Methods and devices for risk assessment of pests and diseases in facility agriculture

Abstract

This invention discloses a method and device for risk assessment of pests and diseases in facility agriculture, belonging to the field of agricultural monitoring technology. The invention comprises five stages: device deployment, multi-source heterogeneous data acquisition, creative data processing, creative risk assessment, risk output, and linkage with prevention and control. It integrates an improved ResNet50 and Outlook attention mechanism to construct a high-precision identification model for multiple crop pests and diseases in facility agriculture. Combined with an adaptive fusion algorithm for multi-source heterogeneous data, a dynamic weight adjustment formula based on the pest and disease life cycle and identification accuracy, and a full-cycle risk evolution prediction model, it achieves a complete closed loop of "identification-fusion-assessment-control-feedback-optimization". This invention constructs a dedicated identification model for multiple crop pests and diseases in facility agriculture by integrating an improved ResNet50 and Outlook attention mechanism. Combined with image data enhancement and transfer learning optimization, it improves the accuracy of multi-crop pest and disease identification and the relevance and accuracy of risk assessment, while reducing the cost of manual inspection.

Description

Technical Field

[0001] This invention belongs to the field of agricultural monitoring technology, and in particular relates to a method and device for risk assessment of pests and diseases in facility agriculture, which is applicable to high-precision identification, risk assessment and coordinated prevention and control of multiple crop pests and diseases in facility agriculture scenarios such as greenhouses and sheds. Background Technology

[0002] Facility agriculture, as an important component of modern agriculture, is characterized by strong controllability, high yield, and superior quality, and is widely used to cultivate cash crops such as tomatoes, cucumbers, and strawberries. However, the enclosed / semi-enclosed environment of facility agriculture makes it highly susceptible to the covert occurrence and rapid spread of pests and diseases, severely impacting crop yield and quality. Annual yield losses due to pests and diseases in facility crops can reach [amount missing]. In some severely affected areas, the figures are even higher.

[0003] Existing pest and disease risk assessment technologies for facility agriculture have many shortcomings: First, the accuracy of pest and disease identification is low. Traditional identification relies on human experience, which is inefficient and highly subjective. Existing identification algorithms based on convolutional neural networks are poorly adapted to the identification of pests and diseases in facility environments with small targets (insect eggs, early lesions) and complex backgrounds, and the identification accuracy is difficult to meet actual needs. Second, the multi-source data fusion method is rigid, and fixed weight fusion is often used, which cannot adapt to dynamic changes in the facility environment and differences in data reliability, resulting in low fusion accuracy. Third, the risk assessment model is static, which can only assess current risks and cannot predict the evolution of future risks, resulting in lagging prevention and control. Fourth, the identification, assessment, and prevention and control are disconnected, failing to form a complete closed loop and lacking practicality.

[0004] Therefore, there is an urgent need for a risk assessment solution for pests and diseases in facility agriculture that integrates high-precision identification, dynamic evaluation, full-cycle prediction, and closed-loop prevention and control, to overcome the limitations of existing technologies and meet actual production needs. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for risk assessment of pests and diseases in facility agriculture. By integrating and improving ResNet50 and Outlook attention mechanisms, a high-precision identification model for multiple crop pests and diseases in facility agriculture is constructed. Combined with an adaptive fusion algorithm for multi-source heterogeneous data, a dynamic weight adjustment formula based on the life cycle of pests and diseases and identification accuracy, and a full-cycle risk evolution prediction model, this invention solves the existing technical problems of low pest and disease identification accuracy, lagging risk assessment, rigid data fusion, and disconnect between prevention and control.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a method for assessing the risk of pests and diseases in facility agriculture, comprising the following steps: Step S1, Device Deployment: Deploy multi-source fusion monitoring terminals, edge computing modules, cloud assessment platforms, user interaction terminals, and linkage prevention and control interfaces; complete communication debugging of each module; and build a facility agriculture pest and disease assessment network. Step S2, Data Acquisition: Simultaneously collect environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data in the facility agriculture scenario through a multi-source fusion monitoring terminal; Step S3, Data Processing: The collected data is preprocessed through the edge computing module and cloud evaluation platform, and pest and disease identification based on improved ResNet50 and Outlook attention and adaptive fusion of multi-source heterogeneous data are generated to produce pest and disease feature fusion vectors. Step S4, Risk Assessment: Based on the fusion vector of pest and disease features and the results of pest and disease identification, the current risk index is calculated through dynamic weight adjustment. Combined with the full-cycle risk evolution, a pest and disease risk assessment model is constructed to predict future risks. The changing trend of pest and disease risk over the days; Step S5, Risk Output and Prevention Linkage: Risk assessment and identification results are output through the user interaction terminal. When the risk exceeds the standard, linkage prevention and control are triggered. After prevention and control, the algorithms and models are iteratively optimized through data feedback to form a virtuous cycle.

[0007] Furthermore, in step S2, the specific process of data acquisition is as follows: The environmental data is collected continuously at a frequency of once every 10 minutes. The crop physiological data is recorded synchronously at a frequency of once per hour and linked to the crop growth cycle file. The image data in the pest and disease characteristic data is collected at a fixed size of 224×224 pixels, and multiple sets of images under different lighting conditions are collected simultaneously. The biological characteristic data is collected once per hour. The historical correlation data and real-time collected data are precisely aligned in time and space through spatial coordinates (monitoring node grid positions) and timestamps. All data carries a unique monitoring node identifier, plot grid coordinates and millisecond-level timestamps to ensure the temporal synchronization, spatial correlation and full-process traceability of multi-source data. At the same time, it provides high-precision and high-reliability data support for subsequent adaptive fusion of multi-source data and risk assessment.

[0008] Furthermore, in step S3, the preprocessing of the collected data includes the following steps: The edge computing module prioritizes local preprocessing of the collected raw data. It uses an adaptive median filtering algorithm to reduce noise in environmental data and crop physiological data, and uses Gaussian filtering combined with an edge protection algorithm to reduce noise in image data while preserving pest and disease characteristics. At the same time, it uses linear interpolation to fill in intermittent missing data, thus completing invalid data screening and preliminary noise reduction. The cloud-based evaluation platform performs secondary optimization on the data preprocessed by the edge computing module. It expands the sample diversity by using multi-dimensional enhancement operations on image data, and performs secondary standardization and calibration on environmental, crop physiological and biological characteristic data to eliminate systematic errors generated during edge preprocessing. At the same time, it establishes a preprocessed data quality verification mechanism to remove substandard abnormal data, achieving hierarchical collaboration, precision and efficiency in the preprocessing process. This reduces the pressure on cloud data transmission and ensures the high accuracy of the preprocessed data, providing a high-quality data foundation for subsequent pest and disease identification and multi-source data fusion.

[0009] Furthermore, the specific process for pest and disease identification in step S3 is as follows: Step S31, Basic Network Optimization: Based on ResNet50, retain its residual skip connection structure to alleviate the gradient vanishing and gradient exploding problems in deep network training. In the four network stages of ResNet50, contextual convolutional blocks are fused. The contextual convolutional blocks are implemented using dilated convolution with dilation rates of 1, 2, and 3 respectively. The receptive field of the convolutional kernel is expanded without increasing the actual size of the convolutional kernel. This integrates the contextual information of pests and diseases in the image and improves the recognition accuracy of small target pests and diseases. Step S32, Attention Mechanism Adaptation: An improved Outlook attention mechanism is introduced, and the multi-scale local window partitioning strategy is optimized. The preprocessed pest and disease images are divided into 8×8 and 16×16 multi-scale local windows. Each window serves as an independent attention calculation unit. The input feature map is converted into query (Q), key (K), and value (V) through a linear layer. Attention weights are calculated within the local window and global context information is fused to enhance the model's ability to acquire local salient features of pests and diseases and suppress interference from complex backgrounds in facility agriculture. Step S33, Model parameter fine-tuning: Using a transfer learning strategy, the parameters of the ResNet50 model pre-trained on the ImageNet dataset are transferred to the pest and disease identification model. Fine-tuning is then performed using a multi-crop pest and disease dataset for facility agriculture to adapt to the pest and disease identification scenarios of tomatoes, cucumbers, strawberries, and other crops. Step S34, Output of identification results: The Softmax classifier outputs the type of pests and diseases, the identification accuracy, and the degree of damage. The identification accuracy is not less than 98.5%, which is no less than 3.2 percentage points higher than the standard ResNet50 model. This achieves high-precision identification of pests and diseases in facility agriculture with multiple crops, small targets, and complex backgrounds.

[0010] Furthermore, in step S32, during the attention mechanism adaptation, an improved Outlook attention mechanism is introduced. The preprocessed 224×224 pixel pest image is simultaneously divided into local windows of two scales: 8×8 and 16×16. The 8×8 window is used to capture the detailed features of small targets such as insect eggs and early tiny lesions, while the 16×16 window is used to capture the overall features of larger targets such as lesion spread areas and larval clusters. Each local window serves as an independent attention calculation unit, and the input pest image feature map is converted into a query matrix (Q) and a key matrix (K) through a linear transformation layer. The model uses a value matrix (V), where the dimensions of Q, K, and V are uniformly set to 256 to ensure computational consistency. Within each local window, local attention weights are obtained through the attention weight calculation formula. At the same time, global context features of the image are extracted through global average pooling to generate a global context key matrix (K), and global attention weights are calculated. A local-global weight collaborative adjustment mechanism is introduced, which integrates local attention weights and global attention weights through preset weight coefficients to enhance the model's ability to capture local salient features of pests and diseases (lesion color, egg outline, larval morphology) while suppressing invalid interference from the complex background of facility agriculture.

