An agricultural pest control decision-making method

Abstract

The application discloses an agricultural disease and pest control decision method, comprising: acquiring disease image data of a target crop and corresponding environment data and user constraint data; performing alignment processing on the disease image data and the environment data, and respectively extracting image features and environment time sequence features; performing cross-modal fusion on the image features and the environment time sequence features to obtain fusion features; determining a disease identification result including a disease type and a severity based on the fusion features; matching at least one candidate control scheme according to the disease identification result; establishing a cost model for each candidate control scheme to determine a total cost; constructing a multi-objective optimization model based on the total cost, a drug resistance risk and an environmental toxicity, and optimizing in combination with the user constraint data to obtain at least one recommended control scheme. The application realizes optimization and decision of a control scheme based on a disease identification result.

Description

Technical Field

[0001] This application relates to the fields of smart agriculture and agricultural information technology, and in particular to a decision-making method for the prevention and control of agricultural pests and diseases. Background Technology

[0002] With the development of smart agriculture and agricultural information technology, agricultural pest and disease control technologies based on image recognition, environmental monitoring, and intelligent analysis are increasingly being applied to field management, agricultural service platforms, and mobile terminals. The occurrence of agricultural pests and diseases is typically related not only to the lesion characteristics on crop leaves, stems, or fruits, but also to environmental factors such as temperature, humidity, light, rainfall, and soil moisture. Furthermore, in actual control processes, farmers or managers usually need to consider budget constraints, pesticide usage requirements, and planting management conditions when selecting pest and disease control solutions. Therefore, how to generate reasonable control plans by combining crop disease image information, environmental information, and user constraints has become an important research direction in the field of smart agriculture.

[0003] In related technologies, existing agricultural pest and disease control systems can typically identify crop disease images and integrate environmental monitoring data and agricultural input information. However, most of these systems adopt a "first identify, then fuse" approach. This means they first independently identify pests and diseases based on images, and then use rules to stitch together environmental data, agricultural input price data, or user demand information. This lack of a unified data alignment and cross-modal fusion mechanism makes it difficult to collaboratively analyze image features, environmental temporal characteristics, and constraint information, thus failing to accurately reflect the true disease status in complex field scenarios. Furthermore, the control results provided by existing systems are usually single recommendations or simple ranking lists. Their generation process relies heavily on empirical rules or single indicators, rarely incorporating factors such as plot area, agricultural input prices, labor costs, pesticide resistance risks, environmental toxicity, user budget limits, and prohibited pesticide requirements into a unified decision-making process. This results in insufficient matching between the output control plan and actual constraints, affecting the plan's practical feasibility.

[0004] Therefore, in agricultural pest and disease control decision-making, the insufficient ability to integrate multi-source data and the poor feasibility of control solutions under user constraints have become urgent problems to be solved. Summary of the Invention

[0005] This application provides a decision-making method for agricultural pest and disease control, aiming to solve the problems of insufficient multi-source data fusion capability and poor executability of control schemes under user constraints in the decision-making of agricultural pest and disease control.

[0006] This invention discloses a decision-making method for agricultural pest and disease control, the method comprising: Acquire disease image data of the target crop, as well as the corresponding environmental data and user constraint data of the disease image data; The disease image data and the environmental data are aligned, and image features and environmental temporal features are extracted respectively. The image features and the environmental temporal features are fused across modally to obtain fused features; The disease identification result is determined based on the fusion features, and the disease identification result includes disease type and severity. At least one candidate control scheme is matched based on the disease identification results; For each of the candidate prevention and control measures, a cost model is established based on the plot area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control measures. A multi-objective optimization model is constructed based on the total cost, drug resistance risk, and environmental toxicity of each candidate control scheme. Using the user constraint data as constraints, the candidate prevention and control schemes are optimized based on the multi-objective optimization model to obtain and output at least one recommended prevention and control scheme.

[0007] Optionally, in the above scheme, acquiring the disease image data of the target crop, as well as the environmental data and user constraint data corresponding to the disease image data, includes: The disease image data of the target crop is collected by mobile terminal or agricultural drone, and the timestamp and location information corresponding to the disease image data are obtained. Environmental data of the target crop's location is collected using field environmental sensors. The environmental data includes at least one of temperature, humidity, light intensity, rainfall, and soil moisture. Obtain at least one of the user-defined budget limit, list of banned pesticides, and organic farming requirements as the user constraint data.

[0008] Optionally, in the above scheme, the step of aligning the disease image data and the environmental data, and extracting image features and environmental temporal features respectively, includes: The environmental data is matched based on the timestamps corresponding to the disease image data to obtain the target environmental data corresponding to the disease image data. The disease image data is scaled and normalized to obtain the processed disease image data. The target environment data is interpolated to form a fixed-length environmental time series. The processed disease image data is input into a convolutional neural network to extract the image features; The environmental time series is input into a recurrent neural network to extract the environmental time series features.

