Garden autonomous variable guide type spraying control method based on intelligent algorithm

Abstract

The invention provides a garden autonomous variable guide type spraying control method based on an intelligent algorithm, and the method comprises the steps: S1, collecting the multi-source data of garden vegetation through a sensor, and generating a vegetation health state evaluation result; s2, processing the canopy image by using a semantic segmentation model, identifying a vegetation organ and a pest and disease damage area, and generating a spraying area mask graph; s3, inputting evaluation and segmentation results into a deep Q network, and dynamically adjusting a spraying decision in combination with historical data; s4, dividing a garden area, optimizing an equipment path and task allocation by adopting an ant colony algorithm, and ensuring that the operation is not repeated and omitted; and S5, transmitting the decision to spraying equipment, adjusting nozzle parameters to realize accurate spraying, monitoring the effect in real time, feeding back the effect to the network, and dynamically optimizing the spraying parameters. According to the invention, by fusing multi-source data and deep learning, the vegetation state is accurately evaluated, the spraying parameters are dynamically adjusted, the strategy is optimized, the spraying accuracy and efficiency are improved, and the method adapts to various garden environments and vegetation requirements.

Description

Technical Field

[0001] The invention relates to the field of garden vegetation maintenance and management, and in particular to a garden autonomous variable guided spraying control method based on an intelligent algorithm. Background Art

[0002] The maintenance and management of garden vegetation is an important part of garden landscape maintenance. Traditional garden spraying control methods mainly rely on manual experience or fixed-mode automated equipment. Although they can meet basic vegetation maintenance needs, they are insufficient in accuracy, adaptability, efficiency and real-time feedback. The severity of vegetation infection and pests cannot be visually judged intuitively. Manual reliance on experience to judge the pest-infested areas may lead to insufficient or excessive spraying. Automated equipment usually adopts fixed spraying modes, such as fixed spraying volume, spraying time and spraying range. It is impossible to dynamically adjust according to the actual state of the vegetation, and it is difficult to flexibly adjust the spraying parameters according to environmental conditions. When multiple devices work together, traditional methods lack effective path planning and task allocation mechanisms, which are prone to repeated spraying or missed areas.

[0003] Therefore, in response to the above problems, a garden autonomous variable guided spraying control method based on intelligent algorithm is proposed. Summary of the invention

[0004] In order to overcome the shortcomings of traditional garden spraying control methods that mainly rely on manual experience or fixed-mode automated equipment, such as insufficient accuracy, poor adaptability, low efficiency and insufficient real-time feedback, the present invention provides a garden autonomous variable guided spraying control method based on an intelligent algorithm.

[0005] The technical solution of the present invention is: a garden autonomous variable guided spraying control method based on intelligent algorithm, comprising: Step S1: collect canopy images, meteorological data, leaf surface humidity, chlorophyll content, soil volume water content and nutrient data of garden vegetation through high-definition cameras, environmental sensors, plant physiological sensors and soil sensors, and use Kalman filtering to embed them into a deep learning fusion model to fuse multi-source data to generate a comprehensive assessment result of vegetation health status, including health index, pest and disease risk level and nutritional status score; Step S2: Use the semantic segmentation model to perform pixel-level segmentation on the canopy image, identify vegetation organs and pest and disease areas, generate a mask map of the spraying area, and determine the spraying location and range, including pest and disease areas, nutrient-deficient areas, and abnormal growth areas; Step S3: Input the vegetation health status assessment results and semantic segmentation results into the deep Q network, combine the historical spraying effects and real-time feedback data, dynamically adjust the spraying amount, spraying time and spraying mode, and generate adaptive spraying decisions; Step S4: Divide the garden area into multiple operation units, optimize the path planning and task allocation of multiple devices using the ant colony algorithm, and achieve information sharing and collaboration among devices through wireless communication technology to ensure that there are no duplications or omissions in the operation units; Step S5: Transmit the adaptive spraying decision to the control system of the spraying device, adjust the nozzle angle, flow rate, and pressure, control the movement of the nozzle through the servo motor to achieve precise spraying and cover the target area. At the same time, monitor the spraying effect in real time through high-definition cameras and sensors, and feedback the data to the deep Q network to dynamically adjust the spraying parameters.

[0006] As a preferred embodiment of the present invention, the specific implementation manner of the deep learning fusion model includes: Step S1.1: Use the LSTM network to model the time series data. The LSTM network includes 2 hidden layers, with 64 neurons in each layer and a time step of 10, and is used to process the time series change characteristics of air temperature and humidity, photosynthetically active radiation, wind speed, soil volumetric water content, and nitrogen, phosphorus, and potassium content; Step S1.2: Use the convolutional neural network to extract the features of the canopy image data. The convolutional neural network includes 3 convolutional layers and 2 fully connected layers, and the convolutional kernel size is 3×3, and is used to extract the leaf density, pest and disease patch distribution, chlorophyll distribution, and leaf surface humidity of the vegetation canopy; Step S1.3: Fuse the time series features output by the LSTM network and the spatial features output by the convolutional neural network through the fully connected layer to generate a comprehensive evaluation result of the vegetation health status; Among them, the indicators of the evaluation result are: Growth potential score: Based on canopy coverage, leaf density, and chlorophyll content, the scoring range is 0 - 100, and is distinguished by 60 as the boundary. When the score ≥ 60, the higher the score, the better the growth state; when the score ≤ 60, the lower the score, the worse the growth state; Range of pest and disease infection index: Based on the pest and disease patch distribution and leaf surface humidity, the index range is 0 - 10, and is distinguished by 5 as the boundary. When the index range > 5, the higher the score, the greater the pest and disease risk; when the index range < 5, the lower the score, the smaller the pest and disease risk; Nutritional imbalance level: Based on soil nutrient content and leaf physiological characteristics, the level is divided into low, medium, and high, indicating the nutritional status of the vegetation.

