. Automatically match the kernel function type and kernel function parameters used in the kernel Fisher discriminant analysis model based on the feature combination structure in the Pareto optimal feature subset candidate set;

[0180] S63. Perform kernel mapping on the input compressed feature data stream to generate a high-dimensional kernel mapping feature expression, perform multi-class classification based on the established discriminant hyperplane, and output the vegetation type classification result of the current drone observation area;

[0181] S64. The vegetation type classification results are spatially projected based on the UAV's geolocation information and the image's spatial distribution information to generate a continuous rasterized classification layer. Each pixel is assigned a vegetation type label to describe the spatial distribution pattern of complex vegetation areas.

[0182] S65. Based on the output vegetation type classification layer, based on the nonlinear mapping relationship between the weight structure between category features implicit in the kernel Fisher discriminant analysis model and the input features, combined with the known ecological indicator label information in the ground samples, perform simultaneous inversion processing of ecological indicators in highly complex vegetation areas and output the ecological indicator inversion results;

[0183] S66. The ecological index inversion results include the following ecological indicators:

[0184] Vegetation coverage index: represents the vertical projection ratio of vegetation within the pixel range at each spatial position;

[0185] Canopy density index: Estimates the canopy volume distribution density per unit area based on the structural characteristics of the lidar point cloud;

[0186] Vegetation health index: The normalized vegetation index is calculated based on the ratio of red and near-infrared band reflectance in the hyperspectral characteristics;

[0187] Community vertical structure index: The three-dimensional structure complexity of the community is derived by integrating the point cloud height distribution and spectral variation amplitude;

[0188] Spatial heterogeneity index: reflects the characteristics of vegetation landscape patches based on the spatial continuity and texture variation of the same type of vegetation in the classification layer.

[0189] In this embodiment, the kernel function types include radial basis function kernel, polynomial kernel or hybrid kernel, and the kernel function parameters include kernel width, polynomial order and hybrid kernel weight coefficient, which are specifically determined by the model fitness index corresponding to the feature subset and the online sample distribution pattern.

[0190] Example 1:

[0191] At 8:32 am on March 14, 2025, the operation and maintenance team completed takeoff calibration at an altitude of 1086m on the southeast side of the national nature reserve. The drone's takeoff coordinates were (xx.8702°N,xx.9885°E), with 6520mAh of battery remaining and 9% of the onboard NVIDIA Jetson Xavier NX GPU load. The system called a high-precision digital elevation model and a previously mapped obstacle distribution model, combined with a sunshine angle model to generate a three-dimensional flight feasible domain model (FE). 3D Then the flight state map G is constructed with a voxel resolution of Δx = Δy = Δz = 2m flight , and initialize the first ant population A route (Scale: 200).

[0192] At 8:33:07 AM, after 35 iterations, the system output its first set of eight Pareto trajectory candidates. The optimal trajectory had an energy consumption range of 371–428 Wh, a coverage range of 81.4%–92.6%, and a safety factor range of 0.78–0.86. Based on remaining power, wind farm measurements, and solar constraints, the O&M strategy module selected trajectory P3 from the candidate set, which had an energy consumption of 389 Wh, a coverage rate of 90.1%, and a minimum safety factor of 0.82.

[0193] 8:34:55, the drone entered the first segment of track P3 - Voxel V 12 To voxel V 48 The horizontal distance of this section was 214 meters, the vertical climb was 37 meters, and the real-time energy consumption was 23Wh. At 8:35:03, the onboard multimodal acquisition system completed the simultaneous acquisition of frame ID-14792: a 269×512 pixel hyperspectral cube, a 4000×3000 RGB image, a 43,215-point lidar point cloud, and an inertial navigation quadruple.

[0194] Based on the image processing unit, the hyperspectral feature extraction operator O is activated simultaneously. spec , image texture operator O tex , point cloud structure operator O lidar and navigation spacetime operator O nav The original 307-dimensional features were extracted in 0.118 seconds. Then, the Pareto optimal feature subset finally output by the second ant population was called according to S5 / S57. (104 dimensions), automatically remove redundant 203 dimensions, and form a compressed feature data stream Fcomp .

[0195] 8:35:07, the system starts the Kernel Fisher Discriminant Analysis model KFDA veg The model automatically selects the RBF kernel with a kernel width of σ = 1.35 and loads the weight matrix of the five types of vegetation historical samples. After 0.43 seconds, the vegetation type classification result is output: voxel V 48 Determined to be “primitive broadleaved forest” (confidence 0.91), voxel V 49 It was determined to be "secondary forest" (confidence 0.83). Then, the system used S6 / S65–S66 to invert the ecological indicators: NDVI = 0.812, vegetation coverage Canopy density Vertical Structure Index And write to the raster buffer in real time.

