Agricultural unmanned aerial vehicle pesticide spraying injury control method and system

Abstract

The application provides an agricultural unmanned aerial vehicle pesticide spraying injury control method and system, and belongs to the technical field of agricultural unmanned aerial vehicle pesticide spraying. The method comprises the following steps: collecting original image data of a farmland in real time and performing preprocessing to obtain preprocessed image data; based on the preprocessed image data, a pre-trained deep learning recognition model is used to recognize target crops, non-target crops, sensitive crops, and the boundaries between operation areas and non-operation areas, to generate boundary coordinate data and crop type recognition results; high-precision positioning data is obtained through RTK positioning, and the high-precision positioning data is used to dynamically calibrate the boundary coordinate data to generate calibrated boundary coordinate data; and according to the calibrated boundary coordinate data and the crop type recognition results, the pesticide spraying parameters are dynamically adjusted, and corresponding early warning signals are triggered. Through image recognition and dynamic calibration, the application realizes accurate determination and dynamic response of the operation boundary, and effectively prevents pesticide drift injury.

Description

Technical Field

[0001] This application belongs to the field of agricultural drone spraying technology, specifically relating to a method and system for controlling accidental damage during agricultural drone spraying. Background Technology

[0002] Agricultural drone spraying is widely used for pest and disease control in various farmland scenarios such as field crops, orchards, and tea gardens due to its high efficiency, precision, and adaptability. However, in actual operations, the complex and variable farmland environment, often involving mixed cropping, irregular terrain, and the presence of seedlings and mature crops, as well as insufficient light due to weather changes (such as cloudy days or evenings), pose significant challenges to boundary recognition and crop differentiation during drone spraying.

[0003] In existing technologies, drone spraying systems primarily rely on satellite positioning systems such as GPS or BeiDou for operational path planning and boundary control. However, satellite positioning alone suffers from insufficient accuracy (typically at the decimeter or meter level), susceptibility to signal blockage, and an inability to distinguish crop types. This makes it difficult to meet the demands of accurate identification of operational and non-operational areas, and of target and sensitive crops (such as vegetables, strawberries, grapes, and tobacco) in complex farmland scenarios. Furthermore, existing systems generally lack effective image recognition and dynamic boundary calibration mechanisms, failing to respond in real-time to changes in terrain or drone drift. This results in pesticides easily drifting into neighboring fields or sensitive crop areas during spraying, causing accidental damage and severely impacting agricultural production efficiency.

[0004] To address the aforementioned issues, some existing technologies attempt to incorporate image acquisition modules to assist in identification. However, these methods suffer from simple identification algorithms, insufficient preprocessing capabilities, and a lack of calibration methods, failing to establish a closed-loop control system centered on boundary recognition. Therefore, there is an urgent need for a method to control accidental damage during agricultural drone spraying that can achieve high-precision boundary recognition, crop type differentiation, and dynamic calibration, in order to solve the problems of low boundary recognition accuracy and high risk of accidental damage in existing technologies. Summary of the Invention

[0005] In a first aspect, embodiments of this application provide a method for controlling accidental damage during pesticide spraying by agricultural drones, comprising the following steps: S1. Real-time acquisition of raw image data of farmland and preprocessing to obtain preprocessed image data; S2. Based on the preprocessed image data, a pre-trained deep learning recognition model is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area, generating boundary coordinate data and crop type recognition results; S3. Obtain high-precision positioning data through RTK positioning, and use the high-precision positioning data to dynamically calibrate the boundary coordinate data to generate calibrated boundary coordinate data; S4. Based on the calibrated boundary coordinate data and crop type identification results, dynamically adjust the spraying parameters and trigger corresponding early warning signals.

[0006] Furthermore, the specific steps of step S1 are as follows: S11. An industrial camera with a resolution not lower than the resolution threshold is used to collect raw image data of farmland in real time at a preset acquisition frequency, covering the work area and the surrounding preset radius range. S12. When the ambient light is below a preset threshold, turn on the infrared fill light for auxiliary illumination of the industrial camera; S13. An adaptive filtering algorithm is used to denoise the original image data; S14. Use an image enhancement algorithm to adjust the contrast and brightness of the denoised image; S15. Perform distortion correction on the enhanced image and output the preprocessed image data.

[0007] Furthermore, the specific steps of step S15 are as follows: S151. Obtain the intrinsic parameter matrix and distortion coefficients of the camera, wherein the distortion coefficients include radial distortion coefficients. and tangential distortion coefficient ; S152. For each pixel in the enhanced image Calculate the corresponding normalized coordinates and radius ; S153. Calculate the corrected coordinates according to the following distortion correction formula. :

[0008] S154. Map the corrected coordinates back to the pixel coordinate system, and obtain the distorted image pixel values ​​through interpolation.

