A crop disease intelligent diagnosis and decision method and system in smart agriculture

Abstract

The application provides a crop disease intelligent diagnosis and decision method and system in smart agriculture, comprising: based on the standardized remote sensing image data in the unified multi-source data set, spatial features are extracted by using a convolutional neural network to generate preliminary positioning data of disease spot areas, and the preliminary positioning data of the disease spot areas contain disease spot position coordinates and range boundaries; the key time weighted sequence data and the preliminary positioning data of the disease spot areas are input into a space-time attention mechanism, space position coordinates and time environment features are fused through attention weight distribution to generate comprehensive disease feature vector data; based on the preliminary prevention and control prescription data, a drug resistance evaluation model is used to analyze the frequency of use of pesticides, disease types and crop types in historical pesticide data to generate optimized prevention and control prescription data, and the type or dosage of pesticides is adjusted to reduce the risk of drug resistance.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for intelligent diagnosis and decision-making of crop diseases in smart agriculture. Background Technology

[0002] Agricultural pest and disease control is an important area for ensuring food security and sustainable agricultural development, and its effectiveness is directly related to crop yield and farmers' income.

[0003] With the widespread application of remote sensing technology and the Internet of Things, pest and disease monitoring and decision-making based on multi-source data has become a research hotspot.

[0004] However, existing methods often struggle to effectively integrate spatial and temporal information when processing multi-source heterogeneous data, resulting in insufficient diagnostic accuracy.

[0005] For example, remote sensing images can identify lesions in the field, but it is difficult to combine them with the dynamic changes in environmental data, which can easily lead to missed detection of early diseases.

[0006] Furthermore, decision-making recommendations often rely on single data analysis and lack comprehensive reasoning for complex agricultural scenarios, resulting in recommendations that are often inaccurate or lack interpretability.

[0007] In disease diagnosis, the core challenge lies in how to efficiently integrate multi-source data over time series and improve diagnostic accuracy through collaborative analysis of spatial and temporal dimensions.

[0008] In terms of spatial dimension, the identification of lesion areas in remote sensing images requires precise positioning, but the distribution of diseases in the field is often uneven, and the characteristics of key areas are easily interfered with by noise.

[0009] In terms of time, environmental factors such as temperature, humidity, and rainfall have a significant impact on the occurrence of diseases, but existing methods are unable to dynamically assign higher weight to key time points.

[0010] For example, during the peak season for powdery mildew, a sudden increase in humidity may be a key triggering factor, but existing systems are unable to adjust their analysis focus according to such dynamic changes, resulting in early disease signals being ignored.

[0011] The separation of spatial and temporal information results in incomplete diagnostic results, which in turn affects the pertinence of subsequent decision-making.

[0012] At the decision-making level, the lack of a systematic knowledge integration and reasoning mechanism is another key issue.

[0013] In agricultural settings, there are complex relationships between factors such as crops, pests and diseases, climate, and soil. Existing methods often recommend pesticides based solely on the diagnosis of a single disease, ignoring the combined effects of historical data and environmental conditions.

[0014] For example, repeated use of the same pesticide on a plot of land may lead to pesticide resistance, but existing systems struggle to combine historical pesticide use records with current environmental conditions to generate optimized spraying plans.

[0015] Therefore, how to achieve early and high-precision disease diagnosis through spatiotemporal collaborative analysis, and generate accurate and interpretable prevention and control prescriptions based on comprehensive knowledge reasoning, has become a key issue in the intelligent prevention and control of agricultural pests and diseases. Summary of the Invention

[0016] This invention provides a method for intelligent diagnosis and decision-making of crop diseases in smart agriculture, mainly including: Acquire and process multi-source heterogeneous data from remote sensing images, temperature and humidity sensors, and rainfall records to generate a unified multi-source dataset; Convolutional neural networks are used to process the remote sensing image data in the unified multi-source dataset to extract spatial features and generate preliminary location data of lesion areas; Long Short-Term Memory (LSTM) networks were used to analyze time-series data in the unified multi-source dataset to analyze the dynamic changes in the environment and to filter environmental time-series feature vectors related to the location of lesions. The key time nodes in the environmental time series feature vector are weighted to generate key time weighted sequence data. The key time-weighted sequence data and the preliminary location data of the lesion area are input into a spatiotemporal attention mechanism for fusion to generate a comprehensive disease feature vector. Based on the comprehensive disease feature vector, a random forest algorithm is used for classification to generate disease classification results that include disease type and severity score; Based on the disease classification results, a preliminary prevention and control prescription is generated through knowledge graph reasoning. Based on the preliminary prevention and control prescription, a drug resistance assessment model is used to analyze historical medication data and generate an optimized prevention and control prescription to reduce the risk of drug resistance.

[0017] This invention provides an intelligent crop disease diagnosis and decision-making system for smart agriculture, mainly comprising: The data integration and preprocessing module is used to acquire multi-source heterogeneous data from remote sensing images, temperature and humidity sensors, and rainfall records. Through data cleaning, format conversion, and normalization, a unified multi-source data set is generated, which includes standardized remote sensing image data and temperature, humidity, and rainfall time series data. The spatial feature extraction module is used to extract spatial features based on the standardized remote sensing image data in the unified multi-source dataset, using a convolutional neural network to generate preliminary location data of lesion areas, wherein the preliminary location data of lesion areas includes the location coordinates and extent boundaries of the lesions. The environmental dynamic analysis module is used to analyze the dynamic changes of the environment based on the temperature, humidity and rainfall time series data in the unified multi-source dataset, using a long short-term memory network, and combined with the preliminary location data of the lesion area to filter environmental time series feature vectors related to the lesion location coordinates. The key time weighting module is used to assign higher weight to the environmental time series feature vector if the humidity change exceeds the statistical threshold H based on historical data. The statistical threshold H is the historical humidity mean plus twice the standard deviation. This generates key time weighted sequence data. The spatiotemporal attention fusion module is used to input the key time-weighted sequence data and the preliminary location data of the lesion area into the spatiotemporal attention mechanism, and generate comprehensive disease feature vector data by fusion of spatial location coordinates and temporal environmental features through attention weight allocation; The disease classification and severity assessment module is used to classify diseases based on the comprehensive disease feature vector data using a random forest algorithm, and generate disease classification results. The disease classification results include early disease types and a quantitative severity score S based on the boundary of lesion range and the intensity of environmental features. The score S ranges from 0 to 100. The preliminary prevention and control prescription generation module is used to obtain field application data from the historical application record database, combine the disease type and severity score S in the disease classification results, and generate preliminary prevention and control prescription data through a pre-constructed knowledge graph containing the relationship between crops, diseases and pesticides. The preliminary prevention and control prescription data includes recommended pesticides and dosages. The resistance assessment and prescription optimization module is used to analyze the frequency of pesticide use, disease type and crop type in historical pesticide use data based on the preliminary prevention and control prescription data, using a resistance assessment model to generate optimized prevention and control prescription data and adjust the type or dosage of pesticides to reduce the risk of resistance.

