A pest and disease monitoring method and system based on the Internet of Things

Abstract

The application discloses a kind of based on Internet of Things pest monitoring method and system, belong to agricultural pest monitoring and wisdom agricultural technology field.The application will target monitoring area be divided into several monitoring grid, complete monitoring equipment deployment, space binding and attribute calibration, the pre-processing and feature extraction of monitoring data are completed by edge end, and standardization feature data are transmitted to cloud end;Cloud end constructs node resource dynamic scheduling model with monitoring accuracy maximization and resource consumption minimization as double optimization target, generates and issues operating strategy parameter, simultaneously constructs spatial correlation matrix and pest spatial diffusion prediction model, completes risk distribution output, hierarchical early warning trigger and strategy update.The application realizes edge cloud collaborative management and control, reduces system resource occupation and early warning delay, realizes monitoring resource reasonable allocation, improves the accuracy of pest monitoring and prevention and control preposition.

Description

Technical Field

[0001] This invention relates to the field of agricultural pest and disease monitoring and smart agriculture technology, and in particular to a pest and disease monitoring method and system based on the Internet of Things. Background Technology

[0002] Agricultural pest and disease monitoring is a core component of modern agricultural production, serving as a crucial support for ensuring crop yield and quality, improving economic benefits, and promoting the sustainable development of green agriculture. With the rapid improvement of smart agriculture technology systems, digital technologies such as the Internet of Things (IoT), big data, artificial intelligence (AI), and edge computing are gradually integrating deeply into agricultural production scenarios. IoT-based pest and disease monitoring technology has become the mainstream solution replacing traditional manual field inspections. Currently, various IoT sensing terminals are widely deployed in large-scale planting bases and contiguous field planting areas in China, enabling automated and routine collection of field environmental data and crop growth image data. Combined with machine learning and image recognition technologies, automated extraction and identification of crop pest and disease-related characteristics are now possible. Simultaneously, the large-scale application of cloud-based data management and analysis platforms has enabled centralized aggregation, unified management, and centralized analysis and processing of data collected from multiple terminals. Edge computing technology is also gradually being implemented in agricultural IoT scenarios, providing mature technical support for local processing of terminal data. Related monitoring methods and systems are continuously iterating and optimizing, becoming an important technological foundation for precise pest and disease control in smart agricultural production systems.

[0003] Existing IoT-based pest and disease monitoring technologies still face several limitations in large-scale practical applications. Most current solutions employ a centralized cloud-based architecture for full data processing, with edge terminals handling only basic data acquisition and transmission. All raw monitoring data must be uploaded to the cloud for unified processing, resulting in significant bandwidth consumption and a heavy storage and computing load on the cloud. Furthermore, the inherent workflow of end-to-end data transmission and centralized cloud processing introduces delays in early warning responses to pest and disease anomalies. Additionally, existing monitoring solutions often employ fixed terminal sampling frequencies, data transmission cycles, and computing power allocation strategies, failing to differentiate their deployment and management based on crop planting characteristics, environmental features, and historical pest and disease occurrence patterns within the monitoring area. This makes it difficult to achieve a reasonable balance between monitoring accuracy and system resource consumption. In addition, existing solutions mostly adopt a single-node independent monitoring mode, without conducting spatial correlation analysis on the monitoring data collected from multiple nodes. This makes it impossible to effectively predict the spread trend of pests and diseases in different areas, and it is also difficult to dynamically optimize and adjust the terminal operation strategy based on real-time monitoring results. Consequently, a complete closed-loop management process of monitoring, scheduling, early warning, and optimization cannot be formed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a pest and disease monitoring method and system based on the Internet of Things.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for monitoring pests and diseases based on the Internet of Things is provided, which includes the following steps:

[0007] S1. Divide the target monitoring area into monitoring grids according to crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics. Generate spatial identity identifiers for each monitoring grid, label the basic attribute tags of each monitoring grid, deploy IoT sensing nodes and edge computing units in each monitoring grid, complete the spatial binding of IoT sensing nodes and edge computing units with the corresponding monitoring grids, and label the operating parameter tags of IoT sensing nodes and edge computing units.

[0008] S2. Collect the raw monitoring data of the corresponding monitoring grid. The raw monitoring data includes environmental time series data and crop image data. Preprocess the raw monitoring data and extract features through a lightweight feature extraction model to obtain standardized feature data. Upload the standardized feature data and node operation status data to the cloud.

[0009] S3. Receive standardized feature data and node operation status data, retrieve the basic attribute labels of the corresponding monitoring grid and the real-time monitoring data of adjacent monitoring grids, construct a node resource dynamic scheduling model with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption, and output and distribute operation strategy parameters.

[0010] S4. Construct a spatial correlation matrix based on standardized feature data from all monitoring grids, build a spatial spread prediction model for pests and diseases, output the risk distribution of pest and disease spread in the monitoring area through the spatial spread prediction model for pests and diseases, and complete the hierarchical early warning triggering and operation strategy parameter update.

[0011] Furthermore, step S1 includes:

[0012] S1.1. Divide the target monitoring area into monitoring grids according to crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics, and generate a unique spatial identity for each monitoring grid;

[0013] S1.2. Label the basic attribute tags for each monitoring grid. The basic attribute tags include crop variety attributes, planting environment attributes, pest and disease risk level attributes, and spatial association attributes between adjacent grids.

[0014] S1.3. Deploy IoT sensing nodes and edge computing units within each monitoring grid to complete the spatial binding of IoT sensing nodes and edge computing units with the corresponding monitoring grid;

[0015] S1.4. Label the operating parameter tags of each IoT sensing node and edge computing unit. The operating parameter tags include computing power threshold, remaining battery life threshold, maximum sampling capacity, and data transmission limit.

[0016] Furthermore, step S2 includes:

[0017] S2.1. Collect raw monitoring data for the corresponding monitoring grid according to the current operating strategy. The raw monitoring data includes environmental time series data and crop image data.

[0018] S2.2. Perform outlier removal, data denoising, and time series normalization on environmental time series data; perform invalid region cropping, image enhancement, and format compression on crop image data.

[0019] S2.3. Construct a lightweight feature extraction model, which includes an input layer, a convolutional feature extraction layer, a fully connected layer, and an output layer. The input layer takes preprocessed environmental time-series data and preprocessed crop image data as input. The convolutional feature extraction layer is fully connected to the input layer to extract multi-dimensional features from the input. The fully connected layer is fully connected to the convolutional feature extraction layer to normalize the extracted features. The output layer is fully connected to the fully connected layer to output standardized feature data.

[0020] S2.4. The standardized feature data and node operation status data are encrypted and uploaded to the cloud, while the original monitoring data is stored locally in the edge computing unit.

