Agricultural disease and pest intelligent monitoring system and method fusing big data and generative adversarial big model

Abstract

The invention relates to the technical field of agriculture, in particular to an agricultural disease and pest intelligent monitoring system and method fusing big data and a generative adversarial big model, which can effectively identify and early warn crop diseases and pests. The system comprises a data acquisition unit; the data processing unit comprises a data cleaning module for performing data denoising, enhancement and correction on the related information of the crops; the data labeling module is used for labeling disease and insect pest images; the feature engineering module is used for extracting pest image features and crop environment features; the model training unit comprises a data sample enhancement module which generates a pest and disease simulation image based on a real pest and disease image according to a generative adversarial large model optimization mode, and improves the training stability and visual similarity according to the difference between the two images. The multi-mode identification module is used for fusing the pest image features and the environment features to generate pest identification feature vectors for identifying pests; and the intelligent application unit identifies disease and pest conditions according to the disease and pest identification feature vectors and gives out early warning.

Description

TECHNICAL FIELD

[0001] One or more embodiments of the present application relate to the field of agricultural technology, and in particular to an agricultural pest and disease intelligent monitoring system and method fusing big data and a generative adversarial large model. BACKGROUND

[0002] In the prior art, crop pest and disease monitoring mostly relies on manual patrol, which not only has weak recognition ability but also has low monitoring efficiency, mainly including the following shortcomings:

[0003] 1) Limitations of data collection: Traditional agricultural pest and disease monitoring relies on manual patrol, which is low in efficiency and limited in coverage. Even when sensors are used, the single sensor (such as a camera) lacks sufficient data collection dimensions and lacks multi-modal data fusion capabilities.

[0004] 2) Small sample learning problem: New pest and disease samples are scarce, and traditional machine learning models have significantly reduced recognition accuracy when samples are insufficient, requiring a large amount of labeled data support.

[0005] 3) Real-time and accuracy contradiction: Existing image recognition models (such as convolutional neural networks CNN) have weak generalization ability in complex farmland environments (such as light changes and leaf shielding conditions), and the false positive rate can be as high as 30% or more in real-time monitoring.

[0006] 4) Lack of prevention and control linkage: Existing monitoring systems lack linkage with irrigation and pesticide spraying equipment, and cannot achieve "monitoring - early warning - prevention" closed-loop management. SUMMARY

[0007] One or more embodiments of the present application describe an agricultural pest and disease intelligent monitoring system and method fusing big data and a generative adversarial large model, which can effectively identify and warn crop pests and diseases.

[0008] According to an embodiment of the first aspect of the present application, an agricultural pest and disease intelligent monitoring system fusing big data and a generative adversarial large model is provided, comprising:

[0009] a data collection unit for collecting crop-related information and including a data interface connected to an agricultural database;

[0010] a data processing unit including: a data cleaning module for data denoising, enhancing and correcting the crop-related information; a data labeling module for labeling pest and disease images; and a feature engineering module for extracting pest and disease image features and crop environment features;

[0011] The model training unit comprises: a data sample enhancement module, which generates simulated pest and disease images based on real pest and disease images according to a generative adversarial large model optimization mode, and improves training stability and visual similarity according to the difference between the real pest and disease images and the simulated pest and disease images; and a multi-modal recognition module, which is configured to fuse pest and disease image features and environmental features to generate a pest and disease recognition feature vector for identifying pest and disease.

[0012] The intelligent application unit is configured to identify pest and disease conditions according to the pest and disease recognition feature vector and selectively issue a warning according to the pest and disease conditions.

[0013] Preferably, in any embodiment,

[0014] The data acquisition unit comprises at least one of an optical camera, a spectrum sensor, a temperature and humidity sensor, a soil moisture sensor, a drone, and a multispectral camera.

[0015] The crop-related information comprises at least one of a crop leaf image, an environmental parameter, and soil data, wherein the environmental parameter comprises at least one of temperature, humidity, and light intensity, and the soil data comprises at least one of nitrogen, phosphorus, and potassium content and pH.

[0016] Preferably, in any embodiment,

[0017] The data of the agricultural database comprises at least one of agricultural meteorological data, regional climate data, historical occurrence records of pest and disease, and historical distribution characteristics of pest and disease.

[0018] Preferably, in any embodiment,

[0019] The data cleaning module adopts a median filter method for denoising, a color space conversion method for enhancing feature contrast, and a Z-score standardization technique for unifying data dimensions.