[0009] Furthermore, the specific steps of step S2 are as follows: S21. Input the preprocessed image data into a pre-trained deep learning recognition model; S22. Deep learning recognition models perform inference on input image data: The encoder part of the model uses layer-by-layer convolution and downsampling operations to extract shallow features from deep features in the image; the shallow features include the edge and corner information of crop leaves, and the deep features include the shape, color and texture semantic information of crop leaves. The decoder part of the model uses upsampling and skip connection operations to gradually restore the deep features to the original image size and outputs a multi-channel feature map with the same size as the input image. Each channel corresponds to a class, and the value of each pixel in each channel represents the unnormalized score of the pixel belonging to the corresponding class. A class probability vector is output for each pixel. The categories include target crops, non-target crops, sensitive crops, and background, with weeds and stones being classified as non-target crops or background to achieve filtering. S23. Apply the Softmax function along the channel direction to the multi-channel feature map, convert the channel score of each pixel into a category probability vector, and take the category corresponding to the highest probability as the final category label of this pixel, thereby generating a pixel-level segmentation map; among them, the pixels corresponding to weeds and stones are classified as non-target crops or background to achieve automatic filtering; S24. Perform connected component analysis on the pixel-level segmentation map, extract the connected component boundary contours of the target crop region, convert the pixel coordinates in the connected component boundary contours into a set of boundary coordinate points in the world coordinate system, generate boundary coordinate data, and output the crop type identification result corresponding to each boundary coordinate point, generating crop type identification result data.

[0010] Furthermore, the deep learning model employs a convolutional neural network, and the specific training process is as follows: SS1. Collect farmland images under several lighting and weather scenarios, and annotate each collected farmland image at the pixel level. The labeling categories include target crop, non-target crop, sensitive crop, and background. Use the annotated farmland images and corresponding category labels to construct a training sample set. SS2. Perform data augmentation on the training sample set, the data augmentation including geometric transformation and color transformation, and expand the augmented samples to a preset multiple of the original sample number; SS3. Set training hyperparameters, including batch size, initial learning rate, learning rate decay strategy, and total number of training epochs; SS4. The following cross-entropy loss function is used to calculate the difference between the model prediction and the true label:

[0011] in, Total number of pixels; Total number of categories; Let c be the true label indicator of the i-th pixel. If the true category of the i-th pixel is c, then... The value is 1, otherwise The value is 0; The probability that the i-th pixel belongs to class c as predicted by the model; SS5. The Adam optimizer is used to iteratively update the model parameters; SS6. After each training round, evaluate the model using the validation set and calculate the evaluation metric on the validation set. When the validation set loss no longer decreases after a preset number of early stopping rounds, stop training early.

[0012] Furthermore, the specific steps of step S3 are as follows: S31. The positioning data is acquired by the airborne RTK positioning module at a preset first update frequency, wherein the positioning accuracy of the RTK positioning module is higher than a preset positioning accuracy threshold. S32. Synchronously acquire auxiliary positioning data; S33. The following dynamic boundary calibration algorithm is used to correct the boundary coordinate data:

[0013] in, For the identified boundary coordinate vector, For RTK positioning coordinate vectors, To assist in locating coordinate vectors, The calibrated boundary coordinate vector; S34. Output the calibrated boundary coordinate data and enable manual auxiliary calibration via the ground control terminal.

[0014] Furthermore, step S3 also includes the following steps: S35. During the drone's flight, the boundary coordinates are updated in real time cyclically at a preset second update frequency: The calibrated boundary coordinate data output from the previous loop step S33 is used as the identified boundary coordinate data in the current loop step S33. Steps S31 to S33 are repeated to output the calibrated boundary coordinate data of the current loop and use it as the input for the next loop, thereby realizing continuous dynamic calibration of boundary coordinate data.

[0015] Furthermore, the specific steps of step S4 are as follows: S41. Receive the calibrated boundary coordinate data and crop type identification results; S42. When the current area is identified as a non-operational area or a sensitive crop area, a stop spraying command is sent and an audible and visual warning signal is triggered; S43. When the current area is identified as both the work area and the target crop area, the spraying flow rate, spraying pressure, and spray width are dynamically adjusted according to real-time environmental parameters, and a safety isolation zone is set up inside the work boundary; S44. Perform spraying operations using anti-drift nozzles to ensure that the pesticide solution does not exceed the calibrated boundary coordinate range.