[0018] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a method for precise control of crop diseases based on multi-source heterogeneous data. Addressing the challenges of diverse data sources, complex spatiotemporal characteristics, and high risk of pesticide resistance in crop disease monitoring and control, this method integrates remote sensing imagery, temperature and humidity time-series data, and rainfall data to construct a unified dataset. A convolutional neural network is used to extract spatial features of lesions, a long short-term memory network is used to analyze environmental dynamics, and a spatiotemporal attention mechanism is combined to generate a comprehensive disease feature vector. Then, a random forest algorithm is used for disease classification, outputting disease type and severity scores. Simultaneously, a preliminary control prescription is generated through knowledge graph reasoning, and pesticide selection and dosage are optimized based on a pesticide resistance assessment model. This invention achieves a closed-loop process from multi-source data integration to precise control, significantly improving the accuracy of disease identification and control efficiency, reducing the risk of pesticide resistance, and providing an intelligent solution for agricultural production. Attached Figure Description

[0019] Figure 1 This is a flowchart of a smart agriculture crop disease intelligent diagnosis and decision-making method according to the present invention.

[0020] Figure 2 This is a schematic diagram of a smart agriculture crop disease intelligent diagnosis and decision-making method according to the present invention.

[0021] Figure 3 This is another schematic diagram of a smart crop disease intelligent diagnosis and decision-making method in smart agriculture according to the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1-3 This embodiment of a smart agriculture crop disease intelligent diagnosis and decision-making method and system may specifically include: S101. Obtain multi-source heterogeneous data from remote sensing images, temperature and humidity sensors, and rainfall records. Through data cleaning, format conversion, and normalization processing, generate a unified multi-source data set. The unified multi-source data set includes standardized remote sensing image data and temperature, humidity, and rainfall time series data.

[0024] Multi-source heterogeneous data is acquired from remote sensing imagery, temperature and humidity sensors, and rainfall records. Raw data is extracted through a pre-defined data acquisition interface, yielding data records containing timestamped image pixel values, temperature and humidity values, and rainfall amounts. The acquired multi-source heterogeneous data undergoes cleaning processing. If missing or outlier values ​​exist, interpolation or median filtering is used to fill in the missing values ​​and remove outliers, resulting in cleaned multi-source data. The cleaned multi-source data undergoes format conversion processing: remote sensing imagery is converted to a unified raster format, and temperature, humidity, and rainfall records are converted to time-series formats, resulting in uniformly formatted multi-source data. The uniformly formatted multi-source data is then normalized using a min-max normalization method to map data values ​​to a pre-defined range, resulting in normalized multi-source data. Based on the normalized multi-source data, standardized feature vectors from the remote sensing imagery and time-series features of temperature, humidity, and rainfall are extracted. Principal component analysis is used to reduce the dimensionality of these features, resulting in a dimensionality-reduced feature dataset. Using the dimensionality-reduced feature dataset, a unified multi-source dataset is constructed. The dataset is then grouped using the k-means clustering algorithm, resulting in a unified multi-source dataset containing standardized remote sensing image data and time-series data of temperature, humidity, and rainfall. Time-series consistency is verified on the unified multi-source dataset. If discontinuous timestamps are found in the time-series data, missing time points are supplemented using linear interpolation to obtain a final, time-consistent multi-source dataset.

[0025] Specifically, the processing of multi-source heterogeneous data, such as remote sensing imagery, temperature and humidity sensors, and rainfall records, is a crucial technical process in fields like environmental monitoring and agricultural management. The following sections analyze and illustrate each technical topic through specific scenarios, focusing on the optimization needs of agricultural irrigation, demonstrating implementation methods and beneficial effects.

[0026] For example, in agricultural irrigation scenarios, remote sensing imagery could come from infrared farmland images captured by drones, temperature and humidity sensors deployed at key points in the farmland, and rainfall records provided by weather stations. The data acquisition interface needs to be compatible with multiple formats, such as TIFF format for remote sensing imagery, CSV files from sensors, and JSON format for rainfall records. Assume a data acquisition session yields 1000×1000 pixel remote sensing imagery, hourly temperature and humidity data, and daily rainfall records, with timestamps from June 1st to June 30th, 2025. The interface extracts this raw data via API, ensuring timestamp alignment.

[0027] In one possible implementation, data cleaning addresses missing and outlier values. For example, if a specific hour in the temperature and humidity data is missing, it can be filled by interpolating the average of the two hours before and after it. For instance, if the missing temperature of 25°C at 14:00 on June 10th is found, it can be obtained by interpolating the values ​​of 24°C at 13:00 and 26°C at 15:00. Outliers, such as a sudden increase in rainfall to 1000mm, can be removed using median filtering and replaced with the average of approximately 5mm from nearby dates. This process ensures data integrity and improves the accuracy of subsequent analysis.

[0028] Specifically, the format conversion unifies the remote sensing imagery into GeoTIFF format, preserving spatial resolution such as 10 meters per pixel. Temperature, humidity, and precipitation data are converted to time-series CSV format, with timestamps aligned hourly, such as temperature sequences from 00:00 to 23:00 on June 1st as 20°C, 21°C, etc. This unified format facilitates the fusion of multi-source data and reduces processing complexity.

[0029] For example, the normalization process uses min-max normalization, mapping remote sensing image pixel values ​​(0-255), temperatures (15-35°C), and rainfall (0-50mm) to a range of 0-1. For instance, a temperature of 25°C is normalized to 0.5 within the 15-35°C range. This method eliminates dimensional differences, making different data comparable and improving model training performance.

[0030] In one embodiment, feature extraction and dimensionality reduction are combined with the vegetation index NDVI and time-series features of temperature, humidity, and rainfall from remote sensing imagery, such as average temperature and rainfall frequency. Principal component analysis reduces high-dimensional features to principal components, such as retaining three principal components with 90% variance, reducing computational complexity while preserving key information for subsequent clustering.

[0031] Specifically, k-means clustering groups the dimensionality-reduced dataset, assuming farmland is divided into three categories: arid, moderately humid, and humid. Arid areas have low NDVI, high temperatures, and low rainfall; humid areas are the opposite. The clustering results guide irrigation zoning and optimize water resource allocation.

[0032] For example, if a time series consistency check reveals a missing temperature and humidity record for June 15th, it can be filled in using linear interpolation of June 14th and 16th, such as a smooth transition of temperature from 24°C to 26°C. This step ensures the continuity of the time series, providing reliable data for dynamic monitoring.

[0033] Understandably, the above process, through data cleaning, format conversion, normalization, feature dimensionality reduction, clustering, and temporal consistency verification, constructs a high-quality multi-source dataset. This dataset supports precision agricultural management, such as irrigation optimization and crop health monitoring, significantly improving resource utilization efficiency and decision-making accuracy.

[0034] S102. Based on the standardized remote sensing image data in the unified multi-source data set, a convolutional neural network is used to extract spatial features and generate preliminary location data of lesion areas. The preliminary location data of lesion areas includes the location coordinates and range boundaries of the lesions.

