Grape disease identification and early warning method based on Internet of Things

Abstract

The invention discloses a grape disease recognition and early warning method based on the Internet of Things, and relates to the technical field of plant disease recognition, image acquisition equipment and environment sensing nodes are arranged in a vineyard, and leaf images and corresponding temperature and humidity, illumination and soil moisture parameters are obtained; inputting the image into a neural network fusing dilated convolution and a residual attention mechanism, realizing extraction of a disease spot region and a disease spot variation feature, and generating a preliminary recognition result; constructing a multi-factor evolution sample set in combination with the recognition result and the environment state of the time node; constructing a space-time correlation graph model based on a graph neural network, estimating a disease propagation risk path and a diffusion probability, and performing early warning judgment at a gateway end through a multi-factor gating discrimination algorithm; the method disclosed by the invention is high in recognition precision and strong in response timeliness, has adaptive prediction and targeted treatment capabilities, and remarkably improves the intelligence and precision level of grape disease management.

Description

Technical Field

[0001] The present invention relates to the technical field of plant disease identification, and in particular to a grape disease identification and early warning method based on the Internet of Things. Background Art

[0002] With the continuous advancement of intelligent agriculture, grape cultivation is gradually moving towards high quality and high efficiency. However, traditional grape disease identification and monitoring methods rely heavily on manual inspections and single-point image recognition. These methods suffer from delayed response, limited coverage, and low recognition accuracy, making it difficult to achieve early diagnosis and accurate early warning of vineyard diseases. Existing technologies still lack technical gaps in multi-source sensor data fusion, spatial information modeling, and dynamic transmission path simulation, making them unable to effectively address the complexity and sudden nature of disease transmission.

[0003] In recent years, the application of IoT technology in agricultural pest and disease monitoring has gradually expanded. By deploying multiple sensor nodes to collect environmental parameters and crop growth information, combined with data processing and intelligent algorithms, real-time monitoring of crop health can be achieved. However, there is still a lack of systematic modeling methods for the spatial spread characteristics, environmental induction mechanisms, and multi-factor coupling relationships of grape diseases. In particular, there is a lack of unified evaluation indicators and dynamic prediction mechanisms for spatiotemporal collaborative identification and early warning. There is an urgent need to develop a comprehensive approach that integrates IoT perception, image recognition, and disease evolution modeling to achieve intelligent identification and early warning of grape diseases and improve the proactive and precise disease management. Summary of the Invention

[0004] The purpose of the present invention is to provide a grape disease identification and early warning method based on the Internet of Things to address the shortcomings of the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a grape disease identification and early warning method based on the Internet of Things, comprising:

[0006] Acquire leaf images collected by image acquisition devices deployed in the vineyard and environmental parameters collected by environmental sensing nodes, including temperature, humidity, light, and soil moisture data;

[0007] Inputting the leaf surface image into a disease recognition neural network model, extracting the diseased spot area and its variation characteristics, and generating a preliminary recognition result;

[0008] Fusing the identification results with the environmental parameters of the corresponding time nodes to construct a multi-factor evolution sample set of disease occurrence;

[0009] Build a spatiotemporal correlation graph model based on historical disease samples to calculate the risk path and diffusion probability of disease transmission;

[0010] Perform edge computing processing on the IoT gateway to determine whether the target area meets the disease warning conditions. If so, output disease warning information and mark the suspected disease spread area;

[0011] According to the early warning results, the intelligent sprinkler system is controlled or manual intervention is prompted to achieve precise control of diseases.

[0012] Preferably, the leaf image input disease recognition neural network model includes:

[0013] Perform multi-scale image enhancement and background interference removal on the acquired leaf surface images;

[0014] The preprocessed image is input into a deep neural network model that integrates dilated convolution and residual attention mechanism. The backbone structure of the model is based on the improved U-Net architecture.

[0015] At the network output end, spatial variation features and color distribution abnormality indicators are integrated to construct a lesion discrimination map, and a lesion variation feature set is extracted through a multi-dimensional feature aggregation mechanism.

[0016] Based on the similarity matching and category attribution judgment of the lesion variation feature set and the historical lesion label library, a preliminary recognition result with spatial location information and initial confidence is generated.

[0017] Preferably, the fusion recognition result and the environmental parameters of the corresponding time node include:

[0018] Obtain the spatial location and confidence information of the lesions marked in the preliminary identification results, perform time index matching on them, and establish a time series of lesion evolution;

[0019] The temperature, humidity, light intensity and soil moisture parameters of the corresponding time nodes are extracted synchronously to construct the multi-dimensional environmental state vector when the lesion occurs;

[0020] A feature fusion algorithm based on multivariate correlation entropy is used to jointly map image recognition features with environmental state vectors to construct a unified causal feature representation space.

