Variable rate application method and system based on target leaf area detection

Abstract

This invention discloses a variable-rate pesticide application method and system based on target leaf area detection, belonging to the technical field of smart agriculture based on computational models. The method includes establishing and training a tree semantic segmentation model; collecting tree point cloud data according to time frames; classifying and labeling the point cloud information within each time frame; dividing the segmented point cloud information into a grid based on the time frame and calculating the predicted leaf area; correcting the predicted leaf area to obtain the corresponding true leaf area; constructing a variable-rate spraying model to calculate the pesticide application rate for the corresponding tree area based on the leaf area density; creating a spraying airflow model to determine the outlet wind speed of the pesticide application device; and performing regional pesticide application based on the pesticide application rate and outlet wind speed. This invention accurately calculates the leaf area of ​​each region of the fruit tree canopy, making it more accurate and scientific than existing technologies.

Description

Technical Field

[0001] This invention belongs to the technical field of smart agriculture based on computational models, specifically, it relates to a variable application method and system based on target leaf area detection. Background Technology

[0002] Effective control of orchard pests and diseases can recover nearly 10% of production losses. Currently, the main method of pest and disease control is still chemical control. The commonly used continuous spraying method not only causes serious environmental pollution but also increases the cost of pesticide input for orchard pest and disease control. To improve the pesticide utilization rate of orchard plant protection operations, some scholars have proposed variable application technology based on target detection. Among them, lidar sensors have been widely used in the field of precision application because they can simultaneously obtain information on target location and growth (canopy volume, leaf area, leaf density, etc.) through data processing.

[0003] For example, invention patent CN116530489A discloses a variable spray control system based on lidar detection of tree canopy leaf area. This solution, based on traditional orchard spraying systems, adds hardware centered on lidar, an industrial control computer, a lower-level computer, and other sensors, thereby improving pesticide utilization and reducing pesticide waste during orchard plant protection operations. Invention patent CN116439218A discloses a real-time variable spraying device and method based on lidar detection of crop characteristics. This solution uses a lidar scanning detection device and a variable spraying device, which can adjust the spraying amount according to the characteristics of crops in the field to improve pesticide utilization and reduce pesticide waste during application.

[0004] However, the above solutions still have the following problems: 1. All the above solutions use the most primitive lidar scanning data and use its point cloud data as the tree leaf area. However, the point cloud data scanned by lidar includes information such as tree leaf surface, tree branches, and tree support. It is unscientific to use the point cloud data of lidar without processing as the tree leaf area. The value is inaccurate. The variable application decision model built from this data cannot maximize the reduction of pesticide use and the improvement of effective utilization rate; 2. Lidar can only scan the area of ​​the outermost layer of leaves in the tree crown. Due to the occlusion between the leaves in the crown, it is impossible to accurately obtain the area of ​​the inner leaves, thus affecting the accuracy of pesticide calculation and the control effect; 3. During the process of lidar collecting point cloud data, the tree crown will inevitably vibrate due to wind and other reasons. Therefore, there are discrete points in the point cloud data. The method of directly using the original point cloud data to estimate the tree leaf area in the above solutions has too large an error. Therefore, the application model built on this basis has not fully explored the potential of pesticide reduction and efficacy preservation. Summary of the Invention

[0005] To address the problem that the accuracy of laser-guided spraying for measuring the density of tree canopy leaves is low due to foliage obstruction, resulting in large errors in pesticide dosage calculation and thus affecting the effectiveness of pest and disease control, this invention provides a variable-rate pesticide application method and system based on target leaf area detection.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows: a variable-rate application method based on target leaf area detection, characterized by comprising the following steps:

[0007] Obtain tree sample datasets at different growth stages, and use the tree sample datasets to train a tree semantic segmentation model;

[0008] The spraying device travels between the rows of trees at a predetermined speed and collects tree point cloud data according to time frames. The point cloud information in the tree point cloud data of each time frame is segmented into main root, leaves, branches and tree support using a tree semantic segmentation model.

[0009] Based on time frames, the segmented point cloud information is divided into grids, and the predicted leaf area of ​​each grid cell is calculated. For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. ;

[0010] The grid is divided into different tree regions according to the tree regions. Leaf area density ;

[0011] Construct a variable spraying model based on the leaf area density. The corresponding tree area was calculated. Dosage ;

[0012] Create a spray airflow model based on the current wind speed. Leaf area density and penetration distance x Determine the outlet wind speed of the pesticide application device ;

[0013] Based on the dosage and outlet wind speed Tree area Apply pesticides in designated areas.

[0014] In a further embodiment, the tree sample dataset is divided into a winter tree sample set and a non-winter tree sample set according to the season; wherein, the winter tree sample set is a tree point cloud dataset containing only branches, and the non-winter tree sample set is a tree point cloud dataset containing branches and leaves.