[0020] Preferably, in any embodiment,

[0021] The pest and disease image features extracted by the feature engineering module comprise traditional features including at least one of a histogram of oriented gradients (HOG), a scale-invariant feature transform (SIFT), and a center-symmetric local binary pattern (CLBP), and deep features combined with a convolutional neural network (CNN) output.

[0022] Preferably, in any embodiment,

[0023] The feature engineering module constructs a time-space-climate three-dimensional feature vector reflecting the correlation between the crop environmental features and the occurrence probability of pest and disease according to the crop environmental features.

[0024] Preferably, in any embodiment,

[0025] The data sample enhancement module comprises a generator and a discriminator;

[0026] The generator adopts a UNet architecture, and generates simulated pest and disease images from real pest and disease images through random noise injection and geometric transformation.

[0027] The discriminator constructs a binary classifier based on ResNet18 to distinguish between real pest and disease images and simulated pest and disease images.

[0028] Preferably, in any embodiment,

[0029] The multi-modal recognition module comprises an input layer, a backbone network, a fusion layer, and a classification layer.

[0030] The input layer fuses pest and disease image features and environmental features to construct a cross-modal feature vector.

[0031] The backbone network adopts EfficientNet-B5 as an image feature extractor.

[0032] The fusion layer combines image and environmental features through a Concatenate operation, and outputs the pest and disease recognition feature vector through an attention mechanism.

[0033] The classification layer adopts a Softmax classifier.

[0034] Preferably, in any embodiment,

[0035] The intelligent application unit is coupled to at least one of a pesticide spraying device, an irrigation system, and a temperature control system.

[0036] According to an embodiment of the second aspect of the application, an agricultural pest and disease intelligent monitoring method based on big data and a generative adversarial large model is provided, which is used to monitor the pest and disease conditions of crops through the agricultural pest and disease intelligent monitoring system based on big data and a generative adversarial large model as described above, comprising:

[0037] The data acquisition unit acquires crop-related information and selectively connects to an agricultural database.

[0038] The data processing unit uses a data cleaning module to denoise, enhance, and correct the crop-related information, uses a data labeling module to label pest and disease images, and uses a feature engineering module to extract pest and disease image features and crop environmental features.

[0039] The model training unit generates simulated pest and disease images based on the real pest and disease images according to the generative adversarial large model optimization mode, improves the training stability and visual similarity according to the difference between the real pest and disease images and the simulated pest and disease images, and generates a pest and disease recognition feature vector for identifying the pest and disease by fusing the pest and disease image features and the environment features through the multi-modal recognition module.

[0040] The intelligent application unit identifies the pest and disease condition according to the pest and disease recognition feature vector and selectively issues a warning according to the pest and disease condition.

[0041] According to one or more embodiments of the present application, the agricultural pest and disease intelligent monitoring system and method fusing big data and generative adversarial large models can effectively identify and warn crop pests and diseases. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0043] Figure 1 FIG. 1 is a structural schematic diagram of an agricultural pest and disease intelligent monitoring system fusing big data and generative adversarial large models according to an embodiment of the present application.

[0044] Figure 2 FIG. 2 is a flowchart of an agricultural pest and disease intelligent monitoring method fusing big data and generative adversarial large models according to an embodiment of the present application. DETAILED DESCRIPTION

[0045] The present application describes an agricultural pest and disease intelligent monitoring system and method fusing big data and generative adversarial large models, which can effectively identify and warn crop pests and diseases.

[0046] The technical solutions of the embodiments of the present application will be described below in conjunction with the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all embodiments. Based on the embodiments described in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] According to an embodiment of the first aspect of the present application, an agricultural pest and disease intelligent monitoring system fusing big data and generative adversarial large models is provided, comprising:

[0048] a data acquisition unit for acquiring crop-related information and comprising a data interface connected to an agricultural database;

[0049] a data processing unit comprising a data cleaning module for data denoising, enhancement and correction of the crop-related information, a data labeling module for labeling pest images, and a feature engineering module for extracting pest image features and crop environment features;

[0050] a model training unit comprising a data sample enhancement module for generating pest simulation images based on real pest images according to the optimization method of generative adversarial large models, and improving training stability and visual similarity according to the difference between real pest images and pest simulation images, and a multi-modal recognition module for fusing pest image features and environment features to generate pest recognition feature vectors for identifying pests;

[0051] an intelligent application unit for identifying pest conditions according to the pest recognition feature vectors and selectively issuing warnings according to the pest conditions.