[0021] Generate a multi-factor evolution sample set including lesion labels, image features, environmental status and evolution trend indicators.

[0022] Preferably, the construction of a spatiotemporal correlation graph model based on historical disease samples and the calculation of the risk path and diffusion probability of disease transmission include:

[0023] The lesion occurrence locations, timestamps, and environmental parameters in the historical disease evolution sample set are constructed as heterogeneous nodes to form a multi-level spatiotemporal graph data structure;

[0024] The dynamic graph convolution module in the graph neural network is used to weight the spatial proximity and environmental similarity between nodes to establish spatiotemporal dependencies.

[0025] Enhance the characteristics of potential transmission paths and estimate the transmission probability by combining historical transmission frequency with the current environmental status;

[0026] Output the disease transmission risk path map in the target area and assign a transmission probability distribution value to each node.

[0027] Preferably, based on the estimation of the propagation probability based on the combination of historical propagation frequency and current environmental status, it further includes: constructing a dynamic propagation control function, which integrates the local environmental disturbance sensitivity factor and the historical propagation path entropy value, and adaptively adjusts the propagation probability weight of each path through an adjustable Bayesian update mechanism, wherein the local environmental disturbance sensitivity factor is used to quantify the nonlinear impact of environmental micro-changes on the disease propagation speed, and the propagation path entropy value is used to evaluate the uncertainty and information content of the path.

[0028] Preferably, the performing edge computing processing on the IoT gateway and outputting disease warning information includes:

[0029] A lightweight deep inference module is deployed within the IoT gateway to perform local analysis on the received disease spot identification results and propagation probability maps. A multi-factor gated discrimination algorithm is used to comprehensively determine whether the current area meets the preset disease warning threshold. The threshold is dynamically adjusted based on the growth rate of the number of lesions, the rate of increase in the propagation probability, and the intensity of environmental induction. If the warning conditions are met, the spatial hot zone rendering module is called to combine the GIS raster information to mark the boundaries of the suspected disease spread area. The warning information is then sent to the upper-level control platform or farmer terminal to implement a disease warning response.

[0030] Preferably, the comprehensive judgment of whether the disease warning threshold is met based on the multi-factor gating discrimination algorithm includes: constructing a multi-factor input vector, integrating the current moment's lesion recognition confidence, propagation probability gradient, environmental parameter volatility and historical disease high incidence weights; inputting into a discrimination network with multi-level gating units, the network adopts a hierarchical gating mechanism to weight different factor channels; combining with an asynchronous activation strategy, only when all factors simultaneously meet the set trigger conditions, outputting the discrimination label of "meets the warning"; and binding the discrimination result with the geographic grid index to realize the positioning of risk areas in the spatial dimension.

[0031] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0032] 1. This invention achieves early detection, dynamic perception, and precise response to grape diseases by constructing an IoT-based disease monitoring and early warning system that integrates image recognition, environmental perception, and spatiotemporal propagation modeling. Compared to traditional methods that rely on manual inspections and timed spraying, this proposed technical solution significantly improves lesion identification accuracy, propagation trend prediction precision, and response timeliness. In particular, supported by a multi-factor gating discriminant algorithm and a dynamic propagation control mechanism, early warning triggering is more stable, spraying strategies are more targeted, and pesticide conservation is more effective.

[0033] 2. The present invention adopts edge computing architecture and intelligent linkage control mechanism, has good system scalability and deployment flexibility, adapts to various plot sizes and operation requirements, significantly reduces the frequency of manual intervention and management costs, improves the level of intelligent and precise grape cultivation, and has wide application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0035] Figure 1 This is a mind map of the method of the present invention. DETAILED DESCRIPTION

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0037] Example 1, please refer to Figure 1 As shown, the grape disease identification and early warning method based on the Internet of Things described in this embodiment includes:

[0038] Acquire leaf images collected by image acquisition devices deployed in the vineyard and environmental parameters collected by environmental sensing nodes, including temperature, humidity, light, and soil moisture data;

[0039] Inputting the leaf surface image into a disease recognition neural network model, extracting the diseased spot area and its variation characteristics, and generating a preliminary recognition result;

[0040] Fusing the identification results with the environmental parameters of the corresponding time nodes to construct a multi-factor evolution sample set of disease occurrence;

[0041] Build a spatiotemporal correlation graph model based on historical disease samples to calculate the risk path and diffusion probability of disease transmission;

[0042] Perform edge computing processing on the IoT gateway to determine whether the target area meets the disease warning conditions. If so, output disease warning information and mark the suspected disease spread area;

[0043] According to the early warning results, the intelligent sprinkler system is controlled or manual intervention is prompted to achieve precise control of diseases.