[0015] The training process of the tree semantic segmentation model is as follows: the tree sample dataset is denoised using a point cloud filtering algorithm to obtain a denoised tree sample dataset; the denoised tree sample dataset is labeled according to the categories of main root, leaves, branches and tree support using a point cloud annotation method.

[0016] The network is trained using a lightweight, large-scale point cloud semantic segmentation network, and a local feature aggregation module is used to learn local features.

[0017] In a further embodiment, the predicted leaf area The calculation method is as follows:

[0018] The forward direction of the pesticide application device is defined as the positive X-axis, the horizontal plane perpendicular to the X-axis is defined as the Y-axis, and the vertical upward direction is defined as the positive Z-axis. The center point of the main root of one tree is selected as the origin. The point cloud information in the continuous frame is divided into several parts in the X-axis direction in units of time frames, and then divided into grids in the Y-axis and Z-axis with a predetermined length.

[0019] The grid cells are numbered in the following manner: ;in, Indicates the number of the current mesh volume along the X-axis; This indicates the number of the outermost grid cell of the designated canopy along the Y-axis in the direction from the main root of the tree, where the designated canopy is the canopy closest to the application device; This represents the Z-axis number of the lowest-level grid cell in the tree canopy; the corresponding value range is as follows: ;

[0020] Calculate the leaf area of ​​the point cloud within the mesh. The least squares method is used to calculate the leaf area of ​​the point cloud. The relationship was obtained by fitting the leaf surface properties of a real leaf: , is the fitting function for leaf area.

[0021] In a further embodiment, the predicted leaf area... Corrections are made to obtain the corresponding true leaf area. The steps are as follows:

[0022] The mesh was calculated using the leaf growth characteristic mapping relationship. leaf area Simultaneously acquire the mesh volume Predicted leaf area ;Will and The iterative leaf area is obtained by using the Kalman filter algorithm for iterative calculation. ;

[0023] Traverse all mesh volumes and divide them into outer and inner meshes. The outer mesh is corrected as follows: Let the corrected true leaf area of ​​the outer mesh be the iterated leaf area: ;

[0024] The correction to the inner mesh is as follows: The inner mesh... With the corresponding If a comparative analysis is performed, Then, the occlusion correction coefficient is obtained with the grid as the center. The corrected true leaf area corresponding to the inner mesh is expressed by the following formula: ;like The corrected true leaf area of ​​the inner mesh is then expressed by the following formula: .

[0025] In a further embodiment, the tree area Leaf area density The calculation method is as follows:

[0026] The canopy leaf density within each grid cell is calculated using the following formula. : In the formula, For mesh The actual leaf area inside, For mesh Volume;

[0027] Then, the tree area Leaf area density The calculation formula is as follows: In the formula, Tree area The true leaf area inside the grid The sum, Tree area The total volume of the grid.

[0028] In a further embodiment, the variable spraying model is expressed in the following form:

[0029] ;

[0030] In the formula, This is the maximum amount of pesticide that the equipment can apply. This refers to the leaf density at which the equipment achieves its maximum application rate. Tree area leaf area density .

[0031] In a further embodiment, the spray airflow model is expressed as follows:

[0032] ;

[0033] In the formula, For droplet deposition rate, To penetrate traffic, To optimize parameters; For the corresponding tree area The outlet wind speed of the pesticide application device;

[0034] To determine the optimal outlet wind speed, wind speed constraints are established, and these constraints are expressed as follows: ,in, Drug drift rate, For mesh The required penetration flow rate for the mesh volume where k is the maximum value. For mesh The required penetration flow rate for a mesh cell where k is 1;

[0035] The defined drift rate The calculation formula is as follows: ;in, This represents the total flow rate of mist droplets exiting the tree canopy surface. This represents the total flow rate of mist droplets flowing into the tree canopy surface.

[0036] In a further embodiment, the criteria for dividing the outer and inner meshes are as follows: the outermost mesh containing leaf labels is divided into the outer mesh, and the rest are the inner meshes.

[0037] In a further embodiment, the occlusion correction coefficient The acquisition process is as follows:

[0038] Define the layer containing the current mesh as the current layer. Identify the nine nearest meshes in each layer above the current layer to obtain N occlusion meshes related to the current mesh. Summate the leaf areas of the point clouds of these N occlusion meshes to obtain the occlusion area. The occlusion correction coefficient occlusion The calculation formula is as follows:

[0039] ;

[0040] In the formula, j This indicates the current grid number along the Y-axis. This represents the total number of grid cells with the same i and k values ​​as the current grid cell. A variable-rate dosing system based on target leaf area detection is used to implement the variable-rate dosing method described above, including:

[0041] The transport vehicle is equipped with a spraying actuator at a designated location;

[0042] A lidar is installed in the direction of travel of the transport vehicle; the lidar is configured to collect point cloud data of trees.