[0007] As a preferred embodiment of the present invention, the specific implementation manner of the semantic segmentation model includes: Step S2.1: Use the U-Net model for canopy image segmentation, and add an attention module to the model to make the model pay more attention to important pixel regions; Step S2.2: Utilize the results of multi-source data fusion to capture features at different scales in different data results, and improve the segmentation accuracy of pest spots; Step S2.3: The decoder part uses a transposed convolutional layer for upsampling. During the training process, it relies on the image data collected daily by the high-definition camera for training, and uses the cross-entropy loss function as the optimization objective; Step S2.4: Augment the trained data by rotation, scaling, and flipping, and optimize the segmentation results, including removing small-area noise regions and filling hole regions.

[0008] As a preferred embodiment of the present invention, the attention module includes: A feature extraction sub-module that uses a convolutional neural network to extract image features, extracts multi-scale features from the canopy image, and captures the detailed information of the vegetation; An environment and task perception sub-module that combines environmental data and task requirements to generate a comprehensive condition vector for guiding the generation of attention weights; An attention generation sub-module that generates an attention weight map based on the feature map and the comprehensive condition vector, indicating the regions that need to be processed preferentially; An adaptive adjustment sub-module that dynamically adjusts the attention weights according to real-time feedback As a preferred embodiment of the present invention, the steps for generating an adaptive spraying decision are as follows: Step S3.1: Input the comprehensive evaluation result of the vegetation health status, the spraying area mask map, and environmental data; Step S3.2: Output the spraying area, spraying amount, spraying time, and spraying mode according to the output result; Step S3.3: Analyze the spraying amount and priority of each region based on the calculation result after the participation of the attention module, and adjust the spraying time and mode in combination with environmental data; Step S3.4: Generate an adaptive spraying decision, including the region, amount, time, and mode.

[0009] As a preferred embodiment of the present invention, the specific rules for generating an adaptive spraying decision are as follows: Dynamic adjustment of spraying parameters: Based on the vegetation health status and semantic segmentation results, if the growth potential score is low, increase the spraying amount of growth promoters, and vice versa; if the pest and disease infection index is high, increase the spraying amount of insecticides, and vice versa; if the nutrient imbalance level is high, increase the spraying amounts of the corresponding fertilizers nitrogen, phosphorus, and potassium; if the nutrient imbalance level is medium, appropriately reduce the spraying amount; if the nutrient imbalance level is low, further reduce it based on the medium level. For different nutrient imbalance levels, the operator makes corresponding settings and adjustments according to the vegetation type; Spraying time adjustment: Based on environmental data and semantic segmentation results, if the wind speed > 5 m / s, the spraying time is shortened to avoid drift; if the wind speed is less than 5 m / s, the spraying time is extended. If the light intensity > 1000 μmol / m² / s, spraying is selected in the early morning or evening to reduce evaporation loss; if the light intensity is less than 1000 μmol / m² / s, spraying is selected at noon or in the afternoon to improve the absorption effect. For large areas with pests and diseases, the spraying time is extended to ensure full coverage; for small areas, the spraying time is shortened to improve efficiency. Spraying mode adjustment: Based on the vegetation health status and semantic segmentation results, if the pest and disease infection index is high, the atomization spraying mode is adopted to improve the coverage uniformity; if the nutrient imbalance level is high, the drip irrigation mode is adopted to accurately supplement nutrients. For diseased leaves, the directional spraying mode is adopted to accurately cover the diseased area; for pest areas, the atomization mode is adopted to expand the coverage range.

[0010] As a preference of the present invention, the specific implementation manner of the ant colony algorithm includes: dividing the garden area into multiple operation units, each operation unit is represented as a node, the distance between nodes is calculated based on the equipment movement time, using the ant colony algorithm to optimize the path planning, the initial value of pheromone is 1, the evaporation coefficient is 0.1, and by dynamically adjusting the pheromone concentration, the equipment task allocation and path planning are optimized.

[0011] As a preference of the present invention, the specific implementation manner of the real-time monitoring and feedback includes: collecting the canopy image data of the spraying area in real time through a high-definition camera, the collection frequency is 1 frame per second, using the optical flow method to analyze the movement trajectory of the spraying droplets, calculating the droplet coverage range and uniformity, and feeding the monitoring data back to the deep Q network. The update frequency of the deep Q network is 1 time per minute, and the update method is batch gradient descent, and the batch size is 64.