[0035] Standardized remote sensing image data is obtained from a multi-source dataset. Preprocessing operations are performed to unify image resolution and format, resulting in a standardized image dataset. A convolutional neural network is used to extract spatial features from the standardized image dataset, generating a spatial feature map containing texture and shape features. Pixel distribution is analyzed using the spatial feature map; if pixel feature values ​​exceed a preset threshold, they are identified as potential lesion regions, resulting in a preliminary lesion region set. For this preliminary lesion region set, a region growing algorithm is used to cluster adjacent pixels, determining the lesion location coordinates and extent boundaries, generating lesion localization data. Boundary features are extracted from the lesion localization data, and morphological processing is used to optimize the extent boundaries, resulting in refined lesion region data. Based on the refined lesion region data, a support vector machine classifier is used to distinguish lesion types, generating classification result data. By comparing the classification result data with the spatial feature map, if the classification result matches the feature map, the final lesion region is determined, generating lesion distribution data.

[0036] For example, when acquiring standardized remote sensing image data from a multi-source dataset, image preprocessing can ensure uniformity in resolution and format. Suppose we are processing a set of farmland remote sensing images; the original data may come from different satellites, with resolutions ranging from 0.5 meters to 2 meters. One possible approach is to use a 1-meter resolution as the standard, employing bilinear interpolation to upsample low-resolution images, while downsampling high-resolution images through pixel aggregation, unifying them into GeoTIFF format. This method ensures spatial scale consistency in the image data, facilitating subsequent analysis.

[0037] In one embodiment, a convolutional neural network is used to extract spatial features. For farmland imagery, a network containing multiple layers of convolutional and pooling layers can be designed to extract texture and shape features.

[0038] For example, a network can identify the texture of spots or the edge shape of lesions on crop leaves, generating a spatial feature map. Assuming the image covers a 1000×1000 pixel area, the feature map may contain edge intensity and texture density information for subsequent lesion localization.

[0039] For example, pixel distribution analysis can identify potential lesion areas using a thresholding method. Assuming the pixel value range in the feature map is 0 to 255, a threshold of 200 can be set; pixels with values ​​higher than 200 are considered potential lesions. This method is based on the fact that lesion areas typically have high-contrast characteristics, such as yellowing areas on leaves. The initial set of lesion areas may contain hundreds of scattered pixels, requiring further clustering.

[0040] In one embodiment, a region growing algorithm is used to cluster adjacent pixels. Starting from a certain pixel, the growth condition is set so that the difference between adjacent pixel values ​​is less than 10, and the region is gradually expanded to form a lesion area. Ultimately, the coordinates of the lesion location can be generated, such as the center point of a region being (500, 600) with a boundary of 50×50 pixels. This method can accurately delineate the shape of the lesion, facilitating subsequent boundary optimization.

[0041] For example, when optimizing lesion boundaries through morphological processing, dilation and erosion operations can be used to smooth the boundaries. Suppose a lesion region has jagged edges; by expanding it by 1 pixel with a single dilation operation and then performing an erosion operation, a smoother boundary contour can be obtained. The refined lesion region data boundaries are clearer, and area calculations are more accurate.

[0042] In one embodiment, a support vector machine classifier is used to distinguish between different types of lesions. Assuming that farmland lesions are classified into two categories—fungal infection and viral infection—boundary features such as shape complexity and color distribution can be extracted and input into the classifier.

[0043] For example, fungal lesions typically have irregular borders and a yellowish color, while viral lesions have more regular borders and a grayish color. The classifier can output the lesion type based on these characteristics.

[0044] For example, when comparing the classification results with the spatial feature map, it can be checked whether the classified fungal lesion areas overlap with high-texture-density areas in the feature map. If the matching degree exceeds 90%, the area is confirmed as the final lesion area, and lesion distribution data is generated. This comparison ensures the reliability of the results and reduces misjudgments. The final lesion distribution data can record the location, type, and area of ​​the lesions, providing a basis for precise application of pesticides.

[0045] S103. Based on the temperature, humidity, and rainfall time series data in the unified multi-source data set, the dynamic change characteristics of the environment are analyzed using a long short-term memory network. Combined with the preliminary location data of the lesion area, environmental time series feature vectors related to the lesion location coordinates are selected.

[0046] A Long Short-Term Memory (LSTM) network was used to process time-series data of temperature, humidity, and rainfall from a unified multi-source dataset, extracting environmental dynamic change feature vectors to obtain a time-series feature set. The spatial coordinates of the time-series feature set were matched with the spatial coordinates of the preliminary lesion location data to calculate the correlation between the feature vectors and the lesion location coordinates, resulting in a correlation coefficient matrix. If the correlation coefficient of a feature vector in the correlation coefficient matrix exceeded a preset threshold, the feature vector was retained, and environmental time-series features highly correlated with the lesion location coordinates were selected, resulting in a filtered feature set. Based on the filtered feature set, principal component analysis was used to reduce the dimensionality of the feature vectors, extracting the main environmental dynamic change features, resulting in a dimensionality-reduced feature set. Using the dimensionality-reduced feature set, a support vector machine (SVM) model was used to classify the environmental features of the lesion areas, determining the environmental dynamic change patterns of the lesion areas, and obtaining classification results. If the classification results indicated a specific environmental dynamic change pattern, a spatial mapping relationship between environmental features and lesion distribution was constructed based on the lesion location coordinates, resulting in a spatial distribution feature map. Based on the spatial distribution feature map, cluster analysis was used to group the lesion areas, determining the environmental feature clustering patterns of the lesion areas, and obtaining the final environmental lesion association analysis results.

[0047] For example, based on time-series data of temperature, humidity, and rainfall from a unified multi-source dataset, Long Short-Term Memory (LSTM) networks can be used to extract dynamic environmental change features. LSM networks, through memory units and gating mechanisms, can effectively capture long-term dependencies in time-series data.

[0048] For example, when processing a set of temperature and humidity data from the past 30 days, with daily temperature ranging from 15-25°C, humidity from 60-80%, and rainfall from 0-20 mm, the network can identify the correlation pattern between sudden temperature drops and increased rainfall, and generate feature vectors containing dynamic trends, such as the rate of temperature drop or the cycle of humidity fluctuations.

[0049] In one possible implementation, the time series feature set is matched with the spatial coordinates of the preliminary location data of the lesion area, and the correlation coefficient matrix is ​​calculated.

[0050] For example, if the coordinates of a lesion location in a certain area are (x1, y1), the corresponding time series features include a sudden increase in humidity and the frequency of rainfall. Through correlation analysis, assuming the correlation coefficient of the humidity feature is 0.85, which is higher than the preset threshold of 0.7, this feature is retained, and environmental features highly correlated with the lesions are selected to form a filter feature set. This screening method can focus on key environmental factors and avoid interference from irrelevant variables.

[0051] Specifically, principal component analysis is used to reduce the dimensionality of the selected feature set and extract the main dynamic environmental change features.

[0052] For example, the original feature set contains 10 dimensions such as temperature, humidity, and rainfall. Through principal component analysis, these features can be compressed to 3 dimensions, retaining 90% of the information variance, mainly reflecting the joint changing trend of humidity and rainfall. This dimensionality reduction method simplifies subsequent calculations and improves classification efficiency.

[0053] For example, by using a dimensionality-reduced feature set, a support vector machine (SVM) model can be employed to classify the environmental characteristics of lesion areas and determine the dynamic change patterns of the environment. Suppose a lesion area exhibits characteristics of high humidity and frequent rainfall; the SVM can classify it as a "humid and hot lesion pattern." The classification results can clearly distinguish the characteristics of lesions under different environmental conditions, such as humidity-dominated or rainfall-dominated patterns.