[0052] In this way, the agricultural pest intelligent monitoring system combining big data and generative adversarial large models acquires crop-related information (which can include temperature and humidity, soil, light, and other surrounding information, and crop leaf image information reflecting pest state characteristics, etc.) through the data acquisition unit, and performs pest labeling and feature extraction on the corrected crop-related information (especially leaf image information) through the data processing unit to enhance the system's deep understanding and recognition of pests. The model training unit trains the system's ability to recognize pest features, wherein the data sample enhancement module generates pest simulation images based on real pest images (and the pest features embodied therein) according to the optimization method of generative adversarial large models (GANs), which is particularly useful in cases where there is a lack of pest (especially rare or new pest) samples. That is, a large number of high-similarity pest simulation images can be generated based on a small number of real pest images for system learning and training to improve pest recognition ability. During the learning and training process, the visual similarity of the simulation images can be continuously improved according to the difference between the real pest images and the pest simulation images to improve the training stability.

[0053] Thus, according to the generative adversarial big model (GANs), through the adversarial training of the generator and the discriminator in the data sample enhancement module, high-fidelity disease and pest simulation data can be generated based on a small amount of real disease and pest samples, effectively alleviating the problem of insufficient disease and pest samples in the agricultural field, thereby providing more effective samples for system learning and training to improve the disease and pest recognition capability. Further, in combination with big data preprocessing and multi-modal data fusion technology, the recognition accuracy in a complex scene can be significantly improved. Thus, when the recognition result exceeds the disease and pest threshold during the monitoring process, the early warning mechanism can be triggered to send an early warning message according to the disease and pest condition.

[0054] As can be seen, the agricultural disease and pest intelligent monitoring system fusing big data and generative adversarial big model provided by the embodiment of the present application can effectively identify and warn the crop diseases and pests.

[0055] Figure 1 is a structural schematic diagram of an agricultural disease and pest intelligent monitoring system fusing big data and generative adversarial big model according to an embodiment of the present application.

[0056] In Figure 1 As can be seen in the embodiment shown, an agricultural disease and pest intelligent monitoring system fusing big data and generative adversarial big model includes:

[0057] The data acquisition unit 100 is configured to acquire crop-related information and includes a data interface connected to an agricultural database;

[0058] The data processing unit 200 includes a data cleaning module configured to perform data denoising, enhancement and correction on the crop-related information; a data labeling module configured to label disease and pest images; and a feature engineering module configured to extract disease and pest image features and crop environment features;

[0059] The model training unit 300 includes a data sample enhancement module configured to generate disease and pest simulation images based on real disease and pest images according to the generative adversarial big model optimization method, and to improve training stability and visual similarity according to the difference between the real disease and pest images and the disease and pest simulation images; and a multi-modal recognition module configured to fuse disease and pest image features and environment features to generate a disease and pest recognition feature vector for identifying diseases and pests;

[0060] The intelligent application unit 400 is configured to identify the disease and pest condition according to the disease and pest recognition feature vector and selectively send an early warning according to the disease and pest condition.

[0061] Preferably, in any embodiment,

[0062] The data acquisition unit includes at least one of an optical camera, a spectrum sensor, a temperature and humidity sensor, a soil moisture sensor, a drone, and a multispectral camera.

[0063] The crop-related information includes at least one of a crop leaf image, an environmental parameter, and soil data, wherein the environmental parameter includes at least one of temperature, humidity, and light intensity, and the soil data includes at least one of nitrogen, phosphorus, and potassium content and pH value.

[0064] Optionally, in an embodiment, the optical camera includes a high-definition camera with a resolution of 4K or above.

[0065] Optionally, in an embodiment, the optical camera includes a high-definition camera with a resolution of 4K or above.

[0066] Optionally, in an embodiment, the optical camera includes a high-definition camera with a resolution of 4K or above.

[0067] Optionally, in an embodiment, the crop-related information includes a crop canopy image collected by aerial photography of a farmland by a UAV.

[0068] Optionally, in an embodiment, the crop-related information includes a crop canopy image collected by aerial photography of a farmland by a UAV.

[0069] Optionally, in an embodiment, the regional climate data includes rainfall and accumulated temperature.

[0070] In a preferred embodiment of the present application, the data collection unit includes:

[0071] A sensing system (network) includes a high-definition camera (resolution ≥ 4K), a spectral sensor (covering visible light / near-infrared waveband), a temperature and humidity sensor, and a soil moisture sensor, so as to collect crop leaf images, environmental parameters (such as temperature, humidity, and light intensity), soil data (such as pH value and nitrogen, phosphorus, and potassium content), and the like in real time; and a UAV (UAV) carrying a multi-spectral camera, regularly (for example, daily / weekly) performing aerial photography of a farmland to obtain large-scale crop canopy images with a resolution of up to 0.1 m / pixel.