[0012] By adopting the above technical solutions, the present invention has the following advantages: 1. The present invention accurately evaluates the health status, pest and disease risks and nutrient requirements of vegetation through multi-source data fusion and deep learning models, generates a comprehensive evaluation result of the vegetation health status, and relies on the actual state of the vegetation and environmental conditions to dynamically adjust the spraying parameters, generate an adaptive spraying decision, ensure the accuracy and efficiency of spraying, and continuously optimize the spraying strategy in this process to adapt to different garden environments and vegetation needs, significantly improving the spraying effect.

[0013] 2. The present invention realizes the intelligence and automation of garden spraying based on machine vision and intelligent algorithms, reduces manual intervention and labor costs. At the same time, through technologies such as deep learning, reinforcement learning and path optimization, the system can ensure the coordinated work between devices, avoid repeated spraying and missed areas, improve the overall operation efficiency, and calculate the droplet coverage range and uniformity to ensure the uniformity and coverage of spraying. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 FIG. is a flowchart of the steps of a method for controlling autonomous variable-guided spraying in a garden based on an intelligent algorithm according to the present invention.

[0015] Figure 2 FIG. is a flowchart of the steps of a specific implementation manner of a deep learning fusion model according to the present invention.

[0016] Figure 3 FIG. is a flowchart of the steps of a specific implementation manner of a semantic segmentation model according to the present invention.

[0017] Figure 4 FIG. is a schematic structural diagram of an attention module according to the present invention.

[0018] Figure 5 FIG. is a flowchart of the steps of generating an adaptive spraying decision according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] Reference to embodiments herein means that a particular feature, structure, or characteristic described in connection with the embodiments can be included in at least one embodiment of the present invention. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive of other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0020] A method for controlling autonomous variable-guided spraying in a garden based on an intelligent algorithm, as Figure 1 shown, includes: Step S1: Collect canopy image data of garden vegetation at a frequency of 1 frame per second through a high-definition camera with a resolution of 4K and a wide-angle lens to cover a larger range. Collect meteorological data through environmental sensors, including air temperature and humidity, photosynthetically active radiation, and wind speed data. The temperature sensor has an accuracy of ±0.5°C, the humidity sensor has an accuracy of ±2%RH, the photosynthetically active radiation sensor has a measurement range of 0 - 2000 μmol / m² / s, and the wind speed sensor has a measurement range of 0 - 20 m / s. The data collection frequency is once per minute. Collect leaf surface humidity and relative chlorophyll content data through plant physiological sensors. The leaf surface humidity sensor has a measurement range of 0 - 100%RH, and the chlorophyll sensor uses the spectral reflection method with a measurement accuracy of ±0.1 SPAD. Collect soil volumetric water content and nutrient data through soil sensors. The soil volumetric water content sensor has a measurement depth of 0 - 30 cm and a measurement accuracy of ±2%, and the nutrient sensor can detect the contents of nitrogen, phosphorus, and potassium with a measurement accuracy of ±0.1 mg / kg. Use a multi-modal data fusion algorithm to fuse the canopy image data, meteorological data, plant physiological data, and soil data. Specifically, use the Kalman filter algorithm to filter the noise of real-time data and combine it with a deep learning fusion model to generate a comprehensive evaluation result of the vegetation health status. The evaluation result includes a vegetation health index, a pest and disease risk level, and a nutrition status score; Step S2: Use a semantic segmentation model to perform pixel-level segmentation on the canopy image. The semantic segmentation model uses the U-Net architecture, including 5 layers of encoders and 5 layers of decoders, with a convolution kernel size of 3×3. Identify different organs of the vegetation (such as leaves, stems, and flowers) and pest and disease areas through the semantic segmentation model, and generate a mask map of the spraying area. Determine the specific location and range to be sprayed according to the mask map. The spraying area includes pest and disease areas, nutrient deficiency areas, and growth anomaly areas; Step S3: Input the comprehensive evaluation result of the vegetation health status and the semantic segmentation result into a reinforcement learning model. The reinforcement learning model uses a deep Q network, and the network structure includes 3 layers of fully connected layers, with the number of neurons in each layer being 128, 64, and 32 respectively, a learning rate of 0.001, and a discount factor of 0.9. The reinforcement learning model dynamically adjusts the spraying parameters, including the spraying volume (range: 0 - 10 L / min), spraying time (range: 0 - 60 seconds), and spraying mode (such as atomization, drip irrigation, or directional spraying), according to the historical spraying effect and real-time feedback data, and generates an adaptive spraying decision. The decision content includes the spraying area, spraying volume, spraying time, and spraying mode; Step S4: Divide the garden area into multiple operation units. The division criteria include vegetation types, severity of pests and diseases, and soil nutrient distribution. Assign operation tasks to each spraying device. The task assignment uses the ant colony algorithm, with the number of ants being 50 and the number of iterations being 100. The optimization goal is to minimize the device movement distance and operation time. Achieve information sharing and collaboration among devices through wireless communication technology to ensure that there are no duplicates or omissions in the operation units, and the communication frequency is 1 time per second; Step S5: Transmit the adaptive spraying decision to the control system of the spraying device. The control system adjusts the angle of the nozzle (adjustment accuracy is ±1°), flow rate (control accuracy is ±0.1 L / min), and pressure (control accuracy is ±0.01 MPa) according to the decision to achieve precise spraying. Control the horizontal and vertical movement of the nozzle through a servo motor. The servo motor uses closed-loop control, with the horizontal movement accuracy being ±0.5° and the vertical movement accuracy being ±0.5°. During the spraying process, collect the canopy image data of the spraying area in real time through a high-definition camera, with the collection frequency being 1 frame per second. Use the optical flow method to analyze the movement trajectory of the spraying droplets, calculate the droplet coverage range and uniformity, and feedback the monitoring data to the reinforcement learning model. The model update frequency is 1 time per minute, and the update method is batch gradient descent, with the batch size being 64.