[0021] Furthermore, step S3 includes:

[0022] S3.1. Receive standardized feature data and node operation status data, retrieve the basic attribute labels of the corresponding monitoring grid, historical pest and disease occurrence data and real-time monitoring data of adjacent monitoring grids, and construct the input dataset for the node resource dynamic scheduling model;

[0023] S3.2. Construct a dynamic scheduling model for node resources. The dynamic scheduling model for node resources includes an input layer, a feature fusion layer, a multi-objective optimization layer, and an output layer. The input of the input layer is the input dataset of the dynamic scheduling model for node resources. The feature fusion layer is fully connected to the input layer and performs feature normalization and correlation fusion processing on the input dataset. The multi-objective optimization layer is fully connected to the feature fusion layer and performs calculations with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption. The output layer is fully connected to the multi-objective optimization layer and outputs the running strategy parameters, which include sampling frequency, data transmission cycle, computing power allocation weight, and early warning threshold adjustment parameters.

[0024] S3.3. Differentiate scheduling of IoT sensing nodes and edge computing units in different monitoring grids according to the operation strategy parameters;

[0025] S3.4. Perform routine updates of the operation strategy parameters according to the set cycle. When the pest and disease risk level of the monitoring grid changes or the node operation status becomes abnormal, trigger the real-time update of the operation strategy parameters and send the updated operation strategy parameters to the corresponding edge computing unit and IoT sensing node.

[0026] Furthermore, step S4 includes:

[0027] S4.1. Construct a spatial correlation matrix based on the spatial location relationship and attribute association of all monitoring grids, and map all standardized feature data to the corresponding monitoring grids to form a spatiotemporal distribution dataset of pest and disease risk;

[0028] S4.2. Construct a spatial spread prediction model for pests and diseases. The model includes an input layer, a spatiotemporal feature extraction layer, a spread pattern fitting layer, and an output layer. The input layer takes a dataset of the spatiotemporal distribution of pest and disease risk as input. The spatiotemporal feature extraction layer is fully connected to the input layer to extract spatial correlation features and temporal change features from the input dataset. The spread pattern fitting layer is fully connected to the spatiotemporal feature extraction layer to calculate the spread pattern of pests and diseases based on the extracted features. The output layer is fully connected to the spread pattern fitting layer to output the risk distribution of pest and disease spread in the monitored area.

[0029] S4.3. Identify monitoring blind spots based on the distribution of pest and disease spread risk, input the identification results of monitoring blind spots into the node resource dynamic scheduling model, and complete the optimization of operation strategy parameters;

[0030] S4.4. Based on standardized feature data and the distribution of pest and disease spread risks, trigger local emergency early warnings at the edge and global hierarchical early warnings in the cloud, and generate corresponding prevention and control guidelines.

[0031] Furthermore, in step S2.3, the parameters of the lightweight feature extraction model are iteratively adjusted using updated data distributed from the cloud.

[0032] Furthermore, in step S3.3, when performing differentiated scheduling of IoT sensing nodes and edge computing units in different monitoring grids, for monitoring grids with pest and disease risk levels higher than the set value, the sampling frequency and computing power allocation weight are increased to shorten the data transmission cycle; for monitoring grids with pest and disease risk levels lower than the set value, the sampling frequency and computing power allocation weight are reduced to extend the data transmission cycle.

[0033] Furthermore, in step S4.1, when constructing the spatial correlation matrix, the spatial positional relationship and basic attribute labels of adjacent monitoring grids are synchronously associated to complete the spatial mapping of standardized feature data of multiple monitoring grids.

[0034] Furthermore, in step S4, after the graded early warning is triggered, the corresponding pest and disease verification data and control effect data are collected. Based on the pest and disease verification data and control effect data, the node resource dynamic scheduling model and the pest and disease spatial diffusion prediction model are iteratively optimized, and the optimized model parameters are sent to the corresponding edge computing unit.

[0035] An IoT-based pest and disease monitoring system is provided, comprising a gridded monitoring and sensing module, an edge computing processing module, a cloud-based collaborative scheduling module, a spatial correlation analysis module, and a tiered early warning module. The gridded monitoring and sensing module is used to complete the grid division of the monitoring area, attribute labeling, and monitoring data collection. The edge computing processing module is used to complete data preprocessing, feature extraction, and local early warning triggering. The cloud-based collaborative scheduling module is used to build a dynamic scheduling model for node resources and generate and distribute operating strategy parameters. The spatial correlation analysis module is used to build a spatial correlation matrix and a pest and disease spatial spread prediction model, outputting the spread risk distribution. The tiered early warning module is used to complete tiered early warning triggering and generate prevention and control guidelines.

[0036] The beneficial effects of this invention are:

[0037] (1) By using a collaborative architecture of gridded monitoring area, edge data processing and cloud-based global control, the entire process of pest and disease monitoring is executed in a closed loop, reducing invalid data transmission, reducing cloud resource consumption, and shortening the early warning response delay for abnormal situations.

[0038] (2) A dynamic scheduling model is constructed with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption. Differentiated operation strategies are implemented for equipment in different monitoring grids to achieve accurate matching between monitoring resources and monitoring needs, reduce ineffective resource consumption, and improve the continuous operation capability of equipment.

[0039] (3) Constructing a spatial correlation matrix and a spatial spread prediction model for pests and diseases can identify the spread trend and monitoring blind spots of pests and diseases in advance. Combined with the model iteration and optimization mechanism, it can continuously improve the accuracy of pest and disease monitoring and the proactiveness of prevention and control work. Attached Figure Description

[0040] Figure 1 A flowchart illustrating the steps of an IoT-based pest and disease monitoring method;

[0041] Figure 2 The following is a flowchart illustrating the specific steps of an IoT-based pest and disease monitoring method provided for an embodiment. Detailed Implementation

[0042] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0043] Example 1

[0044] See Figure 1 This embodiment provides a pest and disease monitoring method based on the Internet of Things, which includes the following steps:

[0045] S1. Divide the target monitoring area into monitoring grids according to crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics. Generate spatial identity identifiers for each monitoring grid, label the basic attribute tags of each monitoring grid, deploy IoT sensing nodes and edge computing units in each monitoring grid, complete the spatial binding of IoT sensing nodes and edge computing units with the corresponding monitoring grids, and label the operating parameter tags of IoT sensing nodes and edge computing units.

[0046] S2. Collect the raw monitoring data of the corresponding monitoring grid. The raw monitoring data includes environmental time series data and crop image data. Preprocess the raw monitoring data and extract features through a lightweight feature extraction model to obtain standardized feature data. Upload the standardized feature data and node operation status data to the cloud.

[0047] S3. Receive standardized feature data and node operation status data, retrieve the basic attribute labels of the corresponding monitoring grid and the real-time monitoring data of adjacent monitoring grids, construct a node resource dynamic scheduling model with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption, and output and distribute operation strategy parameters.

[0048] S4. Construct a spatial correlation matrix based on standardized feature data from all monitoring grids, build a spatial spread prediction model for pests and diseases, output the risk distribution of pest and disease spread in the monitoring area through the spatial spread prediction model for pests and diseases, and complete the hierarchical early warning triggering and operation strategy parameter update.