[0072] A data interface can be connected to an agricultural meteorological database and a historical pest occurrence record database to obtain regional climate data (such as rainfall and accumulated temperature) and historical distribution characteristics of pests.

[0073] Preferably, in any embodiment, the data cleaning module adopts a median filter method for denoising (to improve image quality), adopts a color space conversion method (such as RGB→HSV / Lab) to enhance feature contrast (such as to enhance the contrast of leaf and pest features), and adopts a Z-score standardization technique to unify data dimensions.

[0074] Optionally, in one embodiment, the data cleaning module detects outliers based on a sliding window algorithm for data from the sensor, and fills in missing values using a linear interpolation method.

[0075] Optionally, in one embodiment, the data labeling module labels pest images using at least one of manual labeling and automatic labeling.

[0076] Optionally, in one embodiment, during the manual labeling, a LabelMe tool is used to perform pixel-level labeling on the pest images.

[0077] Optionally, in one embodiment, during the automatic labeling, a YOLOv8 pre-trained model is used to perform preliminary labeling on the pest images, and then an expert reviews the preliminary labeling to complete the labeling of the pest images. In this way, the labeling efficiency of the pest images can be improved by more than 30%.

[0078] Optionally, in one embodiment, the labeling of the pest images includes labeling the types of pests, such as aphids, rust, etc.

[0079] Optionally, in one embodiment, the labeling of the pest images includes labeling the severity of the pests, such as mild, moderate, severe, etc.

[0080] Preferably, in any embodiment, the feature engineering module extracts features of the pest images, including traditional features such as at least one of a histogram of oriented gradients (HOG), a scale-invariant feature transform (SIFT), and a center-symmetric local binary pattern (CLBP), and deep features combined with the output of a convolutional neural network (CNN).

[0081] Optionally, in one embodiment, the deep features include the conv4_1 layer features of ResNet50.

[0082] Preferably, in any embodiment, the feature engineering module constructs a time-space-climate three-dimensional feature vector that reflects the correlation between the crop environment features and the occurrence probability of pests and diseases based on the crop environment features.

[0083] Optionally, in one embodiment, the crop environment features include at least one of daily mean temperature, diurnal temperature range, and soil moisture content.

[0084] In a preferred embodiment of the application, the data processing unit comprises:

[0085] a data cleaning module for data denoising, enhancement and correction of the crop-related information, which can use median filtering to remove noise for collected crop image data, enhance the contrast of leaf and pest characteristics through color space conversion (RGB→HSV / Lab), detect outliers based on a sliding window algorithm for sensor data, fill in missing values using linear interpolation, and unify data dimensions through Z-score standardization;

[0086] a data labeling module for labeling pest images, which can use the LabelMe tool for pixel-level labeling of pest images under manual labeling, mark pest types (such as aphids, rust) and damage levels (mild / medium / severe), and use the YOLOv8 pre-trained model for preliminary labeling of massive images under automatic labeling, and then secondarily review by experts to improve labeling efficiency by more than 30%; and

[0087] a feature engineering module for extracting pest image features and crop environmental features, which can extract traditional features such as HOG (Histogram of Oriented Gradients), SIFT (Scale-Invariant Feature Transform), and CLBP (Center-Symmetric Local Binary Pattern) for image features, and combine deep features output by CNN convolution layers (such as ResNet50 conv4_1 layer features); and can construct a "time-space-climate" three-dimensional feature vector for environmental features, such as the correlation between daily mean temperature, diurnal temperature difference, soil moisture content and pest occurrence probability.

[0088] Preferably, in any embodiment,

[0089] the data sample enhancement module comprises a generator (Generator) and a discriminator (Discriminator);

[0090] wherein the generator adopts a UNet architecture to generate pest simulation images from real pest images through random noise injection and geometric transformation;

[0091] wherein the discriminator constructs a binary classifier based on ResNet18 to distinguish between real pest images and pest simulation images.

[0092] In this way, the data sample enhancement module is iteratively optimized through the adversarial training of the generator and the discriminator, can generate high-fidelity pest and disease simulation data based on a small amount of real pest and disease samples, significantly increases the number of effective samples for the model training unit to train, and thus can effectively alleviate the problem of insufficient pest and disease samples in the agricultural field. A large amount of effective and usable pest and disease simulation data can be generated through the above GANs method, the number of effective samples can be expanded by 10-20 times, the Frechet Inception Distance (FID) score of the generated pest and disease simulation images and real images is less than 30, and the visual similarity can reach more than 90%.