[0021] Specifically, the installation scheme for the high-definition camera and sensors is as follows: Fixed high-definition camera: Installed at high places such as pavilions, corridors, and street lamp poles in the garden to ensure that the camera can cover a large range of vegetation canopies, suitable for open areas or main landscape areas in the garden; For tall tree areas, the camera can be installed on the tree trunk or a special column can be set up for installation, and the height should be slightly higher than the vegetation canopy to ensure that the top and side information of the canopy can be captured; Install cameras on the boundary fences of the garden to ensure that the vegetation in the edge area can be covered, suitable for the edge areas or narrow areas of the garden.

[0022] The installation height of the camera should be adjusted according to the height of the vegetation and the size of the garden area: For low shrubs or lawns, the installation height of the camera is 2 - 3 meters; For medium-height trees, the installation height of the camera is 4 - 6 meters; For tall trees, the installation height of the camera is 8 - 10 meters.

[0023] The installation angle of the camera should ensure that the top and side of the vegetation canopy can be covered: For open areas, the camera can be installed at a top-down angle (30° - 45°); For narrow and long areas, the camera can be installed at a horizontal viewing angle (0° - 15°).

[0024] Mobile camera: For large garden areas or areas that are difficult to cover with fixed cameras, use a drone equipped with a camera for aerial photography; Sensor installation: The air temperature and humidity sensor is installed near the vegetation canopy, at a height of 1 - 2 meters from the ground, to ensure accurate measurement of the air temperature and humidity around the vegetation; The photosynthetically active radiation sensor is installed above the vegetation canopy to ensure direct sunlight reception and measurement of the light intensity received by the vegetation; The wind speed sensor is installed above or near the vegetation canopy to ensure accurate measurement of the wind speed around the vegetation; The plant physiological sensors include a leaf surface humidity sensor and a chlorophyll sensor. Representative leaves are selected for sensor installation in each vegetation area (such as trees, shrubs, lawns). The representative leaves should be selected at the middle position of the vegetation, avoiding leaves with too strong light at the top and insufficient light at the bottom. Multiple sensors are installed in each vegetation area to form a sensor cluster to ensure the comprehensiveness of data.

[0025] The soil volumetric water content sensor is installed in the soil, at a depth of 0 - 30 cm, to ensure measurement of the soil water content in the vegetation root zone; The soil nutrient sensor is installed in the soil, at a depth of 0 - 30 cm, to ensure measurement of the nitrogen, phosphorus, and potassium contents in the soil.

[0026] Specifically, the optical flow method estimates the motion of an object by analyzing the movement of pixels in consecutive frame images. The displacement of pixels in the image between time t and t + 1 is (u, v). The goal of the optical flow method is to solve the displacement vector (u, v) for each pixel. The basic equation of the optical flow method can be expressed as: , where is the brightness of the pixel (x, y) at time t. The specific steps are as follows: Use a high - definition camera to collect canopy image data of the spraying area at a frequency of 1 frame per second, and collect meteorological data through environmental sensors; Use threshold segmentation technology to detect droplets in the image; Use the optical flow method to calculate the movement trajectory of droplets; Map the movement trajectory of droplets onto a two - dimensional plane and calculate the coverage range of droplets; Divide the spraying area into grids and count the uniformity of droplet distribution; Feed the monitoring data back to the reinforcement learning model to dynamically adjust the spraying parameters.

[0027] It should be noted that the deep learning fusion model uses an LSTM network and a convolutional neural network to process time-series data and spatial data. During the process of processing time-series data, the LSTM network can better capture the trends and periodicities of data changes over time and remember information over long time spans. For example, the impact of meteorological changes in the past few days on the current vegetation status. The convolutional neural network can automatically extract spatial features in images through convolutional layers, pooling layers, and fully connected layers, and can efficiently extract local features in images, such as the leaf density of the vegetation canopy, the distribution of pest and disease patches, etc. When the convolutional kernels of the convolutional neural network slide on the image, they share parameters, greatly reducing the number of model parameters and the computational complexity.