[0049] In some embodiments, step S1 includes:

[0050] S1.1. Divide the target monitoring area into monitoring grids according to crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics, and generate a unique spatial identity for each monitoring grid;

[0051] S1.2. Label the basic attribute tags for each monitoring grid. The basic attribute tags include crop variety attributes, planting environment attributes, pest and disease risk level attributes, and spatial association attributes between adjacent grids.

[0052] S1.3. Deploy IoT sensing nodes and edge computing units within each monitoring grid to complete the spatial binding of IoT sensing nodes and edge computing units with the corresponding monitoring grid;

[0053] S1.4. Label the operating parameter tags of each IoT sensing node and edge computing unit. The operating parameter tags include computing power threshold, remaining battery life threshold, maximum sampling capacity, and data transmission limit.

[0054] In some embodiments, step S2 includes:

[0055] S2.1. Collect raw monitoring data for the corresponding monitoring grid according to the current operating strategy. The raw monitoring data includes environmental time series data and crop image data.

[0056] S2.2. Perform outlier removal, data denoising, and time series normalization on environmental time series data; perform invalid region cropping, image enhancement, and format compression on crop image data.

[0057] S2.3. Construct a lightweight feature extraction model, which includes an input layer, a convolutional feature extraction layer, a fully connected layer, and an output layer. The input layer takes preprocessed environmental time-series data and preprocessed crop image data as input. The convolutional feature extraction layer is fully connected to the input layer to extract multi-dimensional features from the input. The fully connected layer is fully connected to the convolutional feature extraction layer to normalize the extracted features. The output layer is fully connected to the fully connected layer to output standardized feature data.

[0058] S2.4. The standardized feature data and node operation status data are encrypted and uploaded to the cloud, while the original monitoring data is stored locally in the edge computing unit.

[0059] In some embodiments, step S3 includes:

[0060] S3.1. Receive standardized feature data and node operation status data, retrieve the basic attribute labels of the corresponding monitoring grid, historical pest and disease occurrence data and real-time monitoring data of adjacent monitoring grids, and construct the input dataset for the node resource dynamic scheduling model;

[0061] S3.2. Construct a dynamic scheduling model for node resources. The dynamic scheduling model for node resources includes an input layer, a feature fusion layer, a multi-objective optimization layer, and an output layer. The input of the input layer is the input dataset of the dynamic scheduling model for node resources. The feature fusion layer is fully connected to the input layer and performs feature normalization and correlation fusion processing on the input dataset. The multi-objective optimization layer is fully connected to the feature fusion layer and performs calculations with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption. The output layer is fully connected to the multi-objective optimization layer and outputs the running strategy parameters, which include sampling frequency, data transmission cycle, computing power allocation weight, and early warning threshold adjustment parameters.

[0062] S3.3. Differentiate scheduling of IoT sensing nodes and edge computing units in different monitoring grids according to the operation strategy parameters;

[0063] S3.4. Perform routine updates of the operation strategy parameters according to the set cycle. When the pest and disease risk level of the monitoring grid changes or the node operation status becomes abnormal, trigger the real-time update of the operation strategy parameters and send the updated operation strategy parameters to the corresponding edge computing unit and IoT sensing node.

[0064] In some embodiments, step S4 includes:

[0065] S4.1. Construct a spatial correlation matrix based on the spatial location relationship and attribute association of all monitoring grids, and map all standardized feature data to the corresponding monitoring grids to form a spatiotemporal distribution dataset of pest and disease risk;

[0066] S4.2. Construct a spatial spread prediction model for pests and diseases. The model includes an input layer, a spatiotemporal feature extraction layer, a spread pattern fitting layer, and an output layer. The input layer takes a dataset of the spatiotemporal distribution of pest and disease risk as input. The spatiotemporal feature extraction layer is fully connected to the input layer to extract spatial correlation features and temporal change features from the input dataset. The spread pattern fitting layer is fully connected to the spatiotemporal feature extraction layer to calculate the spread pattern of pests and diseases based on the extracted features. The output layer is fully connected to the spread pattern fitting layer to output the risk distribution of pest and disease spread in the monitored area.

[0067] S4.3. Identify monitoring blind spots based on the distribution of pest and disease spread risk, input the identification results of monitoring blind spots into the node resource dynamic scheduling model, and complete the optimization of operation strategy parameters;

[0068] S4.4. Based on standardized feature data and the distribution of pest and disease spread risks, trigger local emergency early warnings at the edge and global hierarchical early warnings in the cloud, and generate corresponding prevention and control guidelines.

[0069] In some embodiments, in step S2.3, the parameters of the lightweight feature extraction model are iteratively adjusted using updated data sent from the cloud.

[0070] In some embodiments, during step S3.3, when performing differentiated scheduling of IoT sensing nodes and edge computing units in different monitoring grids, for monitoring grids with a pest and disease risk level higher than a set value, the sampling frequency and computing power allocation weight are increased to shorten the data transmission cycle; for monitoring grids with a pest and disease risk level lower than a set value, the sampling frequency and computing power allocation weight are reduced to extend the data transmission cycle.

[0071] In some embodiments, during step S4.1, when constructing the spatial correlation matrix, the spatial positional relationships and basic attribute labels of adjacent monitoring grids are synchronously associated to complete the spatial mapping of standardized feature data of multiple monitoring grids.

[0072] In some embodiments, in step S4, after the graded early warning is triggered, the pest and disease verification data and control effect data corresponding to the early warning are collected. Based on the pest and disease verification data and control effect data, the node resource dynamic scheduling model and the pest and disease spatial diffusion prediction model are iteratively optimized, and the optimized model parameters are sent to the corresponding edge computing unit.

[0073] An IoT-based pest and disease monitoring system is provided, comprising a gridded monitoring and sensing module, an edge computing processing module, a cloud-based collaborative scheduling module, a spatial correlation analysis module, and a tiered early warning module. The gridded monitoring and sensing module is used to complete the grid division of the monitoring area, attribute labeling, and monitoring data collection. The edge computing processing module is used to complete data preprocessing, feature extraction, and local early warning triggering. The cloud-based collaborative scheduling module is used to build a dynamic scheduling model for node resources and generate and distribute operating strategy parameters. The spatial correlation analysis module is used to build a spatial correlation matrix and a pest and disease spatial spread prediction model, outputting the spread risk distribution. The tiered early warning module is used to complete tiered early warning triggering and generate prevention and control guidelines.