[0016] Furthermore, the specific steps of step S43 are as follows: S431. Acquire real-time environmental parameters, including wind speed data, which is collected by an airborne wind speed sensor; S432. Dynamically adjust spraying pressure based on wind speed data; When the wind speed does not exceed the first wind speed threshold, the spraying pressure is set to the first pressure value; When the wind speed exceeds the first wind speed threshold but does not exceed the second wind speed threshold, the spraying pressure is set to the second pressure value; the second pressure value is greater than the first pressure value. Spraying operations shall be suspended when the wind speed exceeds the second wind speed threshold. S433. Extract the boundary coordinates of sensitive crop areas from the crop type identification results, and extract the operation boundary coordinates from the calibrated boundary coordinate data; S434. Calculate the distance between the boundary of the sensitive crop area and the operation boundary to obtain the distance data between the sensitive crop and the operation boundary; S435. Dynamically set the width of the safety isolation zone based on the distance data; If the distance between the sensitive crop and the work boundary is less than the preset distance threshold, the width of the safety isolation zone is set to the first width. If the distance between the sensitive crop and the work boundary is greater than or equal to a preset distance threshold, the width of the safety isolation zone is set to the second width; the first width is greater than the second width. S436. Based on the adjusted spraying pressure and the set width of the safety isolation zone, control the spraying actuator to perform spraying operations at the corresponding pressure inside the safety isolation zone.

[0017] Secondly, embodiments of this application also provide a control system for accidental damage caused by pesticide spraying by agricultural drones, comprising: The image acquisition and preprocessing module is used to acquire raw image data of farmland in real time and perform preprocessing to obtain preprocessed image data; The deep learning recognition module is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area based on pre-processed image data using a pre-trained deep learning recognition model, and to generate boundary coordinate data and crop type recognition results. The dynamic calibration module is used to acquire high-precision positioning data through RTK positioning, and to use the high-precision positioning data to dynamically calibrate the boundary coordinate data, generating calibrated boundary coordinate data. The spraying control and early warning module is used to dynamically adjust spraying parameters and trigger corresponding early warning signals based on the calibrated boundary coordinate data and crop type identification results.

[0018] As can be seen from the above technical solutions, this application has the following advantages: The agricultural drone spraying accidental damage control method and system provided in this application achieves accurate identification of the operation boundary and crop type by integrating deep learning image segmentation and RTK high-precision positioning, solving the problem of inaccurate boundary identification caused by relying solely on satellite positioning; by incorporating infrared supplementary lighting and adaptive image enhancement algorithms, it achieves stable all-weather operation in low-light environments such as cloudy days and evenings, eliminating the impact of ambient light on identification accuracy; by establishing a dynamic response mechanism based on wind speed and distance to sensitive crops, it achieves real-time adaptive adjustment of spraying parameters, preventing pesticide drift and avoiding the risk of pesticide damage to sensitive crops in neighboring fields. Attached Figure Description

[0019] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for controlling accidental damage during pesticide spraying by agricultural drones according to the present invention.

[0021] Figure 2 This is a schematic diagram of the agricultural drone spraying accidental injury control system of the present invention. Detailed Implementation

[0022] The various embodiments of this disclosure will be described more fully in the following detailed description of the specific steps of the method for controlling accidental damage during agricultural drone spraying. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0023] This embodiment provides a method for controlling accidental damage during agricultural drone spraying. It integrates deep learning recognition and RTK high-precision positioning to achieve accurate determination and dynamic response of the operation boundary, thereby solving the problems of drone spraying drift and boundary accidental damage.

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1The diagram shows a flowchart of a method for controlling accidental damage during pesticide spraying by an agricultural drone in a specific embodiment. The method includes the following steps: S1. Real-time acquisition of raw image data of farmland and preprocessing to obtain preprocessed image data; It should be noted that by acquiring and preprocessing data in real time, the clarity and accuracy of the input data were ensured, which solved the problem of image quality degradation caused by the complex farmland environment. This provided high-quality images for subsequent deep learning recognition and avoided misjudgments caused by image blur. S2. Based on the preprocessed image data, a pre-trained deep learning recognition model is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area, generating boundary coordinate data and crop type recognition results; It should be noted that by using deep learning models, the limitations of traditional geometric boundary delineation are overcome, enabling semantic understanding of farmland scenes. Not only can the geographical boundaries of farmland be simply delineated as in traditional methods, but further analysis of color, texture, and shape features in images can be performed to accurately identify the specific types of crops. By generating pixel-level boundary coordinates and crop type results, fine-grained division of work areas and non-work areas, as well as target crops and sensitive crops, is achieved, thus preventing accidental damage to crops. S3. Obtain high-precision positioning data through RTK positioning, and use the high-precision positioning data to dynamically calibrate the boundary coordinate data to generate calibrated boundary coordinate data; It should be noted that by fusing visual coordinates with geographic coordinates, the high precision of RTK is used to anchor the boundaries of visual recognition. At the same time, real-time calibration during flight solves the problem of inaccurate position recognition caused by vibration and signal fluctuations during the movement of the UAV. This ensures that the UAV can accurately obtain its absolute position in the Earth coordinate system and accurately determine its relative position with respect to the operation boundary. S4. Based on the calibrated boundary coordinate data and crop type identification results, dynamically adjust the spraying parameters and trigger corresponding early warning signals; It should be noted that by triggering early warning signals and dynamically adjusting spraying pressure and flow, a shift from passive obstacle avoidance to proactive prevention of accidental damage has been achieved. In particular, when encountering sensitive crops or non-operational areas, it can cut off the pesticide path or adjust the spray width.