[0044] In one embodiment of the present invention, to accurately identify and effectively warn of grape disease outbreaks, it is necessary to first acquire multi-source sensory information from relevant areas within the vineyard, primarily consisting of image data and environmental parameter data. To this end, the present invention deploys multiple types of sensor nodes within the vineyard according to a predetermined spatial layout. Specifically, these include image acquisition devices and environmental sensing nodes.

[0045] The image acquisition device is preferably a high-resolution visible light camera with night vision capabilities, mounted at canopy height, facing the extended leaf area, to continuously capture images of the grape foliage. To ensure image data stability and clarity, the image acquisition device can integrate auto-exposure and auto-focus modules, and periodically capture images at fixed intervals (e.g., one frame every 30 minutes). All image data is transmitted via wireless communication to an edge processing unit or cloud server for subsequent processing.

[0046] At the same time, in order to fully understand the possible environmental factors that may induce the occurrence of diseases, the present invention configures multiple types of environmental sensing nodes near the image acquisition device. The environmental sensing nodes include temperature and humidity sensors, light intensity sensors, and soil moisture sensors, and their specific functions are as follows:

[0047] Temperature and humidity sensors are used to monitor the temperature and relative humidity of the air surrounding the grape plants. Preferred digital output sensors are SHT series devices. The data collection frequency can be set to every 10 minutes, with the results recorded as a timestamp.

[0048] The light intensity sensor is used to detect changes in light intensity in the target area. It uses a photodiode module with spectral response correction function to reflect the sunlight intensity and changing trends in real time, and is used to analyze the light conditions related to pathogen reproduction.

[0049] The soil moisture sensor is installed 10 to 20 cm deep at the root of the plant to monitor the soil moisture content in the root water absorption area. It uses a capacitive measurement principle combined with a temperature compensation mechanism to ensure the accuracy and stability of soil moisture data.

[0050] All environmental sensing nodes synchronize data with a local gateway via IoT communication protocols (such as LoRa, ZigBee, or NB-IoT). The system uses unique node identifiers and acquisition timestamps to spatially and temporally map environmental parameters to image data, forming a multimodal input dataset that provides raw data support for subsequent disease identification and evolution modeling.

[0051] Through the combined collection of these images and environmental parameters, the present invention achieves multi-source dynamic perception of the individual status of vineyard plants and external inducing conditions, establishes a complete data foundation layer, and ensures the timeliness and reliability of the disease identification and early warning mechanism.

[0052] Based on the acquisition of grape leaf images, the present invention further utilizes a neural network model to identify diseased areas and extract their variational features, generating preliminary identification results that provide a foundation for subsequent disease early warning and disease transmission modeling. This section details the preprocessing techniques, neural network architecture design, feature extraction strategies, and output generation logic involved in this step, ensuring that those skilled in the art can implement it accordingly.

[0053] First, to address the problem of unstable image quality in natural vineyard scenes, the present invention performs multi-scale image enhancement and background interference removal on the acquired leaf images. This step uses a local contrast enhancement algorithm based on guided filtering to enhance the brightness and texture details of low-contrast areas in the image, significantly improving the distinguishability of the edges of the lesions and the mottled texture inside the lesions. Guided filtering can effectively retain edge information and remove high-frequency noise, and is particularly suitable for complex lighting and background occlusion conditions in leaf images. This enhancement module nests a multi-scale filtering structure to uniformly align and fuse features at different scales of the image, thereby improving the recognizability of lesion areas at different sizes.

[0054] Secondly, the preprocessed image is input into a deep neural network model that integrates dilated convolution and residual attention mechanisms. The backbone structure of the model is improved on the basis of U-Net, and has good feature segmentation and semantic extraction capabilities. The U-Net network consists of an encoder, a decoder, and jump connections, which can effectively retain the spatial information of the image. In order to further improve the perception of lesion recognition in small areas, the model introduces a dilated convolution structure, which avoids reducing the size of the feature map while maintaining the expansion of the receptive field, thereby enhancing the ability to capture fine-grained lesion boundaries. At the same time, the model embeds a residual attention module, which combines channel attention and spatial attention mechanisms to enhance the network's response to prominent lesion areas and suppress interference from invalid background areas. It is particularly stable under complex conditions such as leaf vein occlusion and uneven lighting.

[0055] Third, in the network output stage, the present invention designs a lesion discrimination map construction mechanism that integrates spatial variation features and color distribution abnormality indicators. Spatial variation features refer to the degree of gradient change between the grayscale of pixels inside the lesion area and the neighboring pixels. The clarity and structural continuity of the plaque boundary are judged by the consistency analysis of gradient amplitude and direction; the color distribution abnormality measures the degree of aggregation of color abnormal areas through deviation statistics in the HSV color space. Based on the network output feature map, the present invention introduces a multi-dimensional feature aggregation module. This module realizes joint modeling of different feature dimensions through convolution kernel superposition and feature channel fusion, thereby generating a high-precision lesion variation feature set.