[0074] Example 2

[0075] This embodiment provides a specific implementation process for an IoT-based pest and disease monitoring method. This embodiment is implemented based on an edge-cloud collaborative distributed architecture. Through a grid-based monitoring layout, lightweight data processing at the edge, global resource scheduling in the cloud, and pest and disease risk prediction, the entire process of monitoring and controlling pests and diseases in the planting area is completed. Figure 2 As shown, the specific implementation process is as follows:

[0076] Step 1. Monitoring area grid division and equipment deployment:

[0077] Step 1.1. Grid cell division of the monitoring area:

[0078] A monitoring grid refers to a continuous planting monitoring area divided into multiple independent, individually manageable monitoring units according to preset division rules. In this embodiment, the target monitoring area is divided into grids based on crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics. A unique spatial identifier is generated for each monitoring grid obtained after the division. The spatial identifier is a unique code used to distinguish different monitoring grids and can be used to achieve spatial matching and traceability of monitoring data with the corresponding grid.

[0079] Step 1.2. Monitoring grid basic attribute calibration:

[0080] Each monitoring grid is labeled with basic attribute tags, including crop variety attributes, planting environment attributes, pest and disease risk level attributes, and spatial correlation attributes between adjacent grids. The crop variety attribute identifies the type of crop planted within the corresponding monitoring grid; the planting environment attribute identifies environmental information such as topography, light conditions, and irrigation conditions; the pest and disease risk level attribute identifies the probability level of pest and disease occurrence for the corresponding monitoring grid based on historical data; and the spatial correlation attributes between adjacent grids identify the positional relationship and attribute association between the corresponding monitoring grid and its neighboring grids.

[0081] Step 1.3. Monitoring Equipment Deployment and Spatial Binding:

[0082] IoT sensing nodes and edge computing units are deployed within each monitoring grid. IoT sensing nodes are terminal sensing devices capable of acquiring environmental and image data, while edge computing units are edge computing devices capable of local data processing, storage, and transmission. Spatial binding is established between the IoT sensing nodes and edge computing units and their corresponding monitoring grids. Spatial binding involves associating the device identifiers of the IoT sensing nodes and edge computing units with the spatial identifiers of the corresponding monitoring grids, enabling the data collected by the devices to be directly matched to the corresponding monitoring grid.

[0083] Step 1.4. Equipment operating parameter calibration:

[0084] Each IoT sensing node and edge computing unit is labeled with its operating parameter tags, which include computing power threshold, remaining battery life threshold, maximum sampling capacity, and data transmission limit. The computing power threshold identifies the maximum computing power that the corresponding edge computing unit can provide; the remaining battery life threshold identifies the minimum remaining battery power at which the corresponding device can operate normally; the maximum sampling capacity identifies the maximum data acquisition frequency that the corresponding IoT sensing node can achieve; and the data transmission limit identifies the maximum data transmission rate that the corresponding device can achieve.

[0085] In some specific implementations, for large-scale field planting scenarios with contiguous planting areas exceeding 100 hectares, the target monitoring area is divided into basic monitoring grids with units of 1 hectare each. Each monitoring grid corresponds to a unique 16-bit spatial identifier, containing three core information categories: the plot code, crop type code, and initial risk level code. Within each monitoring grid, two sets of IoT sensing nodes and one edge computing unit are deployed. The installation spacing of the IoT sensing nodes is set at 50 meters, and the installation height is set at 0.5 to 1.2 meters above the crop canopy. The edge computing unit is deployed at the center of the monitoring grid, with a maximum transmission distance of no more than 100 meters from the IoT sensing nodes.

[0086] The device's operating parameters were calibrated simultaneously. The edge computing unit's computing power threshold was set to 8 TOPS, the remaining battery life threshold to 20% of full capacity, the maximum sampling capability of the IoT sensing nodes was set to 60 environmental data acquisitions per hour and 1 crop image acquisition every 2 hours, and the overall data transmission limit was set to 5 MB per second. This implementation addresses the problems of insufficient monitoring accuracy due to excessively large grid granularity in large-area planting areas and excessively high equipment costs due to excessively small granularity. By standardizing grid granularity and device deployment density, a balance between monitoring coverage and deployment cost is achieved. Simultaneously, a fixed-format spatial identifier enables accurate traceability and matching of monitoring data, avoiding confusion and mismatch between data from multiple grids.

[0087] In some embodiments, the division of monitoring grids can be adjusted in conjunction with the land ownership boundaries of the monitoring area, so that a single monitoring grid matches the planting plot boundary of a single planting entity, which facilitates the hierarchical push of subsequent monitoring data and the targeted distribution of prevention and control guidelines.

[0088] Step 2. Monitoring data acquisition and feature extraction processing:

[0089] Step 2.1. Raw monitoring data collection:

[0090] According to the current operating strategy, raw monitoring data is collected for the corresponding monitoring grid. The raw monitoring data includes environmental time-series data and crop image data. Environmental time-series data refers to field environment-related data that changes over time and is collected at fixed time intervals. Crop image data refers to image data obtained by taking pictures of crop plants within the monitoring grid. The operating strategy refers to the pre-set or cloud-based rules and parameters used to regulate equipment data collection, data transmission, and computing power allocation.

[0091] Step 2.2. Preprocessing of raw monitoring data:

[0092] The environmental time-series data undergoes outlier removal, data denoising, and time-series normalization. Outlier removal removes invalid data exceeding reasonable limits; data denoising removes noise generated by environmental interference during data acquisition; and time-series normalization converts environmental time-series data of different magnitudes into values ​​within a unified standard range, eliminating the impact of data magnitude differences on subsequent processing. Crop image data undergoes invalid region cropping, image enhancement, and format compression. Invalid region cropping removes background areas without crop plants; image enhancement adjusts contrast and brightness to highlight crop plant details; and format compression reduces image file size without sacrificing key features, minimizing data transmission resource consumption.

[0093] Step 2.3. Standardized Feature Data Extraction:

[0094] A lightweight feature extraction model is constructed, comprising an input layer, a convolutional feature extraction layer, a fully connected layer, and an output layer. The input layer takes preprocessed environmental time-series data and preprocessed crop image data as input. The convolutional feature extraction layer is fully connected to the input layer, performing multi-dimensional feature extraction. This layer contains multiple cascaded convolutional and pooling units. Convolutional units extract local detail features from the input data, while pooling units compress the dimensions of the extracted features, preserving key feature information. The fully connected layer is also fully connected to the convolutional feature extraction layer, normalizing the extracted features and converting features of different dimensions into feature vectors of a uniform format. The output layer is fully connected to the fully connected layer, outputting standardized feature data, including environmental time-series features related to pest and disease occurrence and crop phenotypic features.

[0095] The training process of the lightweight feature extraction model includes the following steps: First, construct a training dataset, which includes preprocessed environmental time-series sample data, preprocessed crop image sample data, and corresponding pest and disease feature annotation data. Second, input the training dataset into the input layer of the lightweight feature extraction model. Feature extraction is performed through a convolutional feature extraction layer, feature normalization is performed through a fully connected layer, and predicted feature data is output through the output layer. Third, calculate the loss value based on the predicted feature data and the annotation data, and iteratively update the parameters of the lightweight feature extraction model using the backpropagation algorithm. Fourth, when the loss value is lower than a set threshold or the number of iterations reaches a set upper limit, the training of the lightweight feature extraction model is complete, and deployable model weight parameters are obtained.