[0043] A central data processing unit is mounted on the chassis of the transport vehicle; the central data processing unit is also electrically connected to the lidar and the spraying actuator.

[0044] The central data processing unit includes:

[0045] The first module is set up to train a tree semantic segmentation model based on a tree sample dataset;

[0046] The second module is set to segment the point cloud information in the tree point cloud data of each time frame into main root, leaves, branches and tree support using a tree semantic segmentation model.

[0047] The third module is configured to divide the segmented point cloud information into grids based on time frames and calculate the predicted leaf area of ​​each grid cell. For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. ;

[0048] The fourth module is configured to divide the mesh into different tree regions. Leaf area density ;

[0049] The fifth module is configured to construct a variable spraying model based on the leaf area density. The corresponding tree area was calculated. Dosage ;

[0050] The sixth module is set up to create a spray airflow model based on the current wind speed. Leaf area density and penetration distance x Determine the outlet wind speed of the pesticide application device ;

[0051] The spraying actuator is configured based on the amount of pesticide applied. and outlet wind speed Tree area Apply pesticides in designated areas.

[0052] The beneficial effects of this invention are as follows: This invention utilizes a trained tree semantic segmentation model to segment point cloud data collected by LiDAR, and then fits the segmented point cloud leaf area with the predicted leaf area. Experimental verification shows that the correlation between the two is high and the reliability is high.

[0053] This invention employs a correction algorithm, based on collected leaf quantity, leaf area, and LiDAR data throughout the entire growth cycle of fruit trees, and trained using a deep learning neural network to obtain the growth characteristics of fruit tree leaves. Based on these characteristics, it predicts the leaf area of ​​each grid cell at each application time using point cloud data and growth cycle data collected during pesticide application. For grid areas with high confidence in the outer point cloud data, a Kalman filter algorithm is used to fuse the leaf area calculated using the point cloud semantic segmentation model with the predicted area based on growth characteristics. For areas with low confidence in the inner point cloud due to occlusion or other reasons, the predicted leaf area based on growth characteristics is corrected according to the degree of occlusion. This leaf correction algorithm can fuse multiple measurement data and predict the leaf area inside the fruit tree canopy, which is difficult for LiDAR to detect. This scheme can accurately calculate the leaf area of ​​each region of the fruit tree canopy, which is more accurate and scientific than the existing method of predicting the canopy volume based on point cloud data to determine the pesticide application rate.

[0054] This invention divides the fruit tree canopy into grids and calculates the leaf density based on the leaf area within each grid. A pesticide application model is then established based on this leaf area density. Since the leaf area data is relatively accurate, the above scheme can calculate the appropriate amount of pesticide based on the leaf area density of different fruit tree areas. Furthermore, by controlling different nozzles, precise pesticide application can be achieved, making pesticide application more accurate while saving a significant amount of pesticides.

[0055] This invention establishes a continuous relationship model between droplet penetration rate, deposition rate, drift rate, wind speed, penetration distance, and leaf volume density, enabling effective prediction of droplet deposition rate and penetration flow rate at various points within the tree canopy. Based on this model, the optimal outlet velocity is given, further enhancing the application effect of pesticides. Attached Figure Description

[0056] Figure 1 This is a flowchart of a variable-rate application method based on target leaf area detection.

[0057] Figure 2 Architecture diagram of the software portion of the variable dosing system.

[0058] Figure 3 Structure of the hardware component of a variable dosing system Figure 1 .

[0059] Figure 4 Structure of the hardware component of a variable dosing system Figure 2 .

[0060] Figures 2 to 4 The components are labeled as follows: 1. Transport vehicle; 2. LiDAR; 3. Spraying actuator; 4. Data processing unit; 5. Fan; 31. Medicine tank; 32. Plunger pump; 33. Adhesive hose; 34. Solenoid valve; 35. Nozzle; 51. Fan housing; 52. Air outlet; 53. Air volume window; 54. Anemometer. Detailed Implementation

[0061] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0062] Example 1

[0063] According to embodiments of the present invention, a variable-rate application method based on target leaf area detection is provided, such as... Figure 1 As shown, it includes the following steps:

[0064] Tree sample datasets from different growth stages are obtained, and these datasets are used to train a tree semantic segmentation model. In practice, the sample data from different growth stages includes winter sample data and spring / summer / autumn sample data, as the branch and leaf state of trees varies with the season. The most obvious difference is that winter trees generally have no leaves and only branches. Therefore, when collecting sample data, using sample data containing only branches accelerates the subsequent training of the tree semantic segmentation model, leading to faster convergence and better recognition results. In other words, the tree sample dataset is divided into a winter tree sample set and a non-winter tree sample set according to the season; the winter tree sample set is a point cloud dataset of trees containing only branches, while the non-winter tree sample set is a point cloud dataset of trees containing both branches and leaves.

