Remote insect condition forecasting system and method based on deep learning

Abstract

The invention discloses a remote insect condition forecasting system and method based on deep learning. The method comprises the following steps: S1, collecting an original insect condition image through a remote ground sensor; s2, using an image enhancement method to generate a virtual insect situation image similar to the insect situation image, expanding an original insect situation image data set and extracting insect pest density features; s3, processing insect pest situation graph structure data through a convolutional neural network, and learning a propagation rule of insect pests among different regions; s4, gradually updating the state of each insect condition sampling node through multi-layer convolution operation, and generating an insect condition prediction model; s5, adopting an octopus algorithm, and dynamically adjusting hyper-parameters in the insect pest situation prediction model by simulating hunting behaviors of octopuses; s6, generating insect condition early warning information, and sending the insect condition early warning information to the user through remote wireless communication; and S7, dynamically adjusting the pest situation prediction model through remote data transmission and real-time monitoring. According to the method, the octopus algorithm, the convolutional neural network and the image processing and data enhancement technology are combined, and intelligent optimization and dynamic adjustment of remote insect condition forecasting are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of insect monitoring and reporting, and in particular to a remote insect monitoring and reporting system and method based on deep learning. Background Art

[0002] To keep abreast of pest occurrence and development trends and ensure the healthy growth of crops, pest monitoring systems are widely used in agricultural plant protection. Currently, commonly used pest monitoring technologies include manual field observation, image recognition-assisted diagnosis, IoT-based data monitoring systems, and prediction methods combined with machine learning models. Traditional manual surveys offer advantages such as high operability and extensive experience, but they face significant limitations such as high labor intensity, delayed response, and insufficient spatial coverage. With the development of IoT technology, ground-based pest monitoring devices are widely deployed in agricultural areas to collect pest images and environmental data, providing the foundation for automated monitoring.

[0003] In recent years, deep learning methods have been widely used in insect infestation identification and prediction. Convolutional neural networks, long-short-term memory networks, and recurrent neural networks have been used to identify insect pest images, learn time series patterns, and predict infestation trends. However, most of these models focus on image classification of insect infestations or simple trend fitting, lacking the ability to model and predict the spatial propagation mechanisms of insect infestations. Furthermore, many current insect infestation prediction methods are based on static training models, which struggle to adapt to environmental changes and regional differences during the prediction process. Dynamic updates of model parameters are also impossible, resulting in limited generalization capabilities and hindering their practical application in multiple scenarios.

[0004] Some existing research has attempted to incorporate graph neural networks for regional modeling of pest patterns. However, these approaches are computationally complex, rely heavily on the connections between pest nodes, and often fail to fully integrate with spatiotemporal transmission trends. Consequently, they lack a complete closed-loop system encompassing pest image acquisition, graph structure modeling, trend prediction, and intelligent early warning feedback. Regarding optimization strategies, while intelligent optimization methods such as particle swarm optimization and genetic algorithms have been used to adjust some model parameters, most approaches are limited to static tuning during the training phase and fail to achieve continuous model evolution as pest transmission characteristics change dynamically.

[0005] Therefore, how to provide a remote insect monitoring system and method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a remote insect monitoring system and method based on deep learning. The present invention combines image processing and data enhancement technology, convolutional neural network and octopus algorithm to construct a dynamic correlation between the insect map structure and the propagation path, drive the intelligent search of the hyperparameter space and the continuous evolution of the prediction model, integrate user feedback to realize the adaptive update of the model, and fully realize the closed-loop linkage of insect monitoring, propagation prediction, risk warning and model optimization. It has significant advantages such as high prediction accuracy, fast response speed, strong regional adaptability and high degree of structuring of warning information. It is particularly suitable for major agricultural production areas and smart agriculture demonstration areas with high insect infestation, dense crops and the need for precise prevention and control.

[0007] A remote insect monitoring method based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1, collect original insect images through remote ground sensors and perform preprocessing;

[0009] S2. Use image enhancement methods to generate virtual insect images similar to insect images, expand the original insect image dataset and extract insect pest density features;

[0010] S3. Convert the extracted pest density features into pest map structure data, process the pest map structure data through a convolutional neural network, use the convolution operation to propagate information between pest sampling nodes, and learn the propagation patterns of pests between different areas;

[0011] S4: Based on the pest spread patterns between different regions, the state of each pest sampling node is gradually updated through multi-layer convolution operations to generate a pest prediction model and output the predicted pest spread path;

[0012] S5. Using the octopus algorithm, the position of each octopus represents the hyperparameter space of the insect prediction model, and the speed of each octopus represents the adjustment range of the hyperparameter of the insect prediction model. By simulating the hunting behavior of the octopus, the hyperparameters in the insect prediction model are dynamically adjusted;

[0013] S6. Generate insect pest warning information based on the hyperparameters in the insect pest prediction model and the predicted insect pest propagation path, and send the insect pest warning information to the user via remote wireless communication;

[0014] S7. Based on the insect warning information, the insect prediction model is dynamically adjusted through remote data transmission and real-time monitoring.

[0015] Optionally, the original insect situation image specifically includes an insect pest type image, an insect pest density image, and an insect pest damage degree image.

[0016] Optionally, the S2 specifically includes:

[0017] S21, flipping and rotating the original insect image using an image enhancement method, and randomly rotating it by an angle θ, θ∈[-π,π], adjusting pixel values ​​according to the boundaries of the original insect image, to obtain a virtual insect image that is different from the original insect image but has similar features;

[0018] S22, adjusting the color of the virtual insect image, and generating a new virtual insect image by adjusting the RGB channels of the virtual insect image;

[0019] S23, combining the new virtual insect image with the original insect image to generate an expanded insect image dataset;

[0020] S24, each virtual insect image in the expanded insect image dataset shares basic features of insect pests with the original insect image;

[0021] S25. Extract pest density features from the expanded pest image dataset:

[0022]

[0023] Among them, D density is the pest density feature, i and j are the indexes of the pest sample points in the pest image, and p i is the pixel position of the i-th pest sample point, p j is the pixel position of the jth pest sample point, ||p i -p j || is the pixel distance between pest sample point i and pest sample point j, ∈ is a constant to prevent division by zero error, α is a hyperparameter to control distance attenuation, is the weighting factor of pest density, n is the total number of pest sample points, and m is the total number of image sampling points.