[0026] This embodiment, by equipping an infrared supplementary lighting camera with an adaptive image enhancement algorithm, combined with RTK dynamic calibration and intelligent spraying control, achieves accurate identification of the work boundary and physical-level prevention of accidental damage, solving the spraying error problems caused by insufficient light, positioning drift and wind interference, and improving the safety and efficiency of agricultural plant protection operations.

[0027] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process in this embodiment, another method for controlling accidental damage during pesticide spraying by agricultural drones is provided. This method includes the following steps: S1. Real-time acquisition of raw image data of farmland and preprocessing to obtain preprocessed image data; The specific steps of step S1 are as follows: S11. Use an industrial camera with a resolution not lower than the resolution threshold (e.g., a resolution threshold of 1080P) to collect raw image data of farmland in real time at a preset acquisition frequency (e.g., 10-15 frames / second), covering the work area and the surrounding preset radius range (e.g., not less than 50m). S12. When the ambient light is below a preset threshold, turn on the infrared fill light for auxiliary illumination of the industrial camera; S13. An adaptive filtering algorithm is used to denoise the original image data;

[0028] in, Given a sequence of input pixel values, For adaptively updated filter coefficients, N This represents the filter order (e.g., 5). These are the pixel values ​​after noise reduction; S14. Use an image enhancement algorithm to adjust the contrast and brightness of the denoised image; The image enhancement algorithm uses linear brightness transformation:

[0029] in, For the input image at pixel position grayscale value, This is the contrast gain coefficient (e.g., 1.2). This is the brightness offset (e.g., 30). The enhanced grayscale value; S15. Perform distortion correction on the enhanced image and output the preprocessed image data; The specific steps of step S15 are as follows: S151. Obtain the intrinsic parameter matrix and distortion coefficients of the camera, wherein the distortion coefficients include radial distortion coefficients. and tangential distortion coefficient ; S152. For each pixel in the enhanced image Calculate the corresponding normalized coordinates and radius ; S153. Calculate the corrected coordinates according to the following distortion correction formula. :

[0030] S154. Map the corrected coordinates back to the pixel coordinate system, and obtain the distorted image pixel values ​​through interpolation (such as bilinear interpolation); S2. Based on the preprocessed image data, a pre-trained deep learning recognition model is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area, generating boundary coordinate data and crop type recognition results; The specific steps of step S2 are as follows: S21. Input the preprocessed image data into a pre-trained deep learning recognition model; S22. Deep learning recognition models perform inference on input image data: The encoder part of the model uses layer-by-layer convolution and downsampling operations to extract shallow features from deep features in the image; the shallow features include the edge and corner information of crop leaves, and the deep features include the shape, color and texture semantic information of crop leaves. The decoder part of the model uses upsampling and skip connection operations to gradually restore the deep features to the original image size and outputs a multi-channel feature map with the same size as the input image. Each channel corresponds to a class, and the value of each pixel in each channel represents the unnormalized score of the pixel belonging to the corresponding class. A class probability vector is output for each pixel. The categories include target crops, non-target crops, sensitive crops, and background, with weeds and stones being classified as non-target crops or background to achieve filtering. S23. Apply the Softmax function along the channel direction to the multi-channel feature map, convert the channel score of each pixel into a category probability vector, and take the category corresponding to the highest probability as the final category label of this pixel, thereby generating a pixel-level segmentation map; among them, the pixels corresponding to weeds and stones are classified as non-target crops or background to achieve automatic filtering; S24. Perform connected component analysis on the pixel-level segmentation map, extract the connected component boundary contour of the target crop region, convert the pixel coordinates in the connected component boundary contour into a set of boundary coordinate points in the world coordinate system, generate boundary coordinate data, and output the crop type identification result corresponding to each boundary coordinate point, generating crop type identification result data. Deep learning models employ convolutional neural networks (such as the U-Net or SegNet architectures), and the specific training process is as follows: SS1. Collect farmland images under various lighting and weather conditions (such as sunny, cloudy, evening, and backlighting), and annotate each collected farmland image at the pixel level. The labeling categories include target crop, non-target crop, sensitive crop, and background. Use the annotated farmland images and corresponding category labels to construct a training sample set. SS2. Perform data augmentation on the training sample set. The data augmentation includes geometric transformation and color transformation (such as random rotation, random scaling, random cropping, horizontal flipping, vertical flipping, brightness adjustment and contrast adjustment) to expand the augmented samples to a preset multiple (such as 5-10 times) of the original sample number. SS3. Set training hyperparameters, including batch size (e.g., 8 or 16), initial learning rate (e.g., 0.001), learning rate decay strategy, and total number of training epochs; For example, the learning rate decay strategy multiplies the learning rate by a preset decay factor (e.g., 0.9) every preset decay epochs (e.g., 10 training epochs). SS4. The following cross-entropy loss function is used to calculate the difference between the model prediction and the true label:

