Agricultural pest prevention and control method and system based on deep learning

Abstract

The invention relates to the technical field of disease and pest prevention and control, and provides an agricultural disease and pest prevention and control method and system based on deep learning. Firstly, multi-source data of a citrus planting area is collected. Then, after space-time alignment and radiation correction are carried out on the collected original data, edge computing nodes are input, and standardized preprocessing data are generated; thirdly, constructing a dynamic agronomic knowledge graph, and performing association modeling on the data; and then, inputting the standardized image data into a Huanglongbing recognition model based on a convolutional neural network, optimizing a disease recognition result through a deep reinforcement learning algorithm, and generating a pesticide application scheme according to the knowledge graph. Finally, according to population density change data collected after pesticide application, environment capacity parameters in the agronomic knowledge graph are updated, and a closed loop feedback optimization mechanism is formed to improve the prevention and control effect. According to the method, accurate disease monitoring, prevention and control can be realized in a complex agricultural environment, and the decision accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of pest control, and more specifically, to a method and system for agricultural pest control based on deep learning. Background Art

[0002] Citrus is one of the most important fruit trees in the world, especially in agricultural production in tropical and subtropical regions. Citrus trees are highly sensitive to environmental changes and pests and diseases, especially Huanglongbing, which has become one of the main factors restricting citrus production and quality. Huanglongbing is a plant disease caused by bacteria, usually transmitted by the transmission medium, Diaphorina citri, which causes the leaves of citrus trees to turn yellow, the branches to dry up, and eventually the entire fruit tree to die. The occurrence of this disease has seriously affected the production and economic benefits of citrus. Since the early symptoms of Huanglongbing are difficult to identify and the treatment methods are limited, early detection and prevention and control have become the key to ensuring the sustainable development of the citrus planting industry. In order to effectively control the spread of Huanglongbing, scientists have conducted a lot of research on early disease monitoring, transmission models and prevention and control programs, striving to make breakthroughs in accurate identification and effective prevention and control. As stated in document 1 (“Research Progress in Deep Learning for Detection and Identification of Plant Leaf Diseases” Shao Mingyue, Zhang Jianhua, et al., 2022), rapid and accurate identification of plant diseases through early diagnosis technology has become the key to improving crop productivity and disease prevention and control.

[0003] At present, with the development of precision agriculture and Internet of Things technology, more and more intelligent means are being applied to the monitoring and prevention of agricultural diseases. Unmanned aerial vehicles, multispectral imaging, environmental sensors and other technical means are widely used in the monitoring of agricultural diseases. These technologies can not only realize the rapid detection of large areas of farmland, but also obtain accurate agricultural data in real time. As mentioned in document 2 ("Application of UAV Multispectral Imaging in Monitoring Damage of Rice Leaf Folder" Guo Mingqi, Bao Yunxuan, et al., 2023), by combining multiple sensor data and image processing technology, good results can be achieved in rice disease monitoring, but how to improve its adaptability in complex environments is still a problem that needs in-depth research.

[0004] In previous studies, deep learning models have shown good potential in the automatic detection and classification of crop diseases. For example, some studies have identified rice diseases by combining convolutional neural networks (CNNs) with multispectral imaging technology, and achieved high classification accuracy. For example, in document 3 (Upadhyay SK, et al., A novel approach for rice plant diseases classification with deep convolutional neural network. International Journal of Information Technology, 2021), a CNN-based rice disease classification method was proposed, which used convolutional neural networks to extract features from rice leaf lesions and achieved a recognition accuracy of 99.7%. However, existing studies mainly focus on the identification of a single crop or disease, and usually do not fully consider the impact of dynamic changes in agronomic background factors such as soil quality and climate change. Therefore, although deep learning has made some progress in disease identification, its practical application in real environments still faces many challenges.

[0005] At present, the application of agronomic knowledge graphs has become a new method to solve agricultural management problems. By constructing a dynamic agronomic knowledge graph, data from different fields can be associated to provide decision support for agricultural production. However, when constructing dynamic knowledge graphs, existing technologies often lack a real-time data feedback mechanism, resulting in insufficient accuracy and timeliness of prevention and control decisions when facing complex environments and changes in growth cycles. In addition, most of the existing deep learning-based prevention and control systems only focus on single disease identification or insect population density monitoring, and have not yet fully combined with dynamic agronomic knowledge graphs for comprehensive analysis, resulting in inaccuracy in prevention and control measures in practice. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for agricultural pest control based on deep learning, which uses an unmanned aerial vehicle equipped with a multispectral imaging device, a depth sensor and an environmental sensor, combined with a dynamic agronomic knowledge graph and a convolutional neural network model, to achieve accurate monitoring and prevention and control decisions for citrus Huanglongbing. The prevention and control decisions are optimized through a deep reinforcement learning algorithm, and a closed-loop optimization mechanism is formed in combination with environmental data feedback to dynamically adjust various parameters in the agronomic knowledge graph. This method solves the problem that the prevention and control system in the prior art lacks real-time feedback, accurate decision-making and dynamic adjustment capabilities, improves the efficiency and accuracy of prevention and control, and ensures the effective prevention and control of citrus Huanglongbing.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for controlling agricultural pests and diseases based on deep learning, characterized in that it comprises the following steps:

[0009] Step S1, collecting multi-source data of the citrus planting area by using a multispectral imaging device, a depth sensor and an environmental sensor of a fixed monitoring station carried by an unmanned aerial vehicle;

[0010] Step S2, input the collected raw data into the edge computing node after time-space alignment and radiation correction, use affine transformation to perform spatial registration of multispectral images and depth images, and generate standardized preprocessed data;

[0011] Step S3, construct a dynamic agronomic knowledge graph, and associate the psyllid population density dynamic model with the Huanglongbing disease transmission probability, the citrus psyllid population density threshold, the safe use range of pesticides, and the citrus growth cycle information through the graph database to build a model, wherein the psyllid population density dynamic model is constructed using the temperature-corrected Logistic equation, and its expression is:

[0012]

[0013] Where N(t) is the current psyllid population density, and the temperature correction coefficient r = k r ·exp(k T (TT 0 )), T is the real-time monitoring temperature value;

[0014] The initial environmental capacity parameter of the psyllid population density dynamic model is calculated by soil organic matter content, canopy density and photosynthetically active radiation, and the calculation formula is:

[0015]

[0016] In the formula, SOM is soil organic matter content; CC is canopy density; PAR is photosynthetically active radiation value, PAR ref is the reference value of photosynthetically active radiation;

[0017] The calculation formula for the Huanglongbing transmission probability is:

[0018]

[0019] Where P HLB (t) represents the probability of Huanglongbing spreading per unit time, β is the transmission rate coefficient, N th is the psyllid population density threshold, N(t) is the current psyllid population density;

[0020] Step S4, inputting the standardized preprocessed data into a Huanglongbing recognition model based on a convolutional neural network, the model dynamically adjusts the convolution kernel weight parameters through a deep reinforcement learning algorithm, the action space of the reinforcement learning algorithm is limited by the pesticide dosage constraint rules in the agronomic knowledge graph, and its reward function is the weighted sum of the lesion recognition accuracy and the knowledge graph rule matching degree;

[0021] Step S5, generating a pesticide application plan based on the output results of the Huanglongbing recognition model and the knowledge graph constraints;

[0022] Step S6, after applying the pesticide, the environmental capacity parameter information in the agronomic knowledge graph is fed back and updated through the collected data on changes in psyllid population density, forming a closed-loop optimization mechanism.

[0023] As a further solution of the present invention, the working band of the multi-spectral imaging device includes three characteristic bands with central wavelengths of 550±5nm, 670±5nm, and 800±5nm. The radiation correction adopts an empirical linear correction method based on a standard diffuse reflection reference plate. The correction formula is:

[0024] L = Gain × DN + Bias;

[0025] Among them, the Gain value is dynamically adjusted according to the ambient light intensity at 12 noon every day, so as to ensure that the multispectral images collected under different environmental conditions have uniform and accurate radiation response characteristics; DN is the original digital signal.

[0026] As a further solution of the present invention, the psyllid population density dynamic model data, Huanglongbing disease transmission probability, pesticide safe use range and citrus growth cycle data obtained in step S3 are introduced as nodes respectively by using graph database technology, and directed edges between nodes are constructed according to quantitative relationships to form a complete citrus prevention and control rule set; during the construction process, the environmental capacity parameter update formula is:

[0027]

[0028] In the formula, γ is the learning rate, ΔN is the change in the density of psyllids before and after the application of pesticides, N is the density of psyllids before the application of pesticides, and PAR ref is the reference value of photosynthetically active radiation, and PAR is photosynthetically active radiation.

[0029] The directed edges between nodes represent the following quantitative relationships:

[0030] The directed edge between the psyllid population density dynamic model node and the Huanglongbing transmission probability node: Based on the transmission probability calculation formula, the attributes of this edge include the transmission rate coefficient and threshold parameter.

[0031] Directed edge between the psyllid population density dynamic model node and the environmental capacity node: Based on the environmental capacity calculation formula, the attributes of this edge include soil organic matter content, canopy closure and photosynthetically active radiation.

[0032] Directed edge between the environmental capacity node and the psyllid population growth node: Based on the temperature-corrected Logistic equation, the attributes of this edge include temperature and growth parameters. The Logistic equation is a basic model for describing population growth.

[0033] Directed edge between the pesticide use specification node and the prevention and control measures node: This edge contains the pesticide use recommendations at different stages of the citrus growth cycle. The attributes of the edge include information such as applicable period, dilution multiple, and mixing rules.

[0034] Directed edge between growth cycle node and prevention and control measure node: This edge describes the prevention and control time window in different growth periods. The attributes of the edge include the start and end time of the time window and the priority.