[0056] In the fourth step, after obtaining the lesion variation feature set, the system matches and classifies the feature set with the pre-built historical lesion label library. The historical lesion label library stores typical lesion feature vectors corresponding to different types of grape diseases (such as powdery mildew, downy mildew, anthracnose, etc.), including information such as patch morphology, color texture, and edge change pattern. The matching process uses a vector similarity measurement algorithm based on cosine similarity, combined with feature weights to calculate the confidence probability of each type of disease, and sorts and outputs the recognition results, retaining the top N types of disease types as candidate results. Finally, the system outputs the optimal recognition result in the form of an initial recognition result, along with the spatial position coordinates of the lesion, the lesion contour mask, and the initial confidence score, for subsequent fusion decision-making and multi-time series change analysis.

[0057] Compared with existing single image segmentation or traditional machine learning recognition methods, this invention uses guided filtering enhancement, void convolution structure, residual attention mechanism and multi-dimensional feature aggregation strategy to significantly improve the recognition accuracy and robustness of small and frequently diseased areas; at the same time, it integrates the historical label library for category attribution judgment, enhances the consistency and adaptability of the model in identifying various grape diseases, and has good generalization ability and deployment practicality.

[0058] In summary, the present invention employs multiple unconventional technical approaches in image preprocessing, neural network structure optimization, and feature fusion, and connects the entire recognition process in a logically clear and compact manner, ensuring a significant improvement in the accuracy and efficiency of lesion recognition. This provides a key foundation for subsequent grape disease transmission modeling and intelligent early warning.

[0059] To further enhance the accuracy of dynamic analysis and early warning for disease identification, this paper proposes a processing mechanism that deeply integrates image recognition results with environmental parameters to construct a multi-factorial evolutionary sample set for disease occurrence. This process not only considers the spatial and temporal evolution of lesions but also fully incorporates the inductive effects of environmental changes on disease occurrence and spread, providing a reliable data foundation for subsequent spatiotemporal modeling and risk prediction.

[0060] First, after identifying lesions in leaf images, the system obtains the spatial location and recognition confidence information contained in the preliminary recognition results. The spatial location of the lesions is determined by the coordinate bounding box or mask output by the recognition model, while the confidence information is based on the classification probability values ​​output by the recognition model for the lesion type. In the data management module, the system assigns a unique time index to each set of lesion recognition results, which is strictly tied to the acquisition timestamp. This constructs a time series of lesion evolution, enabling tracking of lesion status over time on the same leaf or region.

[0061] The system then invokes the Environmental Perception Data Management module to synchronously extract environmental parameters corresponding to the aforementioned time points, including key variables such as temperature, humidity, light intensity, and soil moisture. To ensure the timeliness and accuracy of environmental data, the system employs interpolation and data cleaning algorithms to correct for missing or abruptly altered sensor data points. Ultimately, each set of lesion identification data is associated with a multidimensional environmental state vector, describing the microenvironmental conditions at the time of lesion occurrence.

[0062] After acquiring image features and environmental state data, the present invention proposes a feature fusion algorithm based on multivariate correlation entropy to construct a unified causal feature representation space. This algorithm uses lesion image features (such as texture vectors, color histograms, and edge gradients) and the environmental state vector as input. Using a multivariate mutual information matrix, it calculates the correlation entropy between each feature and determines the causal influence of their joint distribution on disease occurrence. Based on the weighted correlation entropy, low-correlation features are dimensionality compressed, while high-correlation features are retained and normalized, ultimately achieving a joint mapping of image and environmental information.

[0063] After feature fusion, the system integrates these features into a unified structure, generating a multi-factor evolution sample set that includes lesion labels, image features, environmental status, and evolution trend indicators. This sample set is stored in a structured format and can include fields such as sample ID, acquisition time, geographic coordinates, lesion category, recognition confidence, texture feature vector, environmental parameter vector, and change rate indicator. This is used for subsequent model training, propagation path deduction, and early warning rule extraction.

[0064] To accurately model and effectively predict the spread of grape diseases in complex agricultural environments, this paper proposes a method for constructing a spatiotemporal correlation graph model that integrates historical disease evolution samples with environmental information. This model incorporates mechanisms for estimating transmission risk and dynamically regulating transmission probability, forming a comprehensive framework for disease transmission modeling and prediction. This method quantifies the path, direction, and rate of disease spread across time and space, providing highly reliable data support for precise disease early warning and intervention decisions.

[0065] First, based on the multi-factor evolution sample set constructed above, the present invention extracts the spatial location of the lesions, the occurrence timestamp, and the environmental parameter information contained therein, and abstracts them into heterogeneous nodes in the graph structure. Each node represents a disease event, and its attributes include lesion category, spatial coordinates, occurrence time, environmental state vector, etc. Nodes of different attribute types are organized in layers, corresponding to the "spatial node layer", "temporal node layer", and "environmental node layer", respectively, to form a multi-level spatiotemporal graph data structure. This structure allows cross-connections between nodes at different levels, facilitating the modeling of disease transmission paths across spatiotemporal scales.