[0054] In one possible implementation, the spatial mapping relationship between environmental features and lesion distribution is constructed by combining the classification results with the lesion location coordinates.

[0055] For example, in a certain farmland area, the lesion at coordinates (x2, y2) corresponds to a high humidity pattern. Spatial mapping can show the overlap between high humidity areas and densely populated lesion areas, generating a spatial distribution feature map. This mapping can intuitively reflect the spatial impact of environmental factors on lesions.

[0056] For example, based on spatial distribution feature maps, cluster analysis can be used to group lesion areas and determine environmental characteristic clustering patterns. Suppose that K-means clustering is used to divide lesions into three groups: a high humidity group, a frequent rainfall group, and a mixed group. Each group corresponds to different environmental characteristics; for example, lesions in the high humidity group are mostly distributed in low-lying areas. The final environmental lesion association analysis results can reveal the intrinsic relationship between lesion distribution and environmental conditions, providing a basis for precise prevention and control.

[0057] S104. For the environmental time series feature vector, if the humidity change exceeds the statistical threshold H based on historical data, where the statistical threshold H is the historical humidity mean plus twice the standard deviation, then the time node is given a higher weight, and key time weighted sequence data is generated.

[0058] Humidity time-series data is acquired from humidity sensors. Preprocessing is used to extract feature vectors, resulting in an initial feature vector. If the change in the initial feature vector exceeds a preset statistical threshold H, where H is calculated by adding twice the standard deviation to the mean of historical humidity data, this time point is marked as a key node, resulting in a set of key time nodes. Based on this set, a weighting factor is applied to assign higher weights to each key node, generating weighted time-series data. By analyzing the weighted time-series data, the K-means algorithm is used to cluster the data, resulting in humidity change pattern classifications. Based on these classifications, features of significantly changing time nodes are extracted, generating feature-enhanced sequence data. A support vector machine algorithm is used to predict future humidity change trends from the feature-enhanced sequence data. By fusing the weighted time-series data and the predicted humidity change trends, key time-weighted sequence data is generated.

[0059] For example, in the field of agricultural environmental monitoring, humidity sensors are widely used to acquire time-series data on humidity in farmland soil or air to analyze the changing characteristics of the crop growth environment. Humidity sensors record humidity values ​​every 10 minutes, forming continuous time-series data. In the preprocessing stage, feature vectors are extracted using a sliding window method, such as calculating the hourly mean, maximum, and rate of change of humidity, to obtain initial feature vectors. These feature vectors reflect the dynamic change pattern of humidity over time.

[0060] For example, a rapid increase in humidity from 60% to 80% within an hour, indicating a high rate of change, may suggest the presence of rainfall or irrigation events.

[0061] In one possible implementation, the statistical threshold H is calculated based on the statistical characteristics of historical humidity data. Assuming historical data indicates a mean soil moisture of 65% and a standard deviation of 5%, H can be set as 65% + 2 × 5% = 75%. If the initial feature vector at a certain time point shows a humidity change rate exceeding 75%, it is marked as a critical node.

[0062] For example, if the humidity suddenly rises from 70% to 85% at 10:00 AM on a certain day, with a change rate of 15%, exceeding the threshold H, it is marked as a critical node. The set of critical time nodes may include multiple high-change time points within a day, such as the moment of humidity surge after irrigation.

[0063] Specifically, the introduction of weighting factors aims to highlight the importance of key nodes.

[0064] For example, a weight of 2.0 can be assigned to critical nodes, while a weight of 1.0 can be assigned to non-critical nodes. Weighted time series data generates new series by weighting the original humidity data, highlighting humidity changes at critical nodes.

[0065] For example, a humidity value of 85% at 10:00 AM, multiplied by a weight of 2.0, is equivalent to 170%, thus amplifying its influence in subsequent analyses. This weighting method helps to more accurately capture key moments of environmental change.

[0066] For example, when using the K-means algorithm to cluster weighted time series data, the data can be divided into three categories: stable humidity patterns, rapidly rising patterns, and rapidly falling patterns. Suppose that the weighted time series data of a certain farmland shows that the humidity fluctuates between 60% and 70% over a certain period, which is classified as a stable pattern; while after irrigation, the humidity rapidly rises to over 80%, which is classified as a rapidly rising pattern. The clustering results can intuitively reflect the pattern of humidity changes in farmland; for example, the rapidly rising pattern may be highly correlated with irrigation or rainfall.

[0067] In one possible implementation, the time-node features that change significantly are extracted by analyzing the clustering results.

[0068] For example, time points in a rapid increase pattern typically exhibit high rates of change and high humidity values; these features are integrated into feature-enhanced sequence data. Suppose a node has a humidity change rate of 20% and a humidity value of 85%. These enhanced features are then input into a support vector machine (SVM) algorithm. The SVM can predict humidity trends over the next 24 hours; for example, it might predict that the humidity will drop from 75% to 65% the following morning, suggesting a possible lack of rainfall.

[0069] Specifically, when merging weighted time series data and predicted trends to generate key time-weighted series data, the prediction results can be combined with the weighted data.

[0070] For example, forecasts show a future trend of decreasing humidity, while weighted data highlights irrigation events for the day. The key time-weighted sequence data generated by combining the two can be used to guide irrigation planning.

[0071] For example, the system can recommend adding an extra irrigation before a predicted drop in humidity to maintain the necessary moisture for the crops. This integrated approach improves the accuracy of humidity management.

[0072] Using the methods described above, agricultural practitioners can optimize irrigation strategies based on key time-weighted series data. For example, after detecting a rapid increase in humidity, over-irrigation can be avoided by combining predicted trends; while under stable conditions, irrigation frequency can be appropriately reduced. This method ensures the stability of the crop growth environment through multi-dimensional analysis of humidity data.

[0073] S105. Input the key time-weighted sequence data and the preliminary location data of the lesion area into the spatiotemporal attention mechanism, and generate comprehensive disease feature vector data by integrating spatial location coordinates and temporal environmental features through attention weight allocation.

[0074] Key time series data and preliminary lesion area location data are obtained. The data are then standardized and denoised using a preprocessing module to obtain standardized time series data and denoised lesion location data. Based on standardized time series data and denoised lesion location data, a convolutional neural network is used to extract the dynamic features of the time series and the spatial features of the lesion region, resulting in a time dynamic feature set and a spatial feature set. By employing a spatiotemporal attention mechanism, weights are assigned to the temporal dynamic feature set and the spatial feature set, and spatial location coordinates and temporal environmental features are fused to generate a preliminary comprehensive feature vector. If the dimension of the preliminary integrated feature vector exceeds the preset threshold, then the dimension reduction process is carried out by principal component analysis to obtain the dimension-reduced integrated feature vector; Based on the dimensionality-reduced comprehensive feature vector, a support vector machine classifier is used to classify the disease types and obtain the disease classification results; By mapping the disease classification results with spatial location coordinates, spatial visualization data of disease distribution is generated; Based on the spatial visualization data of disease distribution, cluster analysis is used to determine the concentrated areas of disease, and the location data of the concentrated areas of disease are obtained.