[0196] 8:35:09, the airborne meteorological unit detected the voxel edge (V 48 ,V 62 ) When the wind speed increases to 12m / s, the system reduces the pheromone concentration on this side by 15%. 62 The area is not covered yet, and its coverage heuristic factor η cov (V 62 ) automatically increases by 12%. The first ant population immediately uses the new weights to search for the next path. The updated trajectory avoids high wind speed areas, ensuring a safety factor greater than 0.80.

[0197] At 8:38:12, the second ant population completed the feature subset search. m Increased to 8.32, redundancy rate R m Reduced to 0.21, the cost is calculated as T m The system evaluated that the newly added texture directional gradient feature only improved NDVI by 0.3% but increased inference latency by 0.04 seconds. This triggered a culling rule, maintaining a 104-dimensional input dimension and ensuring stable on-device inference within 0.46 seconds.

[0198] 9:47:20, the drone completed three 1km blocks 2 Data collection and real-time reasoning of the sample area, cumulative flight time 32 minutes and 25 seconds, actual energy consumption 378Wh, total coverage rate 92.4%. The system generates three layers of output:

[0199] Vegetation type classification raster — Resolution 0.5m, marking five categories: "primary broad-leaved forest", "secondary forest", "shrubland", "orchard" and "farmland";

[0200] Ecological indicator inversion grid —Contains five bands: NDVI, FVC, canopy density, vertical structure index, and spatial heterogeneity index;

[0201] Track energy consumption and safety report - records the horizontal / vertical displacement, real-time energy consumption, wind field disturbance level and minimum safety distance of each track.

[0202] Comparison results from 75 10m x 10m field plots showed an overall classification accuracy of 89.6% (compared to 76.8% for the traditional full-feature SVM model), with a mean absolute error of ±0.042 for NDVI (compared to ±0.077 for the traditional method). The accuracy of this method for easily confounded primary and secondary forests increased from 72.3% to 88.7%, and for shrublands and orchards from 79.1% to 91.4%.

[0203] Furthermore, compared to an evenly spaced grid track, this method reduces flight time by 18.5%. Compared to full-feature input, model inference speed is increased by 66.7%, and the onboard GPU peak utilization rate is reduced to 47%. All results are transmitted back to the ground control center via 5G within 5 minutes of the mission completion, allowing the protected area management department to directly use them for vegetation health assessment and early detection of pests and diseases.

[0204] The entire process of this embodiment demonstrates the closed loop of three-dimensional track - feature compression - discriminant analysis - ecological inversion - pheromone feedback - path re-regulation, which fully verifies the real-time, accuracy and efficient resource utilization capabilities of the present invention under complex terrain, heterogeneous lighting and multimodal high-dimensional data conditions.

[0205] The present invention introduces an improved three-dimensional environmental perception probability model in flight path planning. The model integrates pheromone concentration, local terrain disturbance index and density of uncovered vegetation areas to form a composite path selection strategy based on three factors: flight cost, mission value and path safety. This strategy not only avoids the limitation of decoupling path planning from information acquisition in traditional algorithms, but also uses a structural feedback loop to dynamically adjust the flight status diagram using real-time flight results, effectively solving the problems of conventional trajectory schemes in vegetation areas, such as path blind spots, excessive energy consumption or risk aggregation.

[0206] The present invention introduces multi-objective ant colony optimization into the multimodal feature compression task in UAV vegetation information collection, and simultaneously optimizes the triple objectives of discriminant divergence, feature redundancy rate and model calculation cost during the feature selection process. The selected subset can significantly compress the model input dimension and maintain a high category discrimination ability. The compressed feature subset further drives the kernel Fisher discriminant analysis model to perform dynamic adaptive matching of kernel function type and kernel parameters, so that the model can maintain stable discrimination performance when processing different objects with the same spectrum or the same objects with different spectra.

[0207] This paper integrates a lightweight kernel Fisher discriminant analysis model into the drone image processing unit and constructs a compressed feature data stream based on a selected Pareto-optimal feature subset, significantly reducing the model's input dimensionality and computational complexity. Furthermore, through the inter-class divergence projection relationship within the model structure, multidimensional ecological indicators such as vegetation coverage, canopy density, and health are derived simultaneously during classification and inference, achieving a closed-loop drone ecological perception system that integrates flight, computation, and analysis.

