A plant community investigation method and device based on multi-modal three-dimensional point cloud

Abstract

The present application belongs to the technical field of ecology and surveying and mapping, and more particularly relates to a plant community investigation method and device based on multi-modal three-dimensional point cloud. The method collects multi-source data through a four-in-one sensor pod carried by a UAV, combines adaptive quadrat division, wind-induced swing artifact removal, self-supervised graph neural network hierarchical segmentation, spectrum-structure coupled biomass calculation and diversity index analysis, and finally outputs a three-dimensional visual report. The present application solves the problems of low efficiency, strong subjectivity and great destructiveness of traditional investigation, and the problems of insufficient data precision, asynchronous multi-modal data and the like in existing three-dimensional scanning technology, and has the advantages of high efficiency, high precision, zero destruction, strong real-time performance and the like, and is suitable for plant community investigation in complex terrain areas.

Description

Technical Field

[0001] This invention belongs to the technical field of the intersection of ecology and surveying, and more specifically, relates to a method and apparatus for investigating plant communities based on multimodal three-dimensional point clouds. Background Technology

[0002] Plant communities are an important component of ecosystems, and plant community surveys are crucial for protecting biodiversity and maintaining ecological balance. Traditional plant community surveys are typically conducted through manual fieldwork. However, in complex terrain areas with high vegetation cover and limited accessibility, traditional fieldwork methods are insufficient. Furthermore, these methods are susceptible to subjective biases and low efficiency due to variations in the expertise of survey personnel, leading to inaccurate results and hindering subsequent research. Additionally, manual biomass surveys often require harvesting some plants and measuring their weight, a method with significant margins of error. If the sample plots contain rare or endangered protected plants, indiscriminate harvesting can damage the ecosystem. Therefore, a more efficient and accurate method for plant community surveys is urgently needed to address these issues.

[0003] 3D laser scanning technology is a high-speed technique for acquiring 3D data of a measured object. It utilizes the principle of laser ranging to rapidly record the surface features of the object, accurately and in real-time matching object texture and size to achieve high-precision 3D imaging. 3D point cloud-based plant community surveys can combine existing knowledge of plant communities with 3D point cloud data, utilizing massive amounts of 3D point cloud data to complete plant community surveys. Although existing research has used lidar for forest resource inventory, the following problems have not been solved:

[0004] a) Online removal of wind-induced sway artifacts;

[0005] b) Synchronous acquisition and processing of spatial-spectral-mechanical multimodal data under the same spatiotemporal reference;

[0006] c) Community-level species identification and online biomass calibration based on self-supervised graph neural networks.

[0007] Therefore, there is an urgent need for a new technical solution that integrates space-spectrum-machine multimodal sensors and features adaptive quadrat, wind disturbance removal, self-supervised learning, and real-time closed-loop optimization. Summary of the Invention

[0008] The present invention aims to overcome at least one of the defects of the prior art and provide a plant community survey method based on multimodal three-dimensional point clouds, so as to solve the problems of low efficiency, strong subjectivity, and high destructiveness of traditional surveys, as well as insufficient data accuracy, asynchronous multimodal data, and difficulty in species identification and biomass calibration in existing three-dimensional scanning technology.

[0009] The present invention also discloses an apparatus loaded with a plant community survey method.

[0010] The detailed technical solution of this invention is as follows:

[0011] A method for investigating plant communities based on multimodal 3D point clouds, the method comprising:

[0012] S1. Use a drone equipped with a four-in-one sensor pod to acquire multi-source data of the target sample plot. The pod includes: a 32-line lidar, a five-band multispectral camera, an ultrasonic wind probe, and an edge computing unit.

[0013] S2. Calculate the adaptive quadrat area S in real time based on the landscape fragmentation index F and the terrain slope T, and... The current quadrat is automatically divided into subquads of fixed size, with a 5m overlap band between each subquad for subsequent point cloud stitching, and the center coordinates of each subquad are pushed to the UAV flight mission.

[0014] S3. During the flight of the UAV, the wind disturbance matrix W(t) output in real time by the ultrasonic wind probe is combined with the IMU attitude angle to construct the "prediction-measurement" residual threshold. When the magnitude and direction of the residual threshold reach a preset threshold, it is judged as a wobbling artifact, and the original point cloud is removed by wobbling artifact removal.