[0075] For example, in the field of agricultural disease monitoring, when acquiring key time-series data and preliminary location data of lesion areas, time-series data such as environmental humidity and temperature can be collected by sensors, and the preliminary spatial coordinates of lesions can be obtained by combining them with UAV imagery. Assuming a farmland monitoring system, a humidity sensor records humidity data hourly, continuously collecting data for a week, resulting in 168 time points; UAV imagery provides the latitude and longitude coordinates of the lesion area with an accuracy of 0.5 meters. The preprocessing module standardizes and denoises this data. Standardization normalizes the humidity data to the range of 0 to 1, while denoising removes abnormal fluctuations in the humidity series using a sliding window averaging method. For example, a sudden drop in humidity from 85% to 10% due to equipment vibration is smoothed to the nearest mean of 75%. The denoised lesion location data can eliminate false lesions caused by light interference in the imagery.

[0076] In one possible implementation, a convolutional neural network (CNN) is used to extract features. Time-series data is input into a one-dimensional CNN to capture the dynamic trend of humidity changes over time; for example, a 10% increase in humidity over three consecutive hours may indicate disease risk. Lesion location data is input into a two-dimensional CNN to extract spatial features, such as lesion shape and distribution density. Assuming a lesion area is approximately 2 meters in diameter, the network can identify its edge features and concentration level. A spatiotemporal attention mechanism then assigns weights to these features.

[0077] For example, time points with a sharp increase in humidity and areas with dense lesions are given higher weights, such as a weight value of 0.8, while time points with stable conditions have a weight of 0.2. After fusion, a preliminary comprehensive feature vector is generated, which contains temporal and spatial information.

[0078] It should be noted that if the feature vector dimension is too high, such as exceeding 100 dimensions, principal component analysis can be used to reduce the dimensionality, retaining 90% of the information and reducing it to 20 dimensions to ensure computational efficiency. The dimensionality-reduced feature vector is then input into a support vector machine classifier to distinguish between disease types, such as powdery mildew or rust. Suppose data from a certain farmland indicates that areas with humidity above 80% and a lesion density exceeding 5 lesions per square meter are predominantly powdery mildew. The classification results are mapped to spatial coordinates to generate visualized data on disease distribution, such as a heat map showing the disease concentrated in the northwest corner of the farmland. Cluster analysis further identifies concentrated disease areas; for example, the K-means algorithm identifies three high-density lesion areas with center coordinates of (100, 200), (150, 180), and (120, 190), facilitating precise pesticide application.

[0079] For example, weight allocation in spatiotemporal attention mechanisms can improve the accuracy of disease prediction, while heatmap visualization intuitively shows disease distribution and guides agricultural management. Dimensionality reduction reduces computational burden and ensures real-time monitoring is feasible. Cluster analysis accurately locates disease-concentrated areas and optimizes resource allocation.

[0080] S106. Based on the comprehensive disease feature vector data, a random forest algorithm is used for classification to generate disease classification results. The disease classification results include early disease types and a quantitative severity score S based on the boundary of lesion range and the intensity of environmental features. The score S ranges from 0 to 100.

[0081] After obtaining the comprehensive characteristics of spatiotemporal fusion, this step uses machine learning algorithms to make the final classification judgment. It not only identifies the "type" of the disease, but also quantifies its "severity", thereby providing direct input for differentiated and refined prevention and control decisions.

[0082] For example, in the field of agricultural disease monitoring, when acquiring comprehensive disease feature vector data, the boundary of lesion extent and the intensity of environmental features can be integrated from multi-source data. The boundary of lesion extent can be obtained through aerial images taken by UAVs, including the geometry and boundary coordinates of the lesions; the intensity of environmental features includes data such as soil moisture, temperature, and light intensity. During data preprocessing and cleaning, median filtering can be used to remove noise points in the images, such as removing pixel outliers caused by uneven lighting, to obtain a standardized feature dataset. Suppose that after cleaning, the standardized dataset of lesion images of a certain farmland contains 1000 sample points, each sample point including the boundary coordinates of the lesion and environmental feature values, such as soil moisture of 60% and temperature of 25°C.

[0083] In one possible implementation, the random forest algorithm can be used to train a classification model based on a standardized feature dataset. Random forests construct multiple decision trees to comprehensively determine the type of disease.

[0084] For example, when dealing with wheat rust and powdery mildew, the model can use 100 decision trees to vote, outputting classification results and feature weights, determining that the disease type is rust with an 80% probability. Feature weights show that soil moisture contributes the most to the classification, with a weight of 0.45. This method effectively captures non-linear relationships in the data, improving classification accuracy.

[0085] Specifically, when extracting the boundary features of lesion range, the Canny edge detection algorithm can be used to calculate the lesion area ratio.

[0086] For example, after processing an aerial image and detecting the boundaries of lesions, the area of ​​the lesions is calculated to represent 15% of the total farmland area. This quantitative value directly reflects the extent of disease spread. The calculation of the environmental characteristic weighted value incorporates environmental impact weights.

[0087] For example, if the weights of soil moisture are preset to 0.6, temperature to 0.3, and light intensity to 0.1, the weighted influence coefficient of the environment on the disease is calculated to be 0.75, indicating that environmental factors have a strong influence on the disease.

[0088] For example, if the proportion of lesion area exceeds a preset threshold of 10%, a comprehensive severity score S can be calculated using a linear regression algorithm, incorporating environmental characteristic weighting values. Assuming an input lesion area proportion of 15% and an environmental weighting value of 0.75, the linear regression model outputs a score S of 85, indicating a high degree of disease severity. Based on this score and earlier classification results, a comprehensive disease classification report is generated, clearly identifying the disease type as rust and the severity score as 85. This report provides a basis for precise pesticide application.

[0089] In one possible implementation, the disease distribution is grouped using the K-means clustering algorithm based on the final classification output.

[0090] For example, dividing farmland into 10 regions and cluster analysis revealed three concentrated disease areas, all located in the southwest corner of the field. This spatial distribution pattern can guide targeted control measures, such as increasing the frequency of spraying in the concentrated areas. Such grouping analysis clearly reveals the spatial patterns of diseases, facilitating the optimization of resource allocation.

[0091] S107. Obtain plot application data from the historical application record database, combine the disease type and severity score S in the disease classification results, and generate preliminary prevention and control prescription data by reasoning through a pre-constructed knowledge graph containing the relationship between crops, diseases and pesticides. The preliminary prevention and control prescription data includes recommended pesticides and dosages.

[0092] After diagnosing the disease and assessing its severity, the next step is to generate a scientific control plan. This step utilizes the powerful tool of knowledge graphs to structure expert knowledge in the agricultural field, enabling logical reasoning based on the diagnostic results to generate preliminary, targeted pesticide recommendations.

[0093] For example, in retrieving pesticide application data from a historical pesticide application record database, a database query language can be used to extract a data table containing plot identifiers, crop types, and historical pesticide application records. Suppose plot A is planted with rice, and historical records show that fungicide X and insecticide Y were used twice each in the past three years, with dosages of 1.5 liters / ha and 2 liters / ha respectively. Through data cleaning, records with missing dosages or incomplete timestamps are removed, resulting in a structured plot pesticide application dataset. This dataset provides a reliable foundation for subsequent analysis, ensuring that recommended pesticides are consistent with historical application practices and avoiding the repeated use of pesticides that may lead to resistance.