[0011] Further, in step S3, during the adaptive fusion of multi-source heterogeneous data, the preprocessed environmental data (E), crop physiological data (P), pest and disease characteristic data (D), and historical correlation data (H) are normalized to obtain standardized data; based on the correlation between each data source and pest and disease risk and the identification accuracy, the fusion weights of the environmental data (E), crop physiological data (P), pest and disease characteristic data (D), and historical correlation data (H) are dynamically allocated. And the sum of the fusion weights of environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data is 1 ( ).

[0012] Furthermore, the specific process of risk assessment in step S4 is as follows: Step S41: Using the pest and disease feature fusion vector and pest and disease identification results (including identification type and identification accuracy) generated in Step S3, calculate the dynamic evaluation weight vector at the current moment using a dynamic weight adjustment strategy based on the linkage between the pest and disease life cycle and identification accuracy; Step S42: Based on the dynamic evaluation weight vector and pest and disease feature fusion vector, and combined with the risk correction coefficient and the damage severity correction coefficient, calculate the current pest and disease risk index, where the risk correction coefficient is set in conjunction with the historical pest and disease occurrence probability, and the damage severity correction coefficient is set according to the classification of the damage severity (mild, moderate, severe) of pests and diseases to crops; Step S43: Integrating the Future This study uses a full-cycle risk evolution prediction model to predict future trends in pest and disease risk by analyzing changes in environmental data, crop physiological data, pest and disease characteristic data, and identification accuracy over a given period. The model assigns appropriate adaptation coefficients based on the degree of influence of changes in each data source, dynamically predicting risk increases and decreases. This risk assessment strategy addresses the shortcomings of existing risk assessment technologies, such as static assessment, strong lag, and inability to predict future risks. It achieves synergy between current status assessment and future prediction, enabling proactive risk assessment. The weather forecast provides precise support for the early prevention and control of pests and diseases in facility agriculture.

[0013] In step S41, the adjustment of the dynamic evaluation weight vector is based on the preset initial weight of the corresponding crop-pest, the life cycle correction coefficient of the pest, the environmental adaptation coefficient, and the identification accuracy correction coefficient. Among them, the life cycle correction coefficient of the pest is set according to the growth stage of the pest; the environmental adaptation coefficient is dynamically calculated based on the standardized environmental data; and the identification accuracy correction coefficient is set according to the classification of the identification accuracy of the pest, with a higher correction coefficient for a higher identification accuracy.

[0014] Furthermore, in step S5, the specific process of linking risk output with prevention and control is as follows: Step S51: Conduct tiered and precise risk output. Based on the risk assessment results, generate differentiated output content according to risk level, pest / disease type, and degree of damage. Low-risk output includes identification results and daily prevention and control tips; medium-risk output includes risk causes, identification accuracy, and targeted prevention and control suggestions; high-risk output includes real-time early warning information and emergency prevention and control plans. Output channels are simultaneously covered to user interaction terminals (mobile APP, web page), supporting visual data dashboard display, custom report generation, and historical data backtracking. At the same time, it adapts to different user needs (growers, plant protection personnel, managers) by providing differentiated display interfaces. Step S52: Construct a hierarchical linkage prevention and control mechanism, and preset linkage trigger thresholds for different risk levels; among them, low risk triggers manual prevention and control reminders, medium risk triggers semi-automatic prevention and control linkage (intelligent devices stand by, and start after manual confirmation), and high risk automatically triggers linkage prevention and control interfaces, which can be seamlessly connected with intelligent insecticidal lamps, spray robots, and natural enemy release devices to automatically execute prevention and control operations according to preset prevention and control plans, and support users to manually adjust prevention and control parameters, pause or restart the prevention and control process through interactive terminals to ensure the flexibility of prevention and control; Step S53: Establish a dynamic feedback optimization closed loop. After the prevention and control measures are implemented, the multi-source fusion monitoring terminal continuously collects crop growth data, pest and disease decline data, and prevention and control effect data, and feeds them back to the cloud evaluation platform in real time. Based on the feedback data, the platform iteratively optimizes the pest and disease identification model, multi-source data fusion algorithm, dynamic weight adjustment strategy, and full-cycle risk evolution prediction model, and corrects relevant parameters to adapt to the changes in the scenario after prevention and control. This achieves a virtuous cycle of "risk output - prevention and control implementation - data feedback - model optimization", improves the accuracy of subsequent risk assessment, identification, and prevention and control, and adapts to the dynamic prevention and control needs of multiple crops and multiple scenarios in facility agriculture.

[0015] This invention is a risk assessment device for pests and diseases in facility agriculture, comprising a multi-source fusion monitoring terminal, an edge computing module, a cloud assessment platform, a user interaction terminal, and a linkage prevention and control interface; The multi-source fusion monitoring terminal is deployed in a grid pattern according to agricultural facility plots. Each monitoring node integrates an optical sensing module, a biosensing module, a high-resolution macro image sensing module, an environmental sensing module, and a miniature positioning module. It is self-powered by solar energy and has a small size. It can be attached to crop leaves / branches to simultaneously collect environmental data, crop physiological data, and pest and disease characteristic data, and send the data to the edge computing module. The size of the collected image data is uniformly 224×224 pixels to adapt to subsequent recognition algorithms. It realizes multi-dimensional data synchronous collection, focuses on optimizing the shooting parameters of the image sensing module, and adapts to the scene of uneven lighting and complex background in greenhouses, solving the problems of insufficient coverage, partial data and poor image quality of existing single monitoring devices. The edge computing module is deployed at the facility agriculture site, with one module configured in each greenhouse. It has a built-in data preprocessing unit, a preliminary feature extraction algorithm, and a lightweight recognition sub-model. It is used to locally filter, reduce noise, preprocess, and identify pests and diseases in the collected raw data, reduce invalid data transmission, and provide support for high-precision recognition in the cloud. The cloud-based assessment platform is deployed on a cloud server to integrate cross-domain historical data, run core identification algorithms, multi-source heterogeneous data adaptive fusion algorithms, dynamic weight risk assessment algorithms, and full-cycle risk evolution prediction models to achieve high-precision identification, risk calculation, and evolution prediction of pests and diseases; the core identification algorithm is a multi-crop pest and disease identification algorithm based on improved ResNet50 and Outlook attention. The user interaction terminal includes a mobile APP and a web page, which are used to receive risk assessment results, pest and disease identification results, early warning information and control suggestions, and support users to manually upload crop images and supplement field data. The linkage and control interface is used to connect with intelligent insecticidal lamps, spraying robots, and natural enemy release devices. When the risk exceeds the standard, it triggers automatic prevention and control linkage to form a closed loop of "identification-assessment-control".

[0016] The present invention has the following beneficial effects: (1) This invention constructs a special identification model for multiple crop diseases and pests in facility agriculture by integrating the improved ResNet50 and Outlook attention mechanism. Combined with image data enhancement and transfer learning optimization, it effectively improves the identification accuracy of small target diseases and pests (insect eggs, early small lesions) and the identification accuracy in complex backgrounds, thereby improving the identification accuracy of multiple crop diseases and pests.

[0017] (2) This invention integrates multi-dimensional data such as environment, crop physiology, pest and disease characteristics and historical correlation through a multi-source heterogeneous data adaptive fusion algorithm, which breaks the limitations of single sensor or image data in the existing technology, eliminates data redundancy and dimensional influence, improves the accuracy of data fusion, provides high-quality and comprehensive feature data for risk assessment, and realizes the spatiotemporal alignment and efficient utilization of multi-source data, avoids the problem of data silos, and lays a solid data foundation for accurate risk assessment.

[0018] (3) This invention replaces the traditional fixed weight assessment mode by designing a dynamic weight adjustment formula based on the life cycle of pests and diseases and the accuracy of identification. It can adapt to the dynamic changes of the closed / semi-closed environment of facility agriculture and the accuracy of pest and disease identification results, improve the pertinence and accuracy of risk assessment, solve the technical pain points of traditional static assessment being lagging behind and out of touch with actual scenarios, and make the risk assessment results more in line with the actual planting needs in the field.