[0093] Optionally, in an embodiment, the geometric transformation method adopted by the generator includes at least one of rotation, scaling and translation.

[0094] Optionally, in an embodiment, the discriminator adopts WGAN-GP (Wasserstein GAN with Gradient Penalty) as a loss function to improve the training stability.

[0095] Preferably, in any embodiment,

[0096] The multi-modal recognition module includes an input layer, a backbone network, a fusion layer and a classification layer.

[0097] The input layer fuses pest and disease image features (extracted by a CNN, for example) and environmental features (input by a fully connected layer, for example) to construct a cross-modal feature vector.

[0098] The backbone network adopts EfficientNet-B5 as an image feature extractor.

[0099] The fusion layer combines image and environmental features through a Concatenate operation, and outputs the pest and disease recognition feature vector through an attention mechanism (Self-Attention) weighting.

[0100] The classification layer adopts a Softmax classifier.

[0101] Optionally, in an embodiment, in the backbone network, the environmental features are reduced in dimension through 3 layers of fully connected layers (each layer has, for example, 256 / 128 / 64 neurons).

[0102] In this way, through the multi-modal data fusion technology, the classification of more than 50 types of pests and diseases can be supported, and the recognition probability and damage level can be output.

[0103] In a preferred embodiment of the application, the model training unit includes:

[0104] The data sample enhancement module generates simulated pest and disease images based on real pest and disease images according to a generative adversarial large model (GANs) optimization mode, and improves training stability and visual similarity according to the difference between the real pest and disease images and the simulated pest and disease images, and can include: a generator (Generator) which can adopt a UNet architecture, input a small amount of real pest and disease images (for example, only 50 new disease samples), and generate simulated pest and disease images through random noise injection and geometric transformation (rotation / scaling / translation) to output a size of, for example, 512x512 pixels; a discriminator (Discriminator) which can construct a binary classifier based on ResNet18 to distinguish between real images and generated images, and a loss function can adopt WGAN-GP (Wasserstein GAN with Gradient Penalty) to improve training stability. After GANs generation, the effective sample size can be expanded to 10-20 times, the FID (Frechet Inception Distance) score of the generated simulated image and the real image is less than 30, and the visual similarity is more than 90%, thereby achieving excellent data enhancement effect;

[0105] The multi-modal recognition module is used for fusing pest and disease image features and environmental features to generate a pest and disease recognition feature vector for identifying pest and disease, and can include: an input layer: fusing image features (CNN extraction) and environmental features (full connection layer input) to construct a cross-modal feature vector; a backbone network: adopting EfficientNet-B5 as an image feature extractor to output a 1024-dimensional feature vector, wherein the environmental features are reduced in dimension through 3 layers of full connection layers (each layer of neurons has 256 / 128 / 64); a fusion layer: merging the image and environmental features through a Concatenate operation, weighting through an attention mechanism (Self-Attention), and outputting a final pest and disease recognition feature vector; a classification layer: which can adopt a Softmax classifier, thereby being able to support classification of more than 50 categories of pest and disease, and outputting identification probability and damage level.

[0106] Preferably, in any embodiment, the intelligent application unit is coupled to at least one of a pesticide spraying device, an irrigation system, and a temperature control system.

[0107] Optionally, in one embodiment, the intelligent application unit automatically adjusts the irrigation frequency of the irrigation system when needed (for example, during a high-temperature and high-humidity period when diseases are prone to occur, such as during the rainy season), to weaken the disease breeding environment.

[0108] Optionally, in one embodiment, the pesticide spraying device includes a drone or an intelligent field sprayer.

[0109] Optionally, in an embodiment, the intelligent application unit comprises an edge computing terminal (e.g. NVIDIA Jetson Nano edge computing device).

[0110] Optionally, in an embodiment, the edge computing terminal is built-in with a lightweight recognition model (e.g. model size < 100MB, inference speed > 20FPS), and can support millisecond-level recognition of real-time camera stream data.

[0111] Optionally, in an embodiment, the edge computing terminal can display the pest and disease recognition results in real time (e.g. aphids, confidence 95%, damage level moderate), and preferably further superimpose time, geographic location, etc. information to generate monitoring logs.

[0112] Optionally, in an embodiment, the intelligent application unit selectively issues a warning according to the warning threshold of the pest and disease damage level.

[0113] Optionally, in an embodiment, when the leaf infection reaches mild damage (≥ 5%), a yellow warning is triggered, when it reaches moderate damage (≥ 20%), an orange warning is triggered, and when it reaches severe damage (≥ 50%), a red warning is triggered.