[0066] Next, the dynamic graph convolution module in the dynamic graph neural network is used to model the connections between nodes. Specifically, the system calculates the initial connection weights based on the spatial position distance of the nodes and the similarity of the environmental state. Spatial proximity is modeled using a Gaussian kernel function, and environmental similarity is measured using a cosine similarity function. Dynamic graph convolution operations are then used to periodically update the edge weights and node states in the graph structure to achieve temporal evolution modeling of the graph structure. In this module, new environmental data and recognition result nodes are introduced at each time step, dynamically expanding the graph structure, and updating the activity of historical propagation paths through a gated update mechanism.

[0067] After the graph structure is constructed, the present invention further enhances the features of potential transmission paths, extracting characteristics such as environmental change trends, lesion type consistency, and historical transmission frequency of consecutive nodes along the path. This constructs a set of transmission path feature vectors, which are then fed into a transmission probability assessment module. This module comprehensively considers the historical transmission success rates of diseases under similar climatic conditions and the physical proximity between the current nodes. It outputs a context-based transmission probability estimate, assigning an initial transmission probability to each path.

[0068] To improve the dynamic adaptability of the propagation risk modeling, the present invention further introduces a dynamic propagation control function based on the above estimation mechanism. This function is driven by two core factors:

[0069] Local environmental perturbation sensitivity factor: This factor measures the nonlinear impact of small environmental changes (such as a 1°C temperature increase or a 5% humidity fluctuation) on propagation speed. The system performs regression analysis on the difference in propagation speed before and after the perturbation in historical samples to construct a perturbation response curve, which serves as a library of sensitivity factors. The control function multiplies the current perturbation index by the sensitivity factor to obtain the perturbation weighting term.

[0070] Propagation Path Entropy: Based on information theory, it is used to assess the uncertainty of a propagation path. A higher path entropy indicates a highly variable historical propagation outcome and a greater dispersion of information distribution. Entropy, as a control parameter, modifies propagation probability, suppressing the propagation weight of highly uncertain paths.

[0071] The propagation probability is ultimately iteratively corrected via an adjustable Bayesian update mechanism. The system uses the initial propagation probability as a prior, combines the perturbation weighting term with the path entropy value to construct a likelihood function, and outputs the updated posterior propagation probability. This Bayesian update mechanism features a sensitivity threshold adjustment channel, which dynamically adjusts the control intensity based on user settings or system self-learning results.

[0072] After completing the spread estimation, the system outputs a disease transmission risk map within the target area, annotating high-risk transmission pathways and transmission probability density hotspots on a GIS layer. Simultaneously, it assigns a transmission probability distribution value to all nodes in the map, forming a probabilistic field view of disease transmission that facilitates subsequent early warning strategy matching and intervention scheduling.

[0073] Compared with existing approaches that use static threshold models or fixed propagation formulas, the combination of multi-layer spatiotemporal graph + dynamic graph neural network + Bayesian propagation control mechanism constructed in this invention has the following significant technical advantages:

[0074] Ability to dynamically update propagation trends on a continuous time scale, adapting to the high variability of agricultural environments;

[0075] By modeling propagation path entropy and disturbance response, we can effectively distinguish between predictable and unpredictable propagation channels and improve modeling robustness.

[0076] It can be deployed on edge gateways or cloud platforms to meet the low-latency and high-accuracy early warning requirements in large-scale field scenarios.

[0077] In summary, the propagation graph modeling and dynamic probability control technology of the present invention fully combines the advantages of spatial information science, graph learning and Bayesian reasoning, and innovatively realizes the transition of disease propagation modeling from static rules to dynamic adaptation.

[0078] To improve the timeliness and processing efficiency of the grape disease identification and response system and avoid the communication delays and data congestion problems caused by the transmission of large-scale image and sensor data to remote servers, this paper proposes a gateway-side edge computing processing method in the Internet of Things system architecture and designs a supporting multi-factor gating discrimination algorithm to dynamically determine whether to trigger a disease warning response, thereby achieving efficient, low-latency intelligent warning of diseases in the vineyard.

[0079] In an embodiment of the present invention, a lightweight deep inference module is first deployed within an IoT gateway device within the vineyard. This module integrates a pruned and quantized optimized neural network model, enabling low-power operation on resource-constrained edge hardware. This module receives real-time lesion recognition results (including lesion type, spatial location, and recognition confidence) from image recognition nodes. It also receives disease propagation probability maps (rasterized images containing propagation risk paths and propagation probabilities at each node location) output by the spatiotemporal propagation model module, and performs data integration and analysis locally on the gateway.