[0015] S4. Input the point cloud after removing sway artifacts into the Self-Supervised Graph Neural Network (Self-GNN-LC) and perform hierarchical-instance segmentation: First, use the upper layer point cloud (greater than 5m) for tree segmentation; then, use multispectral NDVI and laser echo density to perform shrub / herb segmentation on the lower layer point cloud (less than or equal to 5m), obtaining independent tree instances and shrub / herb patch instances; finally, accurately connect the upper layer trees and the lower layer vegetation through trunk-ground geometric consistency to complete cross-layer matching.

[0016] S5. Extract the diameter at breast height (DBH), tree height (H), crown diameter, and crown area from the point cloud of each tree in the example, and substitute them into the spectral-structural coupling biomass equation to calculate the biomass;

[0017] S6. Computational Point Cloud Native Diversity Index: Spectral-Shannon Index Structural-Simpson index The Phylo-Point rarity index (P-Rarity) was calculated, and a community diversity statistics table was generated.

[0018] S7. Transmit the results of steps S4-S6 back to the ground station in real time to complete the output of the 3D WebGL visualization report.

[0019] According to a preferred embodiment of the present invention, the landscape fragmentation index F in S2 refers to the Shannon landscape diversity index based on 0.1m GSD orthophotos; the terrain slope T refers to the average slope calculated based on a 5m resolution digital surface model (DSM); and the formula for the adaptive quadrat area S is as follows:

[0020] (1).

[0021] According to a preferred embodiment of the present invention, the wind disturbance matrix W(t) is formulated as follows:

[0022] (2)

[0023] In equation (2), , , The Euler angles are obtained by complementary filtering of the triaxial instantaneous angular velocities measured by the ultrasonic wind probe. 、 、 These are rotation matrices for rotations about the z-axis, y-axis, and x-axis, respectively.

[0024] The residual threshold The formula is as follows:

[0025] (3)

[0026] In equation (3), Represents the three-dimensional coordinates of the measured points of the lidar. , This represents the predicted coordinates based on the static model of the previous frame.

[0027] According to a preferred embodiment of the present invention, the Self-Supervised Graph Neural Network (Self-GNN-LC) uses EdgeConv as its backbone network, superimposed with a spectral-wind disturbance attention gating module and a cross-layer consistency regularization module.

[0028] Its nodal characteristics are: x, y, z, intensity, echo count, 5-band reflectivity, and wind disturbance residual. ; 10 dimensions in total;

[0029] Edges are constructed as follows: k=20 nearest neighbors, Euclidean distance. ; border rights as follows:

[0030] (4)

[0031] In equation (4), and These are the spectral reflectances of the i-th and j-th points in the point cloud data, respectively, acquired from data collected by a five-band multispectral camera; and These are the wind disturbance residuals of the i-th point and the j-th point in the point cloud data, respectively. , , Indicates the weighting coefficient;

[0032] The loss function used is as follows:

[0033] (5)

[0034] In equation (5), For InfoNCE contrastive learning loss, For wind disturbance consistency loss, For biomass uniformity loss with measured diameter at breast height (DBH) / tree height as a weak label, and These are the weighting coefficients. Optimized using Bayesian methods.

[0035] According to a preferred embodiment of the present invention, the segmentation process of the upper-layer tree in S4 is as follows:

[0036] Region cropping: Point cloud elevation based on the Ground Digital Elevation Model (DTM) Cut out a 5m section, denoted as ;

[0037] Point cloud downsampling: 0.1m voxel grid filtering was used to retain... ;

[0038] Self-GNN-LC Inference: Input Output: semantic tags ; instance embedding ;

[0039] Tree trunk detection: Construct a RANSAC cylinder with a radius of 0.15m at a height of 1.3m to extract candidate tree trunks; then, interpolate the cylinder's center coordinates with... Nearest neighbor matching, preserving confidence. The tree trunk;

[0040] Crown instantiation: using the center of each trunk as a seed, based on... Perform density clustering to generate individual tree instances;

[0041] The process for dividing lower-level shrubs / herbs in S4 is as follows:

[0042] Region clipping: This modifies the elevation data. A 5m point cloud is denoted as ;

[0043] Normalized Difference Vegetation Index (NDVI) Calculation: For For each point, the reflectivity is calculated using the five-band wavelengths:

[0044] (6)

[0045] In equation (6), Indicates the spectral reflectance in the near-infrared band. Indicates the spectral reflectance in the red band;

[0046] Laser echo density calculation: Statistical point density within a cubic voxel with a side length of 0.2m. ,in, This refers to the number of lidar echo points counted within a cube voxel with a side length of 0.2m. This represents the volume of a cube voxel with a side length of 0.2m;