[0208] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A data acquisition system based on remote collection of environmental data by drones, characterized in that: Includes the following modules: The flight mission modeling module is used to construct a three-dimensional flight feasible domain model of the target vegetation area and establish a flight status diagram; The path optimization module is used to initialize the first ant population in the flight state graph, perform multi-target path search based on the improved three-dimensional environment perception probability model, and output a set of Pareto trajectory candidates; The feature selection module is used to initialize the second ant population, build a feature selection probability model based on the multimodal vegetation observation data sample set, perform feature subset search based on three objectives: discriminant divergence, feature redundancy rate, and computational cost, and generate a Pareto optimal feature subset candidate set; The feature extraction module is deployed on the UAV image processing unit and is used to perform multi-scale feature fast extraction on the collected multi-modal vegetation observation data stream and compress it according to the Pareto optimal feature subset candidate set to generate a compressed feature data stream; The classification and inversion module is used to input the compressed feature data stream into the kernel Fisher discriminant analysis model, perform online inference based on the optimal kernel function type and parameters, and generate a vegetation type classification layer.

2. A data acquisition method based on long-distance environmental data acquisition by drones, applied to the data acquisition system based on long-distance environmental data acquisition by drones according to claim 1, characterized in that: The following steps are involved: S1. Obtain a digital elevation model and obstacle distribution model of the target vegetation area, and generate a 3D flight feasible range model based on the solar altitude angle and terrain type. S2. Initialize a first ant population within the three-dimensional flight feasible domain model and use a multi-objective ant colony optimization algorithm to search for flight energy consumption targets, mission area coverage targets, and track safety targets to generate a set of candidate tracks. S3. Initialize a second ant population based on a set of multimodal vegetation observation data samples obtained from ground and historical sampling. Use a multi-objective ant colony optimization algorithm to search for discriminant divergence, feature redundancy, and computational cost objectives to generate a candidate set of feature subsets. S4. Based on the remaining battery level, real-time wind field information, and lighting conditions, the drone selects the current optimal track from the Pareto track candidate set. The drone then flies along this optimal track, continuously collecting hyperspectral image data, visible light image data, lidar point cloud data, and inertial navigation data during flight to form a multimodal vegetation observation data stream. S5. Perform multi-scale feature fast extraction on the multimodal vegetation observation data stream on the image processing unit of the UAV, and retain only the corresponding features in the Pareto feature subset candidate set to obtain a compressed feature data stream; S6. Input the compressed feature data stream into the kernel Fisher discriminant analysis model to generate vegetation type classification results and ecological indicator inversion results.

3. The data acquisition method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 2, characterized in that: Said S1 comprises the following steps: S11. The three-dimensional terrain information of the target vegetation area is modeled using remote sensing images, and a digital elevation model (DEM) is constructed to describe the elevation values corresponding to each two-dimensional geographic location within the vegetation area. veg , where the elevation value corresponding to each spatial point position (x, y) is h(x, y), and the spatial point position belongs to the spatial definition domain Ω of the target vegetation area veg ; S12. Obtain an obstacle distribution model OBS for representing the positions and heights of all obstacles in the vegetation area veg , each obstacle is identified by its two-dimensional position and vertical height For three-dimensional expression, there are N obs obstacle sample points constitute a complete obstacle set; S13. Construct the illumination constraint factor set IL according to the solar altitude angle information of the flight mission period veg , which is used to characterize the imaging illumination conditions of the target vegetation area at different times and different geographical locations. The illumination constraint factor of each spatial point position (x, y) at time t is determined by the solar incident angle θ at that point. s (x, y, t) is jointly determined with the ground object albedo weight factor μ(x, y, t) to reflect the comprehensive weight of the current point imaging illumination quality; S14. Based on the digital elevation model DEM veg , obstacle distribution model OBS veg and the lighting constraint factor set IL veg , combined with the mission area terrain type to build a three-dimensional flight feasible domain model FE 3D In the three-dimensional flight feasible domain model, the flight altitude range corresponding to each two-dimensional spatial point position (x, y) in the vegetation area is the ground elevation value h(x, y) at that position plus the minimum flight safety distance δ min To the maximum allowed flight height z affected by lighting constraints max The continuous height interval between (x, y), the three-dimensional flight feasible area space definition domain should exclude the obstacle area Ω defined by the obstacle projection obs ; S15. 3D flight feasible domain model FE 3D Perform spatial discretization and divide it into multiple cubic three-dimensional voxel units V i , each 3D voxel unit is defined as a cubic area with fixed spatial resolution Δx, Δy, Δz in the x, y, and z directions. The collection of all 3D voxel units together constitutes the entire 3D flight feasible domain model; Among them, Δx, Δy, Δz are the spatial resolutions of voxel division, N vox is the total number of voxels in the 3D flight feasible domain model, V i is the i-th three-dimensional voxel unit.