[0080] After analyzing the data, the present invention uses a multi-factor gated discrimination algorithm to dynamically determine whether the current area meets the disease warning conditions. The core process of the algorithm is as follows:

[0081] Constructing a multi-factor input vector: The system extracts key indicators within the current sampling period from the gateway cache, including:

[0082] The mean and standard deviation of lesion identification confidence are used to reflect the recognition reliability;

[0083] The growth rate of the number of lesions, that is, the increase in the area of ​​newly identified lesions per unit time;

[0084] The spread probability gradient, that is, the rate of increase in the weight of high-risk paths in the disease spread graph;

[0085] Environmental parameter volatility, which measures the degree of variation in key environmental variables such as temperature, humidity, and light intensity during the current period;

[0086] The weight of the historical high-incidence area of ​​diseases is used to identify whether the current plot is in a historical high-incidence area.

[0087] The input is fed into a discriminant network with multiple levels of gating units: This discriminant network is a model structure that integrates a multi-layer perceptron (MLP) network with a gating mechanism. Each factor input channel corresponds to a saliency gating unit, which dynamically assigns weights to different input factors through an attention mechanism (such as an SE-block). For example, if humidity changes little over a period of time while the propagation path gradient increases rapidly, the model will reduce the weight of environmental factors and increase the discriminant weight of propagation factors.

[0088] Integrating an asynchronous activation strategy: To prevent a single indicator from falsely triggering an alert, this invention employs an asynchronous activation discrimination mechanism. This strategy requires that multiple key factors simultaneously meet independent threshold conditions and maintain this state within a set time window before outputting a final label of "alert satisfied."

[0089] Specifically, the mechanism consists of Boolean logic and weight fusion, such as:

[0090] If the lesion number growth rate is >15%, and the spread probability gradient is >0.3, and the humidity change rate is >10%;

[0091] And the above conditions are met in two consecutive time windows;

[0092] Then the output is "Satisfied warning".

[0093] Binding with geographic grid indexes enables spatialized warning annotation: To enhance the spatial specificity of warning information, this method binds the identification results to the GIS plot grid number. Based on the high-risk area outlines in the transmission risk map and the distribution density of the currently identified lesion center points, the system invokes the spatial hotspot rendering module to annotate the boundaries of suspected disease spread areas with a color gradient. The shape of these areas can be constructed using the Alpha Shape algorithm or the cluster outline algorithm, supporting visualization on a two-dimensional map.

[0094] When disease warning conditions are determined to be met, the system immediately triggers a response mechanism, sending warning information to a higher-level control platform or farmer's mobile device via local communication protocols (such as MQTT, CoAP, or NB-IoT). Warning information includes: risk area number, predicted disease type, confidence interval, spread direction trend, and recommended treatment measures. End users can use this information to implement measures such as manual inspections, drone re-inspections, or automated spraying.

[0095] Compared with traditional disease identification systems based on centralized processing on remote servers, the edge computing and multi-factor gating judgment method proposed in this invention has the following technical advantages:

[0096] Fast response time: The edge inference model reduces data return latency, making it particularly useful during the critical window before a disease outbreak.

[0097] High discrimination accuracy: Dynamically assigning factor weights through a gating network effectively avoids false alarms caused by short-term environmental anomalies or recognition errors;

[0098] High spatial resolution: By binding to the geographic grid, the warning information has high spatial accuracy, facilitating precise on-site positioning;

[0099] Strong adaptability: The gate threshold can be self-learning and adjusted to meet the differentiated warning needs under different grape varieties, geographical regions and management strategies.

[0100] In summary, the present invention integrates edge intelligent reasoning and a multi-factor early warning discrimination algorithm on the IoT gateway side to construct a highly efficient and high-precision intelligent identification and early warning mechanism for grape diseases, providing solid technical support for precise disease management in modern orchards.

[0101] In this invention, achieving closed-loop intelligent control of grape diseases requires not only lesion identification, disease spread modeling, and early warning assessment, but also coordinated control of the intelligent sprinkler system or prompting for manual intervention when an early warning is triggered, thus forming a complete intelligent management process of "identification-alert-response." This component forms the core of the system's decision-making and execution modules, significantly improving the timeliness and targeted nature of disease treatment.

[0102] First, after the edge gateway or cloud platform determines that the target area meets the disease warning conditions, the system will immediately generate a structured response instruction containing the following key fields:

[0103] Warning area number (bound to GIS grid coordinates);

[0104] Disease type and identification confidence;

[0105] Recommended response mode (automatic execution or manual confirmation);

[0106] Control target device ID (such as corresponding sprinkler valve, plot number);

[0107] Recommendations on spraying parameters (such as spray flow rate, pesticide concentration, and action time).