[0028] As Figure 2 shown, the specific implementation method of the deep learning fusion model includes: Step S1.1: Use the LSTM network to model the time-series data. The LSTM network contains 2 hidden layers, with 64 neurons in each layer and a time step of 10, which is used to process the time-series change characteristics of air temperature and humidity, photosynthetically active radiation, wind speed, soil volumetric water content, and nitrogen, phosphorus, and potassium content; Step S1.2: Use the convolutional neural network to extract features from the canopy image data. The convolutional neural network contains 3 convolutional layers and 2 fully connected layers, and the convolutional kernel size is 3×3, which is used to extract the leaf density of the vegetation canopy, the distribution of pest and disease patches, chlorophyll distribution, and leaf surface humidity; Step S1.3: Fuse the time-series features output by the LSTM network and the spatial features output by the convolutional neural network through a fully connected layer to generate a comprehensive evaluation result of the vegetation health status; Among them, the indicators of the evaluation result are: Growth potential score: Based on canopy coverage, leaf density, and chlorophyll content, the scoring range is 0-100, and it is distinguished by 60 as the boundary. When the score ≥ 60, the higher the score, the better the growth state. When the score ≤ 60, the lower the score, the worse the growth state; Range of pest and disease infection index: Based on the distribution of pest and disease patches and leaf surface humidity, the index range is 0-10, and it is distinguished by 5 as the boundary. When the index range > 5, the higher the score, the greater the pest and disease risk. When the index range < 5, the lower the score, the smaller the pest and disease risk; Nutritional imbalance level: Based on soil nutrient content and leaf physiological characteristics, the level is divided into low, medium, and high, indicating the nutritional status of the vegetation.

[0029] It should be noted that as Figure 3 shown, the specific implementation method of the semantic segmentation model includes: Step S2.1: Use the U-Net model for canopy image segmentation, and add an attention module to the model to make the model pay more attention to important pixel regions; Step S2.2: Utilize the results of multi-source data fusion to capture features at different scales in different data results, and improve the segmentation accuracy of pest spots; In step S2.3, the decoder part uses a transposed convolutional layer for upsampling. During the training process, it relies on the image data collected daily by a high-definition camera for training, and uses the cross-entropy loss function as the optimization objective; Step S2.4: Augment the trained data by rotation, scaling, and flipping, and optimize the segmentation results, including removing small-area noise regions and filling hole regions.

[0030] Specifically, the core objective of the attention module is to dynamically identify the regions in the canopy image that need to be processed first, and combine the environmental conditions and spraying tasks to autonomously adjust the focus of the model. For example, Figure 4 As shown, the attention module can be divided into the following sub-modules: Feature extraction sub-module: Use a lightweight convolutional neural network to extract image features, extract multi-scale features from the canopy image, and capture the detailed information of the vegetation (such as leaves, branches, and pest and disease areas). The shallow features are used to identify details such as leaf edges and pest and disease spots, and the deep features are used to distinguish vegetation from the background; Environment and task perception sub-module: Combine environmental data and task requirements to generate a comprehensive condition vector, which is used to guide the generation of attention weights. Input environmental data (such as light and humidity), and generate an environmental feature vector through a fully connected layer. For example, when the light intensity is low, the model should pay more attention to the vegetation in the shaded area. Input task requirements (such as fertilization and pest control). For example, in the pest control task, the model should pay more attention to the areas with severe pest and disease damage, and generate a task feature vector through an embedding layer; Attention generation sub-module: Generate an attention weight map based on the feature map and the comprehensive condition vector, indicating the regions that need to be processed first. Perform global average pooling on each channel of the feature map to generate a channel descriptor, combine the channel descriptor with the comprehensive condition vector, and generate channel weights through a fully connected layer. For example, in the pest control task, the weights of the channels related to pests and diseases will be higher. Then perform max pooling and average pooling on the spatial dimension of the feature map to generate a spatial descriptor, combine the spatial descriptor with the comprehensive condition vector, and generate a spatial weight map through a convolutional layer. For example, when the light is weak, the weight of the shaded area will be higher. Then multiply the channel weights and the spatial weights to generate the final attention weight map; Adaptive adjustment sub-module: Dynamically adjust the attention weights according to real-time feedback (such as spraying effect and vegetation growth status); It should be noted that the application of the attention module in the system is as follows: dynamically evaluate the importance of image regions according to factors such as vegetation health status, pest and disease distribution, and canopy density, and adjust the attention weights in combination with the current environment (such as light and humidity) and spraying tasks (such as fertilization, pest control, and watering).

[0031] Specifically, the attention generation sub-module can be applied in the following scenarios: Suppose it is necessary to spray insecticide on a fruit tree in a garden. Use a high-definition camera to take an image of the fruit tree canopy. The module generates an attention weight map according to the current light intensity and the insecticide spraying task. The weight map indicates the areas with severe pest and disease damage, that is, the positions where the leaves are yellowing and the wormholes are obvious. The spraying equipment sprays the areas with severe pest and disease damage preferentially according to the weight map, and dynamically adjusts the parameters of the attention module by monitoring the spraying effect through sensors.