4. The data acquisition method based on long-distance environmental data acquisition by an unmanned aerial vehicle according to claim 3 is characterized in that: The S2 comprises the following steps: S21. In the three-dimensional flight feasible domain model FE 3D The flight state graph G with dynamic constraints is constructed based on the voxel unit division results. flight ,The nodes of the flight state graph are the voxel unit centers, and the connectivity between adjacent voxel units is dynamically adjusted based on the vegetation height difference, wind field disturbance coefficient, and energy consumption threshold; S22. In flight state diagram G flight Initialize the first ant population A route , the first ant population contains N a ant individuals, each ant individual corresponds to an initial drone flight path; S23. During each iteration, each ant in the first ant population selects the next pixel unit for extending the drone's flight path based on the improved 3D environment perception probability model. The improved 3D environment perception probability model is defined as: Among them, τ(V i ,V j ) represents node V i To node V j The path pheromone concentration, τ(V i ,V′ j ) represents node V i To candidate node V′ j The path pheromone concentration, η env (V i ,V j ) represents node V i To node V j The environmental adaptation heuristic factor is combined with the vegetation height difference, terrain complexity and wind field disturbance coefficient between the two nodes. For the task of extracting environmental information in the vegetation area, the flight path of the UAV with small wind disturbance and stable flight is preferred, η env (V i ,V′ j ) represents node V i To candidate node V′ j Environmental adaptation heuristic factor, η cov (V j ) represents node V j The area coverage heuristic factor is the higher the uncovered ratio of the voxel unit where the node is located, the larger its value is, which encourages the UAV to give priority to areas where vegetation information is not collected, η cov (V′ j ) represents the candidate node V′ j The area coverage heuristic factor, N(V i ) represents node V i The set of adjacent nodes that meet the flight constraints, α, β, and γ are the importance weights of pheromone, environmental adaptation heuristic factor, and area coverage heuristic factor, respectively; S24. Calculate the multi-objective fitness index of each UAV flight path, including the flight energy consumption index E k , Mission area coverage index C k and track safety index S k ; S25. After each round of iteration, the multi-objective fitness index (E k ,C k ,S k ) Perform non-dominated sorting and congestion evaluation to select the Pareto optimal path set S26. Based on the actual performance of each node in the selected Pareto optimal path set, reversely update the heuristic factors and pheromone concentrations in the improved three-dimensional environment perception probability model; S27. Repeat steps S23 to S26 for several rounds of iterations until convergence or the maximum number of iterations is reached, and finally output the Pareto track candidate set with dynamic perception enhancement 5. The data acquisition method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 4 is characterized in that: The flight energy consumption index E k The comprehensive energy consumption index represents the comprehensive energy consumption required by the UAV to complete the entire flight path. The comprehensive energy consumption index consists of two parts: the first is the basic flight energy consumption caused by the horizontal projection distance between adjacent flight points in the path, which is weighted by the horizontal flight energy consumption coefficient; the second is the climbing or descending energy consumption caused by the vertical height difference between adjacent flight points in the path. The climbing or descending energy consumption is weighted by the vertical flight energy consumption coefficient and a canopy density adjustment factor is introduced to reflect the energy consumption difference caused by the change in flight altitude under different vegetation canopy densities. The mission area coverage index C k Indicates the degree of coverage of the current flight path for the effective information in the vegetation area that has not been collected. The mission area coverage rate index is calculated by accumulating the effective vegetation information area of all voxel units in the path that has not been covered by the historical path and then calculating the ratio with the total area of the remaining uncollected vegetation information. It measures the actual contribution of the current path to improving the completion of the mission. The track safety index S k Represents the overall safety performance of the flight path. The track safety index is composed of the flight safety factors of all voxel units in the path. The flight safety factor of each voxel unit is the safety distance score between it and the obstacle set, obtained after normalization by the minimum flight safety distance and evaluated in combination with the obstacle distribution model. The track safety index reflects whether the path can continue to maintain in the safe flight channel and whether it is away from high-risk obstacle areas.