[0108] The system determines whether to enter the automatic linkage mode based on the user's preset strategy or historical response rules. When the device is in linkage mode, the system will call the intelligent sprinkler control module, which will connect with the disease warning information to achieve accurate scheduling of the intelligent sprinkler equipment. The control process is as follows:

[0109] Locate the execution area: Determine the specific plot or sprinkler unit number based on the suspected disease spread boundary marked on the early warning map and combined with the GIS raster index.

[0110] Loading the spraying parameter library: The system extracts the corresponding spraying plan from the local pesticide strategy database based on the disease type, and prioritizes pesticide types and dilution concentrations that are highly targeted and have low environmental impact.

[0111] Dynamic sprinkler control: Using agricultural IoT control protocols (such as Modbus, 485, or wireless PLC), control commands are sent to solenoid valves, variable-frequency pumps, and other equipment in the target area, enabling quantitative, zoned, and timed pest control. To prevent overspraying or misspraying, the system incorporates redundant confirmation mechanisms and feedback channels, ensuring real-time feedback and closed-loop confirmation of equipment execution status.

[0112] When the system is in the manual intervention mode (e.g., insufficient recognition confidence or an area that does not meet the conditions for automatic management), intervention suggestions are automatically generated and pushed to managers through the platform interface or mobile terminals. The push information includes:

[0113] Disease grade assessment report;

[0114] Prediction map of suspected transmission paths;

[0115] Recommend manual treatment measures (such as cutting off diseased leaves, local application of pesticides, etc.);

[0116] Environmental precautions (e.g. high wind speed, not suitable for spraying, etc.);

[0117] Comparative analysis of historical governance effects in this region.

[0118] Based on the push information, managers can conduct manual inspections, confirm diagnoses, or manually trigger irrigation responses. The system also records response behaviors and treatment effect parameters (such as the trend of lesion changes after treatment) and feeds these into the model training module to optimize subsequent warning thresholds and response strategies.

[0119] Example 2: To verify the effectiveness of the "Internet of Things-based grape disease identification and early warning method" described in the present invention in intelligent disease identification, dynamic early warning, and precise disease control, this applicant conducted a 30-day field deployment and comparative experiment at a grape planting base, establishing a typical vineyard information perception and intelligent prevention and control system.

[0120] The experiment selected two grape planting demonstration areas A (experimental group) and B (control group) of similar size, each with an area of ​​about 3 mu, and the same grape variety, "Kyoho".

[0121] Deployed in Area A of the experimental group: 12 sets of visible light image acquisition equipment, distributed in the main branch layer and canopy layer; 10 groups of environmental perception nodes, including temperature, humidity, light intensity and soil moisture sensors; a LoRa networked IoT data acquisition module; 1 edge computing gateway with an integrated lightweight inference module; and 6 sets of intelligent sprinkler solenoid valve controllers, linked to the GIS raster positioning module.

[0122] Control group B relied solely on manual inspections and conventional timed spraying, without image recognition and data-driven early warning control.

[0123] During the entire experimental period, area A of the experimental group collected image data every 30 minutes and uploaded environmental parameters simultaneously; the propagation map was updated every 6 hours based on the recognition results, and edge reasoning was performed; the system recorded the number of triggered warnings, response behaviors, number of manual processing times, and the final lesion control effect.

[0124] All disease conditions (including powdery mildew, downy mildew, and anthracnose) are evaluated on-site by a team of agricultural experts before and after the start of the trial period to confirm the number and area of ​​the disease spots.

[0125] As shown in the table, the test results and effect analysis:

[0126]

[0127]

[0128] It can be seen from the experimental data that: the system of the present invention can realize the identification and early warning of lesions 20 hours in advance on average, greatly improving the timeliness of disease response; under the premise of reducing the spraying frequency by 32%, the efficiency of disease spread control is improved by about 64%, effectively reducing the risk of disease spread; intelligent sprinkler response control accurately locates the execution area according to the identification results, effectively avoiding waste of pesticides and environmental pollution; the multi-factor gated early warning model shows strong stability and reliability, with a false alarm rate of less than 5%, meeting actual operation and maintenance needs; the overall manual intervention intensity of the system is reduced by about 75%, significantly reducing the burden on fruit farmers and improving the level of intelligent planting.

[0129] In summary, the above field deployment and data comparison verification fully demonstrate that the Internet of Things-based grape disease identification and early warning method proposed in the present invention is superior to existing traditional methods in terms of identification accuracy, response efficiency, control effect and system intelligence, and has good prospects for promotion and application.