[0032] It should be noted that as Figure 5 shown, the steps to generate an adaptive spraying decision are as follows: Step S3.1: Input the comprehensive evaluation result of the vegetation health status, the spraying area mask map, and the environmental data; Step S3.2: Output the spraying area, spraying volume, spraying time, and spraying mode according to the output result; Step S3.3: Analyze the spraying volume and priority of each area according to the calculation result after the participation of the attention module, and adjust the spraying time and mode in combination with the environmental data; Step S3.4: Generate the final spraying decision, including the area, volume, time, and mode.

[0033] It should be noted that the specific rules for generating an adaptive spraying decision are as follows: Dynamically adjust the spraying parameters based on the vegetation health status and the semantic segmentation result: Growth promoter spraying volume: If the growth potential score of the leaves < 60, the spraying volume should be 5 - 8 L / min, and for every 10-point reduction, an additional 0.5 L / min is added; If the growth potential score of the leaves is between 60 - 80, the spraying volume is 3 - 5 L / min, and the spraying volume remains stable; If the growth potential score of the leaves > 80, the spraying volume is 1 - 3 L / min, and for every 10-point increase, 0.5 L / min is reduced; Insecticide spraying volume: If the pest and disease infection index > 5, the spraying volume is 7 - 10 L / min, and for every 1 increase, 1 L / min is added; If the pest and disease infection index is between 3 - 5, the spraying volume is 4 - 7 L / min, and the spraying volume remains stable; If the pest and disease infection index < 3, the spraying volume is 2 - 4 L / min, and for every 1 reduction, 1 L / min is reduced; Fertilizer spraying amount: If the nutrient imbalance level is high, spray according to the lacking fertilizer. The spraying amount of nitrogen fertilizer is 4 - 6 L / min, the spraying amount of phosphate fertilizer is 3 - 5 L / min, and the spraying amount of potassium fertilizer is 2 - 4 L / min; If the nutrient imbalance level is medium, spray according to the lacking fertilizer. The spraying amount of nitrogen fertilizer is 2 - 4 L / min, the spraying amount of phosphate fertilizer is 1 - 3 L / min, and the spraying amount of potassium fertilizer is 1 - 2 L / min; If the nutrient imbalance level is low, spray according to the lacking fertilizer. The spraying amount of nitrogen fertilizer is 1 - 2 L / min, the spraying amount of phosphate fertilizer is 0.5 - 1 L / min, and the spraying amount of potassium fertilizer is 0.5 - 1 L / min; For different nutrient imbalance levels, the specific parameters are set and adjusted by the operator according to the vegetation type.

[0034] Adjust the spraying time based on environmental data and semantic segmentation results: When the wind speed > 5 m / s, set the spraying time at 5 - 10 seconds, and reduce 1 second for every 1 m / s increase in wind speed. When the wind speed < 5 m / s, set the spraying time at 10 - 20 seconds, and increase 2 seconds for every 1 m / s decrease in wind speed; If the light intensity > 1000 μmol / m² / s, choose to spray in the early morning (6:00 - 8:00) or evening (16:00 - 18:00), 10 - 15 seconds each time, and reduce 1 second for every 200 μmol / m² / s increase in light intensity; if the light intensity < 1000 μmol / m² / s, choose to spray at noon (12:00 - 14:00) or afternoon (14:00 - 16:00), 15 - 20 seconds each time, and increase 1 second for every 200 μmol / m² / s decrease in light intensity; Adjust the spraying mode based on the vegetation health status and semantic segmentation results under the following specific conditions: When the pest and disease infection index > 5, adopt the atomization spraying mode, with the atomization particle size of 50 - 100 μm, and reduce 10 μm for every 1 - level increase; When the nutrient imbalance level is high, adopt the drip irrigation mode, with the drip irrigation flow rate of 2 - 4 L / min; For the diseased leaf area, adopt the directional spraying mode, with the spraying angle of 30° - 45°, and adjust the angle according to the leaf area, increasing 5° for every 10 ㎡ increase in the diseased leaf area; For the pest - infested area, adopt the atomization spraying mode, with the atomization particle size of 100 - 150 μm, and increase 10 μm for every 5 ㎡ increase in the pest - infested area.

[0035] The deep Q-network is used to solve decision-making problems. It can dynamically adjust spraying parameters according to the environmental state and real-time feedback data, so as to achieve adaptive decision-making. The specific implementation method is as follows: The input is the comprehensive evaluation result of the vegetation health status and the semantic segmentation result, and the output is the spraying parameters (spraying volume, spraying time, spraying mode). The experience replay mechanism is used to store historical data. The size of the replay buffer is 10,000 records, and 128 records are randomly selected for each training. The double-network structure of the target network and the online network is used. The target network is updated every 100 steps, and the online network is updated every step. And the deep deterministic policy gradient is embedded to directly process the continuous actions sent by the system to the device, improving the action accuracy.

[0036] The specific implementation method of the Kalman filter algorithm includes: using the Kalman filter to filter the noise of the sensor data. The state vector includes air temperature, humidity, photosynthetically active radiation, and wind speed, and the observation vector is the original sensor data. The state transition matrix and the observation matrix of the Kalman filter are trained based on historical data.