[0065] The spraying device travels at a predetermined speed between the rows of trees and collects tree point cloud data according to time frames. For the point cloud information in the tree point cloud data of each time frame, the tree semantic segmentation model is used to segment it into main roots, leaves, branches and tree supports.

[0066] Based on time frames, the segmented point cloud information is divided into grids, and the predicted leaf area of ​​each grid cell is calculated. For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. ;

[0067] The grid is divided into different tree regions according to the tree regions. Leaf area density ;

[0068] Construct a variable spraying model based on the leaf area density. The corresponding tree area was calculated. Dosage ;

[0069] Create a spray airflow model based on the current wind speed. Leaf area density and penetration distance x Determine the outlet wind speed of the pesticide application device ;

[0070] Based on the dosage and outlet wind speed Tree area Apply pesticides in designated areas.

[0071] In this embodiment, the training process of the tree semantic segmentation model is as follows: the tree sample dataset is denoised using a point cloud filtering algorithm to obtain a denoised tree sample dataset; the denoised tree sample dataset is labeled according to the categories of main root, leaves, branches and tree support using a point cloud annotation method; a lightweight large-scale point cloud semantic segmentation network is used for training, and a local feature aggregation module is used to learn local features.

[0072] This embodiment uses fruit trees as an example to further describe the training process of the tree semantic segmentation model. The specific process is as follows:

[0073] Point cloud filtering algorithms are used to remove outliers, noise points, and point cloud data of fruit trees far from the collected rows from the tree sample dataset. The point cloud filtering algorithm in this embodiment includes the following process: First, a pass-through filtering method is used to specify the value range of the X-axis of the collected point cloud data (the determination of the X-axis is described below) to remove outlier data outside the canopy of a single fruit tree. Then, a statistical filtering method is used to traverse the point cloud data and remove outliers and noise points in areas with low point cloud density.

[0074] The acquired point cloud data includes valid information such as branches and leaves, as well as some noise points caused by canopy vibration. Therefore, data preprocessing of the fruit tree canopy point cloud data is crucial, providing a solid guarantee for subsequent semantic segmentation and leaf area calculation. Dataset construction: After collecting several preprocessed fruit tree canopy point cloud data sets, including those of fruit trees with only trunks and branches in winter and those of fruit trees with leaves growing in different periods of spring, summer, and autumn, point cloud annotation software was used to annotate the points, dividing the information into fruit tree main roots, leaves, branches, and fruit tree supports. Point cloud semantic segmentation model training: The lightweight, large-scale point cloud semantic segmentation network RandLA-Net was used for training. To avoid losing a large amount of branch and leaf feature information during training, the Local Feature Aggregation (LFA) module was used to learn local features. The RandLA-Net semantic segmentation framework uses random point sampling instead of other more complex sampling methods, such as farthest point sampling (FPS), which has high computational complexity and poor scalability, and inverse density importance sampling (IDIS), which is easily affected by noise point information. Random point sampling ensures the speed of network training and segmentation prediction, and achieves good real-time performance in practical applications.

[0075] The RandLA-Net semantic segmentation framework described above is a standard Encoder-decoder structure. During training, the input canopy point cloud data will output the segmentation prediction results of each part of the canopy point cloud at the output end. The error between the prediction results and the true labels is trained until the model converges or reaches the expected result.

[0076] To avoid losing a significant amount of useful branch and leaf feature information during random sampling, this network uses a Local Feature Aggregation (LFA) module to learn local canopy features. This module mainly consists of three neural units: a local spatial encoding unit, an attention pooling unit, and an extended residual block unit.

[0077] The role of the Local Spatial Encoding Unit (LSI) is to, given the position and features of a point P, aggregate and calculate the positions and features of its neighboring points to obtain the local geometric structure features of that point. This allows the LSI to explicitly observe local geometric structure features, ultimately facilitating the entire network's effective learning of complex local structures. The specific operation of the LSI is as follows: for the q-th coordinate... The point is characterized by Use the KNN algorithm to find the set of K neighbor points for this point: and the corresponding feature set: ;for Each adjacent point is encoded in relative position using the following formula. :

[0078] ;

[0079] Encode the features of each neighboring point with its corresponding relative position. By concatenating the data, we obtain the local code for each adjacent point: .

[0080] Secondly, the attention pooling unit is used to aggregate the local encodings obtained from the local spatial encoding unit to obtain features. Combining the above description, the weight parameters W are trained using an MLP method to learn the score of each nearest neighbor. : The score As a mask for feature selection, it is used to remove unimportant features, where feature selection... as follows: .