[0096] In some specific implementations, to address the computational limitations of edge computing units, the lightweight feature extraction model is constructed with an 8-layer network structure, including one input layer, four convolutional feature extraction layers, two fully connected layers, and one output layer. In the four-layer structure of the convolutional feature extraction layers, the first convolutional unit uses 32 3×3 convolutional kernels with a stride of 1 and padding of 1, and the corresponding pooling unit uses 2×2 max pooling with a stride of 2. The second convolutional unit uses 64 3×3 convolutional kernels with a stride of 1 and padding of 1, and the corresponding pooling unit uses 2×2 max pooling with a stride of 2. The third and fourth convolutional units each use 128 3×3 depthwise separable convolutional kernels with a stride of 1 and padding of 1, without separate pooling units.

[0097] The first layer of the fully connected layer has 256 neurons, the second layer has 64 neurons, and the output layer has 16 neurons, corresponding to 16-dimensional standardized feature data. During model training, the batch size is set to 32, the initial learning rate is set to 0.001, the learning rate decay coefficient is set to decrease by 0.1 every 10 iterations, the Adam optimizer is used, the mean squared error loss function is used, and the total number of iterations is set to 100. Training is terminated early when the loss value decreases by less than 0.0001 for 10 consecutive iterations. In this implementation, after preprocessing, the resolution of the original crop image data is compressed from 4096×3072 to 1024×768, and the file size is compressed from an average of 8MB to an average of 500KB. At the same time, most of the key features related to lesions and insect bodies are retained. The time taken for the edge computing unit to complete a single feature extraction is relatively low, achieving efficient processing at the edge, significantly reducing the amount of invalid data transmission, and solving the problems of excessive bandwidth consumption and long cloud processing latency caused by uploading the full amount of original high-definition images.

[0098] Step 2.4. Data Upload and Local Storage:

[0099] Standardized feature data and node operational status data are encrypted and uploaded to the cloud. The node operational status data includes real-time computing power usage, remaining power, and data acquisition status information of IoT sensing nodes and edge computing units. Raw monitoring data is stored locally on the edge computing unit and is not uploaded to the cloud; it is only accessed locally for subsequent model optimization or data verification.

[0100] In some specific implementations, the standardized feature data and node operation status data uploaded by the edge computing unit to the cloud are encrypted using the AES-256 symmetric encryption algorithm. During encryption, the key length is set to 256 bits, the block length is set to 128 bits, the encryption mode is CBC mode, and the initialization vector is set to a 16-bit random number. The initialization vector for each data packet is unique. The data packet structure uploaded by the edge computing unit is set to a fixed format. Each data packet contains a 16-byte header, a 32-byte encryption key checksum, 128 bytes of standardized feature data, 32 bytes of node operation status data, and a 16-byte checksum. The total size of a single data packet is fixed at 224 bytes. The edge computing unit uploads data packets within the corresponding period to the cloud in batches according to the set transmission cycle. The number of data packets uploaded in a single batch does not exceed 1024, and the total data volume in a single batch does not exceed 230KB.

[0101] Meanwhile, the raw monitoring data stored locally on the edge computing unit is encrypted using the same encryption algorithm. The storage partition is set as a separate encrypted partition with a capacity of 32GB, capable of storing at least six months of raw monitoring data. Decryption and retrieval of the corresponding data packet are only permitted when a data verification request is initiated from the cloud. This implementation, through a fixed-format data packet design and high-strength encryption, ensures the security of data transmission, preventing the leakage of production data from the planting entity. Simultaneously, by controlling the standardized data packet size, it significantly reduces the bandwidth consumption of a single data upload, solving the problems of inconsistent data formats, high bandwidth consumption, and insufficient data security in traditional data transmission processes.

[0102] In some embodiments, the convolutional feature extraction layer of the lightweight feature extraction model can use a depthwise separable convolutional structure to replace the conventional convolutional structure, further reducing the resource consumption of the model computation and improving the feature extraction processing speed of the edge computing unit.

[0103] Step 3. Generation and distribution of dynamic scheduling strategies for node resources:

[0104] Step 3.1. Construction of the input dataset for the scheduling model:

[0105] The system receives standardized feature data and node operational status data, retrieves basic attribute labels, historical pest and disease occurrence data, and real-time monitoring data from adjacent monitoring grids for the corresponding monitoring grid, and constructs the input dataset for the node resource dynamic scheduling model. Historical pest and disease occurrence data refers to records related to the occurrence time, type, and extent of pests and diseases in past periods for the corresponding monitoring grid. The input dataset contains all feature data related to node scheduling, providing the data foundation for the calculation of the node resource dynamic scheduling model.

[0106] Step 3.2. Construction and calculation of the dynamic scheduling model for node resources:

[0107] A dynamic scheduling model for node resources is constructed, comprising an input layer, a feature fusion layer, a multi-objective optimization layer, and an output layer. The input layer takes the input dataset as its input. The feature fusion layer is fully connected to the input layer, performing feature normalization and correlation fusion on the input dataset. It uses an attention mechanism to weight different features in the input dataset, highlighting key features related to pest and disease monitoring and resource consumption, thus achieving the correlation fusion of multi-source heterogeneous features. The multi-objective optimization layer is fully connected to the feature fusion layer, solving for the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption. This multi-objective optimization layer employs a multi-objective optimization algorithm, a currently available algorithm for solving multi-objective optimization problems, which can find a balanced solution between two constrained optimization objectives. The output layer is fully connected to the multi-objective optimization layer, outputting the operating strategy parameters, including sampling frequency, data transmission cycle, computing power allocation weights, and early warning threshold adjustment parameters.

[0108] The training process of the node resource dynamic scheduling model includes the following steps: First, a training dataset is constructed, comprising historical monitoring grid attribute data, historical standardized feature data, historical node operating status data, and corresponding scheduling effect annotation data, including monitoring accuracy data and resource consumption data. Second, the training dataset is input into the input layer of the node resource dynamic scheduling model. Feature fusion processing is performed through a feature fusion layer, dual-objective solution calculation is performed through a multi-objective optimization layer, and predicted operating strategy parameters are output through an output layer. Third, the multi-objective loss value is calculated based on the predicted operating strategy parameters and the annotation data, and the parameters of the node resource dynamic scheduling model are iteratively updated using the backpropagation algorithm. Fourth, when the multi-objective loss value falls below a set threshold or the number of iterations reaches a set upper limit, the training of the node resource dynamic scheduling model is completed, and deployable model weight parameters are obtained.

[0109] Step 3.3. Execution of Differentiated Node Scheduling:

[0110] Based on operational strategy parameters, IoT sensing nodes and edge computing units in different monitoring grids are scheduled in a differentiated manner. For monitoring grids with pest and disease risk levels higher than the set value, the sampling frequency and computing power allocation weight are increased, and the data transmission cycle is shortened. For monitoring grids with pest and disease risk levels lower than the set value, the sampling frequency and computing power allocation weight are reduced, and the data transmission cycle is extended. The set value refers to a pre-set critical threshold used to distinguish different pest and disease risk levels.