[0019] (4) This invention constructs a closed-loop system of “data collection-identification-processing-evaluation-output-prevention” by deploying a multi-source fusion monitoring terminal, an edge computing module, a cloud evaluation platform and a linkage prevention and control interface. The edge computing module realizes local data preprocessing, reduces invalid data transmission and improves the real-time performance of processing. The linkage prevention and control interface can be connected to intelligent prevention and control equipment to realize automatic linkage prevention and control when the risk exceeds the standard, which greatly improves the level and efficiency of pest and disease prevention and control in facility agriculture and reduces the cost of manual inspection.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for assessing the risk of pests and diseases in facility agriculture according to the present invention; Figure 2 This is a schematic diagram of the structure of a facility agriculture pest and disease risk assessment device according to the present invention; Figure 3 A schematic diagram of a network model for fusing contextual convolutional blocks; Figure 4 This is a schematic diagram illustrating the principle of contextual convolutional block expansion in this invention. Figure 5 This is the loss curve for the training iterations in Example 3; Figure 6 This is the accuracy curve for the training iteration in Example 3. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0025] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1-6The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0026] See Figure 1 As shown, this invention is a method for risk assessment of pests and diseases in facility agriculture, and its overall implementation scheme is as follows: Step S1, Device Deployment: Deploy multi-source fusion monitoring terminals, edge computing modules, cloud assessment platforms, user interaction terminals, and linkage prevention and control interfaces; complete communication debugging of each module; and build a facility agriculture pest and disease assessment network. Step S2, Data Acquisition: Simultaneously collect environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data in the facility agriculture scenario through a multi-source fusion monitoring terminal; Step S3, Data Processing: The collected data is preprocessed through the edge computing module and cloud evaluation platform, and pest and disease identification based on improved ResNet50 and Outlook attention and adaptive fusion of multi-source heterogeneous data are generated to produce pest and disease feature fusion vectors. Step S4, Risk Assessment: Based on the fusion vector of pest and disease features and the results of pest and disease identification, the current risk index is calculated through dynamic weight adjustment. Combined with the full-cycle risk evolution, a pest and disease risk assessment model is constructed to predict future risks. The changing trend of pest and disease risk over the days; Step S5, Risk Output and Prevention Linkage: Risk assessment and identification results are output through the user interaction terminal. When the risk exceeds the standard, linkage prevention and control are triggered. After prevention and control, the algorithms and models are iteratively optimized through data feedback to form a virtuous cycle.

[0027] In step S1, prior to deployment, preliminary planning needs to be completed based on the scale of the facility agriculture site, crop type, planting density, and existing infrastructure to provide a basis for subsequent deployment. This includes: Clearly define the total area, zoning, crop planting layout (row spacing, plant spacing), existing power supply (mains / solar power), and communication (WiFi / 4G / 5G) coverage of the facility site (greenhouse / greenhouse). Mark monitoring blind spots, areas with weak communication, and the optimal deployment locations for edge computing modules and linked prevention and control equipment. Determine the deployment density of multi-source fusion monitoring terminals based on crop type (tomato, cucumber, strawberry, etc.) (generally based on 50...). One monitoring node; high-density planting scenarios can be adjusted to each 1 unit); determine the number of edge computing modules to deploy based on the number of site zones (at least 1 per independent greenhouse / greenhouse); determine the communication method for each module based on site communication conditions (preferably WiFi 6, with 4G / 5G modules for remote scenarios). Inventory all hardware equipment and accessories for each module, including multi-source fusion monitoring terminals (including optical, biological, image, environmental, and positioning sensor modules and solar power modules), edge computing modules (including host, heat dissipation accessories, and communication modules), cloud servers (or cloud service accounts), user interaction terminals (mobile phones, computers), linkage prevention and control interfaces (including adapters and communication cables), and installation accessories (brackets, fixing clips, waterproof seals, etc.), ensuring that the equipment is undamaged and all accessories are complete.

[0028] The multi-source fusion monitoring terminal is the core of data acquisition. Its deployment needs to take into account the comprehensiveness, accuracy, and stability of data acquisition. The specific implementation is as follows: According to the pre-planned monitoring points, use dedicated brackets or fixing clips to attach the monitoring terminals to the back of the upper and middle leaves of the crop (to avoid direct sunlight affecting image acquisition and sensing accuracy), or fix them to the crop branches at the same height as the crop canopy (from the ground). To ensure the sensor module can properly collect environmental, crop physiological, and pest and disease characteristic data, the monitoring terminals should be evenly spaced to avoid blind spots, with slightly denser monitoring points in edge areas. A solar-powered self-powered system should be used, with solar panels fixed above the monitoring terminals in an unobstructed area (ensuring sufficient daily sunlight). (Hours), connect the power supply interface to the monitoring terminal and ensure waterproof sealing (use waterproof tape to wrap the interface to avoid short circuits caused by the high humidity environment inside the greenhouse); reserve a spare mains power interface for emergency power supply during continuous rainy days or insufficient sunlight.

[0029] Start the monitoring terminal, adjust the parameters of each sensor module, and ensure the environmental sensor modules (temperature, humidity, light, etc.) are functioning correctly. The sampling range and sampling frequency (concentration, etc.) meet the preset requirements (sampling frequency: environmental data once every 10 minutes, crop physiological data once every hour, and pest and disease image data once every 2 hours); the image sensing module is adjusted to shoot parameters (focal length, exposure time) to adapt to uneven lighting in the greenhouse, ensuring that the collected images are uniformly 224×224 pixels, and clearly capturing small target features such as lesions and insect eggs; the positioning module is adjusted to ensure that the spatial coordinates of the monitoring nodes can be accurately obtained (positioning accuracy ±1m) and correspond to the grid position of the plot.

[0030] The edge computing module is responsible for local data preprocessing. Its deployment needs to be close to the monitoring terminal to reduce data transmission latency. The specific implementation is as follows: Deployed at the entrance or central area of ​​the facility site for easy connection to all monitoring terminals, and in a well-ventilated, dry, and unobstructed location, avoiding direct sunlight and rain immersion, with sufficient space for heat dissipation (distance from obstacles around the module). Mount the edge computing module on a dedicated bracket and connect it to a power supply (220V AC, equipped with a voltage regulator). Connect the communication module (WiFi 6 / 4G / 5G) and ensure stable signal transmission. Connect the communication line to the monitoring terminal (wireless communication requires pairing; wired communication requires dedicated cabling with waterproof protection). Install the data preprocessing unit, lightweight identification sub-model, and communication protocol in the edge computing module. Debug the module's operating parameters, set the local data storage capacity (at least 7 days' worth of local data storage), and configure the data transmission threshold (transmit only valid data, discarding invalid and abnormal data). Ensure the module can normally receive the raw data collected by the monitoring terminal and complete preliminary processing.

[0031] The cloud-based assessment platform is responsible for core algorithm operation, data integration, and risk assessment. Deployment can utilize cloud servers or cloud services, and the specific implementation is as follows: If using a local cloud server, it must be deployed in a dedicated data center (temperature). ,humidity Connect to the external network and configure server parameters (CPU). Core, memory ,harddisk ,bandwidth If using cloud services (such as Alibaba Cloud or Tencent Cloud), you need to register a cloud service account, activate the corresponding cloud server configuration, and complete the server environment setup (installing the operating system, database, and software required for algorithm operation). Deploy the core identification algorithm (a multi-crop pest and disease identification algorithm based on an improved ResNet50 and Outlook attention), a multi-source heterogeneous data adaptive fusion algorithm, a dynamic weight risk assessment algorithm, and a full-cycle risk evolution prediction model on the cloud server; import cross-domain historical data (historical pest and disease data, crop growth cycle data, concurrent meteorological data, plant protection and control case data, etc.) to complete database setup and data initialization. Configure the communication protocol between the cloud evaluation platform and the edge computing module to ensure that the edge computing module can upload the pre-processed data to the cloud in real time, and that the cloud can issue parameter adjustment commands to the edge computing module; test the cloud data reception, storage, and processing efficiency to ensure no data loss or delay (data transmission delay). s).

[0032] The user interaction terminal is used for data viewing and command issuance. Its deployment needs to be adapted to the needs of different users (growers, plant protection personnel, and managers). The specific implementation is as follows: When deploying the mobile app, install a dedicated app on the user's mobile phone (supports Android 10.0 and above, iOS 14.0 and above), complete registration, login, and permission allocation (growers can only view data for their own plots, while managers can view data for all plots and issue instructions); test the communication between the app and the cloud assessment platform to ensure real-time receipt of risk assessment results, early warning information, and prevention and control suggestions, and support manual uploading of crop images and supplementation of field data. When deploying the web version, deploy the web version on a cloud server, set the access address, and allow users to access it through a computer browser (Chrome, Edge, etc.). After logging in, users can view visualized data dashboards, historical data, generate custom reports, and issue prevention and control instructions; test the web version's response speed to ensure a smooth interface and timely data updates.

[0033] The linkage prevention and control interface is used to connect intelligent prevention and control equipment to achieve automatic linkage when the risk exceeds the standard. The specific implementation is as follows: The integrated control interface is hardware-interfaced with existing intelligent control equipment (intelligent insecticidal lamps, spraying robots, natural enemy release devices, etc.) within the facility, using standardized adapters to ensure interface compatibility. The interface communication protocol is debugged to ensure the cloud-based assessment platform can issue control commands to the control equipment (such as spraying robots applying pesticides or activating insecticidal lamps), and the control equipment can report its operational status back to the cloud. Integrated control trigger thresholds are set in the cloud-based assessment platform (preset based on crop type and pest / disease type, such as the current risk index). Or predict risk index Automatic control is triggered in real time; configure the operating parameters of the control equipment (such as the spraying robot's application path and dosage, the insecticidal lamp's turn-on time, etc.) to ensure accurate and efficient control operations.