[0047] Joint feature construction: NDVI appended to node features 12 dimensions in total;

[0048] Clustering and segmentation: HDBSCAN is used for clustering of 12-dimensional features; semantic labels. Output shrub / herb patch instances;

[0049] The cross-layer matching process in S4, achieved through trunk-ground geometric consistency, is as follows:

[0050] Trunk center projection: Project the center of each tree trunk obtained earlier onto the ground plane to obtain planar coordinates. ;

[0051] Lower patch projection: Project the lowest point of each shrub / herb patch obtained above onto the ground plane to obtain... ;

[0052] Nearest neighbor matching: if Then, a cross-level index is created; the cross-level consistent table is output. This is used for subsequent community-level statistics; among which:

[0053] TreeID: refers to the unique identifier assigned to each tree after it has been divided into individual trees;

[0054] UnderstoryID: refers to the unique identifier assigned to each patch after the lower shrub / herb patch is segmented.

[0055] According to a preferred embodiment of the present invention, the spectral-structural coupled biomass equation is as follows:

[0056] (7)

[0057] In equation (7), PointDensity is the point density within the tree canopy; and 1 represents the average reflectance; 'a' represents the comprehensive scale factor, which is related to species and region; 'b' represents the diameter at breast height-tree height-biomass index; 'c' represents the Red-Edge / NIR spectral sensitivity index, which is the weight of the absorption / reflectance difference on biomass; 'd' represents the point density index, which reflects the contribution of the complexity of the canopy internal structure. 'a', 'b', 'c', and 'd' are updated and optimized online using a Bayesian NUTS sampler, updated every 30 seconds.

[0058] According to a preferred embodiment of the present invention, the Spectral-Shannon index The formula is as follows:

[0059] (8)

[0060] In equation (8), This indicates that the probability distribution of each point cloud is obtained by performing a k-means clustering algorithm on 5 bands, where k=3;

[0061] The Structural-Simpson index The 5-dimensional structural features were calculated by voxelization at 0.2m, and the bulldozer distance EMD between community voxel distributions was calculated and normalized to [0,1].

[0062] The Phylo-Point rarity index, P-Rarity:

[0063] Single-wood grade: (9)

[0064] In equation (9), Let i be the total number of point clouds for the i-th tree. The Euclidean distance between the 128-dimensional CNN embedding vector of this plant's point cloud and the average embedding vector of the same family;

[0065] Community level: (10)

[0066] In equation (10), N is the number of identified species in the community.

[0067] In another aspect of the present invention, an apparatus for implementing the above-described plant community survey method is provided, the apparatus comprising:

[0068] Unmanned aerial vehicle (UAV) platforms are used to schedule UAV missions.

[0069] The four-in-one sensor pod uses carbon fiber 3K twisted tubing. 7075-T6 aviation aluminum CNC one-piece molding, weight 1.2 kg, can be mounted on a general-purpose UAV platform, including a 32-line lidar, a five-band multispectral camera, and an ultrasonic wind probe, connected to the UAV gimbal through an M12 quick-release interface, supporting roll angle 45°, pitch angle Continuous rotation;

[0070] An edge computing unit for processing and analyzing the collected point cloud data and finally outputting the calculated point cloud native diversity index.

[0071] In another aspect of the present invention, an electronic device is further provided, including:

[0072] At least one processor; and

[0073] A memory, the memory stores instructions, when the instructions are executed by the at least one processor, the at least one processor is caused to execute the plant community survey method as described above.

[0074] In another aspect of the present invention, a machine-readable storage medium is further provided, which stores executable instructions, and when the instructions are executed, the machine is caused to execute the plant community survey method as described above.

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] (1) The present invention realizes large-scale and rapid data collection through a UAV carrying multi-modal sensors, and adaptive quadrat division and real-time data processing reduce manual intervention and improve the survey efficiency.

[0077] (2) The present invention does not require harvesting plants, and realizes non-destructive biomass measurement through a spectral-structure coupled biomass equation, protects rare species and the ecological environment, and achieves zero damage.

[0078] (3) The present invention eliminates wind disturbance artifacts through residual threshold control, and then realizes accurate identification of "trees-shrubs-grasses" based on hierarchical-instance segmentation of self-supervised graph neural networks. Self-supervised learning reduces the dependence on labeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 Is a flowchart of the plant community survey method described in the present invention. DETAILED DESCRIPTION

[0080] The following further describes the present disclosure in conjunction with the drawings and embodiments.