[0037] The specific implementation method of the ant colony algorithm includes: dividing the garden area into multiple operation units, each operation unit is represented as a node, and the distance between nodes is calculated based on the device movement time. The ant colony algorithm is used to optimize the path planning. The initial value of the pheromone is 1, and the evaporation coefficient is 0.1. By dynamically adjusting the pheromone concentration, the device task allocation and path planning are optimized to ensure that all operation units are covered and the total movement distance is the shortest. Specifically, the garden area is divided into multiple operation units, each operation unit is represented as a node, and each node contains the position coordinates (x, y), the vegetation health status (growth potential score, pest and disease infection index, nutrient imbalance level), and the spraying task requirements (spraying volume, spraying time, spraying mode). Record the position coordinates, vegetation health status, and spraying task requirements of each node, initialize the pheromone matrix, set the number of ants, evaporation coefficient, enhancement coefficient, and the number of iterations according to the number of devices, use the ant colony algorithm to generate the device movement path, select the shortest path as the optimal solution, and assign tasks to each device to ensure that all operation units are covered.

[0038] In summary, through technologies such as multi-source data fusion, deep learning, reinforcement learning, path optimization, and real-time monitoring, the present invention realizes the precise spraying control of garden vegetation. It can dynamically adjust the spraying parameters according to the health status, pest and disease risks, and nutrient requirements of the vegetation, ensuring the accuracy and efficiency of spraying. At the same time, through the real-time monitoring and feedback mechanism, the system can continuously optimize the spraying decision-making to adapt to different garden environments and vegetation requirements. This method has the advantages of intelligence, automation, and efficient resource utilization, and can significantly improve the efficiency and quality of garden management.

[0039] The above embodiments are provided for those skilled in the art to implement or use the present invention. Those skilled in the art can make various modifications or changes to the above embodiments without departing from the spirit of the present invention. Therefore, the protection scope of the present invention is not limited by the above embodiments, but should be the maximum scope that conforms to the innovative features mentioned in the claims.

Claims

1. A garden autonomous variable-guided spraying control method based on intelligent algorithms, characterized in that, It includes: Step S1: Collect the canopy images, meteorological data, leaf surface humidity, chlorophyll content, soil volumetric water content and nutrient data of garden vegetation through high-definition cameras, environmental sensors, plant physiological sensors and soil sensors. Use the Kalman filter embedded in the deep learning fusion model to fuse the multi-source data and generate a comprehensive evaluation result of the vegetation health status, including a health index, a pest and disease risk level and a nutrition status score. Step S2: Use a semantic segmentation model to perform pixel-level segmentation on the canopy images, identify the vegetation organs and pest and disease areas, generate a mask map of the spraying area, and determine the spraying position and scope, including pest and disease areas, nutrient deficiency areas and growth anomaly areas. Step S3: Input the vegetation health status evaluation result and the semantic segmentation result into the deep Q network, combine the historical spraying effect and the real-time feedback data, dynamically adjust the spraying amount, spraying time and spraying mode, and generate an adaptive spraying decision. Step S4: Divide the garden area into multiple operation units, use the ant colony algorithm to optimize the path planning and task allocation of multiple devices, and achieve information sharing and collaboration between devices through wireless communication technology to ensure that there are no repetitions and omissions in the operation units. Step S5: Transmit the adaptive spraying decision to the control system of the spraying device, adjust the nozzle angle, flow rate and pressure, control the movement of the nozzle through the servo motor to achieve precise spraying and cover the target area. At the same time, monitor the spraying effect in real time through the high-definition camera and sensors, and feedback the data to the deep Q network to dynamically adjust the spraying parameters.

2. A garden autonomous variable-guided spraying control method based on the intelligent algorithm according to claim 1, characterized in that, The specific implementation method of the deep learning fusion model includes: Step S1.1: Use the LSTM network to model the time series data. The LSTM network contains 2 hidden layers, with 64 neurons in each layer and a time step of 10, which is used to process the time series change characteristics of air temperature and humidity, photosynthetically active radiation, wind speed, soil volumetric water content and nitrogen, phosphorus and potassium content. Step S1.2: Use a convolutional neural network to extract the feature of the canopy image data. The convolutional neural network contains 3 convolutional layers and 2 fully connected layers, and the convolutional kernel size is 3×3, which is used to extract the leaf density, pest and disease patch distribution, chlorophyll distribution and leaf surface humidity of the vegetation canopy. Step S1.3: Fuse the time series features output by the LSTM network and the spatial features output by the convolutional neural network through the fully connected layer to generate a comprehensive evaluation result of the vegetation health status. Among them, the indicators of the evaluation result are: Growth potential score: Based on the canopy coverage rate, leaf density and chlorophyll content, the scoring range is 0-100, and it is distinguished by 60 as the boundary. In the state where the score ≥ 60, the higher the score, the better the growth state; in the state where the score ≤ 60, the lower the score, the worse the growth state. Pest and disease infection index range: Based on the pest and disease patch distribution and leaf surface humidity, the index range is 0-10, and it is distinguished by 5 as the boundary. In the state where the index range > 5, the higher the score, the greater the pest and disease risk; in the state where the index range < 5, the lower the score, the smaller the pest and disease risk. Nutritional imbalance level: Based on soil nutrient content and leaf physiological characteristics, the level is divided into low, medium, and high, indicating the nutritional status of vegetation.