[0081] The MLP method refers to a feedforward artificial neural network model. An MLP consists of multiple layers, including an input layer, one or more hidden layers, and an output layer. Each layer contains multiple neurons (or nodes), which are connected by weight parameters W.

[0082] Finally, the extended residual block unit combines the local spatial coding unit and the attention pooling unit, integrating the local spatial coding unit, the attention pooling unit, and the shared MLP into an extended residual block to increase the receptive field. The upsampling part of the network uses more efficient nearest neighbor interpolation, further improving the algorithm's efficiency.

[0083] Furthermore, this embodiment uses a six-fold cross-training method to validate the tree semantic segmentation model. This involves dividing the entire dataset into six roughly equal subsets, typically each containing the same number of samples. Assuming the dataset contains N samples, each subset contains approximately N / 6 samples. Six rounds of training and validation are then performed. In each round, one subset is selected as the validation set, and the remaining five subsets are used as the training set. The accuracy, precision, recall, F1 score, etc., obtained from each round are summarized, and the average or other statistical measures are calculated to derive the final model performance evaluation metric.

[0084] In a further embodiment, leaf area is predicted. The calculation method is as follows: the forward direction of the pesticide application device is defined as the positive X-axis, the horizontal plane perpendicular to the X-axis is defined as the Y-axis, and the vertical upward direction is defined as the positive Z-axis; the center point of the main root of one tree is selected as the origin, and the point cloud information in the continuous frame is divided into several parts in the X-axis direction in terms of time frames, and then divided into grids in the Y-axis and Z-axis with a predetermined length respectively;

[0085] The grid cells are numbered in the following manner: ;in, Indicates the number of the current mesh volume along the X-axis; This indicates the number of the outermost grid cell of the designated canopy along the Y-axis in the direction from the main root of the tree, where the designated canopy is the canopy closest to the application device; This represents the Z-axis number of the lowest-level grid cell in the tree canopy; the corresponding value range is as follows: ;

[0086] Calculate the leaf area of ​​the point cloud within the mesh. The least squares method is used to calculate the leaf area of ​​the point cloud. The relationship was obtained by fitting the leaf surface properties of a real leaf: , is the fitting function for leaf area.

[0087] It's worth noting that in the weighted minimum residual least squares method, points closer to the straight line have a higher weight during fitting, while points farther away have a lower weight. This can effectively suppress the influence of outliers and yield better correlation results.

[0088] Table 1. Canopy information of 10 kiwifruit trees

[0089]

[0090] In this embodiment, canopy information of 10 kiwifruit trees was collected as shown in Table 1. Various segmentation methods were compared, and it was found that the RandLA-Net semantic segmentation framework described above had a better effect on leaf separation and a higher overall accuracy, as shown in Tables 2 and 3.

[0091] Table 2. Separation results of canopy data across six models

[0092]

[0093] Table 3. Performance of canopy segmentation in different studies

[0094]

[0095] Considering that lidar can only scan the area of ​​the outermost leaves of a tree canopy, and due to occlusion between leaves, the area of ​​the inner leaves cannot be accurately obtained, thus affecting the accuracy of pesticide application calculation and the control effect. Therefore, this embodiment also includes leaf area correction, the specific process of which is as follows: A grid is calculated using the leaf growth characteristic mapping relationship. leaf area Simultaneously acquire the mesh volume Predicted leaf area ;Will and The iterative leaf area is obtained by using the Kalman filter algorithm for iterative calculation. It is worth mentioning that the fruit tree variety is obtained, and the leaf growth characteristics of that variety are preset in the system. These leaf growth characteristics include the leaf growth cycle and the mapping relationship between leaf area and time within the growth cycle. Furthermore, the leaf area distribution follows a normal distribution.

[0096] Traverse all mesh volumes and divide them into outer and inner meshes. The outer mesh is corrected as follows: Let the corrected true leaf area of ​​the outer mesh be the iterated leaf area: ;

[0097] The correction to the inner mesh is as follows: The inner mesh... With the corresponding If a comparative analysis is performed, Then, the occlusion correction coefficient is obtained with the grid as the center. The corrected true leaf area corresponding to the inner mesh is expressed by the following formula: ;like The corrected true leaf area of ​​the inner mesh is then expressed by the following formula: .

[0098] The criteria for dividing the outer and inner meshes are as follows: the outermost mesh containing leaf labels is designated as the outer mesh, and the rest are designated as the inner mesh.

[0099] Occlusion correction factor The acquisition process is as follows: Define the layer where the current grid is located as the current layer; determine the nine grids in each layer above the current layer that are close to the current grid to obtain N occlusion grids about the current grid; sum the leaf areas of the point cloud of the N occlusion grids to obtain the occlusion area. The occlusion correction coefficient occlusion The calculation formula is as follows:

[0100] In the formula, j This indicates the current grid number along the Y-axis. This represents the total number of grid cells that have the same i and k values ​​as the current grid cell.