[0024] Optionally, the S3 specifically includes:

[0025] S31, converting the extracted pest density features into graph structure data, using each pest sampling point in the pest image as a node in the graph structure data, and defining the spatial relationship between the pest nodes as an edge of the graph structure data;

[0026] S32. Construct an adjacency matrix for the graph structure data. The adjacency matrix defines the relationship between each pest sample point. By calculating the pixel distance between the pest sample points, the pest sample points that are close to each other are connected as edges in the graph.

[0027] S33, using convolutional neural network to process graph structure data, propagating and aggregating the fixed features of pest sample points in graph structure data through convolution operation, and obtaining the updated state of each pest sample point in the neighborhood. The convolution operation is calculated through the adjacency matrix of graph structure data, combined with the pest density feature Ddensity :

[0028]

[0029] Among them, H (l+1) is the updated image sampling point feature matrix of the l+1th layer, i and j are the indexes of the pest sample points in the insect image, A ij is the connection weight between pest sample point i and pest sample point j, X j is the characteristic vector of the j-th pest sample point, W (l) is the convolution kernel of the first layer, D density,j is the pest density feature of the jth pest sample point, n is the total number of pest sample points, m is the total number of image sampling points, ||p i -p j || is the pixel distance between pest sample point i and pest sample point j, ∈ is a constant to prevent division by zero error, β is the distance decay hyperparameter, log2(D density,j ) is the logarithm of the pest density feature of the jth node;

[0030] S34, using convolution operations to propagate information between insect sampling nodes, and gradually propagating the state of each node in the insect graph through multi-layer convolution;

[0031] S35. In the convolution operation, through the convolution layer in the convolutional neural network, the information between the pest sampling nodes is propagated from one pest sample point to the adjacent pest sample points in the graph structure data, and the propagation pattern of pests between different areas is learned.

[0032] Optionally, the S4 specifically includes:

[0033] S41. The state of each pest sampling node is gradually updated through multi-layer convolution operations. The updating process propagates the pest sampling node features in the graph structure data through graph convolution operations.

[0034] S42, gradually updating the state of each insect condition sampling node in the graph structure data through multi-layer convolution operations, and combining the results of each layer of convolution operations to generate a final insect condition graph structure;

[0035] S43, generating an insect situation prediction model according to the state of each insect situation sampling node in the final insect situation graph structure;

[0036] S44. The pest prediction model predicts the spatial distribution and future spread trends of pests. It combines historical data, pest node status, and propagation rules to predict future propagation paths:

[0037]

[0038] Among them, P predis the predicted pest propagation path of the pest prediction model, ∈ is a constant to prevent division by zero error, T k is the insect epidemic state at time node k, a and γ are attenuation factors, is the index of each insect sampling node in the final insect graph structure, k is the index of the time node, T is the number of time spans, is the number of insect sampling nodes in the final insect graph structure, For the The more characteristic matrix of a node at the l+1 layer, ||·|| is the Euclidean distance.

[0039] Optionally, the S5 specifically includes:

[0040] S51, initialize the octopus population, set the population size to 30, and represent each octopus individual as a position vector in the hyperparameter space of the insect prediction model. The position vector includes multiple components, each component corresponds to a value of a hyperparameter, and the dimension of the position vector is equal to the number of hyperparameters to be optimized;

[0041] S52. Assign an initial velocity vector to each individual octopus. Each dimension of the initial velocity vector corresponds to an adjustment amplitude on a hyperparameter dimension. The initial velocity value is randomly generated within a given maximum velocity range of [-0.1, 0.1] to control the search step length of the individual octopus in the search space.

[0042] S53. Calculate the fitness value of each octopus individual based on the error between the predicted insect propagation path output by the insect prediction model and the validation set label. The fitness value is the inverse function of the model prediction error. An error less than 0.02 indicates high fitness. Individuals with high fitness will be preferentially retained during the search process.

[0043] S54. Simulate the hunting behavior of an octopus and use an update strategy with nonlinear dynamic adaptability to update the position vector and velocity vector of the octopus individual:

[0044]

[0045] in, is the velocity of the i-th individual in the j-th dimension of the t+1-th generation, is the dynamic inertia weight, η is the local learning factor, λ is the global guidance factor, t is the current iteration round index, d is the hyperparameter dimension of the insect prediction model, j, is the hyperparameter dimension index, i is the octopus individual index, is the number of octopus individuals i in the tth iteration The search speed of hyperparameter dimensions, For the insect prediction path projection onto the hyperparameter dimension, is the optimal hyperparameter value obtained by octopus individual i in the jth dimension in the historical iteration, is the position of the i-th octopus individual in the j-th hyperparameter dimension in the t-th round of optimization, is the global optimal position of all octopus individuals in the jth dimension, g best is the optimal position vector of the octopus group, ||·|| is the Euclidean distance, ∈ is a constant to prevent division by zero error, P pred The insect spread path predicted by the insect prediction model;

[0046] S55, using the position vector of each generation of octopus individuals as a hyperparameter configuration of the insect infestation prediction model, retraining the insect infestation prediction model and calculating the fitness on the validation set, and updating the optimal value of the current octopus individual and the overall optimal value;

[0047] S56. Repeat the speed vector and position vector update and the insect prediction model fitness evaluation until the set maximum number of iterations of 100 times is reached or the fitness converges to the preset threshold of 0.98, output the optimal combination of hyperparameters, and dynamically optimize the insect prediction model.

[0048] Optionally, the hyperparameters in the insect pest prediction model specifically include the final insect pest graph structure adjacency threshold, convolution kernel size, number of convolution layers, propagation attenuation factor, insect pest density fusion weight, prediction window length and learning rate.