[0035] As a further solution of the present invention, step S4 includes the following steps:

[0036] Step S41, inputting the obtained standardized pre-processed image data into a Huanglongbing recognition model based on a convolutional neural network, and outputting a feature vector of each lesion area;

[0037] Step S42, matching the lesion feature vector with the data in the citrus prevention and control rule set in the knowledge graph, using a deep reinforcement learning method to optimize the prevention and control decision, and introducing the knowledge graph rule matching degree when updating the reward function Q value. The reward function update formula is:

[0038]

[0039] In the formula, n is the number of model training iterations, Q(s,a) is the reward function, G(s,A) is the knowledge graph rule matching degree, and λ 0 , 1 and τ are preset constants, satisfying λ 0 >λ 1 >0 and τ is a positive real number;

[0040] Step S43: The output of the prevention and control decision optimization is the optimal pesticide application decision parameters made by the deep reinforcement learning network according to the current state.

[0041] The reward function takes the form of a weighted combination:

[0042] Q(S,A)=v·Acc(s)+(1-v)·G(s,a);

[0043] Where v is the weight coefficient, which is used to balance the importance of disease identification accuracy and compliance with prevention and control rules; Acc(s) is the accuracy of lesion identification, and G(s,a) is the knowledge graph rule matching degree.

[0044] The accuracy of lesion recognition is calculated using the confusion matrix evaluation method:

[0045]

[0046] Where TP is the number of pixels correctly identified as lesions, TN is the number of pixels correctly identified as healthy tissues, FP is the number of healthy tissue pixels misidentified as lesions, and FN is the number of pixels missed as lesions.

[0047] The knowledge graph rule matching degree G(s,a) adopts a multi-index evaluation method. Specifically, G(s,a) considers four key constraint indicators: Huanglongbing transmission probability matching degree, psyllid population density matching degree, pesticide dosage safety degree, and growth cycle suitability. Its calculation formula is:

[0048]

[0049] Among them, the propagation probability matching term exp(-|P HLB(t) -P th |):P HLB(t) is the predicted probability of Huanglongbing spreading under the current state; P th is the propagation probability threshold set in the knowledge graph; the Gaussian kernel function is used to calculate the similarity between the currently predicted Huanglongbing propagation probability and the propagation probability threshold set in the knowledge graph, which can naturally map the probability difference to the [0,1] interval;

[0050] Insect population density matching item N(t) is the current psyllid population density; N th is the psyllid population density threshold specified in the knowledge graph; an exponential kernel form is used to evaluate the closeness of the current psyllid population density to the threshold specified in the knowledge graph, wherein relative error is used for normalization, which can effectively eliminate the impact of different orders of magnitude;

[0051] R(a,t) is a pesticide safety evaluation function with a value range of [0,1], which is used to evaluate whether the pesticide application scheme a selected at time t meets the pesticide safety use specifications stored in the knowledge graph. Specifically, the calculation of R(a,t) comprehensively considers whether the type, concentration and mixing combination of the pesticide in the application scheme meet the pesticide use specifications of the current growth period.

[0052] L(a,t) is a growth cycle suitability evaluation function with a value range of [0,1], which is used to evaluate whether the application scheme a meets the prevention and control requirements of the growth period at time t. This function is based on the citrus growth cycle information stored in the knowledge graph, divides the whole year into different periods, and specifies the corresponding recommended medication scheme for each period.

[0053] As a further solution of the present invention, the closed-loop feedback optimization is performed according to the following steps:

[0054] Step S61, after the application of pesticides is completed, the data of the density of psyllids and environmental parameters before and after the application of pesticides are continuously collected through fixed monitoring stations and drones to form a continuous time series;

[0055] Step S62, weighted processing is performed on the collected data using a fixed 30-day sliding window, each data point is assigned an attenuation weight, and smoothing filtering is performed to eliminate short-term fluctuations;

[0056] Step S63: recursively update using the environmental capacity parameter update formula.

[0057] As a further solution of the present invention, a deep learning-based agricultural pest control system includes:

[0058] Data collection module, including multispectral imaging devices carried by drones, depth sensors and environmental sensors at fixed monitoring stations for collecting multi-source data from citrus growing areas;

[0059] The data preprocessing module is set in the edge computing node to perform spatiotemporal alignment and radiation correction on the collected raw data, and use affine transformation to perform spatial registration of multispectral images and depth images to generate standardized preprocessed data;

[0060] The knowledge graph construction module is used to construct a dynamic agronomic knowledge graph, which associates the psyllid population density dynamic model with the probability of Huanglongbing transmission, the citrus psyllid population density threshold, the safe use range of pesticides, and the citrus growth cycle information;

[0061] A Huanglongbing recognition module is used to input the standardized pre-processed data into a Huanglongbing recognition model based on a convolutional neural network, and dynamically adjust the convolution kernel weight parameters through a deep reinforcement learning algorithm, wherein the action space of the reinforcement learning algorithm is limited by the pesticide dosage constraint rules in the agronomic knowledge graph;

[0062] The prevention and control plan generation module is used to generate a pesticide application plan based on the output results of the Huanglongbing recognition model and the knowledge graph constraints;

[0063] The feedback optimization module is used to feedback and update the environmental capacity parameter information in the agronomic knowledge graph based on the psyllid population density change data collected after pesticide application, forming a closed-loop optimization mechanism.