[0114] In this way, the pesticide spraying range and dosage can be adjusted according to the warning level defined by the warning threshold. For example, full-field pesticide spraying is triggered at a red warning, and precise pesticide application is only targeted at the infected area at a yellow warning.

[0115] Optionally, in an embodiment, the intelligent application unit issues a warning notice to the farmer through at least one of the following ways: short message, mobile phone message push, and agricultural field broadcast.

[0116] Optionally, in an embodiment, the intelligent application unit uploads the warning message and related information to the agricultural Internet of Things platform.

[0117] Optionally, in an embodiment, the model training unit adopts model compression technology.

[0118] Optionally, in an embodiment, in the model compression technology, redundant convolution layers are removed by pruning technology, and the weight precision is reduced from FP32 to INT8 by quantization. In this way, the model size can be compressed by 80%, and the inference speed can be improved by 2 times.

[0119] Optionally, in an embodiment, the TensorRT acceleration engine is adopted in the model training unit. In this way, model inference acceleration can be realized on the NVIDIA Jetson platform, with a delay of < 50ms.

[0120] Optionally, in an embodiment, the model training unit employs a Federated Learning architecture, where each farm node trains the model locally and uploads only the model parameter updates to protect the privacy of the farmer data.

[0121] Optionally, in an embodiment, the central server of the model training unit periodically aggregates the global model, addressing the uneven distribution of pest data across regions.

[0122] Optionally, in an embodiment, the intelligent application unit includes a visualization monitoring screen (e.g., a web-based monitoring dashboard) that displays various pest information in real-time, such as at least one of the following: a pest distribution heat map, a trend prediction curve (e.g., the probability of disease occurrence in the next 7 days), a prevention and control effect statistics (e.g., the area of pesticide spraying, the proportion of cost savings), and the like.

[0123] Optionally, in an embodiment, the intelligent application unit supports historical data query and comparative analysis, providing farmers with decision-making reference for pest prevention and control.

[0124] In a preferred embodiment of the present application, the intelligent application unit includes:

[0125] An edge computing terminal, where a NVIDIA Jetson Nano edge computing device is deployed, with a lightweight recognition model (model size < 100MB, inference speed > 20FPS) built-in, supporting millisecond-level recognition of camera real-time stream data, capable of displaying the recognition results in real-time (e.g., "aphids, confidence 95%, hazard level: moderate") and superimposing time and geographic location information to generate monitoring logs.

[0126] In a preferred embodiment, a warning mechanism based on the intelligent application unit can be employed, where,

[0127] Threshold setting: set warning thresholds according to pest hazard levels (e.g., yellow warning triggered when mild hazard > 5% leaf infection, orange warning triggered when moderate hazard > 20%, red warning triggered when severe hazard > 50%);

[0128] Multi-channel notification: notify farmers in real-time through SMS, APP push, and field broadcasting, while uploading warning data to the agricultural Internet of Things platform.

[0129] In a preferred embodiment, a prevention and control linkage based on the intelligent application unit can be employed, where it can be connected with intelligent pesticide spraying drones and field intelligent sprayers, automatically adjusting the pesticide spraying range and dosage according to the warning level (e.g., full-field pesticide spraying triggered when red warning, precise pesticide spraying only on infected areas when yellow warning); in addition, it can be linked with the irrigation system, automatically adjusting the irrigation frequency during high-temperature and high-humidity periods (e.g., rainy season) to reduce the disease breeding environment.

[0130] According to an embodiment of the second aspect of the present application, there is provided an agricultural pest and disease intelligent monitoring method fusing big data and generative adversarial large model, for monitoring crop pest and disease conditions through the agricultural pest and disease intelligent monitoring system fusing big data and generative adversarial large model as described above, comprising:

[0131] collecting crop related information through a data acquisition unit and selectively connecting to an agricultural database;

[0132] Through a data processing unit, using a data cleaning module to denoise, enhance and correct the crop related information, using a data labeling module to label pest and disease images, and using a feature engineering module to extract pest and disease image features and crop environment features;

[0133] Through a model training unit, using a data sample enhancement module to generate pest and disease simulation images based on real pest and disease images according to the generative adversarial large model optimization method and to improve training stability and visual similarity according to the difference between real pest and disease images and pest and disease simulation images, and using a multi-modal recognition module to fuse pest and disease image features and environment features to generate pest and disease recognition feature vectors for identifying pest and disease;

[0134] Through an intelligent application unit, identifying pest and disease conditions according to the pest and disease recognition feature vectors and selectively issuing warnings according to the pest and disease conditions.