[0031] in, Total number of pixels; Total number of categories; Let c be the true label indicator of the i-th pixel. If the true category of the i-th pixel is c, then... The value is 1, otherwise The value is 0; The probability that the i-th pixel belongs to class c as predicted by the model; SS5 uses the Adam optimizer to iteratively update the model parameters. The iterative update process is as follows: Define the number of iterations Where T is the total number of training rounds multiplied by the number of batches per iteration; initialize the first-order moment estimate. Second-order moment estimation : In step t, calculate the gradient of the current batch. Then update the first-order moment estimate and the second-order moment estimate according to the following formula:

[0032] in, The exponential decay rate is estimated by the first moment (e.g., 0.9). The exponential decay rate (e.g., 0.999) is the estimate of the second moment. Calculate the estimate after bias correction:

[0033] Update the model parameters according to the following formula. :

[0034] in, The learning rate (such as the initial learning rate set in S53). Numerical stability constant (e.g.) ); SS6. After each training round, evaluate the model using the validation set and calculate the evaluation metrics on the validation set (such as average intersection-over-union ratio and overall recognition accuracy). Stop training early when the validation set loss no longer decreases after a preset number of early stop rounds (such as 5 rounds). S3. Obtain high-precision positioning data through RTK positioning, and use the high-precision positioning data to dynamically calibrate the boundary coordinate data to generate calibrated boundary coordinate data; The specific steps of step S3 are as follows: S31. The positioning data is obtained by the airborne RTK positioning module at a preset first update frequency (e.g., not less than 10Hz), and the positioning accuracy of the RTK positioning module is higher than the preset positioning accuracy threshold (e.g., 10cm). S32. Synchronously acquire auxiliary positioning data (such as GPS / BeiDou dual-mode positioning data); S33. The following dynamic boundary calibration algorithm is used to correct the boundary coordinate data:

[0035] in, For the identified boundary coordinate vector, For RTK positioning coordinate vectors, For auxiliary positioning coordinate vectors (such as GPS / BeiDou dual-mode positioning coordinate vectors). The calibrated boundary coordinate vector; S34. Output the calibrated boundary coordinate data and enable manual auxiliary calibration via the ground control terminal; S35. During the flight of the UAV, the boundary coordinates are updated in real time cyclically at a preset second update frequency (e.g., not less than 20Hz): The calibrated boundary coordinate data output from the previous loop step S33 is used as the identified boundary coordinate data in the current loop step S33. Steps S31 to S33 are repeated to output the calibrated boundary coordinate data of the current loop and use it as the input for the next loop, thereby realizing continuous dynamic calibration of boundary coordinate data. S4. Based on the calibrated boundary coordinate data and crop type identification results, dynamically adjust the spraying parameters and trigger corresponding early warning signals; The specific steps of step S4 are as follows: S41. Receive the calibrated boundary coordinate data and crop type identification results; S42. When the current area is identified as a non-operational area or a sensitive crop area, a stop spraying command is sent and an audible and visual warning signal is triggered; S43. When the current area is identified as both the work area and the target crop area, the spraying flow rate, spraying pressure, and spray width are dynamically adjusted according to real-time environmental parameters, and a safety isolation zone is set up inside the work boundary; The specific steps of step S43 are as follows: S431. Acquire real-time environmental parameters, including wind speed data, which is collected by an airborne wind speed sensor; S432. Dynamically adjust spraying pressure based on wind speed data; When the wind speed does not exceed the first wind speed threshold (e.g., 2 m / s), the spraying pressure is set to the first pressure value (e.g., 0.3 MPa). When the wind speed exceeds the first wind speed threshold but does not exceed the second wind speed threshold (e.g., 4 m / s), the spraying pressure is set to the second pressure value (e.g., 0.4 MPa); the second pressure value is greater than the first pressure value. Spraying operations shall be suspended when the wind speed exceeds the second wind speed threshold. S433. Extract the boundary coordinates of sensitive crop areas from the crop type identification results, and extract the operation boundary coordinates from the calibrated boundary coordinate data; S434. Calculate the distance between the boundary of the sensitive crop area and the operation boundary to obtain the distance data between the sensitive crop and the operation boundary; S435. Dynamically set the width of the safety isolation zone based on the distance data; If the distance between the sensitive crop and the work boundary is less than the preset distance threshold (e.g., 3m), the width of the safety isolation zone is set to the first width (e.g., 10 meters). If the distance between the sensitive crop and the work boundary is greater than or equal to a preset distance threshold (e.g., 3m), the width of the safety isolation zone is set to the second width (e.g., 5 meters); the first width is greater than the second width. S436. Based on the adjusted spraying pressure and the set width of the safety isolation zone, control the spraying actuator to perform spraying operations at the corresponding pressure inside the safety isolation zone; S44. Perform spraying operations using anti-drift nozzles to ensure that the pesticide solution does not exceed the calibrated boundary coordinate range.