[0064] Compared with the prior art, the agricultural pest control method and system based on deep learning of the present invention has the following beneficial effects:

[0065] The present invention dynamically adjusts the convolution kernel weight parameters of the convolutional neural network through a deep reinforcement learning algorithm, and combines the dosage constraint rules in the agronomic knowledge graph to optimize the prevention and control decision. This technical feature enables the prevention and control decision to be dynamically adjusted according to real-time data and agronomic background, thereby ensuring the accuracy and timeliness of the prevention and control plan. Compared with the prior art, most of the existing prevention and control systems only rely on static rules or simple machine learning models, which cannot effectively respond to changes in the environment and growth cycle, resulting in insufficient accuracy and adaptability of prevention and control decisions. Therefore, the present invention can better solve the problems of low accuracy and poor real-time performance of prevention and control decisions in traditional technologies by introducing a dynamic feedback mechanism and a deep fusion of agronomic knowledge graphs, and achieve accurate prevention and control of Huanglongbing. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is the flow chart of the second-order detection algorithm for plant disease detection in document 1.

[0067] Figure 2 This is the deep learning network diagram for plant disease recognition in document 1.

[0068] Figure 3 This is the flight path diagram of the drone in document 2.

[0069] Figure 4 This is the flow chart of plant disease classification in document 3.

[0070] Figure 5 The figure is a flowchart of an agricultural pest control method based on deep learning.

[0071] Figure 6 This is a prevention and control decision-making flowchart for an agricultural pest control method based on deep learning.

[0072] Figure 7 This is a graph showing the relationship between the dynamic changes in citrus psyllid population density and environmental parameters.

[0073] Figure 8 This is the distribution map of the infection degree of citrus Huanglongbing. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] Example 1

[0076] A method for agricultural pest control based on deep learning, comprising the following steps:

[0077] Step S1, collecting multi-source data of the citrus planting area by using a multispectral imaging device, a depth sensor and an environmental sensor of a fixed monitoring station carried by an unmanned aerial vehicle.

[0078] In step S2, the collected raw data is input into the edge computing node after time-space alignment and radiation correction, and affine transformation is used to perform spatial registration of the multispectral image and the depth image to generate standardized preprocessed data.

[0079] Step S3, construct a dynamic agronomic knowledge graph, and associate the psyllid population density dynamic model with the Huanglongbing disease transmission probability, the citrus psyllid population density threshold, the safe use range of pesticides, and the citrus growth cycle information through the graph database to build a model, wherein the psyllid population density dynamic model is constructed using the temperature-corrected Logistic equation, and its expression is:

[0080]

[0081] Where N(t) is the current psyllid population density, and the temperature correction coefficient r = k r ·exp(k T (TT 0 )), T is the real-time monitored temperature value.

[0082] The initial environmental capacity parameter of the psyllid population density dynamic model is calculated by soil organic matter content, canopy density and photosynthetically active radiation, and the calculation formula is:

[0083]

[0084] In the formula, SOM is soil organic matter content; CC is canopy density; PAR is photosynthetically active radiation value, PAR ref It is the reference value of photosynthetically active radiation.

[0085] The calculation formula for the Huanglongbing transmission probability is:

[0086]

[0087] Where P HLB (t) represents the probability of Huanglongbing spreading per unit time, β is the transmission rate coefficient, N th is the psyllid population density threshold, N(t) is the current psyllid population density; this expression can quantitatively describe the spread trend of citrus Huanglongbing under different psyllid population densities.

[0088] Step S4, inputting the standardized preprocessed data into a Huanglongbing recognition model based on a convolutional neural network, which dynamically adjusts the convolution kernel weight parameters through a deep reinforcement learning algorithm. The action space of the reinforcement learning algorithm is restricted by the dosage constraint rules in the agronomic knowledge graph, and its reward function is the weighted sum of the lesion recognition accuracy and the knowledge graph rule matching degree.

[0089] Step S5, generating a pesticide application plan based on the output results of the Huanglongbing recognition model and the knowledge graph constraints.

[0090] Step S6, after applying the pesticide, the environmental capacity parameter information in the agronomic knowledge graph is fed back and updated through the collected data on changes in psyllid population density, forming a closed-loop optimization mechanism.