[0081] Example 1,

[0082] Refer Figure 1 , this embodiment provides a plant community survey method based on multi-modal three-dimensional point clouds, and the method includes:

[0083] S1. Use a drone equipped with a four-in-one sensor pod to acquire multi-source data of the target sample plot. The pod includes: a 32-line lidar, a five-band multispectral camera, an ultrasonic wind probe, and an edge computing unit.

[0084] The pod uses the IEEE-1588 PTP precision clock protocol, and is unified in time synchronization with a 32-line lidar, a five-band multispectral camera, an IMU, and an ultrasonic wind probe, ensuring high time synchronization accuracy. 1μs; Meanwhile, a "pulse-timestamp" dual buffering mechanism is implemented inside the FPGA (Xilinx Zynq Ultrascale+) to ensure that the point cloud coordinates and spectral reflectance are aligned at the nanosecond level;

[0085] The 32-line lidar uses a Riegl miniVUX-3UAV, 1000kHz PRR. The collected raw data is transmitted to the edge computing unit via gigabit Ethernet;

[0086] The five-band multispectral camera is a MicaSense RedEdge-MX, with 5 bands (Blue 475nm, Green 560nm, Red 668nm, Red Edge 717nm, NIR 840nm), global shutter, and is coaxially mounted with the lidar, achieving a high field-of-view overlap rate. 95%;

[0087] The ultrasonic wind probe uses an FT742-SM ultrasonic anemometer, which outputs three-dimensional wind speeds u, v, w at 50Hz.

[0088] The edge computing unit uses an NVIDIA Jetson Xavier NX processor with 8GB of LPDDR4x memory and runs Ubuntu 20.04 + CUDA 11.4. It is responsible for running the Self-Supervised Graph Neural Network (Self-GNN) inference in real time.

[0089] S2. Calculate the adaptive quadrat area S in real time based on the landscape fragmentation index F and the terrain slope T, and... The current quadrat is automatically divided into subquads of fixed size, with a 5m overlap band between each subquad for subsequent point cloud stitching, and the center coordinates of each subquad are pushed to the UAV flight mission.

[0090] The landscape fragmentation index F refers to: using a drone equipped with a five-band multispectral camera at an altitude of 80-120m, capturing visible light-near infrared images with 80% heading and 60% lateral overlap; using RTK / PPK positioning to output orthophotos at 0.1m GSD, acquiring 0.1m GSD orthophotos as raw data, and then using the landscape pattern analysis software Fragstats 4.2 to calculate the Shannon Landscape Diversity Index (SHDI). The Shannon Landscape Diversity Index is an indicator that measures the richness and evenness of patch types in a landscape. Its value reflects the degree of landscape fragmentation; a higher value usually means a higher degree of landscape fragmentation.

[0091] The terrain slope T is calculated as the average slope of the 5m resolution digital surface model (DSM) collected by the 32-line lidar during the pre-flight process.

[0092] The formula for the adaptive quadrat area S is as follows:

[0093] (1).

[0094] S3. During the flight of the UAV, the wind disturbance matrix W(t) output in real time by the wind sensor is combined with the attitude angle of the IMU to construct the "prediction-measurement" residual threshold. ,when And the angle between this direction and the instantaneous wind speed If the artifact is found to be wobbly, then the original point cloud is removed.

[0095] The formula for the wind disturbance matrix W(t) is as follows:

[0096] (2)

[0097] In equation (2), , , The Euler angles are obtained by complementary filtering of the triaxial instantaneous angular velocities measured by the ultrasonic wind probe. 、 、 These are rotation matrices for rotations about the z-axis, y-axis, and x-axis, respectively.

[0098] The formula for the residual threshold δ is as follows:

[0099] (3)

[0100] In equation (3), Represents the three-dimensional coordinates of the measured points of the lidar. , Represents the predicted coordinates based on the static model of the previous frame;

[0101] The FPGA implementation of the wobble artifact removal: The residual threshold is calculated in parallel at a 100 MHz clock on the Zynq PL terminal. ,Delay By using AXI4-Stream to output the culling marker bits and point cloud in the same frame, subsequent CPU resources are saved by 38%.