[0049] Optionally, the S6 specifically includes:

[0050] S61. Obtaining the predicted insect spread path output by the insect prediction model and the optimal combination of hyperparameters after optimization of the insect prediction model;

[0051] S62, mapping the predicted locations in the predicted pest transmission path with the geographic information to generate a pest spatial distribution map, which describes the probability and density of pest occurrence in each region at different time periods;

[0052] S63. Identify areas in the pest spatial distribution map where pest density exceeds a set threshold, where the threshold is set to a number of insects greater than or equal to 12 per 100 square centimeters, and extract pest occurrence points with early warning value. Each pest occurrence point with early warning value includes an estimated occurrence time, geographic location, pest type, and severity level.

[0053] S64. Based on the prediction characteristics reflected by the optimization hyperparameters of the pest prediction model and the pest type, corresponding pest control measures are automatically matched to generate pest early warning information;

[0054] S65, converting the insect warning information into a data format recognizable by remote communication, and sending the data to the user terminal via a wireless network;

[0055] S66. Receiving, parsing and visualizing early warning information are realized on the user terminal, and the expected time, location, type and prevention and control suggestions of pests are provided.

[0056] Optionally, the S7 specifically includes:

[0057] S71. Receive pest warning information and transmit it back to the monitoring platform via a remote data transmission channel. The pest warning information includes the actual occurrence time, spatial location, type, level of the pest, and the user's response result.

[0058] S72. Combine the latest pest images continuously collected by remote ground sensors to construct a real-time input data stream and compare it with the predicted pest transmission path. Identify time periods and geographical areas where the error lasts for more than 6 hours and output the error analysis results.

[0059] S73. Calculate an insect infestation prediction error map based on the error analysis results, where the insect infestation prediction error map is indexed by time and spatial location;

[0060] S74, dynamically constructing a feedback weight map based on the areas with significant deviations in the insect prediction error map, wherein each position value in the feedback weight map is adaptively set according to the error analysis results, and the insect prediction model for the corresponding area is adjusted;

[0061] S75. Introducing the feedback weight map as an additional update factor into the insect prediction model to perform weighted adjustment on the insect sampling node status at each layer;

[0062] S76. Redeploy the insect infestation prediction model after feedback correction to the remote end, while maintaining continuous reception and error analysis of real-time data, to form a closed-loop mechanism for remote insect infestation monitoring and dynamic updating of the insect infestation prediction model.

[0063] According to an embodiment of the present invention, a system for suppressing interference in medium and short wave signal transmission based on wireless broadcast television includes the following modules:

[0064] The insect image acquisition module is used to collect images of insect pest types, insect pest density, and insect pest damage extent through remote ground sensors;

[0065] The image enhancement and expansion module is used to perform geometric transformation and color perturbation on the insect image dataset to generate virtual insect images and fuse them with the original insect images to generate an expanded insect image dataset;

[0066] Graph structure data module, used to convert the extracted pest density features into graph structure data;

[0067] An intelligent optimization module, which uses the Octopus algorithm to perform global search and optimization of hyperparameters in the insect prediction model;

[0068] The pest early warning module is used to generate pest early warning information based on the prediction characteristics reflected by the pest type and the pest prediction model hyperparameter optimization;

[0069] The dynamic feedback update module is used to receive remote data transmission and real-time monitoring, identify error areas, and realize dynamic adjustment and continuous optimization of the insect prediction model.

[0070] The beneficial effects of the present invention are:

[0071] The present invention's remote insect monitoring and forecasting system and method based on deep learning overcomes prominent problems in existing insect monitoring and forecasting technologies, such as imperfect modeling, poor dynamic response capabilities, and lack of intelligent feedback mechanisms, and has significant beneficial technical effects. First, the present invention constructs a complete closed-loop system from remote sensing data acquisition, image enhancement, graph structure modeling, spatial propagation prediction, to model optimization and intelligent early warning. Compared with traditional prediction methods that rely on static image classification or time series fitting, it more accurately captures the propagation paths and development trends of insect pests in different regions, significantly improving the spatial resolution and temporal foresight of insect prediction.

[0072] By using convolutional neural networks to learn features and model pest spread patterns from the structured data of pest maps, the system effectively extracts the impact relationships between regional nodes and achieves a high-level understanding and modeling of pest spread behavior. Furthermore, the system uses image enhancement to generate an expanded dataset of virtual pest images, improving the model's robustness in data-scarce scenarios, enhancing the comprehensiveness of feature extraction, and effectively improving the generalization capabilities of the prediction model.

[0073] The present invention innovatively introduces the octopus algorithm to perform global search and dynamic optimization of the hyperparameter space of the prediction model. By simulating the hunting behavior of individual octopuses and combining the model prediction output results, the model parameter configuration is automatically adjusted. This not only solves the problem of traditional hyperparameter tuning relying on manual experience, but also achieves a deep coupling of model performance and insect spread trends, thereby improving the stability and practicality of the prediction accuracy. The early warning information generated based on the insect prediction results is highly structured, including key content such as time, location, type, level and prevention and control recommendations. It can be accurately pushed to the user terminal, providing agricultural managers with timely, clear and executable prevention and control guidance.

[0074] Furthermore, the system supports online adjustment and dynamic updating of the insect infestation prediction model. By introducing a feedback mechanism, user responses and real-time monitoring data are compared with the model's predictions. Error graphs and feedback weight graphs are constructed based on deviations, and the model training process is then corrected to achieve continuous improvement in prediction accuracy. This "data-model-feedback" closed-loop mechanism ensures that the system maintains excellent predictive capabilities and adaptability in agricultural scenarios with frequent insect infestations and complex environmental dynamics, significantly enhancing the system's intelligence, stability, and practicality.