[0091] The working band of the multi-spectral imaging device in the embodiment of the present invention includes three characteristic bands with central wavelengths of 550±5nm, 670±5nm, and 800±5nm, and the spatial resolution of the imaging sensor is not less than 2cm / pixel. The radiation correction adopts an empirical linear correction method based on a standard diffuse reflection reference plate, and the correction formula is:

[0092] L = Gain × DN + Bias;

[0093] Among them, the Gain value is dynamically adjusted according to the ambient light intensity at 12 noon every day, so as to ensure that the multispectral images collected under different environmental conditions have uniform and accurate radiation response characteristics; DN is the original digital signal.

[0094] In the embodiment of the present invention, the graph database technology is used to introduce the psyllid population density dynamic model data, the Huanglongbing disease transmission probability, the safe use range of pesticides and the citrus growth cycle data obtained in step S3 as nodes, and the directed edges between the nodes are constructed according to the quantitative relationship to form a complete citrus prevention and control rule set; during the construction process, the environmental capacity parameter update formula is:

[0095]

[0096] In the formula, γ is the learning rate, ΔN is the change in the density of psyllids before and after the application of pesticides, N is the density of psyllids before the application of pesticides, and PAR ref is the reference value of photosynthetically active radiation, and PAR is photosynthetically active radiation.

[0097] Through the above steps, the constructed knowledge graph clearly and quantitatively records the strict relationship between the psyllid population density dynamic model, the transmission law of Huanglongbing, the citrus psyllid population density threshold, the safe use range of pesticides and the citrus growth cycle information, providing a quantitative basis for subsequent prevention and control decisions.

[0098] The step S4 in the embodiment of the present invention includes the following steps:

[0099] Step S41, inputting the obtained standardized pre-processed image data into a Huanglongbing recognition model based on a convolutional neural network, and outputting a feature vector of each lesion area;

[0100] Step S42, matching the lesion feature vector with the data in the citrus prevention and control rule set in the knowledge graph, using a deep reinforcement learning method to optimize the prevention and control decision, and introducing the knowledge graph rule matching degree when updating the reward function Q value. The reward function update formula is:

[0101]

[0102] In the formula, n is the number of model training iterations, Q(s,a) is the reward function, G(s,a) is the knowledge graph rule matching degree, and λ 0 , 1 and τ are preset constants, satisfying λ 0 >λ 1 >0 and τ is a positive real number;

[0103] Step S43: The output of the prevention and control decision optimization is the optimal pesticide application decision parameters made by the deep reinforcement learning network according to the current state.

[0104] The reward function in the embodiment of the present invention adopts the form of weighted combination:

[0105] Q(s,a)=v·Acc(s)+(1-v)·G(s,a);

[0106] Where v is the weight coefficient, which is used to balance the importance of disease identification accuracy and compliance with prevention and control rules; Acc(s) is the accuracy of lesion identification, and G(s,a) is the knowledge graph rule matching degree.

[0107] The calculation of the accuracy of spot recognition in the embodiment of the present invention adopts the confusion matrix evaluation method:

[0108]

[0109] Where TP is the number of pixels correctly identified as lesions, TN is the number of pixels correctly identified as healthy tissues, FP is the number of healthy tissue pixels misidentified as lesions, and FN is the number of pixels missed as lesions.

[0110] The knowledge graph rule matching degree G(s,a) in the embodiment of the present invention adopts a multi-index evaluation method. Specifically, G(s,a) takes into account four key constraint indicators: the matching degree of Huanglongbing transmission probability, the matching degree of psyllid population density, the safety degree of pesticide dosage, and the suitability of the growth cycle. Its calculation formula is:

[0111]

[0112] Among them, the propagation probability matching term exp(-|P HLB(t) -P th |):P HLB(t) is the predicted probability of Huanglongbing spreading under the current state; P th is the propagation probability threshold set in the knowledge graph; the Gaussian kernel function is used to calculate the similarity between the currently predicted Huanglongbing propagation probability and the propagation probability threshold set in the knowledge graph, which can naturally map the probability difference to the [0,1] interval;

[0113] Insect population density matching item N(t) is the current psyllid population density; N th is the psyllid population density threshold specified in the knowledge graph; an exponential kernel form is used to evaluate the closeness of the current psyllid population density to the threshold specified in the knowledge graph, wherein relative error is used for normalization, which can effectively eliminate the impact of different orders of magnitude;

[0114] R(a,t) is a pesticide safety evaluation function with a value range of [0,1], which is used to evaluate whether the pesticide application scheme a selected at time t meets the pesticide safety use specifications stored in the knowledge graph. Specifically, the calculation of R(a,t) comprehensively considers whether the type, concentration and mixing combination of the pesticide in the application scheme meet the pesticide use specifications of the current growth period.

[0115] L(a,t) is a growth cycle suitability evaluation function with a value range of [0,1], which is used to evaluate whether the application scheme a meets the prevention and control requirements of the growth period at time t. This function is based on the citrus growth cycle information stored in the knowledge graph, divides the whole year into different periods, and specifies the corresponding recommended medication scheme for each period.