[0036] The following is a detailed explanation of the technical solution of this application through a specific field operation. The plot of land in a rice-growing area is irregularly shaped and covers approximately 50 mu (about 3.3 hectares). Adjacent plots are planted with the sensitive crop strawberries. Weather on the day of the operation: sunny turning cloudy, afternoon wind speed approximately 3 m / s, and weak sunlight in the evening.

[0037] The drone is a six-rotor agricultural drone equipped with a 1080P high-definition camera (with infrared illumination), an RTK positioning module (positioning accuracy of 2cm), a wind speed sensor, anti-drift nozzles, and a spraying actuator.

[0038] Homework process: Perform image acquisition and preprocessing: The drone flies along the planned route at a speed of 5 m / s; the camera captures images at 12 frames per second, covering the work area and a surrounding 60m range; when the light sensor detects that the illuminance is below 200 lux, it automatically turns on the infrared supplementary light; An adaptive filtering algorithm (filter order N=5) is used to remove noise, followed by linear brightness transformation (…). =1.2, =30) Enhance contrast.

[0039] Perform distortion correction processing, given the camera distortion coefficients. =-0.28, =0.12, =0.001, =0.0005; Correct each pixel according to the formula in step S153, and then reconstruct the distortion-free image using bilinear interpolation. Output the preprocessed image data; Perform deep learning recognition: The preprocessed images were input into the U-Net model (which has been trained for 100 epochs on 80,000 labeled images including sunny, cloudy, dusk, and backlighting scenes, with a batch size of 8, an initial learning rate of 0.001, and a final validation accuracy of 98.5%). The model inference is performed, the encoder extracts features such as the edge, shape and color of rice leaves, and the decoder outputs a multi-channel feature map; after softmax, a segmentation map is obtained, in which rice is green, strawberries are red, weeds are yellow, and the background is black; Connectivity analysis was used to extract the boundary contour of the rice-growing region and convert it into a set of boundary coordinate points in the world coordinate system. Simultaneously output the mask coordinates of the strawberry region; Perform dynamic calibration: The RTK module outputs the current position at 20Hz. (Accuracy 2cm), GPS / BeiDou output (Accuracy 50cm); Perform calibration at each boundary point: ; Simultaneously, a cyclic update is initiated, using the boundary from the previous calibration as the current identification boundary, and continuously updating at a frequency of 20Hz, achieving an accuracy of 3cm after calibration; when the drone flies close to the corner of the plot, the ground control terminal operator manually fine-tunes the two coordinates of the boundary point to further correct the irregular boundary; Implement spraying control and early warning: The control module receives the calibrated boundary and crop type data; When the drone flies to within 3 meters of the strawberry area boundary, the system identifies the sensitive crop, immediately stops spraying, and issues an audible and visual alarm, such as a flashing red LED light and a "beep-beep-beep" sound from the buzzer. Once the drone enters the rice paddy area, the spraying pressure is adjusted to 0.4 MPa based on the real-time wind speed (e.g., 3 m / s), with a first pressure threshold of 2 m / s and a second pressure threshold of 4 m / s. Meanwhile, because the strawberry boundary distance is only 2.8m (less than the preset threshold of 3m), the width of the safety isolation zone is set to 10m (otherwise it is 5m); the spraying equipment sprays pesticide through the anti-drift nozzle at a pressure of 0.4MPa inside the isolation zone, with a droplet diameter of about 150μm. The actual measured drift distance is no more than 2m and it does not enter the strawberry field. After the entire rice paddy was treated, there were no incidents of pesticides accidentally damaging crops in neighboring fields, and the efficiency of the operation was improved.

[0040] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0041] like Figure 2 As shown, the following are embodiments of the agricultural drone spraying accidental injury control system provided in this disclosure. This system and the agricultural drone spraying accidental injury control methods in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the agricultural drone spraying accidental injury control system, please refer to the embodiments of the above agricultural drone spraying accidental injury control methods.