[0075] In summary, the present invention realizes a fundamental transformation of insect infestation prediction from "passive identification" to "active modeling and dynamic adjustment". By integrating multi-source information perception, graph structure deep modeling, intelligent algorithm optimization and real-time feedback correction, a highly intelligent and adaptable insect infestation monitoring system is constructed, which significantly improves the automation level and scientific decision-making of agricultural plant protection work, and has outstanding beneficial technical effects and wide promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 This is a flow chart of the remote insect monitoring method based on deep learning proposed by the present invention;

[0078] Figure 2 This is a schematic diagram of the remote insect monitoring method based on deep learning proposed in the present invention;

[0079] Figure 3 This is a structural diagram of the deep learning-based remote insect monitoring system proposed in the present invention. DETAILED DESCRIPTION

[0080] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0081] refer to Figure 1-2 The remote insect monitoring method based on deep learning includes the following steps:

[0082] S1, collect original insect images through remote ground sensors and perform preprocessing;

[0083] S2. Use image enhancement methods to generate virtual insect images similar to insect images, expand the original insect image dataset and extract insect pest density features;

[0084] S3. Convert the extracted pest density features into pest map structure data, process the pest map structure data through a convolutional neural network, use the convolution operation to propagate information between pest sampling nodes, and learn the propagation patterns of pests between different areas;

[0085] S4: Based on the pest spread patterns between different regions, the state of each pest sampling node is gradually updated through multi-layer convolution operations to generate a pest prediction model and output the predicted pest spread path;

[0086] S5. Using the octopus algorithm, the position of each octopus represents the hyperparameter space of the insect prediction model, and the speed of each octopus represents the adjustment range of the hyperparameter of the insect prediction model. By simulating the hunting behavior of the octopus, the hyperparameters in the insect prediction model are dynamically adjusted;

[0087] S6. Generate insect pest warning information based on the hyperparameters in the insect pest prediction model and the predicted insect pest propagation path, and send the insect pest warning information to the user via remote wireless communication;

[0088] S7. Based on the insect warning information, the insect prediction model is dynamically adjusted through remote data transmission and real-time monitoring.

[0089] This method uses insect-infestation graph structure modeling and a convolutional propagation mechanism to extract spatial propagation characteristics between infestation areas. It also introduces the Octopus algorithm to simulate hunting behavior in a hyperparameter space, enabling dynamic optimization and adjustment of the insect-infestation prediction model. Image enhancement expands sample diversity and improves the accuracy of insect density feature extraction. By establishing a prediction-optimization-feedback closed loop, the insect-infestation prediction path is dynamically coupled with the model configuration, enhancing prediction accuracy and early warning response capabilities.

[0090] In this embodiment, the original insect situation image specifically includes an insect pest type image, an insect pest density image, and an insect pest damage degree image.

[0091] This method collects images of pest species, density, and damage severity to construct a multidimensional pest feature input space, enabling comprehensive perception of pest status. This method integrates pest type identification, density analysis, and damage severity assessment to enhance the hierarchical and accurate nature of pest modeling, providing a richer and more accurate feature foundation for subsequent pest map construction and propagation path prediction.

[0092] In this embodiment, S2 specifically includes:

[0093] S21, flipping and rotating the original insect image using an image enhancement method, and randomly rotating it by an angle θ, θ∈[-π,π], adjusting pixel values ​​according to the boundaries of the original insect image, to obtain a virtual insect image that is different from the original insect image but has similar features;

[0094] S22, adjusting the color of the virtual insect image, and generating a new virtual insect image by adjusting the RGB channels of the virtual insect image;

[0095] S23, combining the new virtual insect image with the original insect image to generate an expanded insect image dataset;

[0096] S24, each virtual insect image in the expanded insect image dataset shares basic features of insect pests with the original insect image;

[0097] S25. Extract pest density features from the expanded pest image dataset:

[0098]

[0099] Among them, D density is the pest density feature, i and j are the indexes of the pest sample points in the pest image, and p i is the pixel position of the i-th pest sample point, p j is the pixel position of the jth pest sample point, ||p i -p j || is the pixel distance between pest sample point i and pest sample point j, ∈ is a constant to prevent division by zero error, α is a hyperparameter to control distance attenuation, is the weighting factor of pest density, n is the total number of pest sample points, and m is the total number of image sampling points.

[0100] This method generates virtual pest images by rotating, flipping, and color-perturbing original pest images, constructing a diverse enhanced dataset and extracting pest density features, expanding the sample space and refining feature representation. By modeling pest density distribution using pixel-level positional relationships and weighting factors, the method improves sensitivity and recognition accuracy for high-density pest areas, providing high-quality input for graph structure modeling and propagation law learning, and enhancing the model's adaptability and generalization capabilities to diverse pest scenarios.

[0101] In this embodiment, S3 specifically includes:

[0102] S31, converting the extracted pest density features into graph structure data, using each pest sampling point in the pest image as a node in the graph structure data, and defining the spatial relationship between the pest nodes as an edge of the graph structure data;

[0103] S32. Construct an adjacency matrix for the graph structure data. The adjacency matrix defines the relationship between each pest sample point. By calculating the pixel distance between the pest sample points, the pest sample points that are close to each other are connected as edges in the graph.

[0104] S33, using convolutional neural network to process graph structure data, propagating and aggregating the fixed features of pest sample points in graph structure data through convolution operation, and obtaining the updated state of each pest sample point in the neighborhood. The convolution operation is calculated through the adjacency matrix of graph structure data, combined with the pest density feature D density :

[0105]

[0106] Among them, H (l+1) is the updated image sampling point feature matrix of the l+1th layer, i and j are the indexes of the pest sample points in the insect image, A ij is the connection weight between pest sample point i and pest sample point j, X j is the characteristic vector of the j-th pest sample point, W (l) is the convolution kernel of the first layer, D density,j is the pest density feature of the jth pest sample point, n is the total number of pest sample points, m is the total number of image sampling points, ||p i -p j || is the pixel distance between pest sample point i and pest sample point j, ∈ is a constant to prevent division by zero error, β is the distance decay hyperparameter, log2(D density,j ) is the logarithm of the pest density feature of the jth node;

[0107] S34, using convolution operations to propagate information between insect sampling nodes, and gradually propagating the state of each node in the insect graph through multi-layer convolution;

[0108] S35. In the convolution operation, through the convolution layer in the convolutional neural network, the information between the pest sampling nodes is propagated from one pest sample point to the adjacent pest sample points in the graph structure data, and the propagation pattern of pests between different areas is learned.

[0109] This method converts pest density features into graph-structured data, constructs spatial connections between nodes in the pest map, and generates an adjacency matrix by calculating pixel distances to model the topological structure of pest-infested regions. A convolutional neural network is used to perform multi-layer convolution propagation and aggregation on the graph data, gradually updating the status of pest nodes. This effectively learns the spread paths and intensity variations of pests across regions, improving the model's ability to depict pest spread trends and enhancing the spatial perception and regional adaptability of pest predictions.