[0102] S4. Input the point cloud after removing wobble artifacts into the Self-Supervised Graph Neural Network (Self-GNN) and perform hierarchical-instance segmentation: First, ... The 5m upper-layer point cloud was used for single-tree segmentation; then... The 5m lower-layer point cloud was used to perform shrub / herb segmentation by joint clustering of multispectral NDVI and laser echo density, resulting in independent tree instances and shrub / herb patch instances. Finally, through trunk-ground geometric consistency, the upper-layer trees and lower-layer vegetation were accurately linked to complete cross-layer matching. This step can solve the problem of missed detection caused by the vertical intersection of "tree-shrub-grass", and achieve accurate segmentation in two dimensions: "single tree level" + "species functional layer", providing traceable instance objects for subsequent calculation of diameter at breast height, tree height, crown width, biomass and biodiversity index.

[0103] a. Self-GNN-LC network structure

[0104] (1) Overall network framework:

[0105] Basic backbone: Based on EdgeConv as the basic backbone network, a spectral-wind disturbance attention gating module and a cross-layer consistency regularization module are superimposed to form the Self-GNN-LC (Layer Consistency) unique to this patent.

[0106] Node characteristics: x, y, z, intensity, echo count, 5-band reflectivity, wind disturbance residual ; 10 dimensions in total.

[0107] Edge construction: k=20 nearest neighbors, Euclidean distance ; border rights as follows:

[0108] (4)

[0109] In equation (4), and These are the spectral reflectances of the i-th and j-th points in the point cloud data, respectively, acquired from data collected by a five-band multispectral camera; and These are the wind disturbance residuals of the i-th point and the j-th point in the point cloud data, respectively. , , Indicates the weighting coefficient;

[0110] Encoder: 4-layer EdgeConv, each layer outputs a 64-dimensional embedding;

[0111] Dual-head decoding:

[0112] Semantic header: 5 categories of Softmax, namely canopy / subcrown / shrub / herb / ground.

[0113] Example Header: Comparison with the learning head, using InfoNCE loss, temperature .

[0114] b. Loss Function

[0115] (5)

[0116] In equation (5), For InfoNCE contrastive learning loss, For wind disturbance consistency loss, For biomass uniformity loss with measured diameter at breast height (DBH) / tree height as a weak label, and These are the weighting coefficients. Optimized using Bayesian methods.

[0117] c. Hierarchical reasoning

[0118] (1) The process of dividing individual trees in the upper layer is as follows:

[0119] Region cropping: Point cloud elevation based on the Ground Digital Elevation Model (DTM) Cut out a 5m section, denoted as .

[0120] Point cloud downsampling: 0.1m voxel grid filtering was used to retain... .

[0121] Self-GNN-LC Inference: Input Output: semantic tags ; instance embedding .

[0122] Tree trunk detection: Construct a RANSAC cylinder with a radius of 0.15m at a height of 1.3m to extract candidate tree trunks; then, interpolate the cylinder's center coordinates with... Nearest neighbor matching, preserving confidence. The tree trunk.

[0123] Crown instantiation: using the center of each trunk as a seed, based on... Perform density clustering (DBSCAN ε=0.2m, MinPts=50) to form individual tree instances. Each individual tree instance includes: ID, point cloud, diameter at breast height (DBH), tree height (H), crown width, and location of each tree.

[0124] (2) The process for dividing the lower shrubs and herbs is as follows:

[0125] Region clipping: This modifies the elevation data. Point cloud record .

[0126] Normalized Difference Vegetation Index (NDVI) Calculation: For For each point, the reflectivity is calculated using the five-band wavelengths:

[0127] (6)

[0128] In equation (6), Indicates the spectral reflectance in the near-infrared band. The values ​​represent the spectral reflectance in the red band, both of which were collected from a five-band multispectral camera.

[0129] Laser echo density calculation: statistical point density within a 0.2m voxel. ,in, This refers to the number of lidar echo points counted within a cube voxel with a side length of 0.2m. This represents the volume of a cube voxel with a side length of 0.2m;

[0130] Joint feature construction: NDVI appended to node features There are 12 dimensions in total.

[0131] Clustering segmentation: HDBSCAN ( , Clustering of 12-dimensional features; semantic labels Output shrub and herb patch instances, which include: ID, point cloud, mean NDVI, and coverage of each shrub and herb patch.

[0132] (3) The process of completing cross-layer matching through trunk-ground geometric consistency is as follows:

[0133] Trunk center projection: Project the center of each tree trunk obtained earlier onto the ground plane to obtain planar coordinates. .

