Intelligent early warning system for garden greening pest based on machine vision and AI large model

Abstract

The application belongs to the technical field of disease and pest early warning, and particularly relates to a landscape greening disease and pest intelligent early warning system based on machine vision and an AI large model, which comprises an edge target detection module, a cloud multi-modal large model verification module, a confidence degree arbitration module and an early warning decision module.The edge target detection module uses an improved YOLO network to perform real-time detection on landscape vegetation images, and outputs a suspected disease and pest area and a first confidence degree; the cloud multi-modal large model verification module performs fine-grained identification and harm grade evaluation on the suspected area through fine-tuning of landscape disease and pest professional corpus, and outputs a second confidence degree; the confidence degree arbitration module performs consistency judgment and collaborative decision according to the double confidence degrees; and the early warning decision module generates graded early warning information.Through the two-stage detection architecture of edge-cloud collaboration, combined with multi-source data fusion and a graded early warning mechanism, the application realizes high-precision identification and intelligent early warning of landscape diseases and pests.

Description

Technical Field

[0001] This invention relates to the field of pest and disease early warning technology, and in particular to an intelligent early warning system for garden and greening pests and diseases based on machine vision and AI big data models. Background Technology

[0002] Urban greening is an essential component of the urban ecosystem, playing an irreplaceable role in improving the living environment, regulating urban microclimate, and maintaining biodiversity. However, urban vegetation faces long-term threats from various pests and diseases during its growth. Large-scale outbreaks of pests such as aphids, longhorn beetles, and powdery mildew not only damage the landscape and degrade ecological functions but may also lead to irreversible losses of precious resources such as ancient and famous trees. With the intensification of climate change and the increasing complexity of urban ecosystems, the traditional "post-event prevention" model is no longer sufficient to meet the needs of modern urban landscaping's refined management.

[0003] Currently, the monitoring and early warning of pests and diseases in landscaping mainly relies on the following technical means: First, the traditional manual inspection method, in which landscaping maintenance personnel regularly conduct visual observations along inspection routes and judge the types and severity of pests and diseases based on personal experience; second, the use of physical monitoring equipment such as insect-attracting lamps and pheromone traps to collect insect information at fixed points, assisting in manual counting and classification; third, automatic insect monitoring and reporting systems based on image recognition technology, which use fixed monitoring equipment to capture images of insects and call recognition algorithms to identify species; and fourth, some smart landscaping projects introduce Internet of Things (IoT) sensor networks to centrally display multi-dimensional data such as insect infestation, weather, and soil conditions, and provide simple threshold alarms.

[0004] However, the aforementioned existing technologies have the following shortcomings when practically applied to the monitoring and early warning of pests and diseases in landscaping: 1. The recognition accuracy and generalization ability of image recognition models are limited, making it difficult to meet the application needs in complex garden scenarios; 2. Existing technologies lack a collaborative architecture design between large models and dedicated detection models; 3. Existing monitoring and early warning systems lack the ability to deeply integrate and intelligently analyze multi-source heterogeneous data. Summary of the Invention

[0005] 1. Technical problems to be solved The purpose of this invention is to solve the problems in the prior art and proposes an intelligent early warning system for garden pests and diseases based on machine vision and AI large model. It realizes a complete closed loop of garden pests and diseases from automatic identification and intelligent early warning to solution generation, and has the ability to perform multi-source data fusion analysis and interpretable reasoning.

[0006] 2. Technical Solution To achieve the above objectives, the present invention adopts the following technical solution: Firstly, this application provides an intelligent early warning system for garden pests and diseases based on machine vision and AI large-scale models, comprising: an edge target detection module, deployed at edge computing nodes in the garden monitoring area, used to perform real-time target detection on garden vegetation images acquired by front-end acquisition devices, and output detection results of suspected pest and disease areas and a first confidence level; a cloud multimodal large-scale model verification module, deployed on a cloud server, used to receive image cropping data from the suspected pest and disease area detection results, perform fine-grained recognition and hazard level assessment, and output a second confidence level; a confidence level arbitration module, used to perform consistency judgment and collaborative decision-making based on the first confidence level and the second confidence level; and an early warning decision module, used to generate graded early warning information based on the output of the confidence level arbitration module; the edge target detection module adopts an improved YOLO network structure, and the cloud multimodal large-scale model verification module adopts a multimodal large-scale language model finely tuned with professional corpus on garden pests and diseases.