[0110] In this embodiment, the S4 specifically includes:

[0111] S41. The state of each pest sampling node is gradually updated through multi-layer convolution operations. The update process propagates the pest sampling node features in the graph structure data through graph convolution operations.

[0112] S42, gradually updating the state of each insect condition sampling node in the graph structure data through multi-layer convolution operations, and combining the results of each layer of convolution operations to generate a final insect condition graph structure;

[0113] S43, generating an insect situation prediction model according to the state of each insect situation sampling node in the final insect situation graph structure;

[0114] S44. The pest prediction model predicts the spatial distribution and future spread trends of pests. It combines historical data, pest node status, and propagation rules to predict future propagation paths:

[0115]

[0116] Among them, P pred is the predicted pest propagation path of the pest prediction model, ∈ is a constant to prevent division by zero error, T k is the insect epidemic state at time node k, α and γ are attenuation factors, is the index of each insect sampling node in the final insect graph structure, k is the index of the time node, T is the number of time spans, is the number of insect sampling nodes in the final insect graph structure, For the The more characteristic matrix of a node at the l+1 layer, ||·|| is the Euclidean distance.

[0117] This method iteratively updates the status of pest sampling nodes through multi-layer graph convolution operations, integrating pest characteristics with spatial connectivity to construct a dynamically evolving pest graph structure. Combining historical pest data with node propagation characteristics, it generates a pest prediction model that accurately simulates the spatial spread paths and future spread trends of pests. By introducing a time dimension and propagation attenuation mechanisms, the model enables high-precision pest trend predictions across multiple time periods, significantly improving the system's forward-looking ability to assess regional pest evolution and its timely response.

[0118] In this embodiment, the S5 specifically includes:

[0119] S51, initialize the octopus population, set the population size to 30, and represent each octopus individual as a position vector in the hyperparameter space of the insect prediction model. The position vector includes multiple components, each component corresponds to a value of a hyperparameter, and the dimension of the position vector is equal to the number of hyperparameters to be optimized;

[0120] S52. Assign an initial velocity vector to each individual octopus. Each dimension of the initial velocity vector corresponds to an adjustment amplitude on a hyperparameter dimension. The initial velocity value is randomly generated within a given maximum velocity range of [-0.1, 0.1] to control the search step length of the individual octopus in the search space.

[0121] S53. Calculate the fitness value of each octopus individual based on the error between the predicted insect propagation path output by the insect prediction model and the validation set label. The fitness value is the inverse function of the model prediction error. An error less than 0.02 indicates high fitness. Individuals with high fitness will be preferentially retained during the search process.

[0122] S54. Simulate the hunting behavior of an octopus and use an update strategy with nonlinear dynamic adaptability to update the position vector and velocity vector of the octopus individual:

[0123]

[0124] in, is the velocity of the i-th individual in the j-th dimension of the t+1-th generation, is the dynamic inertia weight, η is the local learning factor, λ is the global guidance factor, t is the current iteration round index, d is the hyperparameter dimension of the insect prediction model, j, is the hyperparameter dimension index, i is the octopus individual index, is the number of octopus individuals i in the tth iteration The search speed of hyperparameter dimensions, For the insect prediction path projection onto the hyperparameter dimension, is the optimal hyperparameter value obtained by octopus individual i in the jth dimension in the historical iteration, is the position of the i-th octopus individual in the j-th hyperparameter dimension in the t-th round of optimization, is the global optimal position of all octopus individuals in the jth dimension, g best is the optimal position vector of the octopus group, ||·|| is the Euclidean distance, ∈ is a constant to prevent division by zero error, P pred The insect spread path predicted by the insect prediction model;

[0125] S55, using the position vector of each generation of octopus individuals as a hyperparameter configuration of the insect infestation prediction model, retraining the insect infestation prediction model and calculating the fitness on the validation set, and updating the optimal value of the current octopus individual and the overall optimal value;

[0126] S56. Repeat the speed vector and position vector update and the insect prediction model fitness evaluation until the set maximum number of iterations of 100 times is reached or the fitness converges to the preset threshold of 0.98, output the optimal combination of hyperparameters, and dynamically optimize the insect prediction model.

[0127] This paper uses the octopus algorithm to simulate swarm hunting behavior and conducts global search optimization within the multidimensional hyperparameter space of the insect infestation prediction model. By constructing a speed-position update mechanism and fitness evaluation criteria, individuals are guided to approach the solution domain with the lowest prediction error, improving the model's generalization ability. Dynamic hyperparameter configuration is adjusted using propagation path feedback, enabling intelligent optimization of parameters such as the learning rate, convolutional structure, and propagation attenuation factor. This results in a prediction model with higher accuracy and robustness across different insect infestation scenarios.

[0128] In this embodiment, the hyperparameters in the insect pest prediction model specifically include the final insect pest graph structure adjacency threshold, convolution kernel size, number of convolution layers, propagation attenuation factor, insect pest density fusion weight, prediction window length and learning rate.

[0129] The present invention constructs a high-dimensional hyperparameter optimization space by setting multiple key hyperparameters in the pest prediction model, including the graph structure adjacency threshold, convolution kernel size and number of layers, propagation attenuation factor, density fusion weight, prediction window length, and learning rate. These parameters work synergistically to influence the model's modeling depth and response sensitivity to pest spread. By dynamically optimizing these parameter combinations, the model possesses enhanced feature extraction and propagation trend matching capabilities in multiple pest scenarios, improving the accuracy and stability of pest prediction.

[0130] In this embodiment, S6 specifically includes:

[0131] S61. Obtaining the predicted insect spread path output by the insect prediction model and the optimal combination of hyperparameters after optimization of the insect prediction model;

[0132] S62, mapping the predicted locations in the predicted pest transmission path with the geographic information to generate a pest spatial distribution map, which describes the probability and density of pest occurrence in each region at different time periods;

[0133] S63. Identify areas in the pest spatial distribution map where pest density exceeds a set threshold, where the threshold is set to a number of insects greater than or equal to 12 per 100 square centimeters, and extract pest occurrence points with early warning value. Each pest occurrence point with early warning value includes an estimated occurrence time, geographic location, pest type, and severity level.