6. The data collection method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 4 is characterized in that: The S3 includes the following steps: S31. Multimodal vegetation observation data sample set D obtained based on ground-truth measurements and historical sampling veg Initialize the second ant population A feature , the second ant population contains N f ant individuals, each ant individual corresponds to a feature subset candidate solution; S32. Each ant individual is based on the multimodal vegetation observation data sample set D veg The feature subset selection is performed in the multi-scale feature space of the feature subset, and the feature subset selection probability model is defined as: Among them, f i The i-th candidate feature corresponds to a specific spectral, texture, structural or spatiotemporal feature component in the multimodal vegetation observation data sample set, P(f i ) represents the selection of candidate features f i The probability of being adopted by the current ant individual, τ f (f i ) represents the candidate feature f i The corresponding characteristic pheromone concentration is used to reflect the candidate feature f i The contribution degree to the high discriminant performance subset in the first few iterations. The higher the concentration, the better the performance of the feature in the historical collection. η div (f i ) represents the candidate feature f i The discriminant divergence heuristic factor is used to measure the degree of improvement in the ability to distinguish between vegetation categories after the introduction of this feature. The larger the value, the more effective the feature is in amplifying the differences between different categories. red (f i ) represents the candidate feature f i The redundancy suppression heuristic factor is used to measure the redundancy between the feature and the selected feature set. The larger the value, the less information overlap between the feature and the existing features, and the stronger the complementarity. f represents the characteristic pheromone concentration τ f (f i ) is used to control the influence of pheromone concentration on feature selection probability, β f Denotes the discriminant divergence heuristic factor η div (f i ), which is used to regulate the weight of the contribution of the feature to the class separability in the selection probability, γ f represents the redundancy suppression heuristic factor η red (f i ) is used to control the role of complementarity between features in feature selection probability; S33. Calculate the multi-objective fitness index for each feature subset candidate solution, including the discriminant divergence index J m , feature redundancy index R m and the computational cost index T m ; S34. After each round of iteration, the multi-objective fitness index (J m ,R m ,T m ) to perform non-dominated sorting and crowding evaluation, and select the Pareto optimal feature subset candidate set of multimodal vegetation observation data S35. Based on the actual performance of each feature in the Pareto optimal feature subset candidate set, reversely update each heuristic factor and feature pheromone concentration in the feature selection probability model; S36. Repeat steps S32 to S35 for several rounds of iterations until convergence or the maximum number of iterations is reached, and finally output the Pareto optimal feature subset candidate set with dynamic perception enhancement 7. The data collection method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 6, characterized in that: The discriminant divergence index J m It is used to measure the ability of feature subsets to distinguish different vegetation categories in a multimodal vegetation observation data sample set, and the discriminant divergence index J m It is defined as the ratio of the trace of the inter-class scatter matrix between vegetation categories to the trace of the intra-class scatter matrix within the category in the kernel Fisher discriminant analysis mapping space. The trace of the inter-class scatter matrix is used to measure the degree of dispersion of the center distribution between different categories, and the trace of the intra-class scatter matrix is used to measure the degree of aggregation of sample points within the same category. The feature redundancy index R m It is used to measure the degree of information duplication between all features within the feature subset, and the feature redundancy rate index R m It is defined as the average value of the correlation coefficients between all pairs of features in the feature subset, where the correlation coefficient of each pair of features is used to represent the linear similarity between the two in a statistical sense. The number of features contained in the current feature subset is recorded as M, and there are M(M-1) / 2 pairs of feature combinations whose correlations need to be calculated. The average of all correlation values is taken as the redundancy rate value; The calculation cost index T m It is used to measure the computational overhead required for inference of feature subsets in the subsequent kernel Fisher discriminant analysis model, and calculate the cost index T m It consists of two parts: the first part is the fixed basic inference time overhead t base , which represents the normal time required for model initialization and infrastructure calculation regardless of the feature subset selected; the second part is the linear growth term related to the number of features, indicating that each new feature will introduce a unit of computational overhead t unit , the number of features contained in the current feature subset is M, then the total computational cost index T m The product of the basic inference time plus the unit cost and the number of features, that is, t base With t unit ·M Joint decision.