[0009] Optionally, in the above scheme, the cross-modal fusion of the image features and environmental temporal features to obtain fused features includes: The image features and the environmental temporal features are input into a cross-modal attention mechanism, and the environmental temporal features are weighted based on the correlation between the image features and the environmental temporal features. The weighted environmental temporal features are fused with the image features to obtain the fused features.

[0010] Optionally, in the above scheme, determining the disease identification result based on the fusion features includes: The fused features are input into the disease identification model; The disease type is output through the classification branch of the disease identification model; The severity of the disease is output by the evaluation branch of the disease identification model, corresponding to the disease type. The confidence level is determined based on the predicted probability of each disease type by the disease identification model. The disease type, severity, and confidence level are correlated to generate disease identification results.

[0011] Optionally, in the above scheme, the step of establishing a cost model based on land area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control scheme includes: Obtain the current unit price of agricultural inputs involved in each candidate prevention and control plan; Obtain the land area and labor cost parameters of the target plot; Obtain the number of applications corresponding to each candidate control program; Based on the current unit price, the land area, the labor cost parameters, and the number of applications, a cost model is established for each candidate prevention and control scheme.

[0012] Optionally, in the above scheme, establishing a cost model corresponding to each candidate prevention and control scheme based on the current unit price, the land area, the labor cost parameter, and the number of applications includes: Establish a cost function C = a × Pppp + b × Plab + c × A × D, where C represents the total cost corresponding to the candidate control scheme, Ppppppppplab ... Obtain the agricultural input price, labor cost, plot area, and number of applications for each candidate prevention and control scheme; Substitute the agricultural input price, labor cost, plot area, and number of applications corresponding to each candidate control scheme into the cost function to determine the total cost corresponding to each candidate control scheme.

[0013] Optionally, in the above scheme, the construction of a multi-objective optimization model based on the total cost, drug resistance risk, and environmental toxicity of each candidate control scheme includes: Construct a multi-objective optimization function with the objectives of minimizing total cost, minimizing drug resistance risk, and minimizing environmental toxicity; The resistance risk is determined based on the pesticide rotation coefficient, and the environmental toxicity is determined based on the pesticide median lethal dose and residual period.

[0014] Optionally, in the above scheme, the step of using the user constraint data as a constraint condition to optimize the candidate prevention and control scheme based on the multi-objective optimization model, and obtaining and outputting at least one recommended prevention and control scheme, includes: The constraints of ensuring that the total cost does not exceed the user's budget and that banned pesticides are not used are input into the multi-objective optimization model. The NSGA-II algorithm is used to iteratively solve the candidate prevention and control schemes to obtain the Pareto optimal solution set; At least one recommended prevention and control scheme is determined from the Pareto optimal solution set, and the recommended prevention and control scheme is output through the application interface or web interface.

[0015] Optionally, the method further includes: Receive user feedback on the implementation effect of the recommended prevention and control plan; The disease image data, the environmental data, the user constraint data, and the execution effect feedback are used as new samples; Based on the new samples, the local model parameters used for disease identification and / or prevention decisions are updated through federated learning, and model gradients are generated. The model gradients are encrypted and uploaded to the cloud for aggregation.