[0101] As described above, the leaf area correction algorithm can estimate the leaf area of ​​each region of the fruit tree canopy at the time of the current application based on leaf growth characteristics, the time interval between the current application and the actual growth stage of the fruit tree, and then calculate the leaf density. Combined with the point cloud data collected this time, the Kalman filter algorithm is used for data fusion to finally calculate a more accurate leaf area and leaf area density. In particular, the leaf area correction algorithm can first predict the leaf area of ​​areas inside the fruit tree canopy that cannot be detected by lidar using fruit tree growth characteristics, and then correct it using an occlusion correction coefficient. It can accurately predict the leaf area of ​​the inner layer of the canopy, providing more accurate data support for precision application.

[0102] In a further embodiment, the tree area Leaf area density The calculation method is as follows:

[0103] The canopy leaf density within each grid cell is calculated using the following formula. : In the formula, For mesh The actual leaf area inside, For mesh Volume;

[0104] Then, the tree area Leaf area density The calculation formula is as follows: In the formula, Tree area The true leaf area inside the grid The sum, Tree area The total volume of the grid.

[0105] Correspondingly, the variable spraying model is expressed in the following form: In the formula, This is the maximum amount of pesticide that the equipment can apply. This refers to the leaf density at which the equipment achieves its maximum application rate. Tree area leaf area density .

[0106] The application model of this invention makes decisions based on the leaf density within each grid cell, which can accurately calculate the amount of pesticide required for each grid cell, thereby driving the application device to spray the area precisely. Therefore, it can save pesticides to the greatest extent while ensuring the spraying effect.

[0107] Meanwhile, it is also considered that during the process of collecting point cloud data with lidar, the tree canopy will inevitably vibrate due to wind and other reasons.

[0108] This embodiment also creates a spray airflow model, which is expressed as follows: ;

[0109] In the formula, For droplet deposition rate, To penetrate traffic, To optimize parameters; For the corresponding tree area The outlet wind speed of the pesticide application device;

[0110] To determine the optimal outlet wind speed, wind speed constraints are established, and these constraints are expressed as follows: ,in, Drug drift rate, For mesh The required penetration flow rate for the mesh volume where k is the maximum value. For mesh The required penetration flow rate for a mesh cell where k is 1;

[0111] The defined drift rate The calculation formula is as follows: ;in, This represents the total flow rate of mist droplets exiting the tree canopy surface. This represents the total flow rate of mist droplets flowing into the tree canopy surface.

[0112] Example 2

[0113] This embodiment discloses a variable spraying method system based on target leaf area detection, used to implement the variable spraying method described in Embodiment 1, including: a transport vehicle, on which a spraying actuator is mounted at a designated location;

[0114] A lidar is installed in the direction of travel of the transport vehicle; the lidar is configured to collect point cloud data of trees.

[0115] The central data processing unit is mounted on the chassis of the transport vehicle; such as Figure 2 As shown, the central data processing unit is electrically connected to both the lidar and the spraying actuator.

[0116] The central data processing unit includes:

[0117] The first module is set up to train a tree semantic segmentation model based on a tree sample dataset;

[0118] The second module is set to segment the point cloud information in the tree point cloud data of each time frame into main root, leaves, branches and tree support using a tree semantic segmentation model.

[0119] The third module is configured to divide the segmented point cloud information into meshes based on time frames and calculate the true leaf area. For the actual leaf area The actual leaf area S within the grid is obtained by making corrections;

[0120] The fourth module is configured to divide the grid into different tree regions. i Leaf area density ;

[0121] The fifth module is configured to construct a variable spraying model based on the leaf area density. The corresponding tree area was calculated. i Dosage ;

[0122] The sixth module is set up to create a spray airflow model based on the current wind speed. Leaf area density and penetration distance x Determine the outlet wind speed of the pesticide application device ;

[0123] The spraying actuator is configured based on the amount of pesticide applied. and outlet wind speed Tree area i Apply pesticides in designated areas.

[0124] like Figure 3 and Figure 4 As shown, a fan 5 is provided behind the spraying actuator 3. The fan 5 includes a fan housing 51, which is an inverted cone shape. Multiple air outlets 52 are arranged on both the left and right sides of the fan housing 51 along the vertical direction. Air volume windows 53 are provided inside each air volume window 53. An anemometer 54 is provided on the side of each air volume window 53 away from the fan housing 5. The spraying actuator 3 includes a medicine tank 31. The inlet of the plunger pump 32 is connected to the outlet of the medicine tank 31 via a rubber water pipe 33. The outlet of the plunger pump 32 is connected to multiple solenoid valves 34 via the rubber water pipe 33. The other end of the solenoid valve 34 is connected to a nozzle 35. The nozzles 35 are arranged in two vertical and uniform rows on both sides of the rear of the transport vehicle 1. The solenoid valves 34 are electrically connected to the central data processing unit 4. The lidar 2 of this invention collects point cloud data at the front and transmits it to the central data processing unit 4 for point cloud data processing to obtain characteristics of the fruit tree, including leaf area density of each part. Based on this, it calculates data such as spraying amount and air volume, and controls the spraying actuator 3 and the fan 5 to spray.