[0042] The system includes: The image acquisition and preprocessing module is used to acquire raw image data of farmland in real time and perform preprocessing to obtain preprocessed image data; The deep learning recognition module is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area based on pre-processed image data using a pre-trained deep learning recognition model, and to generate boundary coordinate data and crop type recognition results. The dynamic calibration module is used to acquire high-precision positioning data through RTK positioning, and to use the high-precision positioning data to dynamically calibrate the boundary coordinate data, generating calibrated boundary coordinate data. The spraying control and early warning module is used to dynamically adjust spraying parameters and trigger corresponding early warning signals based on the calibrated boundary coordinate data and crop type identification results.

[0043] This embodiment achieves all-weather, precise operation of UAVs in complex farmland environments by leveraging the interactive collaboration of image acquisition and preprocessing modules, deep learning recognition modules, dynamic calibration modules, and spraying control and early warning modules, as well as the integration of deep learning image segmentation and RTK high-precision positioning, combined with an environment-adaptive dynamic response mechanism. This avoids accidental damage caused by spraying drift.

[0044] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for controlling accidental damage during pesticide spraying by agricultural drones, characterized in that, Includes the following steps: S1. Real-time acquisition of raw image data of farmland and preprocessing to obtain preprocessed image data; S2. Based on the preprocessed image data, a pre-trained deep learning recognition model is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area, generating boundary coordinate data and crop type recognition results; S3. Obtain high-precision positioning data through RTK positioning, and use the high-precision positioning data to dynamically calibrate the boundary coordinate data to generate calibrated boundary coordinate data; S4. Based on the calibrated boundary coordinate data and crop type identification results, dynamically adjust the spraying parameters and trigger corresponding early warning signals.

2. The method for controlling accidental damage during pesticide spraying by agricultural drones according to claim 1, characterized in that, The specific steps of step S1 are as follows: S11. An industrial camera with a resolution not lower than the resolution threshold is used to collect raw image data of farmland in real time at a preset acquisition frequency, covering the work area and the surrounding preset radius range. S12. When the ambient light is below a preset threshold, turn on the infrared fill light for auxiliary illumination of the industrial camera; S13. An adaptive filtering algorithm is used to denoise the original image data; S14. Use an image enhancement algorithm to adjust the contrast and brightness of the denoised image; S15. Perform distortion correction on the enhanced image and output the preprocessed image data.

3. The method for controlling accidental damage during agricultural drone spraying according to claim 2, characterized in that, The specific steps of step S15 are as follows: S151. Obtain the intrinsic parameter matrix and distortion coefficients of the camera, wherein the distortion coefficients include radial distortion coefficients. and tangential distortion coefficient ; S152. For each pixel in the enhanced image Calculate the corresponding normalized coordinates and radius ; S153. Calculate the corrected coordinates according to the following distortion correction formula. : S154. Map the corrected coordinates back to the pixel coordinate system, and obtain the distorted image pixel values ​​through interpolation.

4. The method for controlling accidental damage during agricultural drone spraying according to claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Input the preprocessed image data into a pre-trained deep learning recognition model; S22. Deep learning recognition models perform inference on input image data: The encoder part of the model uses layer-by-layer convolution and downsampling operations to extract shallow features from deep features in the image; the shallow features include the edge and corner information of crop leaves, and the deep features include the shape, color and texture semantic information of crop leaves. The decoder part of the model uses upsampling and skip connection operations to gradually restore the deep features to the original image size and outputs a multi-channel feature map with the same size as the input image. Each channel corresponds to a class, and the value of each pixel in each channel represents the unnormalized score of the pixel belonging to the corresponding class. A class probability vector is output for each pixel. The categories include target crops, non-target crops, sensitive crops, and background, with weeds and stones being classified as non-target crops or background to achieve filtering. S23. Apply the Softmax function along the channel direction to the multi-channel feature map, convert the channel score of each pixel into a category probability vector, and take the category corresponding to the highest probability as the final category label of this pixel, thereby generating a pixel-level segmentation map; among them, the pixels corresponding to weeds and stones are classified as non-target crops or background to achieve automatic filtering; S24. Perform connected component analysis on the pixel-level segmentation map, extract the connected component boundary contours of the target crop region, convert the pixel coordinates in the connected component boundary contours into a set of boundary coordinate points in the world coordinate system, generate boundary coordinate data, and output the crop type identification result corresponding to each boundary coordinate point, generating crop type identification result data.