[0130] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A grape disease identification and early warning method based on the Internet of Things, characterized by: include: Acquire leaf images collected by image acquisition devices deployed in the vineyard and environmental parameters collected by environmental sensing nodes, including temperature, humidity, light, and soil moisture data; Inputting the leaf surface image into a disease recognition neural network model, extracting the diseased spot area and its variation characteristics, and generating a preliminary recognition result; Fusing the identification results with the environmental parameters of the corresponding time nodes to construct a multi-factor evolution sample set of disease occurrence; Build a spatiotemporal correlation graph model based on historical disease samples to calculate the risk path and diffusion probability of disease transmission; Edge computing is performed on the IoT gateway to determine whether the target area meets the disease warning conditions. If so, the disease warning information is output and the suspected disease spread area is marked. According to the early warning results, the intelligent sprinkler system is controlled or manual intervention is prompted to achieve precise control of diseases.

2. The method for identifying and warning grape diseases based on the Internet of Things according to claim 1, characterized in that: The leaf image input disease recognition neural network model includes: Perform multi-scale image enhancement and background interference removal on the acquired leaf surface images; The preprocessed image is input into a deep neural network model that integrates dilated convolution and residual attention mechanism. The backbone structure of the model is based on the improved U-Net architecture. At the network output end, spatial variation features and color distribution abnormality indicators are integrated to construct a lesion discrimination map, and a lesion variation feature set is extracted through a multi-dimensional feature aggregation mechanism. Based on the similarity matching and category attribution judgment of the lesion variation feature set and the historical lesion label library, a preliminary recognition result with spatial location information and initial confidence is generated.

3. The grape disease identification and early warning method based on the Internet of Things according to claim 1 is characterized in that: The fusion recognition result and the environmental parameters of the corresponding time node include: Obtain the spatial location and confidence information of the lesions marked in the preliminary identification results, perform time index matching on them, and establish a time series of lesion evolution; The temperature, humidity, light intensity and soil moisture parameters of the corresponding time nodes are extracted synchronously to construct the multi-dimensional environmental state vector when the lesion occurs; A feature fusion algorithm based on multivariate correlation entropy is used to jointly map image recognition features with environmental state vectors to construct a unified causal feature representation space. Generate a multi-factor evolution sample set including lesion labels, image features, environmental status and evolution trend indicators.

4. The grape disease identification and early warning method based on the Internet of Things according to claim 1, characterized in that: The construction of a spatiotemporal correlation graph model based on historical disease samples and the calculation of the risk path and diffusion probability of disease transmission include: The lesion occurrence locations, timestamps, and environmental parameters in the historical disease evolution sample set are constructed as heterogeneous nodes to form a multi-level spatiotemporal graph data structure; The dynamic graph convolution module in the graph neural network is used to weight the spatial proximity and environmental similarity between nodes to build edges and establish spatiotemporal dependencies; Enhance the characteristics of potential transmission paths and estimate the transmission probability by combining historical transmission frequency with the current environmental status; Output the disease transmission risk path map in the target area and assign a transmission probability distribution value to each node.

5. The method for identifying and warning grape diseases based on the Internet of Things according to claim 4, characterized in that: On the basis of estimating the propagation probability by combining the historical propagation frequency with the current environmental status, it further includes: constructing a dynamic propagation control function, which integrates the local environmental disturbance sensitivity factor and the historical propagation path entropy value, and adaptively adjusts the propagation probability weight of each path through an adjustable Bayesian update mechanism, wherein the local environmental disturbance sensitivity factor is used to quantify the nonlinear impact of environmental micro-changes on the disease propagation speed, and the propagation path entropy value is used to evaluate the uncertainty and information content of the path.

6. The grape disease identification and early warning method based on the Internet of Things according to claim 1 is characterized in that: The edge computing processing and output of disease warning information at the IoT gateway end includes: A lightweight deep inference module is deployed within the IoT gateway to perform local analysis on the received disease spot identification results and propagation probability maps. A multi-factor gated discrimination algorithm is used to comprehensively determine whether the current area meets the preset disease warning threshold. The threshold is dynamically adjusted based on the growth rate of the number of lesions, the rate of increase in the propagation probability, and the intensity of environmental induction. If the warning conditions are met, the spatial hot zone rendering module is called to combine the GIS raster information to mark the boundaries of the suspected disease spread area. The warning information is then sent to the upper-level control platform or farmer terminal to implement a disease warning response.

7. The method for identifying and warning grape diseases based on the Internet of Things according to claim 6, characterized in that: The multi-factor gating discriminant algorithm comprehensively determines whether the disease warning threshold is met, including: constructing a multi-factor input vector that integrates the current lesion identification confidence, propagation probability gradient, environmental parameter volatility, and historical disease high-incidence weights; inputting this vector into a discriminant network equipped with multi-level gating units, which uses a hierarchical gating mechanism to weight different factor channels; combining it with an asynchronous activation strategy to output a "meet warning" discriminant label only when all factors simultaneously meet the set trigger conditions; and binding the discriminant results to a geographic grid index to locate risk areas in the spatial dimension.

Images