[0034] Example 1: In this example, the area selected is... Using a tomato greenhouse as an experimental setting, the facility agriculture pest and disease risk assessment device described in this invention was deployed. The specific deployment steps are as follows: Multi-source fusion monitoring terminal deployment: per The standard for one monitoring node is 10 nodes evenly deployed within the greenhouse. The nodes are attached to the underside of the upper and middle leaves of the tomato plants (to avoid direct sunlight affecting image acquisition). Each node integrates: an optical sensing module (model: OPT-100), a biosensor module (model: BIO-200), and a high-resolution macro image sensing module (model: CAM-300, 12 megapixels, shooting parameters: focal length 5cm, exposure time 1 / 100s, adapted to the light intensity inside the greenhouse). ), Environmental sensing module (model: ENV-400, temperature and humidity collection range) , Light intensity , concentration The system includes a miniature positioning module (model: GPS-100, positioning accuracy ±1m), which uses a 5V solar self-powered module and measures 18cm×12cm×8cm to ensure stable operation. Edge computing module deployment: One edge computing module (model: ECM-500, CPU: Intel Core i5-12400, memory: 16G, storage: 128G) is deployed at the entrance of the greenhouse. It has a built-in data preprocessing unit and a lightweight identification sub-model (built on MobileNetV3). It connects to 10 monitoring nodes through wireless communication (WiFi 6) to realize real-time local data processing. Deployment of the cloud-based evaluation platform: It adopts Alibaba Cloud ECS server (configuration: 16 CPU cores, 32G memory, 1T hard disk, 100M bandwidth) to deploy core identification algorithms, fusion algorithms and evaluation algorithms. It communicates with the edge computing module through 4G / 5G network to integrate historical data (tomato greenhouse pest and disease data, tomato growth cycle data, local meteorological data of the same period, and plant protection and control case data of the past 3 years). User interaction terminal deployment: Install a dedicated mobile APP (supports Android 10.0 and above, iOS 14.0 and above), and deploy the web version on a cloud server to enable data viewing and command issuance functions; Deployment of the linkage prevention and control interface: Connect the linkage prevention and control interface with the existing intelligent insecticidal lamps (model: IL-200), spraying robots (model: SR-300), and natural enemy release devices (model: NR-100) in the greenhouse, and set the linkage trigger threshold (current risk index). Or predict risk index After debugging, it was put into use.

[0035] In some embodiments, device deployment also requires collaborative debugging and acceptance testing. For example, during single-module testing, the operational status of each module is tested separately to ensure normal data acquisition by the multi-source fusion monitoring terminal, effective preprocessing by the edge computing module, stable operation of the cloud-based assessment platform algorithm, normal data display on the user interaction terminal, and smooth interface connection for joint prevention and control. During collaborative testing, all modules are started to simulate the entire process of data acquisition, transmission, processing, assessment, and prevention and control, testing the collaborative working effect of each module. The focus is on testing the real-time performance of data transmission, the accuracy of risk assessment, and the timeliness of prevention and control linkage, while troubleshooting communication failures, data loss, and command delays. When each module operates stably and without faults, and the accuracy of data acquisition by the monitoring terminal is high... Data transmission delay Local preprocessing efficiency of edge computing modules Accuracy of risk assessment on cloud-based assessment platform Response time of the joint prevention and control interface When the user interaction terminal data is updated in a timely manner and the interface is smooth, the deployment and acceptance are complete.

[0036] In step S2, before data collection, it is necessary to combine the conditions of the facility agriculture site, crop type, and key pest and disease monitoring points to complete the setting of collection parameters, equipment debugging, and data planning, laying the foundation for standardized collection. Specifically, based on crop type (tomato, cucumber, strawberry, etc.) and pest and disease monitoring needs, determine the collection frequency, collection range, and accuracy requirements for various types of data, and clarify the labeling specifications for timestamps and spatial coordinates (timestamps use millisecond level, and spatial coordinates use monitoring node grid coordinates, corresponding one-to-one with the actual location of the plot); conduct comprehensive debugging of each sensor module of the multi-source fusion monitoring terminal, including environmental sensor module, crop physiological sensor module, image sensor module, biological sensor module, and positioning module, to ensure that each module operates normally and that the accuracy and resolution of the collected data meet the preset requirements without deviation; clarify the storage format and identification rules for various types of data, assign a unique identifier to each type of data, and associate it with the monitoring node number, crop type, and collection time to ensure data traceability; at the same time, establish a temporary local storage mechanism to avoid data loss during the collection process, and synchronously upload it to the edge computing module and cloud evaluation platform after collection is completed.

[0037] Environmental data is a crucial factor influencing the occurrence and spread of pests and diseases. The core focus is on collecting key environmental parameters related to pest and disease growth within the facility agriculture setting, ensuring continuous and real-time data collection. The specific implementation is as follows: The collected parameters are adjusted according to the type of greenhouse crop and the habits of pests and diseases. The core collected indicators include temperature, humidity, and light intensity. Concentration, soil moisture, and in some scenarios, supplementary indicators such as soil pH and ventilation volume can be included; the sampling frequency is set to once every 10 minutes, continuous and uninterrupted; the sampling accuracy meets the requirements for temperature and humidity. , Light intensity , concentration Soil moisture The environmental sensing module of the multi-source fusion monitoring terminal collects environmental parameters around the monitoring node in real time. The sensing module is embedded inside the monitoring terminal, and deployment is ensured to be unobstructed and free from interference, including temperature and humidity. The sensor module is located at the distance from the crop canopy. Soil moisture sensor module inserted into soil To ensure accuracy, the data collection process avoids soil clumping and moisture buildup. Each environmental data point includes a millisecond-level timestamp, the spatial grid coordinates of the corresponding monitoring node, and the monitoring node number. It also indicates the greenhouse ventilation status (on / off) and shading status (on / off) at the time of collection, facilitating subsequent analysis of the impact of environmental factors on pests and diseases. During data collection, if sudden changes in environmental parameters occur (such as a rapid rise or fall in temperature and humidity), the monitoring terminal automatically marks the abnormal data and activates redundant sensor modules for secondary data collection to verify data validity and avoid invalid data due to sensor malfunction.

[0038] Crop physiological data is used to reflect crop growth status and the degree of damage. The core data collection focuses on physiological indicators related to pest and disease infection, establishing a correlation between crop growth status and pest and disease risk. The specific implementation is as follows: The sampling indicators are adapted according to crop type. For tomatoes and cucumbers, the focus is on collecting leaf chlorophyll content, leaf water content, stem thickness, and growth rate; for strawberries, the focus is on collecting leaf thickness, fruit sugar content, and root activity. The sampling frequency is set to once per hour, synchronously linked to the crop growth cycle (sowing period, transplanting period, growing season, and fruiting period). The sampling focus can be adjusted appropriately for different growth cycles. Data is collected through the optical and physiological sensing modules of the multi-source fusion monitoring terminal. The optical sensing module is aimed at the functional leaves of the crop (avoiding aging and diseased leaves) to ensure a consistent sampling angle (perpendicular to the leaf) and avoid direct sunlight affecting sampling accuracy. The physiological sensing module uses non-contact sampling to avoid damaging the crop. Each physiological data point carries a millisecond-level timestamp, monitoring node spatial coordinates, crop type, and growth cycle stage. It also labels the collected leaves / parts (e.g., the third functional leaf of a tomato, or the leaf of a strawberry runner) to ensure that the data is accurately correlated with the crop's growth status. The collected physiological data is stored in conjunction with environmental data from the same period to form an "environment-physiology" time-series data pair, which provides support for subsequent analysis of the impact of environmental factors on the crop's physiological state and for judging early signs of pest and disease infection.

[0039] Pest and disease characteristic data are core data for pest and disease identification and risk assessment. The core data collection focuses on image and biological features of pests and diseases, with a particular emphasis on capturing the characteristics of small targets (insect eggs, early lesions). This data is adapted to the complex context of facility agriculture. The specific implementation is as follows: Data is acquired through a high-resolution macro image sensing module of the multi-source fusion monitoring terminal, with a fixed image size of 224×224 pixels to accommodate subsequent recognition algorithms. The acquisition frequency is once every two hours, simultaneously capturing images under different lighting conditions (8 AM, 12 PM, and 4 PM) to avoid feature loss due to uneven lighting. The acquired objects include crop leaves, stems, and fruits, focusing on capturing features such as lesions, insect eggs, larvae, and adults. Images must be clear, unblurred, and unobstructed, with lesion / egg features comprising at least 10% of the image. Data is also acquired through the biosensor module of the multi-source fusion monitoring terminal, primarily collecting the concentrations of insect pheromones and plant volatiles. Different pests and diseases correspond to different specific biological signals (e.g., aphids correspond to specific insect pheromones, and tomato late blight corresponds to specific plant volatiles). The acquisition frequency is once per hour, ensuring the biosensor module is free of contaminants to avoid affecting acquisition accuracy. Each pest and disease feature data entry includes a millisecond-level timestamp, monitoring node spatial coordinates, crop type, and sampling location. Image data is additionally annotated with shooting lighting conditions and shooting angle, while biological feature data is additionally annotated with the ambient temperature and humidity at the time of collection, facilitating subsequent recognition algorithms to adapt to different scenarios. For small targets such as insect eggs and early tiny lesions, the collection frequency is appropriately increased (once every hour), while typical sample images are manually collected to ensure the diversity and completeness of feature data, thereby improving subsequent recognition accuracy.