[0007] In one possible implementation, the confidence arbitration module performs the following judgment logic: When the first confidence level is greater than or equal to the first preset threshold and the second confidence level is greater than or equal to the second preset threshold, the detection result is determined to be a confirmed diagnosis. When the first confidence level is greater than or equal to the first preset threshold and the second confidence level is less than the second preset threshold, the detection result is determined to be suspicious, and a manual review mark is triggered. When the first confidence level is less than the first preset threshold, the detection result is ignored; wherein the first preset threshold is lower than the second preset threshold.

[0008] In one possible implementation, the improved YOLO network employs an optimized feature pyramid structure to enhance the feature extraction capability of small target pests and diseases, and introduces an attention mechanism to reduce the impact of complex background interference on detection accuracy; the multimodal large model is an open-source multimodal large model that has undergone incremental pre-training and instruction fine-tuning, and has the ability to understand pest and disease images and reason about prevention and control knowledge.

[0009] In one possible implementation, the edge target detection module and the cloud multimodal large model verification module use an asynchronous message queue for data transmission. The cropped data of the suspected pest and disease area image is batch cached at the edge and then sent to the cloud through a compressed transmission protocol. When the cloud network connection is interrupted, the edge target detection module uses the edge detection result as a temporary output and caches the image data locally. After the network is restored, it resends the data to the cloud for secondary verification.

[0010] In one possible implementation, a multi-source data fusion module is also included, the input data of which includes: Pest and disease identification results data, including pest and disease type, damage level, discovery location and timestamp; Meteorological and environmental data, including temperature, humidity, precipitation, wind speed, and light intensity; Historical records of pests and diseases, including data on the occurrence of pests and diseases over the past 3-5 years and the corresponding meteorological conditions; Vegetation growth status data; the multi-source data fusion module performs weighted fusion based on the timeliness and confidence level of various types of data, and outputs a comprehensive risk index for the target area.

[0011] In one possible implementation, the formula for calculating the comprehensive risk index is: in The confidence score for the currently detected pests and diseases. This is a predicted probability of future occurrence based on historical pest and disease records and weather forecast data. Score the vegetation growth status. , , These are the corresponding weighting coefficients; the weighting coefficients are dynamically adjusted based on the environmental characteristics and pest and disease occurrence patterns of the target area.

[0012] In one possible implementation, a pest and disease occurrence probability prediction module is also included. The prediction module adopts a time series analysis and machine learning fusion algorithm, uses historical pest and disease records and corresponding historical meteorological data as training samples, takes future meteorological forecast data as input, and outputs the predicted probability values ​​of various pests and diseases in the target area in the future period.

[0013] In one possible implementation, a tiered early warning module is also included, which triggers tiered early warnings based on the comprehensive risk index. When the comprehensive risk index is greater than or equal to the first warning threshold, a level one warning is triggered and emergency response recommendations are generated. When the comprehensive risk index is greater than or equal to the second warning threshold and less than the first warning threshold, a level-two warning is triggered and preventive action recommendations are generated. When the comprehensive risk index is less than the second warning threshold, the monitoring data is recorded for continuous model training.

[0014] In one possible implementation, a large-scale intelligent decision-making module is also included. After a graded early warning is triggered, the large-scale intelligent decision-making module automatically generates a customized prevention and control plan based on the type of pest and disease, the level of damage, the plant species, and the current environmental information, combined with a professional knowledge base of horticultural pests and diseases. The professional knowledge base contains basic information on various pests and diseases, knowledge of prevention and control methods, and specifications for pesticide use. The prevention and control plan is output in the form of structured text and includes interpretable evidence for the plan.

[0015] In one possible implementation, an intelligent work order dispatching and closed-loop management module is also included. This module automatically generates electronic work orders based on the prevention and control plan and the location information of maintenance personnel. The work orders include the responsible personnel, the content of the treatment, and the completion deadline. After the maintenance personnel complete the treatment, they upload the treatment result information. The system automatically arranges review tasks based on the treatment result information, evaluates the prevention and control effect through image comparison analysis, and completes the closed loop when the prevention and control effect reaches the preset standard and stores the data in the historical knowledge base.