[0134] S64. Based on the prediction characteristics reflected by the optimization hyperparameters of the pest prediction model and the pest type, corresponding pest control measures are automatically matched to generate pest early warning information;

[0135] S65, converting the insect warning information into a data format recognizable by remote communication, and sending the data to the user terminal via a wireless network;

[0136] S66. Receiving, parsing and visualizing early warning information are realized on the user terminal, and the expected time, location, type and prevention and control suggestions of pests are provided.

[0137] This method constructs a spatial distribution map of pest infestations based on the propagation paths and optimal hyperparameter combinations output by the pest prediction model, accurately locating areas of high pest density. By setting density thresholds to filter warning points and automatically matching prevention and control recommendations based on pest type and model characteristics, it enables the intelligent generation and remote delivery of pest warnings. This method improves the relevance and timeliness of warning responses, ensuring that users have timely access to the time, location, and prevention and control plans for pest infestations, enhancing the intelligence and practicality of pest management.

[0138] In this embodiment, the S7 specifically includes:

[0139] S71. Receive pest warning information and transmit it back to the monitoring platform via a remote data transmission channel. The pest warning information includes the actual occurrence time, spatial location, type, level of the pest, and the user's response result.

[0140] S72. Combine the latest pest images continuously collected by remote ground sensors to construct a real-time input data stream and compare it with the predicted pest transmission path. Identify time periods and geographical areas where the error lasts for more than 6 hours and output the error analysis results.

[0141] S73. Calculate an insect infestation prediction error map based on the error analysis results, where the insect infestation prediction error map is indexed by time and spatial location;

[0142] S74, dynamically constructing a feedback weight map based on the areas with significant deviations in the insect prediction error map, wherein each position value in the feedback weight map is adaptively set according to the error analysis results, and the insect prediction model for the corresponding area is adjusted;

[0143] S75. Introducing the feedback weight map as an additional update factor into the insect prediction model to perform weighted adjustment on the insect sampling node status at each layer;

[0144] S76. Redeploy the insect infestation prediction model after feedback correction to the remote end, while maintaining continuous reception and error analysis of real-time data, to form a closed-loop mechanism for remote insect infestation monitoring and dynamic updating of the insect infestation prediction model.

[0145] The present invention receives user-side insect warning feedback and real-time monitoring images, constructs an insect prediction error map, identifies deviation areas, and dynamically generates a feedback weight map to introduce a model update process. Through an error-driven model fine-tuning mechanism, weighted correction and continuous optimization of the predicted node status are achieved. This method constructs a closed-loop "warning-feedback-update" system, significantly improving the model's adaptability and prediction accuracy under actual insect changes, ensuring the continuity of insect monitoring and the efficiency of intelligent response.

[0146] refer to Figure 3 The remote insect monitoring system based on deep learning includes the following modules:

[0147] The insect image acquisition module is used to collect images of insect pest types, insect pest density, and insect pest damage extent through remote ground sensors;

[0148] The image enhancement and expansion module is used to perform geometric transformation and color perturbation on the insect image dataset to generate virtual insect images and fuse them with the original insect images to generate an expanded insect image dataset;

[0149] Graph structure data module, used to convert the extracted pest density features into graph structure data;

[0150] An intelligent optimization module, which uses the Octopus algorithm to perform global search and optimization of hyperparameters in the insect prediction model;

[0151] The pest early warning module is used to generate pest early warning information based on the prediction characteristics reflected by the pest type and the pest prediction model hyperparameter optimization;

[0152] The dynamic feedback update module is used to receive remote data transmission and real-time monitoring, identify error areas, and realize dynamic adjustment and continuous optimization of the insect prediction model.

[0153] This paper constructs a modular system that integrates pest image acquisition, enhancement and expansion, graph structure modeling, intelligent optimization, warning generation, and dynamic feedback. By integrating the multidimensional features of pest images with virtual image expansion, the model's input diversity is enhanced. The Octopus algorithm performs a global hyperparameter search to enhance the robustness of the prediction model. High-precision warning information is generated by combining prediction characteristics with pest type. Real-time error feedback is used to continuously optimize model parameters, achieving an intelligent closed-loop system for pest monitoring, prediction, and response, effectively improving the system's practicality and prediction accuracy.

[0154] Example 1:

[0155] To verify the feasibility of this invention, we applied it to the monitoring and prediction of insect infestations in the main summer corn-producing area of ​​a high-standard farmland smart agriculture demonstration zone. In recent years, major pests such as armyworms, corn borers, and Spodoptera litura have frequently appeared in the region. These outbreaks are characterized by sudden onset, wide migratory ranges, dense distribution across fields, and high spatiotemporal heterogeneity. Traditional methods, which rely on manual inspections and fixed-site early warning systems, struggle to accurately and timely monitor pest developments. This often leads to missed, false, and delayed reports, delayed control measures, and reduced or even complete crop failures.

[0156] During the demonstration deployment, the system initially deployed 16 remote ground-based insect monitoring sensors along the boundaries of different plots and in the fields. These sensors periodically collected images of insect traps and microclimate data, including temperature, humidity, wind speed, and rainfall. Data was collected hourly, with 24 sets of images and environmental data uploaded daily, covering the main cornworm infestation period.

[0157] The system uses an image enhancement module to construct an extended image dataset based on the collected insect infestation image data. The original dataset consists of approximately 11,500 images, which have been expanded to 23,000 images through simulation enhancement. Enhancement methods include biomimetic contrast adjustment, local blur simulation, illumination perturbation simulation, and geometric transformations, ensuring data diversity and closeness to reality. The system utilizes a convolutional neural network model to extract key features from the images, such as pest type, density distribution, and spatial location, and constructs an insect infestation graph structure. By performing multi-layer convolution operations on the graph structure, the system identifies patterns in insect infestation transmission and, combined with historical data, completes the initial training of the insect infestation prediction model.