[0134] Lower patch projection: Project the lowest point of each shrub / herb patch obtained above onto the ground plane to obtain... .

[0135] Nearest neighbor matching: if Then, a cross-level index is created; the cross-level consistent table is output. This is used for subsequent community-level statistics; among which:

[0136] TreeID: refers to the unique identifier assigned to each tree after it has been divided into individual trees. It is used to distinguish different individual trees and facilitates the separate recording, analysis and management of parameters such as diameter at breast height, tree height, and crown width of trees. For example, when counting the number of trees in a plot or the biomass of each tree, TreeID can be used to accurately link to the data of the corresponding individual tree.

[0137] UnderstoryID: refers to the unique identifier assigned to each understory shrub / herb patch after segmentation, used to distinguish different understory vegetation patches, and can record parameters such as mean NDVI and coverage.

[0138] S5. Extract the diameter at breast height (DBH), tree height (H), crown diameter, and crown area from the point cloud data for each tree in the example, and substitute them into the spectral-structural coupling biomass equation:

[0139] (7)

[0140] In equation (7), PointDensity is the point density within the canopy ( ); and 1 represents the average reflectance; 'a' represents the comprehensive scale factor, which is related to species and region; 'b' represents the diameter at breast height-tree height-biomass index, which is theoretically close to 1; 'c' represents the Red-Edge / NIR spectral sensitivity index, which is the weight of the absorption / reflectance difference on biomass; 'd' represents the point density index, which reflects the contribution of the complexity of the canopy internal structure. 'a', 'b', 'c', and 'd' are updated and optimized online using a Bayesian NUTS sampler, updated every 30 seconds.

[0141] S6. Computational Point Cloud Native Diversity Index: Spectral-Shannon Index Structural-Simpson index The Phylo-Point rarity index (P-Rarity) was calculated, and a community diversity statistics table was generated.

[0142] The Spectral-Shannon index The formula is as follows:

[0143] (8)

[0144] In equation (8), This indicates that the probability distribution of each point cloud is obtained by performing a k-means clustering algorithm on 5 bands, where k=3;

[0145] The Structural-Simpson index The 5-dimensional structural features were calculated by voxelization at 0.2m, and the bulldozer distance EMD between community voxel distributions was calculated and normalized to [0,1].

[0146] The Phylo-Point rarity index, P-Rarity:

[0147] Single-wood grade: (9)

[0148] In equation (9), Let i be the total number of point clouds for the i-th tree. The Euclidean distance between the 128-dimensional CNN embedding vector of this plant's point cloud and the average embedding vector of the same family;

[0149] Community level: (10)

[0150] In equation (10), N is the number of identified species in the community.

[0151] S7. Transmit the results of steps S4-S6 back to the ground station in real time and complete the output of the 3D WebGL visualization report within 10 minutes.

[0152] Example 2

[0153] This embodiment provides an apparatus for implementing the above-mentioned plant community survey method. The apparatus includes: an unmanned aerial vehicle platform, a four-in-one sensor pod, and an edge computing unit.

[0154] The drone platform is used to implement drone mission scheduling;

[0155] The edge computing unit uses an NVIDIA Jetson Xavier NX processor with 8GB of LPDDR4x memory and runs Ubuntu 20.04 + CUDA 11.4. It is responsible for running the Self-Supervised Graph Neural Network (Self-GNN) inference in real time, processing and analyzing the collected point cloud data, and finally outputting the calculation of the point cloud native diversity index.

[0156] The four-in-one sensor pod is made of carbon fiber 3K twisted tubing. 7075-T6 aviation aluminum CNC one-piece molding, weight Weighing 1.2kg, it can be mounted on general-purpose drone platforms such as the DJI M300 RTK and Autel Dragonfish.

[0157] All sensors connect to the drone gimbal via an M12 quick-release connector, allowing for roll angle adjustments. 45°, pitch angle Continuous rotation meets the requirements for multi-angle scanning;

[0158] The four-in-one sensor pod adopts the IEEE-1588 PTP precision clock protocol, providing unified time synchronization for the 32-line lidar, five-band multispectral camera, IMU, and ultrasonic wind probe, ensuring high time synchronization accuracy. 1μs; at the same time, a "pulse-timestamp" dual buffering mechanism is implemented inside the FPGA (Xilinx Zynq Ultrascale+) to ensure that the point cloud coordinates and spectral reflectance are aligned at the nanosecond level.

[0159] Example 3

[0160] This embodiment also provides an electronic device, including:

[0161] At least one processor; and

[0162] A memory that stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the plant community survey method as described above.