[0094] In one possible implementation, convolutional neural networks (CNNs) can be used to extract disease features when processing disease images or sensor data. Suppose we input an image of a rice leaf; the network identifies brown spots on the leaf and, based on predefined disease classification rules, identifies it as rice blast, with a score S of 75, exceeding the predefined disease response threshold of 60. The CNN extracts the texture and color features of the lesions through multiple convolutional and pooling operations, ensuring accurate classification results. This method can quickly extract key information from complex images, providing a basis for precise disease control.

[0095] Specifically, when querying candidate pesticides from a crop-disease-pesticide knowledge graph, the knowledge graph can store the relationships between rice, rice blast, and fungicides X and Z. The query results show that both fungicides X and Z are applicable to rice blast, but based on historical application records, fungicide X has a higher usage frequency and more stable efficacy. A weighted ranking algorithm calculates the priority of X as 0.8 and Z as 0.6 based on efficacy scores and usage frequency, recommending fungicide X. This method utilizes the structured relationships of the knowledge graph to improve recommendation efficiency, while also reducing the risk of pesticide resistance by incorporating historical data.

[0096] For example, when predicting pesticide dosage, a linear regression algorithm can predict a dosage of fungicide X of 1.8 liters per hectare based on a severity score S=75 and the crop characteristics of rice. The algorithm references historical dosage data and environmental factors, such as the impact of humidity on rice blast, to ensure the dosage is appropriate for the current disease state. This method optimizes pesticide use, reducing waste and environmental pollution.

[0097] In one possible implementation, after generating preliminary prevention and control prescription data, its completeness and consistency need to be verified through data validation rules. For example, the system checks whether the prescription data for plot A includes the plot identifier, disease type, recommended pesticide, and dosage, and whether the dosage value is within a reasonable range (e.g., 0.5-2.5 liters / hectare). If a dosage is found to be missing, the system marks it as invalid and regenerates it. This validation mechanism ensures the reliability of the prescription data and facilitates actual pesticide application. It is understandable that the prevention and control prescription dataset can be stored using a relational database to ensure data traceability.

[0098] For example, once the prescription data for plot A is stored, it can be used for subsequent disease tracking and effectiveness evaluation. This storage method facilitates data management and analysis, providing support for long-term disease prevention and control.

[0099] Preferably, the above method integrates historical data, image analysis, and knowledge graphs to achieve a closed-loop process from disease identification to pesticide recommendation. Each step supports the others, ensuring that the recommended solutions are accurate and efficient, and applicable to agricultural disease control scenarios.

[0100] S108. Based on the preliminary prevention and control prescription data, the drug resistance assessment model is used to analyze the frequency of drug use, disease type and crop type in the historical drug use data, generate optimized prevention and control prescription data, and adjust the type or dosage of drug to reduce the risk of drug resistance.

[0101] The initial prevention and control prescription may not have fully considered the problem of drug resistance caused by long-term use. This step is the last critical line of defense in the entire decision-making process. By analyzing historical drug use patterns, it proactively assesses and avoids the risk of drug resistance, optimizes the prescription, and reflects the concept of sustainable agricultural production.

[0102] In one possible implementation, the process of obtaining pesticide application frequency, disease type, and crop type from historical pesticide application data and constructing a structured dataset can be based on database queries and data cleaning.

[0103] For example, suppose a plot of land is planted with rice. Historical records show that pesticide A was used 10 times and pesticide B was used 5 times in the past three years, and the diseases affected included rice blast and sheath blight. The dataset can be organized into a table, including plot number, rice type, pesticide name, number of applications, and corresponding disease type. This structured approach facilitates the analysis of pesticide application patterns and intuitively reflects pesticide usage preferences and disease occurrence patterns.

[0104] For example, when using the random forest algorithm to classify pesticide application patterns, pesticide application frequency, disease type, and crop type can be used as feature inputs, and the output is a resistance risk level. Assuming the input data is a rice paddy with a high incidence of rice blast and pesticide A being used 70% of the time, the random forest model, through analysis of multiple decision trees, determines that pesticide A may pose a high risk of resistance. The model will then integrate historical pesticide effects and disease characteristics, outputting a "high" risk level and triggering the selection of alternative pesticides.

[0105] In one embodiment, when screening alternative agents from a drug database, screening criteria can be set according to the disease type and crop type.

[0106] For example, for rice blast disease, the database might contain two alternative options: pesticide C and pesticide D. During the screening process, pesticides with different mechanisms of action than pesticide A should be given priority to reduce the risk of resistance. After the candidate pesticide list is generated, further analysis based on historical control effects is necessary.

[0107] Specifically, linear regression models can be used to evaluate the control efficacy of candidate pesticides. Assume that pesticide C achieved control success rates of 85%, 80%, and 90% against rice blast in the past three years, while pesticide D achieved 70%, 75%, and 80%. Through linear regression analysis, combined with rice characteristics and the severity of rice blast, the model predicts that pesticide C has a superior control effect and recommends it as the optimal pesticide.

[0108] In one possible implementation, when optimizing drug dosage using a grid search algorithm, a dosage range can be set and iterative testing can be performed using historical data.

[0109] For example, the standard dosage of pesticide C is 100 ml / mu. Grid search, by trying dosage schemes of 80, 100, and 120 ml / mu, combined with historical control effects and rice growth stages, determined that 100 ml / mu was the optimal dosage. This method ensures that the dosage is both effective and economical.

[0110] For example, when generating optimized control prescription data, the plot number, rice blast disease information, pesticide C, and a dosage of 100 ml / acre are integrated to form a complete prescription. The prescription data can be presented through a visual interface for easy reference by farmers. When evaluating the control effect, historical pesticide application data can be combined to analyze the actual effect of pesticide C on similar plots.

[0111] For example, historical records show that Agent C achieved an 88% success rate in controlling rice blast in high-incidence areas, proving the reliability of the recommended approach.

[0112] It should be noted that environmental factors, such as rainfall and temperature, can also be incorporated when assessing the effectiveness of control measures. For example, if a plot of land has recently experienced frequent rainfall, the application time of pesticide C should be adjusted to after the rainfall to ensure efficacy. This comprehensive analysis ensures that the prescription data better reflects actual needs and improves the precision of control measures.