[0116] The closed-loop feedback optimization in the embodiment of the present invention is performed according to the following steps:

[0117] Step S61, after the application of pesticides is completed, the data of the density of psyllids and environmental parameters before and after the application of pesticides are continuously collected through fixed monitoring stations and drones to form a continuous time series;

[0118] Step S62, weighted processing is performed on the collected data using a fixed 30-day sliding window, each data point is assigned an attenuation weight, and smoothing filtering is performed to eliminate short-term fluctuations;

[0119] Step S63: recursively update using the environmental capacity parameter update formula.

[0120] Example 2

[0121] A deep learning-based agricultural pest control system, comprising:

[0122] Data collection module, including multispectral imaging devices carried by drones, depth sensors and environmental sensors at fixed monitoring stations for collecting multi-source data from citrus growing areas;

[0123] The data preprocessing module is set in the edge computing node to perform spatiotemporal alignment and radiation correction on the collected raw data, and use affine transformation to perform spatial registration of multispectral images and depth images to generate standardized preprocessed data;

[0124] The knowledge graph construction module is used to construct a dynamic agronomic knowledge graph, which associates the psyllid population density dynamic model with the probability of Huanglongbing transmission, the citrus psyllid population density threshold, the safe use range of pesticides, and the citrus growth cycle information;

[0125] A Huanglongbing recognition module is used to input the standardized pre-processed data into a Huanglongbing recognition model based on a convolutional neural network, and dynamically adjust the convolution kernel weight parameters through a deep reinforcement learning algorithm, wherein the action space of the reinforcement learning algorithm is limited by the pesticide dosage constraint rules in the agronomic knowledge graph;

[0126] The prevention and control plan generation module is used to generate a pesticide application plan based on the output results of the Huanglongbing recognition model and the knowledge graph constraints;

[0127] The feedback optimization module is used to feedback and update the environmental capacity parameter information in the agronomic knowledge graph based on the psyllid population density change data collected after pesticide application, forming a closed-loop optimization mechanism.

[0128] Example 3

[0129] The directed edges between nodes of the knowledge graph in the embodiment of the present invention are constructed based on the attribute graph model commonly used in graph databases. Specifically, the directed edges between nodes represent the following quantitative relationships:

[0130] The directed edge between the psyllid population density dynamic model node and the Huanglongbing transmission probability node: Based on the transmission probability calculation formula, the attributes of this edge include the transmission rate coefficient and threshold parameter.

[0131] Directed edge between the psyllid population density dynamic model node and the environmental capacity node: Based on the environmental capacity calculation formula, the attributes of this edge include soil organic matter content, canopy closure and photosynthetically active radiation.

[0132] Directed edge between the environmental capacity node and the psyllid population growth node: Based on the temperature-corrected Logistic equation, the attributes of this edge include temperature and growth parameters. The Logistic equation is a basic model for describing population growth.

[0133] Directed edge between the pesticide use specification node and the prevention and control measures node: This edge contains the pesticide use recommendations at different stages of the citrus growth cycle. The attributes of the edge include information such as applicable period, dilution multiple, and mixing rules.

[0134] Directed edge between growth cycle node and prevention and control measure node: This edge describes the prevention and control time window in different growth periods. The attributes of the edge include the start and end time of the time window and the priority.

[0135] The following is a Python code example that builds directed edges between nodes of a knowledge graph based on the commonly used attribute graph model in graph databases. Please note that this example is only a starting point and may need to be adjusted according to actual conditions and device interfaces in actual applications.

[0136]

[0137]

[0138] This code is only used as an example. In actual applications, appropriate modifications and adjustments need to be made according to specific circumstances.

[0139] Example 4

[0140] When the spring shoots sprout 2-5mm, use 800 times of 40% chlorpyrifos or 1000 times of dichlorvos or 1000 times of 40% profenofos + 5000 times of acarid + 1500 times of 10% fenpropimorph to control overwintering adult psyllids, red spiders and scab 1-2 times (the first time at the end of early March, the second time at the end of mid-April).

[0141] During the summer shoot period, use 1500 times of 40% chlorpyrifos or 1000 times of 40% methidathion or 1500 times of 1.8% avermectin + 5000 times of acarid + 1500 times of 10% fenpropimorph to control red spider mites, psyllids, scale insects, stink bugs, leaf miners, sandworm disease and anthrax and other pests and diseases 2 to 3 times (the first time in late May, the second time in mid-June, and the third time in early July).

[0142] During the autumn shoot period, use 5000 times of Mite Danger + 1500 times of 1.8% Avermectin or 1500 times of 10% Imidacloprid or 800 times of 20% Monoformamidine + 1500 times of 10% Difenoconazole to control red spider mites, psyllids, rust mites, sand skin disease and anthracnose and other pests and diseases 1 to 2 times (the first time in late August, the second time in late September).