[0016] 3. Beneficial effects Compared with the prior art, the advantages of this invention are: (1) In this invention, the two-stage collaborative architecture of fast suspected area localization by YOLO at the edge and fine-grained secondary verification of multimodal large model in the cloud, combined with the dual confidence arbitration mechanism, solves the problems of insufficient detection accuracy of single model and high inference latency of large model. The false alarm rate is reduced by about 40%, and the false negative rate is reduced by about 30%.

[0017] (2) In this invention, by deeply integrating multi-dimensional information such as pest and disease identification results, meteorological data, historical occurrence patterns, and vegetation growth status, a comprehensive risk assessment model is constructed, which significantly improves the early warning accuracy compared with the existing single image recognition scheme.

[0018] (3) In this invention, customized prevention and control plans are automatically generated through large language model reasoning. Combined with intelligent dispatching and closed-loop management modules, a complete closed loop of "perception-analysis-decision-execution" is realized, which solves the problem of "heavy monitoring and light decision-making" in the existing system.

[0019] (4) In this invention, the lightweight edge model is responsible for real-time detection, while the large cloud model is responsible for fine verification. This reduces the computing power requirements at the edge while ensuring recognition accuracy, and supports low-cost deployment in large-scale garden areas. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the intelligent early warning system for garden pests and diseases based on machine vision and AI large model proposed in this invention. Figure 2 This is a flowchart illustrating the two-stage collaborative detection and confidence arbitration method proposed in this invention. Figure 3 This is a flowchart of a prevention and control scheme generation method based on large language model reasoning. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Reference Figure 1 The intelligent early warning system for garden pests and diseases based on machine vision and AI large-scale models includes: an edge target detection module, deployed at edge computing nodes in the garden monitoring area, used to perform real-time target detection on garden vegetation images acquired by front-end acquisition devices, and output the detection results of suspected pest and disease areas and the first confidence level; a cloud multimodal large-scale model verification module, deployed on a cloud server, used to receive image cropping data from the suspected pest and disease area detection results, perform fine-grained recognition and hazard level assessment, and output the second confidence level; a confidence level arbitration module, used to make consistency judgments and collaborative decisions based on the first and second confidence levels; and an early warning decision module, used to generate graded early warning information based on the output of the confidence level arbitration module. The edge target detection module adopts an improved YOLO network structure, and the cloud multimodal large-scale model verification module adopts a multimodal large-scale language model finely tuned with professional corpus on garden pests and diseases.

[0023] The edge target detection module and the cloud multimodal large model verification module use an asynchronous message queue for data transmission. Image cropping data of suspected pest and disease areas are batch cached at the edge and then sent to the cloud through a compressed transmission protocol. When the cloud network connection is interrupted, the edge target detection module uses the edge detection results as temporary output and caches the image data locally. After the network is restored, it is resent to the cloud for secondary verification.

[0024] The confidence level arbitration module executes the following judgment logic: When the first confidence level is greater than or equal to the first preset threshold and the second confidence level is greater than or equal to the second preset threshold, the detection result is determined to be a confirmed diagnosis. When the first confidence level is greater than or equal to the first preset threshold and the second confidence level is less than the second preset threshold, the detection result is determined to be suspicious and a manual review mark is triggered. When the first confidence level is less than the first preset threshold, the detection result is ignored; wherein the first preset threshold is lower than the second preset threshold.

[0025] Example 1: A Pest and Disease Identification Method Based on Two-Stage Collaborative Detection This embodiment provides an intelligent early warning system for pests and diseases in landscaping based on machine vision and AI large models. It focuses on describing a two-stage collaborative detection and confidence arbitration method using YOLO target detection and multimodal large models, which solves the problems of insufficient detection accuracy of single models and high inference latency of large models in the prior art.

[0026] Step 1: Build a lightweight target detection module for the edge.

[0027] A lightweight target detection model is deployed on edge computing nodes in the garden monitoring area. The model adopts an improved YOLOv8 network structure. The input of the detection model is garden vegetation images acquired by front-end acquisition devices (including but not limited to fixed insect monitoring cameras, UAV aerial cameras, and handheld shooting devices), and the output is the detection results of suspected pest and disease areas in the images, including: the location and size information of the bounding box of the detected target. Preliminary classification labels and their corresponding confidence scores; Image cropping data of the target area. The improved YOLOv8 network adopts an optimized feature pyramid structure to enhance the feature extraction capability for small target pests and early disease spots, and introduces an attention mechanism to reduce the impact of complex background interference on detection accuracy.