Claims

1. A variable-rate application method based on target leaf area detection, characterized in that, Includes the following steps: Obtain tree sample datasets at different growth stages, which are divided into winter tree sample sets and non-winter tree sample sets according to the season; train a tree semantic segmentation model using the tree sample datasets; The spraying device travels at a predetermined speed between rows of trees and collects tree point cloud data according to time frames. For the point cloud information in the tree point cloud data of each time frame, the tree semantic segmentation model is used to segment it into main root, leaves, branches and tree support; Based on time frames, the segmented point cloud information is divided into grids, and the predicted leaf area of ​​each grid cell is calculated. For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. ; The grid is divided into different tree regions according to the tree regions. Leaf area density ; Construct a variable spraying model based on the leaf area density. The corresponding tree area was calculated. Dosage ; Create a spray airflow model based on the current wind speed. Leaf area density and penetration distance x Determine the outlet wind speed of the pesticide application device ; Based on the dosage and outlet wind speed Tree area Apply pesticides in designated areas; The predicted leaf area The calculation method is as follows: The forward direction of the pesticide application device is defined as the positive X-axis, the horizontal plane perpendicular to the X-axis is defined as the Y-axis, and the vertical upward direction is defined as the positive Z-axis. The center point of the main root of one tree is selected as the origin. The point cloud information in the continuous frame is divided into several parts in the X-axis direction in units of time frames, and then divided into grids in the Y-axis and Z-axis with a predetermined length. The grid cells are numbered in the following manner: ;in, Indicates the number of the current mesh volume along the X-axis; This indicates the number from the outermost grid cell of the specified canopy to the main root of the tree along the Y-axis direction, where the specified canopy is the canopy closest to the application device; This represents the Z-axis number of the lowest-level grid cell in the tree canopy; the corresponding value range is as follows: ; Calculate the leaf area of ​​the point cloud within the mesh. The least squares method is used to calculate the leaf area of ​​the point cloud. The relationship was obtained by fitting the leaf surface properties of a real leaf: , is the fitting function for leaf area; For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. The steps are as follows: The mesh was calculated using the leaf growth characteristic mapping relationship. leaf area Simultaneously acquire the mesh volume Predicted leaf area ;Will and The iterative leaf area is obtained by using the Kalman filter algorithm for iterative calculation. ; Traverse all mesh volumes and divide them into outer and inner meshes. The outer mesh is corrected as follows: Let the corrected true leaf area of ​​the outer mesh be the iterated leaf area: ; The correction to the inner mesh is as follows: The inner mesh... With the corresponding If a comparative analysis is conducted, Then, the occlusion correction coefficient is obtained with the grid as the center. The corrected true leaf area corresponding to the inner mesh is expressed by the following formula: ;like The corrected true leaf area of ​​the inner mesh is then expressed by the following formula: ; The occlusion correction coefficient The acquisition process is as follows: Define the layer containing the current mesh as the current layer. Identify the nine nearest meshes in each layer above the current layer to obtain N occlusion meshes related to the current mesh. Summate the leaf areas of the point clouds of these N occlusion meshes to obtain the occlusion area. The occlusion correction coefficient The calculation formula is as follows: ； In the formula, This indicates the current grid number along the Y-axis. This represents the total number of grids with the same i and k values ​​as the current grid, where N is the number of occluded grids in the current grid.

2. The variable-rate application method based on target leaf area detection according to claim 1, characterized in that, The winter tree sample set is a tree point cloud dataset containing only branches, while the non-winter tree sample set is a tree point cloud dataset containing branches and leaves. The training process of the tree semantic segmentation model is as follows: the tree sample dataset is denoised using a point cloud filtering algorithm to obtain a denoised tree sample dataset; the denoised tree sample dataset is labeled according to the categories of main root, leaves, branches and tree support using a point cloud annotation method. The network is trained using a lightweight, large-scale point cloud semantic segmentation network, and a local feature aggregation module is used to learn local features.

3. The variable-rate application method based on target leaf area detection according to claim 1, characterized in that, The criteria for dividing the outer and inner meshes are as follows: the outermost mesh containing leaf labels is designated as the outer mesh, and the rest are designated as the inner mesh.

4. The variable-rate application method based on target leaf area detection according to claim 1, characterized in that, The variable spraying model is expressed in the following form: ； In the formula, This is the maximum amount of pesticide that the equipment can apply. This refers to the leaf density at which the equipment achieves its maximum application rate. Tree area leaf area density .