[0111] In some specific implementations, the input dataset of the node resource dynamic scheduling model contains eight core feature items: crop variety attributes, planting environment attributes, initial pest and disease risk level attributes, spatial correlation attributes between adjacent grids, real-time standardized feature data, real-time node operating status data, historical pest and disease occurrence data, and real-time monitoring data of adjacent grids. Each feature item is converted into a 16-dimensional normalized feature vector, and the total feature dimension of the input dataset is 128 dimensions. In the structure of the node resource dynamic scheduling model, the feature fusion layer adopts a multi-head attention mechanism with eight attention heads. Weights are assigned to the eight core feature items, with the weights ranging from 0 to 1, and the sum of all weights is 1.

[0112] The multi-objective optimization layer employs the NSGA-II multi-objective optimization algorithm, with a population size of 100, 50 iterations, a crossover probability of 0.9, and a mutation probability of 0.1. The algorithm aims to maximize monitoring accuracy and minimize resource consumption as its dual optimization objectives. During differentiated scheduling, for high-risk monitoring grids with a pest and disease risk level above 0.7, the sampling frequency is increased from the default once per hour to once every 10 minutes, the computing power allocation weight is increased from the default 0.2 to 0.8, and the data transmission cycle is shortened from the default once every 24 hours to once every hour. For low-risk monitoring grids with a pest and disease risk level below 0.3, the sampling frequency is decreased from the default once per hour to once every 6 hours, the computing power allocation weight is decreased from the default 0.2 to 0.05, and the data transmission cycle is extended from the default once every 24 hours to once every 72 hours. This implementation method achieves a reasonable allocation of monitoring resources to high-risk areas by quantifying risk level classification and adjusting corresponding scheduling parameters, while significantly reducing resource consumption in low-risk areas. Actual operation has verified that the sampling density of monitoring data in high-risk areas has been significantly improved, and the power consumption of nodes in low-risk areas has been effectively controlled, solving the problems of missed detections in high-risk areas and resource waste in low-risk areas caused by fixed sampling strategies.

[0113] Step 3.4. Scheduling strategy update and distribution:

[0114] The system performs routine updates of the operational strategy parameters according to a set cycle. When the pest and disease risk level of the monitored grid changes or the node's operational status becomes abnormal, the operational strategy parameters are triggered for real-time updates, and the updated operational strategy parameters are distributed to the corresponding edge computing units and IoT sensing nodes. The set cycle refers to a pre-defined fixed time interval for executing global scheduling strategy updates. An abnormal node operational status refers to situations where the node's remaining battery power is lower than the remaining endurance threshold, the computing power usage exceeds the computing power threshold, or data acquisition is interrupted.

[0115] In some embodiments, the multi-objective optimization layer of the node resource dynamic scheduling model can include a node remaining endurance weight term. For nodes with remaining endurance below a set threshold, the sampling frequency and computing power allocation weight are reduced first to extend the continuous running time of the node.

[0116] Step 4. Risk analysis and early warning implementation for the spread of pests and diseases:

[0117] Step 4.1. Construction of the spatiotemporal distribution dataset of pest and disease risk:

[0118] A spatial correlation matrix is ​​constructed based on the spatial location relationships and attribute associations of all monitoring grids. This matrix represents the spatial location relationships and attribute associations between different monitoring grids, enabling spatial correlation analysis of data from multiple monitoring grids. The spatial location relationships and basic attribute labels of adjacent monitoring grids are synchronously correlated to complete the spatial mapping of standardized feature data from multiple monitoring grids. All standardized feature data are mapped to their corresponding monitoring grids, forming a spatiotemporal distribution dataset of pest and disease risk. This dataset contains pest and disease-related feature data from different monitoring grids and at different time points, and can be used to analyze the occurrence patterns and spread trends of pests and diseases.

[0119] Step 4.2. Prediction of the risk of pest and disease spread:

[0120] A spatial spread prediction model for pests and diseases is constructed, comprising an input layer, a spatiotemporal feature extraction layer, a spread pattern fitting layer, and an output layer. The input layer takes as input a spatiotemporal distribution dataset of pest and disease risk. The spatiotemporal feature extraction layer is fully connected to the input layer, extracting spatial correlation features and temporal variation features from the input dataset. This layer includes spatial and temporal feature extraction branches; the spatial branch extracts spatial correlation features between different monitoring grids, while the temporal branch extracts feature variation patterns for the same monitoring grid at different time points. The spread pattern fitting layer is fully connected to the spatiotemporal feature extraction layer, calculating and fitting pest and disease spread patterns based on the extracted features. This layer fits the spread rate and direction of pests and diseases based on the spatial correlation and temporal continuity of pest and disease transmission. The output layer is fully connected to the spread pattern fitting layer, outputting the pest and disease spread risk distribution for the monitoring area. This risk distribution characterizes the probability level of pest and disease occurrence in different monitoring grids within the monitoring area in future periods.

[0121] The training process of the pest and disease spatial spread prediction model includes the following steps: First, a training dataset is constructed, which includes historical spatiotemporal distribution sample data of pest and disease risks, and corresponding labeled data of actual pest and disease spread. Second, the training dataset is input into the input layer of the pest and disease spatial spread prediction model. Spatiotemporal features are extracted through a spatiotemporal feature extraction layer, diffusion pattern fitting is performed through a diffusion pattern fitting layer, and the predicted pest and disease spread risk distribution data is output through the output layer. Third, the loss value is calculated based on the predicted pest and disease spread risk distribution data and the labeled data, and the parameters of the pest and disease spatial spread prediction model are iteratively updated using the backpropagation algorithm. Fourth, when the loss value is lower than a set threshold or the number of iterations reaches a set upper limit, the training of the pest and disease spatial spread prediction model is completed, and deployable model weight parameters are obtained.

[0122] In some specific implementations, the constructed spatial correlation matrix is ​​an N×N two-dimensional matrix, where N is the total number of monitoring grids. Each element in the matrix corresponds to the spatial correlation degree between two monitoring grids. The spatial correlation degree is calculated based on three dimensions: spatial distance between the two monitoring grids, crop variety similarity, and planting environment similarity. The weight of spatial distance is set to 0.5, the weight of crop variety similarity is set to 0.3, and the weight of planting environment similarity is set to 0.2. The value of spatial correlation degree ranges from 0 to 1, with higher values ​​indicating a stronger correlation between the two monitoring grids in terms of pest and disease transmission. In the structure of the pest and disease spatial diffusion prediction model, the spatial feature extraction branch of the spatiotemporal feature extraction layer uses a graph convolutional neural network with 3 layers and 64 output dimensions for each layer.