[0040] Historical correlation data is used to assist in risk assessment, optimize identification and fusion algorithms. The core data collection is historical data related to the current monitoring scenario, realizing the linkage between historical and real-time data. The specific implementation is as follows: The core collection includes four types of historical data, including historical pest and disease data (recent). Data on the occurrence time, severity, scope of infection, control measures and effectiveness of pests and diseases in the same period of the previous year), and crop growth cycle data (nearly). Annual crop growth cycle parameters, physiological index ranges at each stage), and meteorological data for the same period (near-year). The system integrates environmental data such as temperature, humidity, light, and precipitation from the same period last year, as well as plant protection and pest control case data (corresponding to optimal control plans, pesticide dosages, and control timing for pests and diseases). It also integrates cross-domain historical data through a cloud-based evaluation platform, prioritizing historical monitoring data from local facility agriculture parks and supplementing with publicly available historical data from local plant protection departments and agricultural technology extension centers to ensure data relevance and suitability. The collected historical data is calibrated, eliminating abnormal and invalid data, and standardizing data formats and units to ensure consistency between historical and real-time data. Historical data is also categorized and stored by crop type, pest and disease type, and time dimension for easy comparison with real-time data. Finally, historical data is precisely correlated with real-time collected environmental, crop physiological, and pest and disease characteristic data through spatial coordinates, crop type, and time dimension, providing historical reference for subsequent dynamic weight adjustments and full-cycle risk prediction.

[0041] After data acquisition is complete, the edge computing module performs preliminary calibration on various data types. Environmental data, crop physiological data, and biological characteristic data are calibrated using the mean value method to remove outliers and correct sensor errors. Image data undergoes brightness and contrast calibration to ensure consistent image features under different lighting conditions, facilitating subsequent recognition algorithm processing. Based on the spatial coordinates of monitoring nodes and millisecond-level timestamps, precise spatiotemporal alignment of the four types of heterogeneous data is achieved, ensuring a one-to-one correspondence between environmental data, crop physiological data, and pest and disease characteristic data from the same monitoring node at the same time. Historical correlation data is aligned with real-time data along the time dimension to avoid data misalignment and ensure the synergy of multi-source data. A data quality verification mechanism is established, with verification indicators including data completeness (no missing or omitted data), accuracy (error within a preset range), and synchronization (spatiotemporal alignment error). (seconds), and re-collect or supplement the data that does not meet the standard to ensure the high quality of the collected data.

[0042] Example 2: This example uses the tomato greenhouse (single greenhouse, area) from Example 1. Multi-source heterogeneous data collection was implemented; the preliminary preparations were as follows: a single standardized tomato greenhouse, with a row spacing of 1.2m and a plant spacing of 0.5m, and 10 multi-source fusion monitoring terminals were deployed (every 50 plants). (1 unit), the crop is in its growth period, with a focus on monitoring pests and diseases such as tomato late blight and aphids; environmental data is collected every 10 minutes, crop physiological data is collected every hour, and pest and disease characteristic data (images every 2 hours, biological characteristics every hour); timestamps are in milliseconds, and spatial coordinates correspond to greenhouse grid numbers (1-10), one-to-one with monitoring terminals; the sensor modules of each monitoring terminal are debugged to ensure the accuracy of temperature and humidity in the environmental sensor modules. , The image sensing module captures images at a size of 224×224 pixels, while the biosensing module can accurately capture aphid pheromones.

[0043] Environmental data collection includes temperature, humidity, and light intensity. Concentration and soil moisture were measured every 10 minutes. The temperature and humidity sensor module was positioned 12cm from the tomato canopy, and the soil moisture sensor module was inserted 12cm into the soil. Example data collected: 25℃, 60%RH, light intensity 800 lux. Concentration 800 ppm, soil moisture 60%; each data point is labeled with a timestamp, grid coordinates 1-10, and greenhouse ventilation status (on). Crop physiological data collection included tomato leaf chlorophyll content, leaf water content, stem thickness, and growth rate, collected hourly. The optical sensor module was aimed at the third functional leaf of the tomato plant (non-disease, non-aging leaf) for non-contact data collection; data was labeled with timestamps, grid coordinates, crop growth stage (growth stage), and leaf location. Pest and disease characteristic data collection included image data every 2 hours, photographing tomato leaves and stems, focusing on late blight lesions (green, turbid, water-soaked appearance) and aphids (small wingless insects). Images were simultaneously collected at 8:00 AM, 12:00 PM, and 4:00 PM to ensure clarity. Biological characteristic data collection included hourly collection of aphid-specific pheromones; data was labeled with timestamps, grid coordinates, shooting light conditions, and collection location. Historical data was collected by accessing the cloud-based evaluation platform to retrieve data on the occurrence of late blight and aphids in the greenhouse tomatoes over the past three years (occurrence time, severity, and control measures), physiological parameters of the tomato growth period, environmental data from the same period, and plant protection control cases (late blight application scheme). After calibration, the data was aligned with real-time data along the time dimension. After collection, the edge computing module performed mean calibration on the collected data, removed outliers, and uploaded the data synchronously to the cloud via WiFi 6. Data was stored locally for 7 days to ensure data integrity and synchronization.

[0044] In step S3, the edge computing module deploys a data preprocessing unit, a preliminary feature extraction algorithm, and a lightweight recognition sub-model (built based on MobileNetV3); the cloud evaluation platform deploys the core recognition algorithm (a multi-crop pest and disease recognition algorithm based on improved ResNet50 and Outlook attention) and a multi-source heterogeneous data adaptive fusion algorithm, and completes the initialization of all algorithm parameters.

[0045] The edge computing module performs local preprocessing, removing abnormal mutation values ​​(such as sudden increases or decreases in temperature and humidity exceeding reasonable ranges) from environmental data and crop physiological data, as well as blurry and overexposed / underexposed images from image data based on preset thresholds, retaining valid data. The cloud-based evaluation platform then performs secondary optimization, enhancing the image data by using geometric transformations, color transformations, and noise injection to expand its diversity and improve the robustness of subsequent recognition models. Specific enhancement methods include horizontal flipping and rotation (…). ), cropping, scaling ( Contrast adjustment Brightness adjustment Gaussian noise (intensity 0.01) is injected to expand the original image data by 3 times. All preprocessed effective data (environmental, crop physiology, biological characteristics, and historical correlation data) are mapped to the [0,1] interval to eliminate the influence of dimensions. The core formula is: In the formula, For standardized data values, The original data values, , These represent the minimum and maximum values ​​for the corresponding data types, ensuring that data from different dimensions can be fused and calculated. A verification mechanism is also established, with verification metrics including data integrity (no missing or omitted data), accuracy (errors within a preset range), and synchronization (spatiotemporal alignment error). (seconds), and collect or supplement data that does not meet the standards to ensure high quality of preprocessed data.

[0046] See Figure 3 As shown, a high-precision identification model for multiple crop diseases and pests in facility agriculture is constructed to overcome the limitations of existing algorithms in identifying diseases and pests in complex scenes and small targets. The specific implementation is as follows: Based on ResNet50, its residual skip connection structure is retained to alleviate the gradient vanishing and gradient exploding problems in deep network training. The core formula of the residual connection is: ; In the formula, For residual block output features, The output is the convolution operation of the residual block. Input features into the residual block to achieve efficient feature transfer.

[0047] See Figure 4As shown, in the four network stages of ResNet50, contextual convolutional blocks are fused (implemented using dilated convolution), expanding the receptive field of the convolutional kernel without increasing the actual size of the kernel. This integrates the contextual information of pests and diseases in the image, improving the recognition accuracy of small targets (insect eggs, early tiny lesions). The formula for the dilated convolution kernel is: In the formula, To expand the output features of the convolution, For convolution kernel weights, Expansion rate (preset) (phased adaptation) These are the sampling points of the input feature map.

[0048] The preprocessed image feature map is divided into 8×8 and 16×16 multi-scale local windows, with each window serving as an independent attention calculation unit. This approach balances local features with global contextual information and suppresses interference from the greenhouse background (soil, supports). A linear layer transforms the input feature map into query (Q), key (K), and value (V). Attention weights are calculated within the local window, with the core formula as follows: ; In the formula, for The dimension of the matrix (value 256). For normalization function, Local window attention weight coefficients (adapted to 8×8 and 16×16 multi-scale windows, with a value range of [value range missing]). ), Attention scores are assigned to each region of the image. Let K be the square root of the dimension of the bond matrix (K). For the global context key matrix, This represents the attention score at the global level. The global context attention weight coefficients are used to address the technical shortcomings of existing Outlook attention mechanisms, such as poor adaptability and insufficient background interference suppression. This ensures that the model can accurately extract pest and disease features even in scenarios with multiple crops, complex backgrounds, and small targets, thus providing support for subsequent high-precision identification.