[0135] Figure 2 is a flowchart of an agricultural pest and disease intelligent monitoring method fusing big data and generative adversarial large model according to an embodiment of the present application.

[0136] In Figure 2 the embodiment shown, there is an agricultural pest and disease intelligent monitoring method fusing big data and generative adversarial large model, for monitoring crop pest and disease conditions through the agricultural pest and disease intelligent monitoring system fusing big data and generative adversarial large model as described above, comprising:

[0137] 510: Collecting crop related information through a data acquisition unit and selectively connecting to an agricultural database;

[0138] 520: Through a data processing unit, using a data cleaning module to denoise, enhance and correct the crop related information, using a data labeling module to label pest and disease images, and using a feature engineering module to extract pest and disease image features and crop environment features;

[0139] 530: Through the model training unit, the data sample enhancement module is used to generate pest and disease simulation images based on the real pest and disease images according to the generative adversarial large model optimization mode, and the training stability and visual similarity are improved according to the difference between the real pest and disease images and the pest and disease simulation images, the multi-modal recognition module is used to fuse the pest and disease image features and the environment features to generate a pest and disease recognition feature vector for identifying the pest and disease;

[0140] 540: Through the intelligent application unit, the pest and disease condition is identified according to the pest and disease recognition feature vector, and a warning is selectively sent according to the pest and disease condition.

[0141] Therefore, the agricultural pest and disease intelligent monitoring system and method fusing big data and generative adversarial large model provided by one or more embodiments of the present application can effectively identify and warn the crop pests and diseases.

[0142] In one or more embodiments of the present application, farmland data can be collected by multi-source sensors and unmanned aerial vehicles, a pest and disease feature library is constructed by using big data preprocessing technology, small sample pest and disease image enhancement and intelligent identification are realized in combination with generative adversarial large model (GANs), and related Internet of Things devices are linked to realize pest and disease warning and prevention and control. The agricultural pest and disease intelligent monitoring system and method fusing big data and generative adversarial large model provided by one or more embodiments of the present application can solve the problems of insufficient data, low recognition accuracy and poor real-time performance in traditional monitoring, can accurately identify more than 50 kinds of pests and diseases, the detection accuracy is more than 98%, and effective identification can be provided for new pests and diseases with scarce samples, so that an efficient and low-cost monitoring scheme can be provided for intelligent agriculture.

[0143] The agricultural pest and disease intelligent monitoring system and method fusing big data and generative adversarial large model provided by one or more embodiments of the present application can realize at least one of the following beneficial effects:

[0144] 1) Data-driven accurate identification: through multi-source data fusion and GANs data enhancement, the small sample learning problem can be solved, the average accuracy on a 50-pest and disease test set is 98.2%, which is more than 15% higher than that of the traditional CNN method.

[0145] 2) Real-time monitoring and intelligent prevention and control: millisecond-level identification can be realized through an edge computing terminal, the warning response time is less than 10 seconds, the linkage efficiency of the prevention and control equipment is improved by 50%, and the amount of pesticide used is reduced by more than 30%.

[0146] 3) Low-cost deployment: based on open source framework (PyTorch / TensorFlow) and commercial hardware (NVIDIA Jetson), the deployment cost of a single system is less than 5000 yuan, which is suitable for small and medium-sized farmers to promote.

[0147] 4) Strong scalability: support federated learning, multi-modal data access, easy to extend to crop growth monitoring, weed identification and other scenarios, and build a full-chain intelligent agricultural solution.

[0148] One embodiment of the present application provides a computer readable storage medium, which stores a computer program, and when the computer program is executed in a computer, the computer executes the agricultural pest intelligent monitoring method of fusing big data and a generative adversarial large model in any one of the embodiments of the specification.

[0149] One embodiment of the present application provides a computing device, comprising a memory and a processor, the memory stores executable code, and when the processor executes the executable code, the method of fusing big data and a generative adversarial large model for intelligent monitoring of agricultural pests in any one of the embodiments of the specification is implemented.

[0150] It should be noted that the terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0151] It should be understood that the term "and / or" used herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0152] It can be understood that the structure shown in the embodiments of the present application does not constitute a specific limitation on the device of the embodiments of the present application. In other embodiments of the specification, the above device can include more or fewer components than the illustration, or combine certain components, or split certain components, or different component arrangements. The illustrated components can be implemented in hardware, software, or a combination of software and hardware.