[0028] Step 2: Construct a cloud-based multimodal large model secondary verification module.

[0029] A multimodal large model is deployed in the cloud. The multimodal large model is a large language model that has been fine-tuned with professional corpus of garden pests and diseases. It has image understanding and text reasoning capabilities. Based on the open source multimodal large model, the professional understanding of pest and disease images is enhanced through incremental pre-training and instruction fine-tuning.

[0030] The multimodal large model receives the cropped image data of the suspected pest and disease area output by the edge target detection module and performs the following operations: performs fine-grained identification on the suspected pest and disease area to determine whether pest and disease symptoms are indeed present; Once the presence of pests and diseases is confirmed, the specific type of pest or disease must be determined. The severity of pests and diseases is assessed in a graded manner, and the graded assessment includes at least three levels: mild, moderate and severe. Output the corresponding model confidence score.

[0031] Step 3: Perform dual-engine confidence arbitration and collaborative decision-making.

[0032] The detection confidence score output by the edge target detection module and the model confidence score output by the cloud multimodal large model are uniformly input into the confidence arbitration module for consistency judgment and collaborative decision-making. Specifically, when the target detection confidence score is greater than or equal to the first preset threshold and the large model confidence score is greater than or equal to the second preset threshold, the detection result is determined to be a confirmed case, and the warning processing flow is automatically entered. When the target detection confidence is greater than or equal to the first preset threshold and the large model confidence is less than the second preset threshold, the detection result is judged as suspicious, triggering manual review marking, and the target region image is added to the review queue. When the target detection confidence is less than the first preset threshold, the detection result is ignored and judged as a false alarm. The first preset threshold is lower than the second preset threshold.

[0033] Step 4: Perform load distribution and latency optimization for edge-cloud collaboration.

[0034] The target detection module is deployed at the edge to quickly locate suspected areas locally; the multimodal large model is deployed in the cloud, utilizing cloud GPU computing power for secondary verification. Asynchronous message queues are used for data transmission between the edge and the cloud. The cropped image data of suspected pest and disease areas is batch cached at the edge and then sent to the cloud via a compressed transmission protocol to reduce cloud-edge communication latency.

[0035] Example 2: This embodiment provides an intelligent early warning and decision-making method for garden and greening pests and diseases based on multi-source data fusion. By integrating pest and disease identification results with multi-dimensional information such as meteorological data, historical occurrence patterns, and vegetation growth status, a precise early warning and assessment model is constructed.

[0036] Step 1: Construct a multi-source data acquisition and standardization module.

[0037] The data types collected by the multi-source data acquisition module include: pest and disease identification result data: pest and disease type, hazard level, discovery location and timestamp output by the system in Example 1, and the data format is a structured pest and disease event record; Meteorological environmental data: collected through meteorological sensors deployed in the park area, including temperature, humidity, precipitation, wind speed, and light intensity; Historical Pest and Disease Records: Collect historical pest and disease occurrence data for the target garden area over the past 3-5 years, including pest and disease types, first discovery time, degree of damage, and corresponding meteorological conditions, and construct a historical pest and disease occurrence database; Vegetation growth status data: The vegetation growth status of the target area is assessed by the NDVI vegetation index or similar vegetation health indices.

[0038] Step 2: Construct a prediction model for the probability of pest and disease occurrence.

[0039] Based on historical pest and disease records and corresponding historical meteorological data, a pest and disease occurrence probability prediction model is constructed using a fusion algorithm of time series analysis and machine learning. The prediction model takes the meteorological forecast data for the next 7 days as input and outputs the predicted probability values ​​of various pests and diseases occurring in the target garden area during the future period.

[0040] Step 3: Perform multimodal data fusion and comprehensive risk assessment.

[0041] The results of pest and disease identification, the predicted probability of pest and disease occurrence, meteorological environmental data, and vegetation growth status data are input into the comprehensive risk assessment module. The comprehensive risk assessment module uses a weighted fusion algorithm to calculate the risk index of each monitoring grid. in The confidence score for the currently detected pests and diseases. This is a predicted probability of future occurrence based on historical pest and disease records and weather forecast data. Score the vegetation growth status. , , These are the corresponding weighting coefficients, and each weighting coefficient is optimized and calibrated based on historical data using a grid search method.