[0143] Figure 8The distribution of Huanglongbing in a standard 80m×80m citrus observation area in the early and late stages of three growth periods (spring shoot period, summer shoot period, and autumn shoot period) is shown, forming a total of six distribution characteristics of observation periods (a1, a2, b1, b2, c1, and c2). Each diamond area in the figure uses different gray levels (level IV) to represent the infection level of Huanglongbing. The lightest area (level I) represents a healthy or slightly infected state, and as the gray level deepens, it represents a mild infection (level II), a moderate infection (level III), and a severe infection (level IV). The darkest area (level V) represents a severe infection state. The contour lines in each sub-figure outline the distribution boundaries of the same infection level.

[0144] The reward function in the embodiment of the present invention is in the form of a weighted combination: Q(s,a)=v·Acc(s)+(1-v)·G(s,a), where v is a weight coefficient used to balance the importance of disease identification accuracy and compliance with prevention and control rules.

[0145] The accuracy of lesion recognition is calculated using the confusion matrix evaluation method:

[0146]

[0147] Where TP is the number of pixels correctly identified as lesions, TN is the number of pixels correctly identified as healthy tissues, FP is the number of healthy tissue pixels misidentified as lesions, and FN is the number of pixels missed as lesions.

[0148] The knowledge graph rule matching degree G(s,a) in the embodiment of the present invention adopts a multi-index evaluation method. Specifically, G(s,a) takes into account four key constraint indicators: the matching degree of Huanglongbing transmission probability, the matching degree of psyllid population density, the safety degree of pesticide dosage, and the suitability of the growth cycle. Its calculation formula is:

[0149]

[0150] Among them, the propagation probability matching term exp(-|P HLB(t) -P th |):P HLB(t) is the predicted probability of Huanglongbing spreading under the current state; P th is the propagation probability threshold set in the knowledge graph; the Gaussian kernel function is used to calculate the similarity between the currently predicted Huanglongbing propagation probability and the propagation probability threshold set in the knowledge graph, which can naturally map the probability difference to the [0,1] interval;

[0151] Insect population density matching item N(t) is the current psyllid population density; N this the psyllid population density threshold specified in the knowledge graph; an exponential kernel form is used to evaluate the closeness of the current psyllid population density to the threshold specified in the knowledge graph, wherein relative error is used for normalization, which can effectively eliminate the impact of different orders of magnitude;

[0152] R(a,t) is a pesticide safety evaluation function with a value range of [0,1], which is used to evaluate whether the pesticide application scheme a selected at time t meets the pesticide safety use specifications stored in the knowledge graph. Specifically, the calculation of R(a,t) comprehensively considers whether the type, concentration and mixing combination of the pesticide in the pesticide application scheme meet the pesticide use specifications of the current growth period. For example, a larger value is used when 40% chlorpyrifos 800 times liquid is used in the spring shoot period, while a smaller value is used when other unrecommended pesticides or improper mixing schemes are used.

[0153] L(a,t) is a growth cycle suitability evaluation function with a value range of [0,1], which is used to evaluate whether the application scheme a meets the prevention and control requirements of the growth period at time t. Based on the citrus growth cycle information stored in the knowledge graph, this function divides the whole year into key periods such as the spring shoot period (March-April), summer shoot period (May-July) and autumn shoot period (August-September), and specifies the corresponding recommended medication scheme for each period. For example, when the recommended combination of drugs is used to control overwintering psyllid adults when the spring shoots sprout 2-5mm, L(a,t) takes a larger value; when the medication period or drug selection does not match the characteristics of the growth cycle, a smaller value is taken.

[0154] Weight coefficient w 1 、w 2 、w 3 、w 4 It reflects the relative importance of each constraint and satisfies w 1 +w 2 +w 3 +w 4 =1.

[0155] The following is a Python code example that evaluates the match between the current state s and action a and the expert rules stored in the knowledge graph. Please note that this example is only a starting point and may need to be adjusted in actual applications according to actual conditions and device interfaces.

[0156]

[0157]

[0158]

[0159]

[0160]

[0161]