[0049] The parameters of the pre-trained ResNet50 model (trained on the ImageNet dataset) were transferred to this recognition model. Fine-tuning was performed using a multi-crop pest and disease dataset for facility agriculture (covering over 10 pests and diseases and healthy samples from crops such as tomatoes, cucumbers, and strawberries). Training parameters were set as previously prepared (batch size=96, epoch=60, etc.). The pest and disease recognition results (recognition type, severity of damage) are output through a Softmax classifier, and the recognition accuracy (Acc) is calculated. The core formula is: ; in, A true positive (a correctly identified pest or disease sample). It is a true negative (a correctly identified healthy sample). This is a false positive (a healthy sample was mistakenly identified as a pest or disease). In the case of false negatives (where diseased or pest samples are mistakenly identified as healthy), the accuracy rate of this invention is over 98.5%.

[0050] When performing adaptive fusion of multi-source heterogeneous data, the accuracy (Acc) of pest and disease identification is deeply integrated into the fusion process. The weight ratio of pest and disease feature data is adjusted to generate a pest and disease feature fusion vector (F). The fusion weights are dynamically allocated based on the correlation between each data source (environment E, crop physiology P, pest and disease features D, and historical association H) and pest and disease risk, as well as the identification accuracy (Acc). , , , And satisfy The specific implementation and related formulas are as follows: ; ; ; ; In the formula, The correlation coefficient is... The fusion weights for environmental data in multi-source heterogeneous data. The fusion weights for crop physiological data in multi-source heterogeneous data. The fusion weights for pest and disease feature data in multi-source heterogeneous data. The fusion weights for historically related data in multi-source heterogeneous data. The fusion result after removing the corresponding environmental data, The fusion result is after removing the corresponding crop physiological data. This is the fusion result after removing data related to pests and diseases. To improve the accuracy of pest and disease identification. The fusion result after removing corresponding historical related data, To remove the first After the first type of data source, the remaining three types of data sources are merged to obtain the fusion result; , , Multiplying by a factor of 0.95 is to account for the impact of non-core data sources.

[0051] Example 3: This example also uses the tomato greenhouse (single greenhouse, area) from Example 1. Taking ( ) as an example, the data processing implementation is carried out as follows: During edge local preprocessing, abrupt changes in temperature and humidity (such as a sudden drop from 38℃ to 10℃) and blurred images were removed, while valid data were retained. Adaptive median filtering was used to reduce noise in environmental data (such as temperature 25℃ and humidity 60%RH) and crop physiological data (chlorophyll 50SPAD). Gaussian filtering (σ=0.5) was used to reduce noise in late blight images while preserving lesion features. Linear interpolation was used to fill in missing soil moisture values ​​at one time point (originally missing, with data before and after being 58% and 62% respectively, and after filling, the value was 60%).

[0052] The images undergo further optimization in the cloud, including horizontal flipping, 20° rotation, 15% contrast enhancement, and expansion to three times the sample size. For example, temperature data is then processed. The original temperature was 25℃, and after standardization... The remaining data are standardized using the same formula; only when the data integrity reaches 99.5% and the spatiotemporal alignment error is less than 0.8 seconds does it meet the verification standard.

[0053] See Figure 5 , Figure 6 As shown, by transferring the pre-trained ResNet50 parameters and fine-tuning them using a cucumber disease and pest dataset (28,642 images), the loss value decreased to below 0.025 after 30 training epochs; cucumber downy mildew was identified, with moderate damage, and the accuracy was calculated using the formula. Dynamic weight calculation yields... , , , The weights sum to 1.

[0054] Fusion vector generation: Substitute into the fusion formula, and standardize the data. The calculation yields: A 1×28 dimensional fusion vector is generated with a fusion accuracy of 95.8%; after successful verification, it can be used for subsequent risk assessment.

[0055] Optionally, various network models, such as traditional deep convolutional VGG16, GoogleNet, ResNet34, and ResNet50, were applied to the tomato leaf disease dataset and trained under the same experimental configuration. During training, the same training parameters were set for all different models. This comparative experiment verified the effectiveness of using ResNet50 as the backbone network for identifying pests and diseases in greenhouse agriculture. The experimental results are shown in Table 1. Table 1: Accuracy Comparison Among Different Models As shown in Table 1 of the comparative experimental results, although the recognition accuracy of these four convolutional network models on the pest and disease dataset is all above 90%, there are still differences among the different models in the task of identifying pests and diseases in facility agriculture. Among them, the VGG16 network model has the lowest accuracy, at only 94.53%. The GoogleNet network model has an accuracy of 95.08%, the ResNet34 network model has an accuracy of 95.31%, while the ResNet50 network model ranks first with an accuracy of 95.43%.

[0056] Based on a comprehensive comparison of experimental results, this approach selects the ResNet50 network model with the highest accuracy for improvement. By adding contextual convolution and Outlook attention mechanisms to the model based on transfer learning strategies, the model's feature extraction capability for pest and disease images is significantly enhanced, enabling efficient identification of pests and diseases in facility agriculture.

[0057] In step S4, based on the pest and disease identification results, the pest and disease type, its life cycle stage (egg stage / larval stage / adult stage), and identification accuracy (Acc) are determined, and the corresponding pest and disease type is selected. and Substitute the standardized temperature, humidity, and light data into the formula for the environmental adaptation function to calculate the function value. ; Set the preset initial weight vector Substituting into the dynamic weight adjustment formula, the results are calculated dimension by dimension. Ensure the weight sum is 1 (if there is a slight deviation, perform normalization correction); Verify The range of weight values ​​for each dimension ( The weights are summed to 1, and those that do not meet the standard are recalculated to ensure the effectiveness of the weights.

[0058] It can be understood that the dynamic weight adjustment formula is: In the formula, Let t be the dynamic evaluation weight vector (dimension 1×4), which corresponds to the evaluation weights of environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data, and satisfy the weight sum of 1; The initial weight vector (1×4 dimension) is set according to the crop-disease type, for example: tomato-late blight. Cucumber downy mildew ; The correction coefficient for the life cycle of pests and diseases is selected based on the growth stage of the pests and diseases (0.8 for egg stage, 1.0 for larval stage, and 1.2 for adult stage). The environment adaptation function is calculated based on standardized environmental data (temperature, humidity, light intensity, etc.). The specific calculation formula is as follows: ; In the formula, The average temperature and humidity values ​​are the standardized values. The standardized light intensity is represented by 0.5, which serves as a suitable environmental baseline value. The function value range is... The more suitable the environment is for the growth of pests and diseases, the closer the function value is to 1.2; The accuracy correction function is selected based on the accuracy (Acc) of pest and disease identification (refer to the grading standard in the previous preparation). The higher the identification accuracy, the higher the correction coefficient, which strengthens the weight of high-precision identification results.

[0059] The 28 feature values ​​of the disease and pest feature fusion vector (F) are respectively compared with the dynamic weight vector (F). Multiply the corresponding weight values ​​of each crop and sum them to obtain a weighted sum; combine historical correlation data (historical probability of pest and disease occurrence) and current crop physiological data (crop growth) to determine the risk correction coefficient (S); determine the damage severity correction coefficient (K) based on the damage severity in the pest and disease identification results; substitute the weighted sum, S, and K into the current situation risk assessment formula to calculate the risk correction coefficient (K). ; It can be represented as: ; In the formula, This represents the risk index of the current status of pests and diseases, with a range of values. The higher the value, the higher the risk; is the i-th feature value of the fusion vector (F) of pest and disease features (n=28, that is, the dimension of the fusion vector is 1×28). For dynamic weight vectors ( The i-th weight value of ) and One-to-one correspondence; The sum of the fusion vector and dynamic weights reflects the basic risk level after multi-source data fusion; S is the risk correction coefficient, set based on historical pest and disease occurrence probabilities and current crop growth, with a value range of... The higher the historical probability of occurrence and the weaker the crop growth, the larger the S value; K is the damage severity correction coefficient, which is set according to the damage severity classification in the pest and disease identification results: mild damage K=1.0, moderate damage K=1.2, and severe damage K=1.5. The more severe the damage, the higher the correction coefficient. Based on the preset risk level threshold, the current risk level (low risk / medium risk / high risk) is determined, providing a basis for subsequent prevention and control coordination.

[0060] Integrating the future through a cloud-based assessment platform Based on 10 days of meteorological forecast data (adapted to environmental data), crop growth trend forecast data (adapted to crop physiological data), and the patterns of pest and disease transmission, predict the next k days. Based on the optimization trend of the recognition model, predict... ; Clearly define the number of days to predict k ( Select the corresponding crop-pest prediction coefficients a, b, c, and d; Substituting the predictor coefficient into the exponent term, the result is calculated. The current risk index Substituting the index term into the full-cycle risk evolution prediction model, the results for the next k days are calculated. ;analyze Risk index of the current situation The difference is used to predict the trend of risk (a positive difference indicates rising risk, and a negative difference indicates falling risk). If the risk exceeds the high-risk threshold, issue an early warning.