[0016] Compared with the prior art, this application has at least the following beneficial effects: This application, based on further analysis and research into the problems of existing technologies, recognizes the shortcomings of current technologies in agricultural pest and disease control decision-making, such as insufficient multi-source data fusion capabilities and poor executability of control schemes under user constraints. By acquiring disease image data of the target crop, along with corresponding environmental data and user constraint data, the information required for disease diagnosis is no longer limited to a single image input. Instead, it simultaneously incorporates environmental information related to disease occurrence and development, as well as user constraint information related to control implementation. Furthermore, by aligning disease image data and environmental data and extracting image features and environmental temporal features separately, data from different sources and in different formats can be transformed into feature expressions that can be processed uniformly. Then, by performing cross-modal fusion of image features and environmental temporal features, a fused feature is obtained. This fused feature simultaneously includes both the apparent characteristics of crop lesions and environmental change characteristics. Finally, based on the fused feature, the disease identification result is determined, ensuring that the disease type and severity are no longer determined solely by a single image, but rather based on the collaborative analysis of image and environmental information. Based on this, the present application can solve the problems of insufficient multi-source data fusion capability and difficulty in accurately reflecting the real disease status in complex field scenarios in the background technology. Furthermore, after obtaining the disease identification results, the present application matches at least one candidate control scheme based on the disease identification results, and establishes a cost model for each candidate control scheme based on the plot area, agricultural input price and labor cost to determine the total cost corresponding to the candidate control scheme. Then, based on the total cost of each candidate control scheme, the drug resistance risk and environmental toxicity, a multi-objective optimization model is constructed, and user constraint data is used as constraints. The candidate control scheme is optimized based on the multi-objective optimization model to obtain and output at least one recommended control scheme. Since the candidate control scheme has been jointly screened by cost factors, risk factors and user constraints before output, the output result is no longer a single experience recommendation that is detached from actual use conditions, but a control decision result that matches the actual needs such as user budget and drug use restrictions. This can solve the problem of insufficient matching between existing control schemes and actual constraints in the background technology, resulting in poor scheme executability. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an agricultural pest and disease control decision-making method provided in one embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1As shown, an agricultural pest and disease control decision-making method is provided, including the following steps: Acquire disease image data of the target crop, as well as the corresponding environmental data and user constraint data of the disease image data; The disease image data and the environmental data are aligned, and image features and environmental temporal features are extracted respectively. The image features and the environmental temporal features are fused across modally to obtain fused features; The disease identification result is determined based on the fusion features, and the disease identification result includes disease type and severity. At least one candidate control scheme is matched based on the disease identification results; For each of the candidate prevention and control measures, a cost model is established based on the plot area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control measures. A multi-objective optimization model is constructed based on the total cost, drug resistance risk, and environmental toxicity of each candidate control scheme. Using the user constraint data as constraints, the candidate prevention and control schemes are optimized based on the multi-objective optimization model to obtain and output at least one recommended prevention and control scheme.

[0020] In this embodiment, an agricultural pest and disease control decision-making method includes: acquiring disease image data of a target crop, as well as environmental data and user constraint data corresponding to the disease image data; aligning the disease image data and the environmental data, and extracting image features and environmental temporal features respectively; performing cross-modal fusion on the image features and the environmental temporal features to obtain fused features; determining a disease identification result based on the fused features, the disease identification result including disease type and severity; matching at least one candidate control scheme according to the disease identification result; establishing a cost model for each candidate control scheme based on plot area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate control scheme; constructing a multi-objective optimization model based on the total cost, pesticide resistance risk, and environmental toxicity of each candidate control scheme; using the user constraint data as constraints, optimizing the candidate control scheme based on the multi-objective optimization model to obtain and output at least one recommended control scheme. The method can generally be executed by a server, agricultural service platform, or cloud processing system. Its processing flow corresponds to the process of "data collection and uploading, cross-modal feature fusion and disease identification, disease diagnosis result generation, dynamic cost model triggering, and multi-objective optimization and scheme generation".

[0021] Specifically, when acquiring disease image data, users can use a mobile application to take pictures of crop lesions, or agricultural drones equipped with cameras can collect images of crop leaves, stems, or fruits; simultaneously, the system automatically records the timestamps and GPS location information corresponding to the images. When acquiring environmental data, environmental sensors deployed in the field can collect at least one parameter among temperature, humidity, light intensity, rainfall, and soil moisture, and upload the environmental data for a pre-set time period to the server. When acquiring user constraint data, users can input personalized parameters such as budget limits, lists of prohibited pesticides, and organic farming requirements on the app or web interface, or directly read them from the user constraint database. The aforementioned disease image data, environmental data, and user constraint data constitute the raw input for subsequent identification and optimization processing.

[0022] Furthermore, when aligning disease image data and environmental data, the timestamp of the disease image data can be used as a benchmark to extract data segments from the environmental data that correspond to that timestamp or fall within a preset time window, forming target environmental data that matches the image. When extracting image features, the image can be scaled, cropped, and normalized before being input into a convolutional neural network to obtain image feature vectors. When extracting environmental temporal features, the target environmental data can be interpolated and completed to form a fixed-length temporal sequence before being input into a recurrent neural network to obtain environmental temporal feature vectors. When performing cross-modal fusion of image features and environmental temporal features, both can be input into a cross-modal attention module. The environmental temporal features are weighted according to their correlation with the image features, and then combined with the image features to form fused features. When determining the disease identification result based on the fused features, the fused features can be input into the disease identification model. The classification part outputs the disease type, the evaluation part outputs the severity, and the confidence score is formed by combining the model's predicted probability, resulting in a structured disease identification result.