[0113] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for intelligent diagnosis and decision-making of crop diseases in smart agriculture, characterized in that, The method includes: Acquire and process multi-source heterogeneous data from remote sensing images, temperature and humidity sensors, and rainfall records to generate a unified multi-source dataset; Convolutional neural networks are used to process the remote sensing image data in the unified multi-source dataset to extract spatial features and generate preliminary location data of lesion areas; Long Short-Term Memory (LSTM) networks were used to analyze time-series data in the unified multi-source dataset to analyze the dynamic changes in the environment and to filter environmental time-series feature vectors related to the location of lesions. The key time nodes in the environmental time series feature vector are weighted to generate key time weighted sequence data. The key time-weighted sequence data and the preliminary location data of the lesion area are input into a spatiotemporal attention mechanism for fusion to generate a comprehensive disease feature vector. Based on the comprehensive disease feature vector, a random forest algorithm is used for classification to generate disease classification results that include disease type and severity score; Based on the disease classification results, a preliminary prevention and control prescription is generated through knowledge graph reasoning. Based on the preliminary prevention and control prescription, a drug resistance assessment model is used to analyze historical medication data and generate an optimized prevention and control prescription to reduce the risk of drug resistance. For the environmental time series feature vector, if the humidity change exceeds a statistical threshold H based on historical data, where the statistical threshold H is the historical humidity mean plus twice the standard deviation, then a higher weight is assigned to that time point, generating key time-weighted series data, including: Humidity time series data is acquired from a humidity sensor, and feature vectors are extracted through preprocessing to obtain an initial feature vector; If the change in the initial feature vector exceeds the preset statistical threshold H, where H is calculated by adding twice the standard deviation to the mean of historical humidity data, then the time node is marked as a key node, and a set of key time nodes is obtained. Based on the set of key time nodes, a weighting factor is used to assign higher weights to each key node to generate weighted time series data. By analyzing weighted time series data, the K-means algorithm was used to cluster the data to obtain a classification of humidity change patterns; Based on humidity change patterns, significant time-point features are extracted to generate feature-enhanced sequence data. The support vector machine algorithm is used to predict the future humidity change trend of the feature-enhanced sequence data; By fusing weighted time series data and predicted humidity change trends, key time-weighted series data are generated; Based on the preliminary control prescription data, a resistance assessment model is used to analyze the frequency of pesticide use, disease type, and crop type in historical pesticide application data to generate optimized control prescription data. This involves adjusting pesticide types or dosages to reduce the risk of resistance, including: By extracting pesticide application frequency, disease type, and crop type from historical pesticide application data, a structured dataset is constructed to obtain the distribution of pesticide application patterns. The random forest algorithm is used to classify the distribution of pesticide application patterns, and the pesticide resistance risk level is determined by combining the disease type and crop type. If the resistance risk level is higher than the preset threshold, alternative pesticides are selected from the pre-established pesticide database according to the disease type and crop type to obtain a list of candidate pesticides; The historical control effects of each pesticide in the candidate pesticide list were analyzed by linear regression model, and the optimal pesticide type was determined by combining crop type and disease type. Based on the historical application records of the optimal pesticide type and the disease type, a grid search algorithm is used to optimize the pesticide dosage and obtain the adjusted dosage scheme. Based on the adjusted dosage regimen, and taking into account crop type and disease type, optimized prevention and control prescription data are generated; The types and dosages of drugs are extracted from the optimized prevention and control prescription data, and combined with historical medication data to evaluate the prevention and control effect and obtain the prevention and control effect evaluation results.

2. The intelligent diagnosis and decision-making method for crop diseases in smart agriculture according to claim 1, characterized in that, The process involves acquiring multi-source heterogeneous data from remote sensing images, temperature and humidity sensors, and rainfall records. Through data cleaning, format conversion, and normalization, a unified multi-source data set is generated. This unified multi-source data set includes standardized remote sensing image data and time-series data of temperature, humidity, and rainfall, including: Multi-source heterogeneous data is acquired from remote sensing images, temperature and humidity sensors, and rainfall records. Raw data is extracted through a preset data acquisition interface to obtain data records containing image pixel values ​​with timestamps, temperature and humidity values, and rainfall. The acquired multi-source heterogeneous data is cleaned. If there are missing or outlier values ​​in the data, interpolation or median filtering is used to fill in the missing values ​​and remove the outliers to obtain the cleaned multi-source data. After cleaning, the multi-source data is converted into a unified raster format, and the temperature and humidity data and rainfall record data are converted into a time series format, resulting in multi-source data with a unified format. Normalize multi-source data with uniform format by using the minimum-maximum normalization method to map data values ​​to a preset range, and obtain normalized multi-source data. Based on the normalized multi-source data, the standardized feature vectors of remote sensing images and the time series features of temperature, humidity and rainfall are extracted. Principal component analysis is used to reduce the dimensionality of the features to obtain the dimensionality-reduced feature dataset. Using the dimensionality-reduced feature dataset, a unified multi-source dataset is constructed. The k-means clustering algorithm is used to group the dataset, resulting in a unified multi-source dataset containing standardized remote sensing image data and temperature, humidity and rainfall time series data. A time series consistency check is performed on the unified multi-source dataset. If the timestamps of the time series data are found to be discontinuous, the missing time points are supplemented by linear interpolation to obtain the final time-consistent multi-source dataset.

3. The intelligent diagnosis and decision-making method for crop diseases in smart agriculture according to claim 1, characterized in that, Based on the standardized remote sensing image data in the unified multi-source dataset, a convolutional neural network is used to extract spatial features and generate preliminary location data for lesion areas. This preliminary location data includes the lesion location coordinates and extent boundaries, including: Standardized remote sensing image data is obtained from a multi-source dataset, and the image resolution and format are unified through preprocessing operations to obtain a standardized image dataset. A convolutional neural network is used to extract spatial features from a standardized image dataset, generating a spatial feature map containing texture and shape features. By analyzing the pixel distribution through spatial feature maps, if the pixel feature value exceeds a preset threshold, it is identified as a potential lesion area, and a preliminary set of lesion areas is obtained. For the initial set of lesion regions, a region growing algorithm is used to cluster adjacent pixels to determine the location coordinates and extent boundaries of the lesions, thereby generating lesion location data. Boundary features are extracted from lesion location data, and the range boundary is optimized through morphological processing to obtain refined lesion region data; Based on the refined lesion area data, a support vector machine classifier is used to distinguish lesion types and generate classification result data. By comparing the classification results with the spatial feature map, if the classification results match the feature map, the final lesion area is determined and lesion distribution data is generated.

4. The intelligent diagnosis and decision-making method for crop diseases in smart agriculture according to claim 1, characterized in that, Based on the temperature, humidity, and rainfall time-series data in the unified multi-source dataset, a long short-term memory network is used to analyze the dynamic changes in the environment. Combined with the preliminary location data of the lesion areas, environmental time-series feature vectors related to the lesion location coordinates are selected, including: Long Short-Term Memory (LSTM) networks are used to process temperature, humidity, and rainfall time series data from a unified multi-source dataset, extracting environmental dynamic change feature vectors to obtain a time series feature set. By matching the spatial coordinates of the time series feature set with the preliminary location data of the lesion area, the correlation between the feature vector and the lesion location coordinates is calculated to obtain the correlation coefficient matrix. If the correlation coefficient of a certain feature vector in the correlation coefficient matrix exceeds the preset threshold, then the feature vector is retained, and environmental time series features that are highly correlated with the location coordinates of lesions are selected to obtain the selected feature set. Based on the selected feature set, principal component analysis is used to reduce the dimensionality of the feature vectors and extract the main dynamic environmental change features to obtain the dimensionality-reduced feature set. By using a dimensionality-reduced feature set, a support vector machine model is employed to classify the environmental characteristics of the lesion area, determine the dynamic environmental change pattern of the lesion area, and obtain the classification result. If the classification results indicate a specific dynamic change pattern in the environment, then by combining the coordinates of the lesion location, a spatial mapping relationship between environmental characteristics and lesion distribution is constructed to obtain a spatial distribution feature map; Based on the spatial distribution feature map, cluster analysis was used to group the lesion areas, determine the environmental feature clustering pattern of the lesion areas, and obtain the final environmental lesion association analysis results.