[0158] The Octopus algorithm was introduced to optimize the hyperparameter space of the insect infestation prediction model. In this application, the optimization dimensions included six: the model learning rate, convolution kernel size, density fusion coefficient, propagation attenuation parameter, and time window length. The population size was set to 30, and the maximum number of iterations was 100. The system simulated the search behavior of individual octopuses in the parameter space, using the minimum prediction error as the fitness function. Ultimately, the optimal hyperparameter combination was obtained and applied to the insect infestation propagation path prediction model.

[0159] The generated pest prediction paths are combined with GIS maps to create a heat map of pest spatial distribution and peak pest time periods. The system automatically determines the risk level and generates structured pest warning information, including pest type, time, location, level, and recommended control measures. This information is then pushed to managers' mobile phones via the cloud platform. The entire system, from image acquisition to warning delivery, maintains a latency of less than 30 minutes, ensuring high timeliness.

[0160] Table 1 Optimization effect of remote insect monitoring method based on deep learning

[0161]

[0162] Table 1 shows that during the entire application cycle, the present invention achieved a 91.8% accuracy rate in predicting high-density insect outbreaks, an improvement of approximately 27% compared to manual inspections. The warning response time was shortened by an average of 18 hours, effectively reducing the spread of insect infestations. According to farmer feedback, the timely coverage rate of prevention and control increased by 32%, and the insect population base compression rate exceeded 60%.

[0163] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A remote insect monitoring method based on deep learning, characterized by: The steps include: S1, collect original insect images through remote ground sensors and perform preprocessing; S2. Use image enhancement methods to generate virtual insect images similar to insect images, expand the original insect image dataset and extract insect pest density features; S3. Convert the extracted pest density features into pest map structure data, process the pest map structure data through a convolutional neural network, use the convolution operation to propagate information between pest sampling nodes, and learn the propagation patterns of pests between different areas; S4: Based on the pest spread patterns between different regions, the state of each pest sampling node is gradually updated through multi-layer convolution operations to generate a pest prediction model and output the predicted pest spread path; S5. Using the octopus algorithm, the position of each octopus represents the hyperparameter space of the insect prediction model, and the speed of each octopus represents the adjustment range of the hyperparameter of the insect prediction model. By simulating the hunting behavior of the octopus, the hyperparameters in the insect prediction model are dynamically adjusted; S6. Generate insect pest warning information based on the hyperparameters in the insect pest prediction model and the predicted insect pest propagation path, and send the insect pest warning information to the user via remote wireless communication; S7. Based on the insect warning information, the insect prediction model is dynamically adjusted through remote data transmission and real-time monitoring.

2. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The original insect situation image specifically includes an insect pest type image, an insect pest density image, and an insect pest damage degree image.

3. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The S2 specifically includes: S21, flipping and rotating the original insect image using an image enhancement method, and randomly rotating it by an angle θ, θ∈[-π,π], adjusting pixel values ​​according to the boundaries of the original insect image, to obtain a virtual insect image that is different from the original insect image but has similar features; S22, adjusting the color of the virtual insect image, and generating a new virtual insect image by adjusting the RGB channels of the virtual insect image; S23, combining the new virtual insect image with the original insect image to generate an expanded insect image dataset; S24, each virtual insect image in the expanded insect image dataset shares basic features of insect pests with the original insect image; S25. Extract pest density features from the expanded pest image dataset: Among them, D density is the pest density feature, i and j are the indexes of the pest sample points in the pest image, and p i is the pixel position of the i-th pest sample point, p j is the pixel position of the jth pest sample point, ||p i -p j || is the pixel distance between pest sample point i and pest sample point j, ∈ is a constant to prevent division by zero error, α is a hyperparameter to control distance attenuation, is the weighting factor of pest density, n is the total number of pest sample points, and m is the total number of image sampling points.

4. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The S3 specifically includes: S31, converting the extracted pest density features into graph structure data, using each pest sampling point in the pest image as a node in the graph structure data, and defining the spatial relationship between the pest nodes as an edge of the graph structure data; S32. Construct an adjacency matrix for the graph structure data. The adjacency matrix defines the relationship between each pest sample point. By calculating the pixel distance between the pest sample points, the pest sample points that are close to each other are connected as edges in the graph. S33, using convolutional neural network to process graph structure data, propagating and aggregating the fixed features of pest sample points in graph structure data through convolution operation, and obtaining the updated state of each pest sample point in the neighborhood. The convolution operation is calculated through the adjacency matrix of graph structure data, combined with the pest density feature D density : Among them, H (l+1) is the updated image sampling point feature matrix of the l+1th layer, i and j are the indexes of the pest sample points in the insect image, A ij is the connection weight between pest sample point i and pest sample point j, X j is the characteristic vector of the jth pest sample point, W (l) is the convolution kernel of the first layer, D density,j is the pest density feature of the jth pest sample point, n is the total number of pest sample points, m is the total number of image sampling points, ||p i -p j || is the pixel distance between pest sample point i and pest sample point j, ∈ is a constant to prevent division by zero error, β is the distance decay hyperparameter, log2(D density,j ) is the logarithm of the pest density feature of the jth node; S34, using convolution operations to propagate information between insect sampling nodes, and gradually propagating the state of each node in the insect graph through multi-layer convolution; S35. In the convolution operation, through the convolution layer in the convolutional neural network, the information between the pest sampling nodes is propagated from one pest sample point to the adjacent pest sample points in the graph structure data, and the propagation pattern of pests between different areas is learned.

5. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The S4 specifically includes: S41. The state of each pest sampling node is gradually updated through multi-layer convolution operations. The update process propagates the pest sampling node features in the graph structure data through graph convolution operations. S42, gradually updating the state of each insect condition sampling node in the graph structure data through multi-layer convolution operations, and combining the results of each layer of convolution operations to generate a final insect condition graph structure; S43, generating an insect situation prediction model according to the state of each insect situation sampling node in the final insect situation graph structure; S44. The pest prediction model predicts the spatial distribution and future spread trends of pests. It combines historical data, pest node status, and propagation rules to predict future propagation paths: Among them, P pred is the predicted pest propagation path of the pest prediction model, ∈ is a constant to prevent division by zero error, T k is the insect epidemic state at time node k, a and γ are attenuation factors, is the index of each insect sampling node in the final insect graph structure, k is the index of the time node, T is the number of time spans, is the number of insect sampling nodes in the final insect graph structure, is the more characteristic matrix of the i-th node at the l+1 layer, and ||·|| is the Euclidean distance.

6. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The S5 specifically includes: S51, initialize the octopus population, set the population size to 30, and represent each octopus individual as a position vector in the hyperparameter space of the insect prediction model. The position vector includes multiple components, each component corresponds to a value of a hyperparameter, and the dimension of the position vector is equal to the number of hyperparameters to be optimized; S52. Assign an initial velocity vector to each individual octopus. Each dimension of the initial velocity vector corresponds to an adjustment amplitude on a hyperparameter dimension. The initial velocity value is randomly generated within a given maximum velocity range of [-0.1, 0.1] to control the search step length of the individual octopus in the search space. S53. Calculate the fitness value of each octopus individual based on the error between the predicted insect propagation path output by the insect prediction model and the validation set label. The fitness value is the inverse function of the model prediction error. An error less than 0.02 indicates high fitness. Individuals with high fitness will be preferentially retained during the search process. S54. Simulate the hunting behavior of an octopus and use an update strategy with nonlinear dynamic adaptability to update the position vector and velocity vector of the octopus individual: in, is the velocity of the i-th individual in the t+1th generation in the i-th dimension, is the dynamic inertia weight, η is the local learning factor, λ is the global guidance factor, t is the current iteration round index, d is the hyperparameter dimension of the insect prediction model, j, is the hyperparameter dimension index, i is the octopus individual index, is the number of octopus individuals i in the tth iteration The search speed of hyperparameter dimensions, For the insect prediction path projection onto the hyperparameter dimension, is the optimal hyperparameter value obtained by octopus individual i in the i-th dimension in the historical iteration, is the position of the i-th octopus individual in the i-th hyperparameter dimension in the t-th round of optimization, is the global optimal position of all octopus individuals in the i-th dimension, g best is the optimal position vector of the octopus group, ||·|| is the Euclidean distance, ∈ is a constant to prevent division by zero error, P pred The insect spread path predicted by the insect prediction model; S55, using the position vector of each generation of octopus individuals as a hyperparameter configuration of the insect infestation prediction model, retraining the insect infestation prediction model and calculating the fitness on the validation set, and updating the optimal value of the current octopus individual and the overall optimal value; S56. Repeat the speed vector and position vector update and the insect prediction model fitness evaluation until the set maximum number of iterations of 100 times is reached or the fitness converges to the preset threshold of 0.98, output the optimal combination of hyperparameters, and dynamically optimize the insect prediction model.

7. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The hyperparameters in the insect pest prediction model specifically include the final insect pest graph structure adjacency threshold, convolution kernel size, number of convolution layers, propagation attenuation factor, insect pest density fusion weight, prediction window length and learning rate.

8. The remote insect monitoring method based on deep learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Obtaining the predicted insect spread path output by the insect prediction model and the optimal combination of hyperparameters after optimization of the insect prediction model; S62, mapping the predicted locations in the predicted pest transmission path with the geographic information to generate a pest spatial distribution map, which describes the probability and density of pest occurrence in each region at different time periods; S63. Identify areas in the pest spatial distribution map where pest density exceeds a set threshold, where the threshold is set to a number of insects greater than or equal to 12 per 100 square centimeters, and extract pest occurrence points with early warning value. Each pest occurrence point with early warning value includes an estimated occurrence time, geographic location, pest type, and severity level. S64. Based on the prediction characteristics reflected by the optimization hyperparameters of the pest prediction model and the pest type, corresponding pest control measures are automatically matched to generate pest early warning information; S65, converting the insect warning information into a data format recognizable by remote communication, and sending the data to the user terminal via a wireless network; S66. Receiving, parsing and visualizing early warning information are realized on the user terminal, and the expected time, location, type and prevention and control suggestions of pests are provided.

9. The remote insect monitoring method based on deep learning according to claim 1, characterized in that: The S7 specifically includes: S71. Receive pest warning information and transmit it back to the monitoring platform via a remote data transmission channel. The pest warning information includes the actual occurrence time, spatial location, type, level of the pest, and the user's response result. S72. Combine the latest pest images continuously collected by remote ground sensors to construct a real-time input data stream and compare it with the predicted pest transmission path. Identify time periods and geographical areas where the error lasts for more than 6 hours and output the error analysis results. S73. Calculate an insect infestation prediction error map based on the error analysis results, where the insect infestation prediction error map is indexed by time and spatial location; S74, dynamically constructing a feedback weight map based on the areas with significant deviations in the insect prediction error map, wherein each position value in the feedback weight map is adaptively set according to the error analysis results, and the insect prediction model for the corresponding area is adjusted; S75. Introducing the feedback weight map as an additional update factor into the insect prediction model to perform weighted adjustment on the insect sampling node status at each layer; S76. Redeploy the insect infestation prediction model after feedback correction to the remote end, while maintaining continuous reception and error analysis of real-time data, to form a closed-loop mechanism for remote insect infestation monitoring and dynamic updating of the insect infestation prediction model.

10. A remote insect monitoring system based on deep learning, which implements the remote insect monitoring method based on deep learning according to any one of claims 1 to 9, characterized in that: Includes the following modules: The insect image acquisition module is used to collect images of insect pest types, insect pest density, and insect pest damage extent through remote ground sensors; The image enhancement and expansion module is used to perform geometric transformation and color perturbation on the insect image dataset to generate virtual insect images and fuse them with the original insect images to generate an expanded insect image dataset; Graph structure data module, used to convert the extracted pest density features into graph structure data; An intelligent optimization module, which uses the Octopus algorithm to perform global search and optimization of hyperparameters in the insect prediction model; The pest early warning module is used to generate pest early warning information based on the prediction characteristics reflected by the pest type and the pest prediction model hyperparameter optimization; The dynamic feedback update module is used to receive remote data transmission and real-time monitoring, identify error areas, and realize dynamic adjustment and continuous optimization of the insect prediction model.

Images