8. The data collection method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 6 is characterized in that: The S5 comprises the following steps: S51. Receive, on the image processing unit of the UAV, a multimodal vegetation observation data stream collected in real time by the UAV, the data stream including hyperspectral image data, visible light image data, lidar point cloud data, and inertial navigation data; S52. Build a multi-scale fast feature extraction operator set for multimodal vegetation observation data streams, including a hyperspectral feature extraction operator, an image texture feature extraction operator, a point cloud structure feature extraction operator, and a navigation spatiotemporal feature extraction operator. This operator set performs parallel fast feature extraction for vegetation data of different modalities. S53. Use the hyperspectral feature extraction operator to extract vegetation band features from the hyperspectral image data and obtain the hyperspectral feature vector F used to characterize the spectral reflectance characteristics of vegetation. spec ,The different components in each hyperspectral feature vector represent the spectral reflectance intensity of different bands, reflecting the species information of the target vegetation area; S54. Use the image texture feature extraction operator to extract vegetation texture features from the visible light image data to obtain a texture feature vector F for describing the texture pattern of the vegetation area. tex ,The components of each texture feature vector represent the spatial texture uniformity and heterogeneity characteristics of the vegetation area, reflecting the differences in the community structure of the vegetation area; S55. Use the point cloud structure feature extraction operator to extract the three-dimensional structure features of the lidar point cloud data and obtain the three-dimensional structure feature vector F used to express the vegetation canopy structure. lidar ,The components of each three-dimensional structural characteristic vector represent the vegetation canopy height, canopy density and vertical spatial distribution characteristics, reflecting the three-dimensional structural differences of vegetation; S56. Use the navigation spatiotemporal feature extraction operator to extract the flight attitude and position information features of the inertial navigation data to obtain the navigation feature vector F used to characterize the relationship between the UAV's flight state and spatial position. nav ,The components of each navigation eigenvector represent the UAV’s attitude stability, flight trajectory and ,spatiotemporal information at the observation moment, reflecting the spatial and ,temporal context of vegetation data collection; S57. Obtain the multi-scale feature set F multi ={F spec ,F tex ,F lidar ,F nav All features in} are based on the Pareto optimal feature subset candidate set The feature retention rules determined in the above are used for rapid screening, and only the feature components belonging to the Pareto optimal feature subset candidate set are retained, and the unselected redundant feature components are removed to obtain the compressed feature data stream F comp .

9. The data collection method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 8, characterized in that: The S6 comprises the following steps: S61. compress the feature data stream F comp Input Kernel Fisher Discriminant Analysis Model KFDA veg ,The kernel Fisher discriminant analysis model is a classification and inversion model constructed based on the Pareto optimal feature subset candidate set. ,The kernel Fisher discriminant analysis model uses a nonlinear mapping method to map the original feature space to a high-dimensional ,feature space; S62. Automatically match the kernel function type and kernel function parameters used in the kernel Fisher discriminant analysis model based on the feature combination structure in the Pareto optimal feature subset candidate set; S63. Perform kernel mapping on the input compressed feature data stream to generate a high-dimensional kernel mapping feature expression, perform multi-class classification based on the established discriminant hyperplane, and output the vegetation type classification result of the current drone observation area; S64. The vegetation type classification results are spatially projected based on the UAV's geolocation information and the image's spatial distribution information to generate a continuous rasterized classification layer. Each pixel is assigned a vegetation type label to describe the spatial distribution pattern of complex vegetation areas. S65. Based on the output vegetation type classification layer, based on the nonlinear mapping relationship between the weight structure between category features implicit in the kernel Fisher discriminant analysis model and the input features, combined with the known ecological indicator label information in the ground samples, perform simultaneous inversion processing of ecological indicators in highly complex vegetation areas and output the ecological indicator inversion results; S66. The ecological index inversion results include the following ecological indicators: Vegetation coverage index: represents the vertical projection ratio of vegetation within the pixel range at each spatial position; Canopy density index: Estimates the canopy volume distribution density per unit area based on the structural characteristics of the lidar point cloud; Vegetation health index: The normalized vegetation index is calculated based on the ratio of red and near-infrared band reflectance in the hyperspectral characteristics; Community vertical structure index: The three-dimensional structure complexity of the community is derived by integrating the point cloud height distribution and spectral variation amplitude; Spatial heterogeneity index: reflects the characteristics of vegetation landscape patches based on the spatial continuity and texture variation of the same type of vegetation in the classification layer.

10. The data collection method based on long-distance collection of environmental data by an unmanned aerial vehicle according to claim 9, characterized in that: The kernel function types include radial basis function kernel, polynomial kernel or hybrid kernel, and the kernel function parameters include kernel width, polynomial order and hybrid kernel weight coefficient, which are specifically determined by the model fitness index corresponding to the feature subset and the online sample distribution pattern.

Images