[0042] The weighting coefficient automatically increases when the cumulative rainfall in the target grid area exceeds a preset threshold. The weight values ​​are used to compensate for the decrease in image acquisition quality under high humidity conditions. The impact of accuracy.

[0043] Step 4: Execute tiered early warning triggering and push notifications.

[0044] According to the comprehensive risk index Execute tiered early warning trigger: When When the value is greater than or equal to the first warning threshold, a Level 1 warning (high risk) is triggered, and emergency response suggestions are generated and pushed to the maintenance personnel's terminal. when When the threshold is greater than or equal to the second warning threshold but less than the first warning threshold, a level two warning (attention) is triggered, and preventive action recommendations are generated. when When the value is below the second warning threshold, no warning is triggered, and the monitoring data is recorded for continuous model training.

[0045] Example 3: This embodiment provides a method for generating prevention and control plans based on large language model reasoning, realizing a complete closed-loop management from pest and disease identification to precise treatment.

[0046] Step 1: Construct a professional knowledge base for garden pests and diseases and a prompt word project template.

[0047] The professional knowledge base contains the following structured knowledge: basic information on diseases and pests: names, scientific names, biological characteristics, and host plant ranges of various diseases and pests; Knowledge of prevention and control methods: Recommended prevention and control measures for different pests and diseases and different levels of damage, including physical control methods, biological control methods and chemical control methods; Pesticide Application Guidelines: This includes the applicable targets, application concentrations, application timing, and safety intervals for various pesticides. Data is sourced from plant protection technical regulations issued by the agricultural department. The prompt word template includes the following structured elements: current pest and disease identification results, including pest and disease type, damage level, location of discovery, and detection confidence level. Environmental information for the target area, including current meteorological conditions and vegetation growth status; Output format constraints require large models to output prevention and control plans in a structured format, including at least prevention and control measures, priority order, and precautions.

[0048] Step 2: Perform large-scale model prevention and control scheme reasoning and generation.

[0049] When the graded early warning trigger module outputs early warning information, the big model intelligent decision-making module automatically generates a customized prevention and control plan based on the type of pests and diseases, the level of damage, the plant species and the current environmental information. The generation process includes: retrieving matching prevention and control knowledge entries from the professional knowledge base. Select appropriate prevention and control time windows based on current meteorological conditions; The priority of prevention and control and the intensity of pesticide application should be determined based on the severity of the hazard. The prevention and control plan is output in a structured text format, along with the justification for its interpretability. The justification for the prevention and control measures should include at least the reasons for the recommendations and the sources cited.

[0050] Step 3: Implement intelligent order dispatch and closed-loop processing.

[0051] Based on the prevention and control plan and the current location of the maintenance personnel, the system automatically generates an electronic work order containing the responsible personnel, the content of the treatment, and the completion deadline. The work order is sent to the maintenance personnel's terminal and synchronized to the garden management platform. After the maintenance personnel complete the treatment, they upload the treatment result information to the system through the terminal. The treatment result information includes the treatment time, treatment measures, and on-site images after the treatment.

[0052] The system automatically assigns review tasks based on the treatment results and evaluates the prevention and control effect through image comparison analysis. When the prevention and control effect fails to meet the preset standard, a secondary treatment process is triggered. When the prevention and control effect meets the preset standard, the closed loop is completed and the relevant data is stored in the historical knowledge base for model iteration and optimization.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart early warning system for pests and diseases in landscaping based on machine vision and AI large-scale models, characterized in that, include: The edge target detection module is deployed on the edge computing nodes in the garden monitoring area. It is used to perform real-time target detection on garden vegetation images acquired by the front-end acquisition equipment and output the detection results of suspected pest and disease areas and the first confidence level. The cloud-based multimodal large model verification module, deployed on a cloud server, is used to receive image cropping data from the detection results of the suspected pest and disease areas, perform fine-grained recognition and hazard level assessment, and output a second confidence level. The confidence level arbitration module is used to make consistency judgments and collaborative decisions based on the first confidence level and the second confidence level. The early warning decision module is used to generate graded early warning information based on the output of the confidence arbitration module; the edge target detection module adopts an improved YOLO network structure, and the cloud-based multimodal large model verification module adopts a multimodal large language model that has been fine-tuned with a professional corpus of garden pests and diseases.