[0153] Each embodiment in the present application is described in a progressive manner, and the same and similar parts of each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. Especially for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0154] Those skilled in the art should understand that, in one or more examples described above, the functions described in the present application can be implemented by hardware, software, a plug-in or any combination thereof. When implemented by software, the functions can be stored in a computer readable medium or transmitted as one or more instructions or codes on a computer readable medium.

[0155] The above detailed description sets forth the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above detailed description is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent monitoring system for agricultural pests and diseases that integrates big data and generative adversarial models, characterized in that, include: The data acquisition unit is used to collect crop-related information and includes a data interface for connecting to an agricultural database. The data processing unit includes: a data cleaning module for denoising, enhancing, and correcting the crop-related information; a data annotation module for annotating images of pests and diseases; and a feature engineering module for extracting features from images of pests and diseases and features of the crop environment. The model training unit includes: a data sample augmentation module, which generates simulated pest images based on real pest images according to the generative adversarial large model optimization method, and improves training stability and visual similarity based on the difference between real pest images and simulated pest images; and a multimodal recognition module, which is used to fuse pest image features and environmental features to generate pest recognition feature vectors for identifying pests. The intelligent application unit is used to identify the status of pests and diseases based on the pest and disease identification feature vector and to selectively issue early warnings based on the status of pests and diseases.

2. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models as described in claim 1, characterized in that, The data acquisition unit includes at least one of the following: an optical camera, a spectral sensor, a temperature and humidity sensor, a soil moisture sensor, a drone, and a multispectral camera; The crop-related information includes at least one of the following: crop leaf images, environmental parameters, and soil data. The environmental parameters include at least one of temperature, humidity, and light intensity, and the soil data includes at least one of nitrogen, phosphorus, and potassium content and pH.

3. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models as described in claim 1, characterized in that, The data in the agricultural database includes at least one of the following: agricultural meteorological data, regional climate data, historical records of pest and disease occurrences, and historical distribution characteristics of pests and diseases.

4. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models as described in claim 1, characterized in that, The data cleaning module uses median filtering to remove noise, color space conversion to enhance feature contrast, and Z-score standardization to unify data dimensions.

5. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models as described in claim 1, characterized in that, The features of the pest and disease images extracted by the feature engineering module include: traditional features including at least one of Histogram of Oriented Gradients (HOG), Scale Invariant Feature Transform (SIFT), and Central Symmetric Local Binary Pattern (CLBP), and deep features combined with the output of a Convolutional Neural Network (CNN).

6. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models according to claim 1, characterized in that, The feature engineering module constructs a three-dimensional feature vector of time, space, and climate that reflects the correlation between the crop environmental characteristics and the probability of pest and disease occurrence, based on the crop environmental characteristics.

7. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models as described in claim 1, characterized in that, The data sample enhancement module includes a generator and a discriminator; The generator described above uses the UNet architecture to generate simulated pest images from real pest images by injecting random noise and performing geometric transformations. The discriminator is based on ResNet18 to construct a binary classifier to distinguish between real pest and disease images and simulated pest and disease images.

8. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models according to claim 1, characterized in that, The multimodal recognition module includes an input layer, a backbone network, a fusion layer, and a classification layer; The input layer fuses pest and disease image features and environmental features to construct a cross-modal feature vector; The backbone network uses EfficientNet-B5 as the image feature extractor. The fusion layer merges image and environmental features through a concatenate operation and outputs the pest and disease identification feature vector through a weighted attention mechanism. The classification layer uses the Softmax classifier.

9. The intelligent monitoring system for agricultural pests and diseases integrating big data and generative adversarial models as described in claim 1, characterized in that, The intelligent application unit is connected to at least one of the spraying equipment, irrigation system, and temperature control system.

10. A method for intelligent monitoring of agricultural pests and diseases that integrates big data and generative adversarial models, characterized in that, It is used to monitor crop pest and disease status through the intelligent monitoring system for agricultural pests and diseases that integrates big data and generative adversarial models as described in any one of claims 1-9, including: The data acquisition unit collects relevant crop information and selectively connects it to an agricultural database; The data processing unit uses a data cleaning module to denoise, enhance, and correct the crop-related information, uses a data labeling module to label the pest and disease images, and uses a feature engineering module to extract the features of the pest and disease images and the crop environment features. Through the model training unit, the data sample enhancement module generates simulated pest images based on real pest images according to the generative adversarial large model optimization method. It improves training stability and visual similarity based on the difference between real pest images and simulated pest images. The multimodal recognition module fuses pest image features and environmental features to generate pest recognition feature vectors for identifying pests. The intelligent application unit identifies pest and disease conditions based on the pest and disease identification feature vector and selectively issues early warnings based on the pest and disease conditions.

Images