3. A garden autonomous variable-guided spraying control method based on the intelligent algorithm according to claim 1, characterized in that, The specific implementation method of the semantic segmentation model includes: Step S2.1: Use the U-Net model for canopy image segmentation, and add an attention module to the model to make the model pay more attention to important pixel regions; Step S2.2: Utilize the results of multi-source data fusion to capture features at different scales in different data results, and improve the segmentation accuracy of pest spots; Step S2.3: The decoder part uses a transposed convolutional layer for upsampling. During the training process, rely on the image data collected daily by the high-definition camera for training, and use the cross-entropy loss function as the optimization target; Step S2.4: Use rotation, scaling, and flipping to enhance the trained data and optimize the segmentation results, including removing small-area noise regions and filling hole regions.

4. An intelligent algorithm-based autonomous variable-guided spraying control method for gardens according to claim 3, characterized in that, The attention module includes: Feature extraction sub-module, which uses a convolutional neural network to extract image features, extracts multi-scale features from the canopy image, and captures the detailed information of vegetation; Environment and task perception sub-module, which combines environmental data and task requirements to generate a comprehensive condition vector for guiding the generation of attention weights; Attention generation sub-module, which generates an attention weight map according to the feature map and the comprehensive condition vector, indicating the area that needs to be processed preferentially; Adaptive adjustment sub-module, which dynamically adjusts the attention weights according to real-time feedback A garden autonomous variable-guided spraying control method based on an intelligent algorithm according to claim 1, characterized in that the steps for generating an adaptive spraying decision are as follows: Step S3.1: Input the comprehensive evaluation result of the vegetation health status, the spraying area mask map, and environmental data; Step S3.2: Output the spraying area, spraying amount, spraying time, and spraying mode according to the output result; Step S3.3: Analyze the spraying amount and priority of each area according to the calculation result after the participation of the attention module, and adjust the spraying time and mode in combination with environmental data; Step S3.4: Generate an adaptive spraying decision, including area, amount, time, and mode.

5. A garden autonomous variable-guided spraying control method based on an intelligent algorithm according to claim 5, characterized in that, The specific rules for generating an adaptive spraying decision are as follows: Dynamic adjustment of spraying parameters: Based on the vegetation health status and semantic segmentation results, if the growth potential score is low, increase the spraying amount of growth promoters, otherwise decrease it; if the pest and disease infection index is high, increase the spraying amount of pesticides, otherwise decrease it; if the nutritional imbalance level is high, increase the spraying amounts of corresponding fertilizers nitrogen, phosphorus, and potassium; if the nutritional imbalance level is medium, appropriately reduce the spraying amount; if the nutritional imbalance level is low, further reduce it on the basis of the medium level. For different nutritional imbalance levels, operators make corresponding settings and adjustments according to the vegetation type; Spraying time adjustment: Based on environmental data and semantic segmentation results, if the wind speed > 5 m / s, the spraying time is shortened to avoid drift; if the wind speed is less than 5 m / s, the spraying time is extended. If the light intensity > 1000 μmol / m² / s, spraying is selected in the early morning or evening to reduce evaporation loss; if the light intensity is less than 1000 μmol / m² / s, spraying is selected at noon or in the afternoon to improve the absorption effect. For large areas with pests and diseases, the spraying time is extended to ensure full coverage; for small areas, the spraying time is shortened to improve efficiency. Spraying mode adjustment: Based on the vegetation health status and semantic segmentation results, if the pest and disease infection index is high, the atomization spraying mode is adopted to improve the coverage uniformity; if the nutrient imbalance level is high, the drip irrigation mode is adopted to accurately supplement nutrients. For diseased leaves, the directional spraying mode is adopted to accurately cover the diseased area; for pest areas, the atomization mode is adopted to expand the coverage range.

6. A garden autonomous variable-guided spraying control method based on the intelligent algorithm according to claim 1, characterized in that The specific implementation method of the ant colony algorithm includes: dividing the garden area into multiple operation units, each operation unit is represented as a node, the distance between nodes is calculated based on the equipment movement time, and the ant colony algorithm is used to optimize the path planning. The initial value of pheromone is 1, and the evaporation coefficient is 0.

1. By dynamically adjusting the pheromone concentration, the equipment task allocation and path planning are optimized.

7. A garden autonomous variable-guided spraying control method based on the intelligent algorithm according to claim 1, characterized in that, The specific implementation method of the real-time monitoring and feedback includes: real-time collecting the canopy image data of the spraying area through a high-definition camera, with a collection frequency of 1 frame per second. The optical flow method is used to analyze the movement trajectory of the spraying droplets, calculate the droplet coverage range and uniformity, and feedback the monitoring data to the deep Q network. The deep Q network is updated at a frequency of 1 time per minute, and the update method is batch gradient descent, with a batch size of 64.

Images