5. The method for controlling accidental damage during pesticide spraying by agricultural drones according to claim 4, characterized in that, The deep learning model uses a convolutional neural network, and the specific training process is as follows: SS1. Collect farmland images under several lighting and weather scenarios, and annotate each collected farmland image at the pixel level. The labeling categories include target crop, non-target crop, sensitive crop, and background. Use the annotated farmland images and corresponding category labels to construct a training sample set. SS2. Perform data augmentation on the training sample set, the data augmentation including geometric transformation and color transformation, and expand the augmented samples to a preset multiple of the original sample number; SS3. Set training hyperparameters, including batch size, initial learning rate, learning rate decay strategy, and total number of training epochs; SS4. The following cross-entropy loss function is used to calculate the difference between the model prediction and the true label: in, Total number of pixels; Total number of categories; Let c be the true label indicator of the i-th pixel. If the true category of the i-th pixel is c, then... The value is 1, otherwise The value is 0; The probability that the i-th pixel belongs to class c as predicted by the model; SS5. The Adam optimizer is used to iteratively update the model parameters; SS6. After each training round, evaluate the model using the validation set and calculate the evaluation metric on the validation set. When the validation set loss no longer decreases after a preset number of early stopping rounds, stop training early.

6. The method for controlling accidental damage during agricultural drone spraying according to claim 1, characterized in that, The specific steps of step S3 are as follows: S31. The positioning data is acquired by the airborne RTK positioning module at a preset first update frequency, wherein the positioning accuracy of the RTK positioning module is higher than a preset positioning accuracy threshold. S32. Synchronously acquire auxiliary positioning data; S33. The following dynamic boundary calibration algorithm is used to correct the boundary coordinate data: in, For the identified boundary coordinate vector, For RTK positioning coordinate vectors, To assist in locating coordinate vectors, The calibrated boundary coordinate vector; S34. Output the calibrated boundary coordinate data and enable manual auxiliary calibration via the ground control terminal.

7. The method for controlling accidental damage during pesticide spraying by agricultural drones according to claim 6, characterized in that, Step S3 also includes the following steps: S35. During the drone's flight, the boundary coordinates are updated in real time cyclically at a preset second update frequency: The calibrated boundary coordinate data output from the previous loop step S33 is used as the identified boundary coordinate data in the current loop step S33. Steps S31 to S33 are repeated to output the calibrated boundary coordinate data of the current loop and use it as the input for the next loop, thereby realizing continuous dynamic calibration of boundary coordinate data.

8. The method for controlling accidental damage during pesticide spraying by agricultural drones according to claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Receive the calibrated boundary coordinate data and crop type identification results; S42. When the current area is identified as a non-operational area or a sensitive crop area, a stop spraying command is sent and an audible and visual warning signal is triggered; S43. When the current area is identified as both the work area and the target crop area, the spraying flow rate, spraying pressure, and spray width are dynamically adjusted according to real-time environmental parameters, and a safety isolation zone is set up inside the work boundary; S44. Perform spraying operations using anti-drift nozzles to ensure that the pesticide solution does not exceed the calibrated boundary coordinate range.

9. The method for controlling accidental damage during pesticide spraying by agricultural drones according to claim 8, characterized in that, The specific steps of step S43 are as follows: S431. Acquire real-time environmental parameters, including wind speed data, which is collected by an airborne wind speed sensor; S432. Dynamically adjust spraying pressure based on wind speed data; When the wind speed does not exceed the first wind speed threshold, the spraying pressure is set to the first pressure value; When the wind speed exceeds the first wind speed threshold but does not exceed the second wind speed threshold, the spraying pressure is set to the second pressure value; the second pressure value is greater than the first pressure value. Spraying operations shall be suspended when the wind speed exceeds the second wind speed threshold. S433. Extract the boundary coordinates of sensitive crop areas from the crop type identification results, and extract the operation boundary coordinates from the calibrated boundary coordinate data; S434. Calculate the distance between the boundary of the sensitive crop area and the operation boundary to obtain the distance data between the sensitive crop and the operation boundary; S435. Dynamically set the width of the safety isolation zone based on the distance data; If the distance between the sensitive crop and the work boundary is less than the preset distance threshold, the width of the safety isolation zone is set to the first width. If the distance between the sensitive crop and the work boundary is greater than or equal to a preset distance threshold, the width of the safety isolation zone is set to the second width; the first width is greater than the second width. S436. Based on the adjusted spraying pressure and the set width of the safety isolation zone, control the spraying actuator to perform spraying operations at the corresponding pressure inside the safety isolation zone.

10. A control system for accidental damage caused by pesticide spraying by agricultural drones, characterized in that, include: The image acquisition and preprocessing module is used to acquire raw image data of farmland in real time and perform preprocessing to obtain preprocessed image data; The deep learning recognition module is used to identify target crops, non-target crops, sensitive crops, and the boundaries between the work area and the non-work area based on pre-processed image data using a pre-trained deep learning recognition model, and to generate boundary coordinate data and crop type recognition results. The dynamic calibration module is used to acquire high-precision positioning data through RTK positioning, and to use the high-precision positioning data to dynamically calibrate the boundary coordinate data, generating calibrated boundary coordinate data. The spraying control and early warning module is used to dynamically adjust spraying parameters and trigger corresponding early warning signals based on the calibrated boundary coordinate data and crop type identification results.

Images