[0123] The temporal feature extraction branch employs a Long Short-Term Memory (LSTM) network with 128 neurons in the hidden layer and a time step of 24, corresponding to monitoring data from the past 24 time periods. The diffusion pattern fitting layer uses a fully connected network with three fully connected layers, each with 256, 128, and 64 neurons respectively. The output layer is an N-dimensional vector representing the pest and disease diffusion risk value for the next 72 hours across N monitoring grids, with the risk value ranging from 0 to 1. During model training, the batch size is set to 16, the initial learning rate to 0.0005, the AdamW optimizer is used, the cross-entropy loss function is employed, and the total number of iterations is 200. Training is terminated early when the validation set accuracy improvement is less than 0.1% over 15 consecutive iterations. This implementation, through a prediction model combining a quantized spatial correlation matrix and spatiotemporal features, effectively predicts pest and disease diffusion trends. The model demonstrates high accuracy in predicting high-risk areas across different periods, solving the problem of single-point monitoring failing to predict cross-regional pest and disease diffusion, leading to delayed control measures.

[0124] Step 4.3. Monitoring blind spot identification and scheduling strategy optimization:

[0125] Based on the risk distribution of pest and disease spread, monitoring blind spots are identified. These blind spots refer to areas where effective monitoring is not achieved and there is a risk of missed pest and disease detection. These include areas where nodes are malfunctioning, high-risk areas with insufficient sampling frequency, and boundary areas where no nodes are deployed. The results of the blind spot identification are input into the node resource dynamic scheduling model to optimize the operational strategy parameters.

[0126] Step 4.4. Triggering of Tiered Early Warnings and Generation of Prevention and Control Guidelines:

[0127] Based on standardized feature data and the distribution of pest and disease spread risks, local emergency warnings at the edge and global tiered warnings in the cloud are triggered, generating corresponding prevention and control guidelines. Local emergency warnings are executed locally by edge computing units. When standardized feature data reaches the warning threshold, a local warning action is triggered directly, without waiting for cloud processing. Global tiered warnings in the cloud are executed based on the distribution of pest and disease spread risks across the entire region. Warnings of corresponding levels are triggered according to different risk levels, and prevention and control guidelines matching the risk level and crop variety are generated. These guidelines are used to guide growers in carrying out corresponding pest and disease control operations.

[0128] Step 4.5. Iterative Model Optimization:

[0129] After triggering the tiered early warning system, the system collects corresponding pest and disease verification data and control effect data. Pest and disease verification data refers to the actual pest and disease occurrence data obtained by verifying the data against the corresponding monitoring grid after the early warning is triggered. Control effect data refers to the data on changes in pest and disease occurrence after the implementation of corresponding control measures. Based on the pest and disease verification data and control effect data, the system iteratively optimizes the node resource dynamic scheduling model and the pest and disease spatial diffusion prediction model, and distributes the optimized model parameters to the corresponding edge computing units. Simultaneously, based on the pest and disease verification data and control effect data, the system iteratively adjusts the parameters of the lightweight feature extraction model, updating these parameters using update data distributed from the cloud.

[0130] In some specific implementations, incremental training is used for iterative optimization of the node resource dynamic scheduling model, the pest and disease spatial spread prediction model, and the lightweight feature extraction model. The incremental training dataset consists of newly added pest and disease verification data, control effect data, and standardized feature data, not exceeding 30 days. The number of samples in the incremental training dataset does not exceed 20% of the number of samples in the initial training dataset. During incremental training, the parameters of the bottom feature extraction layer of the model remain unchanged, and only the parameters of the top output layer of the model are fine-tuned. The initial learning rate during fine-tuning is set to 10% of the initial training learning rate. The total number of iterations is set to 20, and the batch size is set to 16. Training is terminated early when the loss value decreases by less than 0.0001 for 5 consecutive iterations. The model iteration optimization cycle is set to a fixed 30 days. Temporary model iteration optimization is triggered when the false alarm rate or false negative rate in the newly added early warning verification data exceeds a set threshold of 10% for 7 consecutive days.

[0131] After optimization, the model parameters are distributed from the cloud to the corresponding edge computing units. The data packets for parameter distribution are encrypted, and the size of a single model parameter file does not exceed 5MB. The edge computing unit completes the model parameter update in no more than 10 seconds, and the update process does not affect the device's basic data acquisition and local early warning functions. This implementation method significantly reduces the computing power consumption of model iterative optimization through incremental training, greatly shortens the model optimization cycle, and ensures the continuous and stable operation of the monitoring system by updating parameters without interrupting business operations. It solves the problems of excessive computing power consumption and system business interruption caused by traditional full model retraining.

[0132] In some embodiments, meteorological forecast data can be added as an auxiliary input to the diffusion pattern fitting layer of the pest and disease spatial diffusion prediction model to improve the prediction accuracy of the pest and disease diffusion risk distribution.

[0133] In some embodiments, the model iterative optimization process can adopt incremental training, which fine-tunes the model parameters only based on newly added verification data and effect data, thereby reducing the computational cost of model training and shortening the model optimization cycle.

[0134] This embodiment achieves refined zoning and control of planting areas through a grid-based monitoring layout. Lightweight data processing and feature extraction at the edge reduce the amount of invalid data transmitted to the cloud, lowering cloud storage and computing resource consumption, and shortening the delay in early warning response to anomalies. This embodiment uses a dynamic node resource scheduling model to match monitoring resources with monitoring needs, enabling differentiated equipment operation strategies for monitoring grids with different risk levels, reducing ineffective resource consumption in low-risk areas and improving the continuous operation capability of equipment. This embodiment uses a pest and disease spatial spread prediction model to predict pest and disease spread trends, identifying high-risk areas and monitoring blind spots in advance, improving the comprehensiveness of monitoring and the proactiveness of prevention and control. This embodiment combines local emergency response with precise global control through an edge-cloud collaborative architecture. Furthermore, a full-process model iteration and optimization mechanism continuously improves the accuracy of monitoring and the rationality of scheduling, adapting to the pest and disease monitoring needs of different planting scenarios.