[0163] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable computing device, consumer electronic device, etc.

[0164] Example 4

[0165] This embodiment also provides a machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the plant community survey method described above.

[0166] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0167] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0168] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0169] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for investigating plant communities based on multimodal 3D point clouds, characterized in that, The method includes: S1. Use a drone equipped with a four-in-one sensor pod to acquire multi-source data of the target sample plot. The pod includes: a 32-line lidar, a five-band multispectral camera, an ultrasonic wind probe, and an edge computing unit. S2. Calculate the adaptive quadrat area S in real time based on the landscape fragmentation index F and the terrain slope T, and then... 400 The current quadrat is automatically divided into subquads of fixed size, with a 5m overlap band between each subquad for subsequent point cloud stitching, and the center coordinates of each subquad are pushed to the UAV flight mission. S3. During the flight of the UAV, the residual threshold is constructed by combining the wind disturbance matrix W(t) output in real time by the ultrasonic wind probe with the IMU attitude angle. When the magnitude and direction of the residual threshold reach a preset threshold, it is judged as a wobbling artifact, and the original point cloud is removed by wobbling artifact removal. The formula for the wind disturbance matrix W(t) is as follows: （1） In equation (1), , , The Euler angles are obtained by complementary filtering of the triaxial instantaneous angular velocities measured by the ultrasonic wind probe. 、 、 These are rotation matrices for rotations about the z-axis, y-axis, and x-axis, respectively. The residual threshold The formula is as follows: （2） In equation (2), Represents the three-dimensional coordinates of the measured points of the lidar. , Represents the predicted coordinates based on the static model of the previous frame; S4. Input the point cloud after removing sway artifacts into the Self-Supervised Graph Neural Network (Self-GNN-LC) and perform hierarchical-instance segmentation: First, use the upper layer point cloud (greater than 5m) for tree segmentation; then, use multispectral NDVI and laser echo density to perform shrub / herb segmentation on the lower layer point cloud (less than or equal to 5m), obtaining independent tree instances and shrub / herb patch instances; finally, accurately connect the upper layer trees and the lower layer vegetation through trunk-ground geometric consistency to complete cross-layer matching. S5. Extract the diameter at breast height (DBH), tree height (H), crown diameter, and crown area from the point cloud of each tree in the example, and substitute them into the spectral-structural coupling biomass equation to calculate the biomass; S6. Computational Point Cloud Native Diversity Index: Spectral-Shannon Index Structural-Simpson index The Phylo-Point rarity index (P-Rarity) was calculated, and a community diversity statistics table was generated. S7. Transmit the results of steps S4-S6 back to the ground station in real time to complete the output of the 3D WebGL visualization report.

2. The method for investigating plant communities based on multimodal three-dimensional point clouds according to claim 1, characterized in that, The landscape fragmentation index F mentioned in S2 refers to the Shannon landscape diversity index based on 0.1m GSD orthophotos; the terrain slope T refers to the average slope calculated based on the 5m resolution digital surface model (DSM); the formula for the adaptive quadrat area S is as follows: （3）。 3. The method for investigating plant communities based on multimodal three-dimensional point clouds according to claim 1, characterized in that, The Self-Supervised Graph Neural Network (Self-GNN-LC) uses EdgeConv as its backbone network and superimposes a spectral-wind disturbance attention gating module and a cross-layer consistency regularization module. Its node characteristics are: , , Intensity, number of echoes, 5-band reflectivity, wind disturbance residual ; 10 dimensions in total; Edges are constructed as follows: k=20 nearest neighbors, Euclidean distance. ; border rights as follows: （4） In equation (4), and These are the spectral reflectances of the i-th and j-th points in the point cloud data, respectively, acquired from data collected by a five-band multispectral camera; and These are the wind disturbance residuals of the i-th point and the j-th point in the point cloud data, respectively. , , Indicates the weighting coefficient; The loss function used is as follows: （5） In equation (5), For InfoNCE contrastive learning loss, For wind disturbance consistency loss, For biomass uniformity loss with measured diameter at breast height (DBH) / tree height as a weak label, and These are the weighting coefficients. Optimized using Bayesian methods.