2. The system according to claim 1, characterized in that, The confidence level arbitration module executes the following judgment logic: When the first confidence level is greater than or equal to the first preset threshold and the second confidence level is greater than or equal to the second preset threshold, the detection result is determined to be a confirmed diagnosis. When the first confidence level is greater than or equal to the first preset threshold and the second confidence level is less than the second preset threshold, the detection result is determined to be suspicious, and a manual review mark is triggered. When the first confidence level is less than the first preset threshold, the detection result is ignored; wherein the first preset threshold is lower than the second preset threshold.

3. The system according to claim 1, characterized in that, The improved YOLO network uses an optimized feature pyramid structure to enhance the feature extraction capability of small target pests and diseases, and introduces an attention mechanism to reduce the impact of complex background interference on detection accuracy; the multimodal large model is an open-source multimodal large model that has undergone incremental pre-training and instruction fine-tuning, and has the ability to understand pest and disease images and reason about prevention and control knowledge.

4. The system according to claim 1, characterized in that, The edge target detection module and the cloud multimodal large model verification module use an asynchronous message queue for data transmission. The cropped data of the suspected pest and disease area image is batch cached at the edge and then sent to the cloud through a compressed transmission protocol. When the cloud network connection is interrupted, the edge target detection module uses the edge detection result as a temporary output and caches the image data locally. After the network is restored, it is resent to the cloud for secondary verification.

5. The system according to claim 1, characterized in that, It also includes a multi-source data fusion module, the input data of which includes: Pest and disease identification results data, including pest and disease type, damage level, discovery location and timestamp; Meteorological and environmental data, including temperature, humidity, precipitation, wind speed, and light intensity; Historical records of pests and diseases, including data on the occurrence of pests and diseases over the past 3-5 years and the corresponding meteorological conditions; Vegetation growth status data; the multi-source data fusion module performs weighted fusion based on the timeliness and confidence level of various types of data, and outputs a comprehensive risk index for the target area.

6. The system according to claim 5, characterized in that, The formula for calculating the comprehensive risk index is as follows: in The confidence score for the currently detected pests and diseases. This is a predicted probability of future occurrence based on historical pest and disease records and weather forecast data. Score the vegetation growth status. , , These are the corresponding weighting coefficients; the weighting coefficients are dynamically adjusted based on the environmental characteristics and pest and disease occurrence patterns of the target area.

7. The system according to claim 5, characterized in that, It also includes a pest and disease occurrence probability prediction module. The prediction module adopts a time series analysis and machine learning fusion algorithm, uses historical pest and disease records and corresponding historical meteorological data as training samples, and uses future meteorological forecast data as input to output the predicted probability values ​​of various pests and diseases in the target area in the future period.

8. The system according to claim 1, characterized in that, It also includes a tiered early warning module, which triggers tiered early warnings based on the comprehensive risk index: When the comprehensive risk index is greater than or equal to the first warning threshold, a level one warning is triggered and emergency response recommendations are generated. When the comprehensive risk index is greater than or equal to the second warning threshold and less than the first warning threshold, a level-two warning is triggered and preventive action recommendations are generated. When the comprehensive risk index is less than the second warning threshold, the monitoring data is recorded for continuous model training.

9. The system according to claim 1, characterized in that, It also includes a large-scale intelligent decision-making module. After a graded early warning is triggered, the large-scale intelligent decision-making module automatically generates a customized prevention and control plan based on the type of pest and disease, the level of damage, the plant species, and the current environmental information, combined with a professional knowledge base of garden pests and diseases. The professional knowledge base contains basic information on various pests and diseases, knowledge of prevention and control methods, and specifications for the use of pesticides. The prevention and control plan is output in the form of structured text and includes interpretable evidence for the plan.

10. The system according to claim 1, characterized in that, It also includes an intelligent work order dispatching and closed-loop management module. The intelligent work order dispatching and closed-loop management module automatically generates electronic work orders based on the prevention and control plan and the location information of maintenance personnel. The work order includes the responsible personnel, the content of the treatment, and the completion deadline. After the maintenance personnel complete the treatment, they upload the treatment result information. The system automatically arranges review tasks based on the treatment result information, evaluates the prevention and control effect through image comparison analysis, and completes the closed loop when the prevention and control effect reaches the preset standard and stores the data in the historical knowledge base.

Images