[0135] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A pest and disease monitoring method based on the Internet of Things, characterized in that, Includes the following steps: S1. Divide the target monitoring area into monitoring grids according to crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics. Generate spatial identity identifiers for each monitoring grid, label the basic attribute tags of each monitoring grid, deploy IoT sensing nodes and edge computing units within each monitoring grid, complete the spatial binding of IoT sensing nodes and edge computing units with the corresponding monitoring grids, and label the operating parameter tags of IoT sensing nodes and edge computing units. The operating parameter tags include computing power threshold, remaining battery life threshold, maximum sampling capacity, and data transmission limit. S2. Collect the raw monitoring data of the corresponding monitoring grid. The raw monitoring data includes environmental time series data and crop image data. Preprocess the raw monitoring data and extract features through a lightweight feature extraction model to obtain standardized feature data. Upload the standardized feature data and node operation status data to the cloud. S3. Receive standardized feature data and node operation status data, retrieve the basic attribute labels of the corresponding monitoring grid and the real-time monitoring data of adjacent monitoring grids, construct a node resource dynamic scheduling model with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption, output and distribute operation strategy parameters, and perform differentiated scheduling of IoT sensing nodes and edge computing units in different monitoring grids. S4. Construct a spatial correlation matrix based on the standardized feature data of all monitoring grids, build a spatial spread prediction model for pests and diseases, output the risk distribution of pests and diseases in the monitoring area through the spatial spread prediction model for pests and diseases, and complete the hierarchical early warning triggering and operation strategy parameter update. Specifically, when differentiating the scheduling of IoT sensing nodes and edge computing units in different monitoring grids, for monitoring grids with a pest and disease risk level higher than the set value, the sampling frequency and computing power allocation weight are increased to shorten the data transmission cycle; for monitoring grids with a pest and disease risk level lower than the set value, the sampling frequency and computing power allocation weight are reduced to extend the data transmission cycle. In constructing the spatial correlation matrix, the spatial location relationship and basic attribute labels of adjacent monitoring grids are synchronously associated to complete the spatial mapping of standardized feature data of multiple monitoring grids.

2. The method according to claim 1, characterized in that, Step S1 includes: S1.

1. Divide the target monitoring area into monitoring grids according to crop type, planting density, historical frequency of pests and diseases, and geographical environmental characteristics, and generate a unique spatial identity for each monitoring grid; S1.

2. Assign basic attribute labels to each monitoring grid. The basic attribute labels include crop variety attributes, planting environment attributes, pest and disease risk level attributes, and spatial association attributes between adjacent grids. S1.

3. Deploy IoT sensing nodes and edge computing units within each monitoring grid to complete the spatial binding of IoT sensing nodes and edge computing units with the corresponding monitoring grid; S1.

4. Label the operating parameters of each IoT sensing node and edge computing unit.

3. The method according to claim 1, characterized in that, Step S2 includes: S2.

1. Collect raw monitoring data for the corresponding monitoring grid according to the current operating strategy. The raw monitoring data includes environmental time series data and crop image data. S2.

2. Perform outlier removal, data denoising, and time series normalization on environmental time series data; perform invalid region cropping, image enhancement, and format compression on crop image data. S2.

3. Construct a lightweight feature extraction model, which includes an input layer, a convolutional feature extraction layer, a fully connected layer, and an output layer. The input layer takes preprocessed environmental time-series data and preprocessed crop image data as input. The convolutional feature extraction layer is fully connected to the input layer to extract multi-dimensional features from the input. The fully connected layer is fully connected to the convolutional feature extraction layer to normalize the extracted features. The output layer is fully connected to the fully connected layer to output standardized feature data. S2.

4. The standardized feature data and node operation status data are encrypted and uploaded to the cloud, while the original monitoring data is stored locally in the edge computing unit.

4. The method according to claim 1, characterized in that, Step S3 includes: S3.

1. Receive standardized feature data and node operation status data, retrieve the basic attribute labels of the corresponding monitoring grid, historical pest and disease occurrence data and real-time monitoring data of adjacent monitoring grids, and construct the input dataset for the node resource dynamic scheduling model; S3.

2. Construct a dynamic scheduling model for node resources. The dynamic scheduling model for node resources includes an input layer, a feature fusion layer, a multi-objective optimization layer, and an output layer. The input of the input layer is the input dataset of the dynamic scheduling model for node resources. The feature fusion layer is fully connected to the input layer and performs feature normalization and correlation fusion processing on the input dataset. The multi-objective optimization layer is fully connected to the feature fusion layer and performs calculations with the dual optimization objectives of maximizing monitoring accuracy and minimizing resource consumption. The output layer is fully connected to the multi-objective optimization layer and outputs the running strategy parameters, which include sampling frequency, data transmission cycle, computing power allocation weight, and early warning threshold adjustment parameters. S3.

3. Differentiate scheduling of IoT sensing nodes and edge computing units in different monitoring grids according to the operation strategy parameters; S3.

4. Perform routine updates of the operation strategy parameters according to the set cycle. When the pest and disease risk level of the monitoring grid changes or the node operation status becomes abnormal, trigger the real-time update of the operation strategy parameters and send the updated operation strategy parameters to the corresponding edge computing unit and IoT sensing node.

5. The method according to claim 1, characterized in that, Step S4 includes: S4.

1. Construct a spatial correlation matrix based on the spatial location relationship and attribute association of all monitoring grids, and map all standardized feature data to the corresponding monitoring grids to form a spatiotemporal distribution dataset of pest and disease risk; S4.

2. Construct a spatial spread prediction model for pests and diseases. The model includes an input layer, a spatiotemporal feature extraction layer, a spread pattern fitting layer, and an output layer. The input layer takes a dataset of the spatiotemporal distribution of pest and disease risk as input. The spatiotemporal feature extraction layer is fully connected to the input layer to extract spatial correlation features and temporal change features from the input dataset. The spread pattern fitting layer is fully connected to the spatiotemporal feature extraction layer to calculate the spread pattern of pests and diseases based on the extracted features. The output layer is fully connected to the spread pattern fitting layer to output the risk distribution of pest and disease spread in the monitored area. S4.

3. Identify monitoring blind spots based on the risk distribution of pest and disease spread, input the identification results of monitoring blind spots into the node resource dynamic scheduling model, and complete the optimization of operation strategy parameters; S4.

4. Based on standardized feature data and the distribution of pest and disease spread risks, trigger local emergency early warnings at the edge and global hierarchical early warnings in the cloud, and generate corresponding prevention and control guidelines.

6. The method according to claim 3, characterized in that, In step S2.3, the parameters of the lightweight feature extraction model are iteratively adjusted using updated data sent from the cloud.

7. The method according to claim 1, characterized in that, In step S4, after the graded early warning is triggered, the corresponding pest and disease verification data and control effect data are collected. Based on the pest and disease verification data and control effect data, the node resource dynamic scheduling model and the pest and disease spatial diffusion prediction model are iteratively optimized, and the optimized model parameters are sent to the corresponding edge computing unit.

8. An Internet of Things-based pest and disease monitoring system, used to perform the method as described in any one of claims 1-7, characterized in that, It includes a grid-based monitoring and sensing module, an edge computing processing module, a cloud-based collaborative scheduling module, a spatial correlation analysis module, and a hierarchical early warning module; the grid-based monitoring and sensing module is used to complete the grid division, attribute labeling, and monitoring data collection of the monitoring area; the edge computing processing module is used to complete data preprocessing, feature extraction, and local early warning triggering; The cloud-based collaborative scheduling module is used to build a dynamic scheduling model for node resources, generate and distribute operational strategy parameters; the spatial correlation analysis module is used to build a spatial correlation matrix and a spatial spread prediction model for pests and diseases, and output the spread risk distribution. The tiered early warning module is used to trigger tiered early warnings and generate prevention and control guidelines.

Images