5. The variable-rate application method based on target leaf area detection according to claim 1, characterized in that, The spray air volume model is expressed as follows: ； In the formula, To penetrate traffic, To optimize parameters; For the corresponding tree area The outlet wind speed of the pesticide application device; To determine the optimal outlet wind speed, wind speed constraints are established, and these constraints are expressed as follows: ,in, Drug drift rate, For mesh The required penetration flow rate for the mesh volume where j is taken as the maximum value. For mesh The required penetration flow rate for a mesh volume with j set to 1; The drug drift rate The calculation formula is as follows: ;in, This represents the total flow rate of mist droplets exiting the tree canopy surface. This represents the total flow rate of mist droplets flowing into the tree canopy surface.

6. A variable-rate dosing system based on target leaf area detection, used to implement the variable-rate dosing method as described in any one of claims 1 to 5; characterized in that, include: The transport vehicle is equipped with a spraying actuator at a designated location; A lidar is installed in the direction of travel of the transport vehicle; the lidar is configured to collect point cloud data of trees. A central data processing unit is mounted on the chassis of the transport vehicle; the central data processing unit is also electrically connected to the lidar and the spraying actuator. The central data processing unit includes: The first module is set up to acquire tree sample datasets at different growth stages, which are divided into winter tree sample datasets and non-winter tree sample datasets according to the season; and to train a tree semantic segmentation model using the tree sample datasets. The second module is configured to be a spraying device that travels between rows of trees at a predetermined speed and collects tree point cloud data according to time frames. For the point cloud information in the tree point cloud data of each time frame, a tree semantic segmentation model is used to segment it into main roots, leaves, branches and tree supports. The third module is configured to divide the segmented point cloud information into grids based on time frames and calculate the predicted leaf area of ​​each grid cell. For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. ; The fourth module is configured to divide the mesh into different tree regions. Leaf area density ; The fifth module is configured to construct a variable spraying model based on the leaf area density. The corresponding tree area was calculated. Dosage ; The sixth module involves creating a spraying airflow model based on the current wind speed. Leaf area density and penetration distance x Determine the outlet wind speed of the pesticide application device ; The spraying actuator is configured to apply the pesticide dosage. and outlet wind speed Tree area Apply pesticides in designated areas; Among them, the predicted leaf area The calculation method is as follows: The forward direction of the pesticide application device is defined as the positive X-axis, the horizontal plane perpendicular to the X-axis is defined as the Y-axis, and the vertical upward direction is defined as the positive Z-axis. The center point of the main root of one tree is selected as the origin. The point cloud information in the continuous frame is divided into several parts in the X-axis direction in units of time frames, and then divided into grids in the Y-axis and Z-axis with a predetermined length. The grid cells are numbered in the following manner: ;in, Indicates the number of the current mesh volume along the X-axis; This indicates the number from the outermost grid cell of the specified canopy to the main root of the tree along the Y-axis direction, where the specified canopy is the canopy closest to the application device; This represents the Z-axis number of the lowest-level grid cell in the tree canopy; the corresponding value range is as follows: ; Calculate the leaf area of ​​the point cloud within the mesh. The least squares method is used to calculate the leaf area of ​​the point cloud. The relationship was obtained by fitting the leaf surface properties of a real leaf: , is the fitting function for leaf area; For the predicted leaf area Corrections are made to obtain the corresponding true leaf area. The steps are as follows: The mesh was calculated using the leaf growth characteristic mapping relationship. leaf area Simultaneously acquire the mesh volume Predicted leaf area ;Will and The iterative leaf area is obtained by using the Kalman filter algorithm for iterative calculation. ; Traverse all mesh volumes and divide them into outer and inner meshes. The outer mesh is corrected as follows: Let the corrected true leaf area of ​​the outer mesh be the iterated leaf area: ; The correction to the inner mesh is as follows: The inner mesh... With the corresponding If a comparative analysis is conducted, Then, the occlusion correction coefficient is obtained with the grid as the center. The corrected true leaf area corresponding to the inner mesh is expressed by the following formula: ;like The corrected true leaf area of ​​the inner mesh is then expressed by the following formula: ; The occlusion correction coefficient The acquisition process is as follows: Define the layer containing the current mesh as the current layer. Identify the nine nearest meshes in each layer above the current layer to obtain N occlusion meshes related to the current mesh. Summate the leaf areas of the point clouds of these N occlusion meshes to obtain the occlusion area. The occlusion correction coefficient The calculation formula is as follows: ； In the formula, This indicates the current grid number along the Y-axis. This represents the total number of grids with the same i and k values ​​as the current grid, where N is the number of occluded grids in the current grid.

Images