5. The intelligent diagnosis and decision-making method for crop diseases in smart agriculture according to claim 1, characterized in that, The key time-weighted sequence data and the preliminary location data of the lesion area are input into a spatiotemporal attention mechanism. Through attention weight allocation, spatial location coordinates and temporal environmental features are fused to generate comprehensive disease feature vector data, including: Key time series data and preliminary lesion area location data are obtained. The data are then standardized and denoised using a preprocessing module to obtain standardized time series data and denoised lesion location data. Based on standardized time series data and denoised lesion location data, a convolutional neural network is used to extract the dynamic features of the time series and the spatial features of the lesion region, resulting in a time dynamic feature set and a spatial feature set. By employing a spatiotemporal attention mechanism, weights are assigned to the temporal dynamic feature set and the spatial feature set, and spatial location coordinates and temporal environmental features are fused to generate a preliminary comprehensive feature vector. If the dimension of the preliminary integrated feature vector exceeds the preset threshold, then the dimension reduction process is carried out by principal component analysis to obtain the dimension-reduced integrated feature vector; Based on the dimensionality-reduced comprehensive feature vector, a support vector machine classifier is used to classify the disease types and obtain the disease classification results; By mapping the disease classification results with spatial location coordinates, spatial visualization data of disease distribution is generated; Based on the spatial visualization data of disease distribution, cluster analysis is used to determine the concentrated areas of disease, and the location data of the concentrated areas of disease are obtained.

6. The intelligent diagnosis and decision-making method for crop diseases in smart agriculture according to claim 1, characterized in that, Based on the comprehensive disease feature vector data, a random forest algorithm is used for classification to generate disease classification results. These results include early disease types and a quantified severity score S based on the boundary of lesion extent and the intensity of environmental features. The score S ranges from 0 to 100 and includes: Obtain comprehensive disease feature vector data, including lesion range boundaries and environmental feature intensity. Through preprocessing and cleaning the data, noise data is removed to obtain a standardized feature dataset. Based on the standardized feature dataset, a random forest algorithm is used to train and construct a classification model, outputting the early disease type classification results and initial feature weights, and determining the disease type identification results; Boundary feature parameters are extracted from the boundary of the lesion range, and the lesion area ratio is calculated by the edge detection algorithm to obtain the quantified value of the lesion range; Based on the intensity of environmental characteristics and combined with the preset environmental impact weights, the weighted impact coefficient of the environment on the disease is calculated to obtain the weighted value of the environmental characteristics. If the quantified value of the lesion area is greater than the preset threshold, the comprehensive severity score S is calculated by combining the environmental feature weighting value and using a linear regression algorithm to obtain the quantified score result. By quantifying the score results and the early disease type classification results, a comprehensive disease classification report containing disease type and severity score S is generated, and the final classification output is determined. Based on the final classification output, a clustering algorithm is used to group the disease distribution to obtain the spatial distribution pattern of the disease.

7. The intelligent diagnosis and decision-making method for crop diseases in smart agriculture according to claim 1, characterized in that, Plot application data is obtained from a historical pesticide application record database. Combined with the disease type and severity score S from the disease classification results, preliminary control prescription data is generated using a pre-constructed knowledge graph containing the relationships between crops, diseases, and pesticides. This preliminary control prescription data includes recommended pesticides and dosages, including: The pesticide application data of the plots is obtained from the historical pesticide application record database, including plot identification, crop type and historical pesticide application records, to obtain a structured plot pesticide application dataset; The disease classification model processes the input disease images or sensor data, uses a convolutional neural network algorithm to extract disease features, and combines the preset disease classification rules to determine the disease type and severity score S. If the severity score S exceeds the preset threshold T, then query the pre-built crop-disease-pesticide knowledge graph for pesticide nodes that match the disease type and crop type to obtain a list of candidate pesticides. Otherwise, mark it as a low-risk disease and generate prescription data without recommended pesticides; Based on the list of candidate drugs, and combined with the drug usage frequency and effect data in historical drug use records, a weighted sorting algorithm is used to calculate the recommendation priority of each drug, and the recommended drug with the highest priority is obtained. By using the pesticide-dosage relationship in the knowledge graph, combined with the severity score S and crop type, a linear regression algorithm is used to predict the appropriate pesticide dosage and generate dosage data. By integrating recommended pesticides and dosage data, preliminary prevention and control prescription data are generated, which includes plot identification, disease type, recommended pesticides and dosage data; Preliminary prevention and control prescription data is stored in a database, and the integrity and consistency of the prescription data are verified using data validation rules to obtain the final prevention and control prescription dataset.

8. A smart crop disease intelligent diagnosis and decision-making system for smart agriculture, applicable to the smart crop disease intelligent diagnosis and decision-making method for smart agriculture as described in claim 1, characterized in that, The system includes: The data integration and preprocessing module is used to acquire multi-source heterogeneous data from remote sensing images, temperature and humidity sensors, and rainfall records. Through data cleaning, format conversion, and normalization, a unified multi-source data set is generated, which includes standardized remote sensing image data and temperature, humidity, and rainfall time series data. The spatial feature extraction module is used to extract spatial features based on the standardized remote sensing image data in the unified multi-source dataset, using a convolutional neural network to generate preliminary location data of lesion areas, wherein the preliminary location data of lesion areas includes the location coordinates and extent boundaries of the lesions. The environmental dynamic analysis module is used to analyze the dynamic changes of the environment based on the temperature, humidity and rainfall time series data in the unified multi-source dataset, using a long short-term memory network, and combined with the preliminary location data of the lesion area to filter environmental time series feature vectors related to the lesion location coordinates. The key time weighting module is used to assign higher weights to time nodes based on the environmental time series feature vector if the humidity change exceeds the statistical threshold H based on historical data. The statistical threshold H is the historical humidity mean plus twice the standard deviation, and the key time weighting sequence data is generated. The spatiotemporal attention fusion module is used to input the key time-weighted sequence data and the preliminary location data of the lesion area into the spatiotemporal attention mechanism, and generate comprehensive disease feature vector data by fusion of spatial location coordinates and temporal environmental features through attention weight allocation; The disease classification and severity assessment module is used to classify diseases based on the comprehensive disease feature vector data using a random forest algorithm, and generate disease classification results. The disease classification results include early disease types and a quantitative severity score S based on the boundary of lesion range and the intensity of environmental features. The score S ranges from 0 to 100. The preliminary prevention and control prescription generation module is used to obtain field application data from the historical application record database, combine the disease type and severity score S in the disease classification results, and generate preliminary prevention and control prescription data through a pre-constructed knowledge graph containing the relationship between crops, diseases and pesticides. The preliminary prevention and control prescription data includes recommended pesticides and dosages. The resistance assessment and prescription optimization module is used to analyze the frequency of pesticide use, disease type and crop type in historical pesticide use data based on the preliminary prevention and control prescription data, using a resistance assessment model to generate optimized prevention and control prescription data and adjust the type or dosage of pesticides to reduce the risk of resistance.

Images