[0162] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for controlling agricultural pests and diseases based on deep learning, characterized in that: The following steps are involved: Step S1, collecting multi-source data of the citrus planting area by using a multispectral imaging device, a depth sensor and an environmental sensor of a fixed monitoring station carried by an unmanned aerial vehicle; Step S2, input the collected raw data into the edge computing node after time-space alignment and radiation correction, use affine transformation to perform spatial registration of multispectral images and depth images, and generate standardized preprocessed data; Step S3, construct a dynamic agronomic knowledge graph, and associate the psyllid population density dynamic model with the probability of Huanglongbing transmission, the citrus psyllid population density threshold, the safe use range of pesticides, and the citrus growth cycle information through a graph database to model the psyllid population density dynamic model, wherein the psyllid population density dynamic model is constructed using a temperature-corrected Logistic equation; the initial environmental capacity parameter of the psyllid population density dynamic model is calculated through soil organic matter content, canopy closure, and photosynthetically active radiation, and the calculation formula is: In the formula, SOM is soil organic matter content; CC is canopy density; PAR is photosynthetically active radiation value, PAR ref is the reference value of photosynthetically active radiation; The calculation formula for the Huanglongbing transmission probability is: Where P HLB (t) represents the probability of Huanglongbing spreading per unit time, β is the transmission rate coefficient, N th is the psyllid population density threshold, N(t) is the current psyllid population density; Step S4, inputting the standardized preprocessed data into a Huanglongbing recognition model based on a convolutional neural network, the model dynamically adjusts the convolution kernel weight parameters through a deep reinforcement learning algorithm, the action space of the reinforcement learning algorithm is limited by the pesticide dosage constraint rules in the agronomic knowledge graph, and its reward function is the weighted sum of the lesion recognition accuracy and the knowledge graph rule matching degree; Step S5, generating a pesticide application plan based on the output results of the Huanglongbing recognition model and the knowledge graph constraints; Step S6, after applying the pesticide, the environmental capacity parameter information in the agronomic knowledge graph is fed back and updated through the collected data on changes in psyllid population density, forming a closed-loop optimization mechanism.

2. The method for agricultural pest control based on deep learning according to claim 1, characterized in that: The working band of the multi-spectral imaging device includes three characteristic bands with central wavelengths of 550±5nm, 670±5nm, and 800±5nm, and the radiation correction adopts an empirical linear correction method based on a standard diffuse reflection reference plate.

3. The method for agricultural pest control based on deep learning according to claim 1, characterized in that: Using graph database technology, the psyllid population density dynamic model data, Huanglongbing transmission probability, pesticide safe use range and citrus growth cycle data obtained in step S3 are introduced as nodes, and directed edges between nodes are constructed according to quantitative relationships to form a complete citrus prevention and control rule set; during the construction process, the environmental capacity parameter update formula is: In the formula, γ is the learning rate, ΔN is the change in the density of psyllids before and after the application of pesticides, N is the density of psyllids before the application of pesticides, and PAR ref is the reference value of photosynthetically active radiation, and PAR is photosynthetically active radiation.

4. The method for agricultural pest control based on deep learning according to claim 1, characterized in that: The step S4 comprises the following steps: Step S41, inputting the obtained standardized pre-processed image data into a Huanglongbing recognition model based on a convolutional neural network, and outputting a feature vector of each lesion area; Step S42, matching the lesion feature vector with the data in the citrus prevention and control rule set in the knowledge graph, using a deep reinforcement learning method to optimize the prevention and control decision, and introducing the knowledge graph rule matching degree when updating the reward function Q value. The reward function update formula is: Where n is the number of model training iterations, Q(s,a) is the reward function, G(s,a) is the knowledge graph rule matching degree, λ0, λ1 and τ are preset constants, satisfying λ0>λ1>0 and τ is a positive real number; Step S43: The output of the prevention and control decision optimization is the optimal pesticide application decision parameters made by the deep reinforcement learning network according to the current state.

5. The method for agricultural pest control based on deep learning according to claim 1, characterized in that: The closed-loop feedback optimization is performed according to the following steps: Step S61, after the application of pesticides is completed, the data of the density of psyllids and environmental parameters before and after the application of pesticides are continuously collected through fixed monitoring stations and drones to form a continuous time series; Step S62, weighted processing is performed on the collected data using a fixed 30-day sliding window, each data point is assigned an attenuation weight, and smoothing filtering is performed to eliminate short-term fluctuations; Step S63: recursively update using the environmental capacity parameter update formula.

6. An agricultural pest control system based on deep learning, characterized in that: A method for controlling agricultural pests and diseases based on deep learning according to any one of claims 1 to 5 is implemented, comprising: Data collection module, including multispectral imaging devices carried by drones, depth sensors and environmental sensors at fixed monitoring stations for collecting multi-source data from citrus growing areas; The data preprocessing module is set in the edge computing node to perform spatiotemporal alignment and radiation correction on the collected raw data, and use affine transformation to perform spatial registration of multispectral images and depth images to generate standardized preprocessed data; The knowledge graph construction module is used to construct a dynamic agronomic knowledge graph, which associates the psyllid population density dynamic model with the probability of Huanglongbing transmission, the citrus psyllid population density threshold, the safe use range of pesticides, and the citrus growth cycle information; A Huanglongbing recognition module is used to input the standardized pre-processed data into a Huanglongbing recognition model based on a convolutional neural network, and dynamically adjust the convolution kernel weight parameters through a deep reinforcement learning algorithm, wherein the action space of the reinforcement learning algorithm is limited by the pesticide dosage constraint rules in the agronomic knowledge graph; The prevention and control plan generation module is used to generate a pesticide application plan based on the output results of the Huanglongbing recognition model and the knowledge graph constraints; The feedback optimization module is used to feedback and update the environmental capacity parameter information in the agronomic knowledge graph based on the psyllid population density change data collected after pesticide application, forming a closed-loop optimization mechanism.

Images