4. The method for investigating plant communities based on multimodal three-dimensional point clouds according to claim 3, characterized in that, The process for dividing individual trees in the upper layer of S4 is as follows: Region cropping: Point cloud elevation based on the Ground Digital Elevation Model (DTM) Cut out a 5m section, denoted as ; Point cloud downsampling: 0.1m voxel grid filtering was used to retain... ; Self-GNN-LC Inference: Input Output: semantic tags ; instance embedding ; Tree trunk detection: Construct a RANSAC cylinder with a radius of 0.15m at a height of 1.3m to extract candidate tree trunks; then, interpolate the cylinder's center coordinates with... Nearest neighbor matching, preserving confidence. The tree trunk; Crown instantiation: using the center of each trunk as a seed, based on... Perform density clustering to generate individual tree instances; The process for dividing lower-level shrubs / herbs in S4 is as follows: Region clipping: This modifies the elevation data. A point cloud of 5m is denoted as ; Normalized Difference Vegetation Index (NDVI) Calculation: For For each point, the reflectivity is calculated using the five-band wavelengths: （6） In equation (6), Indicates the spectral reflectance in the near-infrared band. Indicates the spectral reflectance in the red band; Laser echo density calculation: Statistical point density within a cubic voxel with a side length of 0.2m. ,in, This refers to the number of lidar echo points counted within a cube voxel with a side length of 0.2m. This represents the volume of a cube voxel with a side length of 0.2m; Joint feature construction: NDVI appended to node features 12 dimensions in total; Clustering and segmentation: HDBSCAN is used for clustering of 12-dimensional features; semantic labels. Output shrub / herb patch instances; The cross-layer matching process in S4, achieved through trunk-ground geometric consistency, is as follows: Trunk center projection: Project the center of each tree trunk obtained earlier onto the ground plane to obtain planar coordinates. ; Lower patch projection: Project the lowest point of each shrub / herb patch obtained above onto the ground plane to obtain... ; Nearest neighbor matching: if Then, a cross-level index is created; the cross-level consistent table is output. This is used for subsequent community-level statistics; among which: TreeID: refers to the unique identifier assigned to each tree after it has been divided into individual trees; UnderstoryID: refers to the unique identifier assigned to each patch after the lower shrub / herb patch is segmented.

5. The method for investigating plant communities based on multimodal three-dimensional point clouds according to claim 4, characterized in that, The spectral-structure coupled biomass equation is as follows: （7） In equation (7), PointDensity is the point density within the tree canopy; and 1 represents the average reflectance; 'a' represents the comprehensive scale factor, which is related to species and region; 'b' represents the diameter at breast height-tree height-biomass index; 'c' represents the Red-Edge / NIR spectral sensitivity index, which is the weight of the absorption / reflectance difference on biomass; 'd' represents the point density index, which reflects the contribution of the complexity of the canopy internal structure. 'a', 'b', 'c', and 'd' are updated and optimized online using a Bayesian NUTS sampler, updated every 30 seconds.

6. The method for investigating plant communities based on multimodal three-dimensional point clouds according to claim 5, characterized in that, The Spectral-Shannon index The formula is as follows: （8） In equation (8), This indicates that the probability distribution of each point cloud is obtained by performing a k-means clustering algorithm on five bands, where k=3; The Structural-Simpson index The 5-dimensional structural features were calculated by voxelization at 0.2m, and the bulldozer distance EMD between community voxel distributions was calculated and normalized to [0,1]. The Phylo-Point rarity index, P-Rarity: Single-wood grade: (9) In equation (9), Let i be the total number of point clouds for the i-th tree. The Euclidean distance between the 128-dimensional CNN embedding vector of this plant's point cloud and the average embedding vector of the same family; Community level: (10) In equation (10), N is the number of identified species in the community.

7. An apparatus for implementing the plant community survey method as described in any one of claims 1-6, characterized in that, The device includes: Unmanned aerial vehicle (UAV) platforms are used to schedule UAV missions. The four-in-one sensor pod uses carbon fiber 3K twisted tubing. 7075-T6 aviation aluminum CNC one-piece molding, weight Weighing 1.2kg, it can be mounted on general-purpose drone platforms and includes a 32-line LiDAR, a five-band multispectral camera, and an ultrasonic wind sensor. It connects to the drone gimbal via an M12 quick-release interface and supports roll angle... 45°, pitch angle Continuous rotation; Edge computing units are used to process and analyze the collected point cloud data and ultimately output the native diversity index of the point cloud.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the plant community survey method as described in any one of claims 1 to 6.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores executable instructions that, when executed, cause the machine to perform the plant community survey method as described in any one of claims 1 to 6.

Images