[0061] It can be understood that the formula for the full-cycle risk evolution prediction model is: ; In the formula, For the future sky( The predicted risk index; This represents the risk index of the current situation at this moment. To predict the number of days, take the following values. It can be adjusted according to actual prevention and control needs; This represents the change in environmental data over the next k days (after standardization), which is the difference between the mean of environmental data over the next k days and the mean of current environmental data. A positive value indicates that the environment is more suitable for the growth of pests and diseases, while a negative value indicates that the environment is not suitable. This represents the change in crop physiological data over the next k days (after standardization). A negative difference indicates that the crop is growing poorly and is more susceptible to pests and diseases. This represents the change in pest and disease characteristics over the next k days (after standardization). A positive difference indicates that the pest and disease characteristics are more pronounced and the risk is higher. This represents the change in the accuracy of pest and disease identification over the next k days. A positive difference indicates improved identification accuracy and more precise risk assessment. The predictive coefficient is set based on crop-pest / disease type (general reference value is...). ), which correspond to the influence weight of the changes in each data source; e is the natural constant (approximately 2.718), used to reflect the exponential evolution characteristics of risk (the spread of pests and diseases has an exponential growth trend).

[0062] In practice, based on preset thresholds, the current risk level and the predicted risk level for the next k days are determined, and the risk warning level is clarified (low risk: routine monitoring; medium risk: targeted prevention and control; high risk: emergency prevention and control). The current risk index is compared with the actual occurrence of pests and diseases in the field, and the predicted risk index is compared with the subsequent actual risk situation to verify the accuracy of the assessment (the prediction accuracy rate is required to be ≥92%). If the standard is not met, the model parameters (prediction coefficient, correction coefficient, etc.) are adjusted to ensure the reliability of the assessment.

[0063] Example 4: This example also uses the tomato greenhouse (single greenhouse, area) from Example 1. Taking a single tomato greenhouse as an example, a risk assessment was conducted. Ten multi-source fusion monitoring terminals were deployed, focusing on monitoring tomato late blight. A fusion vector F (dimension 1×28) of pest and disease characteristics was obtained, with a sum of eigenvalues ​​of 4.8. Tomato late blight (larval stage) was identified with an accuracy of Acc=98.8%, indicating mild damage. Standardized data: (Average temperature and humidity) (Light intensity); Preset parameters: , (Larva stage) , (The historical probability of occurrence is moderate, and the crop is growing well). (Mild victimization), predictive coefficient (Predicted for 7 days).

[0064] After determining the basic parameters as follows ; Computational environment adaptation function ; Calculate the dynamic weight vector ; Dimensional calculation: , , , get The weights sum = 1.036 (1.0 after normalization); since the weights of each dimension are all within... Between these values, the weights sum to 1.0, indicating successful verification.

[0065] In practice, the current risk index is applied, and the sum of the eigenvalues ​​of the fusion vector F is 4.8, with a dynamic weight vector. The summation is 1.0, so the weighted sum is 4.8 × 1.0 = 4.8; the correction coefficients are S = 1.0 and K = 1.0; thus, the current situation risk index is obtained. Due to the current risk index at The area is classified as medium risk and requires targeted prevention and control measures.

[0066] Therefore, in the next 7 days, (The environment is more suitable for the spread of late-stage diseases) (Crop growth has declined slightly) (The characteristics of diseases and pests are more obvious) (Recognition accuracy slightly improved); Determine prediction parameters: The exponent is calculated, specifically as follows: Therefore, the risk index for the next 7 days is predicted to be: Since the risk index for the next 7 days is 10.98 (high risk threshold), which is significantly higher than the current risk index of 4.8, it is predicted that the risk will rise to high risk in the next 7 days. Therefore, it is necessary to issue an early warning and start prevention and control preparations.

[0067] In step S5, based on the risk assessment and identification results output by the user interaction terminal, and combined with the risk level (low, medium, high), differentiated output content is designed to ensure that the content accurately matches the prevention and control needs, as detailed below: When it is low risk ( When outputting basic core information, no complex prevention and control operations are required. The focus is on daily monitoring prompts, including: pest and disease identification results (type, identification accuracy), current risk index and level, and daily prevention and control prompts (such as regular ventilation, cleaning diseased leaves, and routine monitoring frequency). There are no early warning information and no prevention and control linkage is triggered.

[0068] When it is considered medium risk ( When [the system] is in operation, it outputs detailed information and targeted prevention and control suggestions to guide users in taking targeted prevention and control measures. Specifically, this includes: basic information (type of pest or disease, identification accuracy, degree of damage, current risk index and level), risk cause analysis (combining environmental data and crop physiological data to explain the key factors contributing to the risk, such as suitable temperature and humidity for pest and disease spread, weak crop growth, etc.), targeted prevention and control suggestions (application plans, recommendations for the use of control equipment, and control timelines), and future [measures]. Daily risk change prediction (brief trend, such as stable risk / slowly rising risk), and manual prevention and control reminders (reminder messages pushed through user interaction terminals).

[0069] When the risk level is high (R≥6), comprehensive information and emergency prevention and control plans are output, triggering coordinated prevention and control. This includes: core early warning information (real-time early warning pop-up, voice reminder, clearly stating the high-risk level and urgency), detailed basic information (type of pest / disease, identification accuracy, degree of damage, current risk index, and future risk level). Risk prediction curve and key nodes), in-depth analysis of risk causes (combining multi-source data to identify the core factors and development trends leading to high risk), emergency prevention and control plan (precise application plan, operating parameters of prevention and control equipment, and suggestions for the configuration of prevention and control personnel), real-time tracking portal for prevention and control progress (supporting users to view the operating status of prevention and control equipment and application progress), and emergency response plan (such as supplementary prevention and control measures when pests and diseases spread rapidly).

[0070] In practice, the output channels simultaneously cover mobile apps and web pages, with customized display interfaces for different user types to ensure efficient information delivery. Specifically: The mobile app targets growers and frontline plant protection personnel, featuring a simple and intuitive interface that highlights core information. It supports push notifications (early warnings, control reminders), manual operations (adjusting control parameters, pausing / restarting control processes, uploading field data), and visual charts (risk change curves, comparison charts of identification results). It is adapted to mobile user habits and supports offline viewing of cached data. The web page targets managers and professional plant protection personnel, requiring a comprehensive and detailed interface that displays the overall risk situation (risk comparisons across multiple greenhouses and crops), detailed data reports (identification accuracy, risk index, control effectiveness data), control execution statistics (control equipment runtime, pesticide dosage), model optimization records, and supports custom report generation, historical data backtracking, and multi-user permission management, adapting to the data analysis needs of desktop computers.

[0071] See Figure 2 As shown, the present invention is a risk assessment device for pests and diseases in facility agriculture, which can be used to perform the embodiments of the present invention. The method described in Example 4 includes: a multi-source fusion monitoring terminal, an edge computing module, a cloud-based assessment platform, a user interaction terminal, and a joint prevention and control interface; The multi-source fusion monitoring terminal is deployed in a grid pattern according to facility agriculture plots. Each monitoring node integrates an optical sensing module, a biosensing module, a high-resolution macro image sensing module, an environmental sensing module, and a miniature positioning module. It is self-powered by solar energy and has a small size. It can be attached to crop leaves / branches to simultaneously collect environmental data, crop physiological data, and pest and disease characteristic data, and send the data to the edge computing module. The size of the collected image data is uniformly 224×224 pixels to adapt to subsequent recognition algorithms. It realizes multi-dimensional data synchronous collection, focuses on optimizing the shooting parameters of the image sensing module, and adapts to the scene of uneven lighting and complex background in greenhouses, solving the problems of insufficient coverage, partial data and poor image quality of existing single monitoring devices. The edge computing module is deployed at the facility agriculture site, with one module configured in each greenhouse. It has a built-in data preprocessing unit, preliminary feature extraction algorithm and lightweight recognition sub-model, which is used to locally filter, reduce noise, preprocess and identify pests and diseases in the collected raw data, reduce invalid data transmission and provide support for high-precision recognition in the cloud. The cloud-based assessment platform is deployed on a cloud server to integrate cross-domain historical data, run core identification algorithms, multi-source heterogeneous data adaptive fusion algorithms, dynamic weight risk assessment algorithms, and full-cycle risk evolution prediction models to achieve high-precision identification, risk calculation, and evolution prediction of pests and diseases; the core identification algorithm is a multi-crop pest and disease identification algorithm based on improved ResNet50 and Outlook attention. The user interaction terminal includes a mobile APP and a web page, which are used to receive risk assessment results, pest and disease identification results, early warning information and control suggestions, and support users to manually upload crop images and supplement field data; The linkage and control interface is used to connect with intelligent insecticidal lamps, spraying robots, and natural enemy release devices. When the risk exceeds the standard, it triggers automatic linkage and control, forming a closed loop of "identification-assessment-control".

[0072] It is worth noting that the various units included in the above system embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0073] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium.

[0074] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for assessing the risk of pests and diseases in facility agriculture, characterized in that, Includes the following steps: Step S1, Device Deployment: Deploy multi-source fusion monitoring terminals, edge computing modules, cloud assessment platforms, user interaction terminals, and linkage prevention and control interfaces; complete communication debugging of each module; and build a facility agriculture pest and disease assessment network. Step S2, Data Acquisition: Simultaneously collect environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data in the facility agriculture scenario through a multi-source fusion monitoring terminal; Step S3, Data Processing: The collected data is preprocessed, pests and diseases are identified, and multi-source heterogeneous data are adaptively fused through the edge computing module and cloud evaluation platform to generate pest and disease feature fusion vectors. Step S4, Risk Assessment: Based on the fusion vector of pest and disease characteristics and the results of pest and disease identification, the current risk index is calculated through dynamic weight adjustment, and a pest and disease risk assessment model is constructed by combining the full cycle risk evolution. Step S5, Risk Output and Prevention Linkage: Output risk assessment results and identification results through user interaction terminal. When the risk exceeds the standard, trigger linkage prevention and control. After prevention and control, iteratively optimize the pest and disease risk assessment model through data feedback.

2. The method for assessing the risk of pests and diseases in facility agriculture according to claim 1, characterized in that, In step S2, the specific process for data acquisition is as follows: The environmental data is collected continuously at a frequency of once every 10 minutes. The crop physiological data is recorded synchronously at a frequency of once per hour and linked to the crop growth cycle file. The image data in the pest and disease characteristic data is collected at a fixed size of 224×224 pixels, and multiple sets of images under different lighting conditions are collected simultaneously. The biological characteristic data is collected once per hour. The historical data and real-time data are aligned in time and space using spatial coordinates and timestamps. All data carries a unique monitoring node identifier, plot grid coordinates, and millisecond-level timestamps.

3. The method for assessing the risk of pests and diseases in facility agriculture according to claim 1, characterized in that, In step S3, the preprocessing of the collected data includes the following steps: The edge computing module prioritizes local preprocessing of the collected raw data, uses an adaptive median filtering algorithm to reduce noise in environmental data and crop physiological data, uses Gaussian filtering combined with an edge protection algorithm to reduce noise in image data while preserving pest and disease characteristics, and uses linear interpolation to fill in intermittent missing data. The cloud-based evaluation platform performs secondary optimization on the data preprocessed by the edge computing module, expands the sample size of the image data through multi-dimensional enhancement operations, performs secondary standardization calibration on the environmental, crop physiological and biological characteristic data, and establishes a preprocessed data quality verification mechanism to remove abnormal data that does not meet the standards.

4. The method for assessing the risk of pests and diseases in facility agriculture according to claim 1, characterized in that, In step S3, the specific process for identifying pests and diseases is as follows: Step S31, Basic Network Optimization: Based on ResNet50, retain its residual skip connection structure, and fuse contextual convolutional blocks in the four network stages of ResNet50; the contextual convolutional blocks are implemented using dilated convolution, with dilation rates set to 1, 2, and 3 respectively, and integrate the contextual information of pests and diseases in the image. Step S32, Attention Mechanism Adaptation: Introduce an improved Outlook attention mechanism and optimize the multi-scale local window partitioning strategy. Divide the preprocessed pest and disease image into 8×8 and 16×16 multi-scale local windows. Each window serves as an independent attention calculation unit. The input feature map is converted into a query, key, and value through a linear layer. Attention weights are calculated within the local window and global context information is fused. Step S33, Model parameter fine-tuning: Using a transfer learning strategy, the parameters of the ResNet50 model pre-trained on the ImageNet dataset are transferred to the pest and disease identification model; Step S34: Output the identification results: Output the pest and disease type, identification accuracy, and severity level through the Softmax classifier.

5. The method for assessing the risk of pests and diseases in facility agriculture according to claim 4, characterized in that, In step S32, during the attention mechanism adaptation, the preprocessed 224×224 pixel pest image is simultaneously divided into local windows of two scales: 8×8 and 16×16. Each local window serves as an independent attention calculation unit, and the input pest image feature map is converted into a query matrix, a key matrix, and a value matrix through a linear transformation layer. Within each local window, the local attention weight is obtained through the attention weight calculation formula. At the same time, the global context features of the image are extracted through global average pooling to generate a global context key matrix, and the global attention weight is calculated. Then, the local attention weights and global attention weights are fused together using preset weight coefficients.

6. The method for assessing the risk of pests and diseases in facility agriculture according to claim 1, characterized in that, In step S3, during the adaptive fusion of multi-source heterogeneous data, the preprocessed environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data are normalized to obtain standardized data. Based on the correlation between each data source and the pest and disease risk and the identification accuracy, the fusion weights of environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data are dynamically allocated, and the sum of the fusion weights of environmental data, crop physiological data, pest and disease characteristic data, and historical correlation data is 1.

7. The method for assessing the risk of pests and diseases in facility agriculture according to claim 1, characterized in that, In step S4, the specific process of risk assessment is as follows: Step S41: Combine the pest and disease feature fusion vector and pest and disease identification results generated in step S3, and use a dynamic weight adjustment strategy based on the linkage between the pest and disease life cycle and identification accuracy to calculate the dynamic evaluation weight vector at the current moment. Step S42: Based on the dynamic evaluation weight vector and the fusion vector of pest and disease characteristics, and combined with the risk correction coefficient and the damage degree correction coefficient, calculate the current pest and disease risk index; wherein, the risk correction coefficient is set in combination with the historical probability of pest and disease occurrence, and the damage degree correction coefficient is set according to the classification of the damage degree of pests and diseases to crops. Step S43: Integrating the Future The changes in environmental data, crop physiological data, pest and disease characteristic data, and identification accuracy over the days are used to predict future pest and disease risk trends using a full-cycle risk evolution prediction model. The full-cycle risk evolution prediction model assigns corresponding adaptation coefficients based on the degree of influence of changes in each data source to dynamically predict risk rise and fall trends.

8. The method for assessing the risk of pests and diseases in facility agriculture according to claim 7, characterized in that, In step S41, the adjustment of the dynamic evaluation weight vector is based on the preset initial weight of the corresponding crop-pest, the life cycle correction coefficient of the pest, the environmental adaptation coefficient, and the identification accuracy correction coefficient. Among them, the life cycle correction coefficient of the pest is set according to the growth stage of the pest; the environmental adaptation coefficient is dynamically calculated based on the standardized environmental data; and the identification accuracy correction coefficient is set according to the classification of the identification accuracy of the pest.

9. The method for risk assessment of pests and diseases in facility agriculture according to claim 1, characterized in that, In step S5, the specific process for linking risk output with prevention and control is as follows: Step S51: Conduct tiered and precise risk output. Based on the risk assessment results, generate differentiated output content according to risk level, pest and disease type, and degree of damage. Among them, low-risk output includes identification results and daily prevention and control tips; medium-risk output includes risk causes, identification accuracy, and targeted prevention and control suggestions; and high-risk output includes real-time early warning information and emergency prevention and control plans. The output channels simultaneously cover user interaction terminals. Step S52: Construct a hierarchical linkage prevention and control mechanism, and preset linkage trigger thresholds for different risk levels; among them, low risk triggers manual prevention and control reminders, medium risk triggers semi-automatic prevention and control linkage, and high risk automatically triggers linkage prevention and control interfaces, which are connected to intelligent insecticidal lamps, spray robots, and natural enemy release devices. Step S53: Establish a dynamic feedback optimization closed loop. After the prevention and control measures are implemented, the multi-source fusion monitoring terminal continuously collects crop growth data, pest and disease decline data, and prevention and control effect data, and feeds them back to the cloud evaluation platform in real time. Based on the feedback data, the platform iteratively optimizes the pest and disease identification model, the multi-source data fusion algorithm, the dynamic weight adjustment strategy, and the full-cycle risk evolution prediction model.

10. A facility agriculture pest and disease risk assessment device, used to implement the facility agriculture pest and disease risk assessment method as described in any one of claims 1-9, characterized in that: This includes a multi-source fusion monitoring terminal, an edge computing module, a cloud-based assessment platform, a user interaction terminal, and a joint prevention and control interface; The multi-source fusion monitoring terminal is deployed in a grid pattern according to the facility agriculture plots. Each monitoring node integrates an optical sensing module, a biological sensing module, a high-resolution macro image sensing module, an environmental sensing module, and a micro positioning module to synchronously collect environmental data, crop physiological data, and pest and disease characteristic data, and send the data to the edge computing module. The edge computing module is deployed at the facility agriculture site and has a built-in data preprocessing unit, preliminary feature extraction algorithm and lightweight recognition sub-model, which are used to perform local screening, noise reduction, preprocessing and pest identification on the collected raw data. The cloud-based assessment platform is deployed on a cloud server and is used to integrate cross-domain historical data and run pest and disease identification algorithms, multi-source heterogeneous data adaptive fusion algorithms, dynamic weight risk assessment algorithms, and full-cycle risk evolution prediction models. The user interaction terminal includes a mobile APP and a web page, which are used to receive risk assessment results, pest and disease identification results, early warning information and control suggestions, and support users to manually upload crop images and supplement field data. The linkage and control interface is used to connect with intelligent insecticidal lamps, spraying robots, and natural enemy release devices to trigger automatic prevention and control linkage when the risk exceeds the standard.

Images