[0023] During the candidate control scheme generation stage, at least one candidate control scheme can be matched from the control knowledge base based on the disease type in the disease identification results. For each candidate control scheme, the system obtains the real-time price of the corresponding pesticide, biological agent, or fertilizer from the agricultural input price database, obtains the plot area from the GIS system or user input, obtains labor cost parameters from the labor cost database or default configuration, and establishes a cost model based on the number of applications for the candidate scheme to calculate the total cost of the candidate scheme. On this basis, the system further constructs a multi-objective optimization model based on the total cost, resistance risk, and environmental toxicity of each candidate control scheme; where resistance risk can be calculated by the pesticide rotation coefficient, and environmental toxicity can be obtained by combining the pesticide median lethal dose (LD50) and residual period. Finally, user constraint data is used as hard constraints input into the multi-objective optimization model to optimize and solve each candidate control scheme, obtain at least one recommended control scheme, and output it on the APP or Web platform.

[0024] This embodiment integrates disease image data, environmental data, and user constraint data into the same decision-making chain, and sequentially completes identification, cost calculation, and multi-objective optimization, so that the output of the prevention and control plan has a clear data source, processing flow, and constraint basis.

[0025] In this embodiment, acquiring the disease image data of the target crop, as well as the corresponding environmental data and user constraint data, includes: The disease image data of the target crop is collected by mobile terminal or agricultural drone, and the timestamp and location information corresponding to the disease image data are obtained. Environmental data of the target crop's location is collected using field environmental sensors. The environmental data includes at least one of temperature, humidity, light intensity, rainfall, and soil moisture. Obtain at least one of the user-defined budget limit, list of banned pesticides, and organic farming requirements as the user constraint data.

[0026] This embodiment further defines "acquiring disease image data of the target crop, as well as the corresponding environmental data and user constraint data of the disease image data". When collecting disease image data of the target crop through a mobile terminal or agricultural drone, the mobile terminal can be a device with camera function such as a mobile phone or tablet, and the agricultural drone can be a plant protection drone or inspection drone equipped with a camera; during collection, diseased areas of crop leaves, stems, ears, or fruits can be photographed, and the image format can be JPG or PNG, with a resolution of not less than 1080P. After the collection is completed, the system automatically reads the timestamp of the shooting time and GPS positioning information, and uploads them together with the image file.

[0027] When collecting environmental data of the target crop area using field environmental sensors, IoT sensor nodes deployed in the field can be used to periodically sample at least one of temperature, humidity, light intensity, rainfall, and soil moisture. The sensors can upload the sampling results to a server via LoRa, NB-IoT, or 4G / 5G communication. The server can cache and archive the environmental data for a preset time period prior to the image capture time, such as the last 72 hours, for subsequent matching and retrieval.

[0028] When retrieving at least one of the user-defined budget limit, prohibited pesticide list, and organic farming requirements as user constraint data, it can be manually entered by the user in the app or web page, or automatically retrieved from the user's historical configuration file. The budget limit can be represented as the highest acceptable total cost for this pest control effort; the prohibited pesticide list can be represented as the set of pesticide names that the user prohibits from using; and the organic farming requirements can be represented as allowing only candidate solutions that comply with organic farming standards to be output. This user constraint data can be stored in a user constraint database for use in subsequent optimization phases.

[0029] This embodiment clarifies the acquisition sources, acquisition methods, and storage methods of three types of input data: disease images, environmental parameters, and user constraints, providing a directly implementable input foundation for subsequent data alignment, feature extraction, and constraint optimization.

[0030] In this embodiment, the alignment processing of the disease image data and the environmental data, and the extraction of image features and environmental temporal features respectively, includes: The environmental data is matched based on the timestamps corresponding to the disease image data to obtain the target environmental data corresponding to the disease image data. The disease image data is scaled and normalized to obtain the processed disease image data. The target environment data is interpolated to form a fixed-length environmental time series. The processed disease image data is input into a convolutional neural network to extract the image features; The environmental time series is input into a recurrent neural network to extract the environmental time series features.

[0031] In this embodiment, the phrase "aligning the disease image data and the environmental data, and extracting image features and environmental temporal features respectively" is specifically defined. When matching the environmental data based on the timestamp corresponding to the disease image data, the system first reads the capture time of the disease image, and then searches the environmental database for environmental sampling records corresponding to that time point or falling within a preset time window. If the environmental data sampling frequency is higher than the image acquisition frequency, a continuous environmental sequence before the image time point can be extracted; if there are missing environmental data, linear interpolation or nearest neighbor interpolation can be used to fill in the gaps, ultimately obtaining the target environmental data corresponding to the disease image data.

[0032] When scaling and normalizing disease image data, the original image can be uniformly scaled to 224×224 pixels, and the pixel values ​​can be normalized, for example, mapped to the [0,1] interval or subjected to mean-variance standardization, to obtain the processed disease image data. When the processed disease image data is input into a convolutional neural network, a ResNet18 backbone network can be used to perform convolution, pooling, and feature encoding on the disease image, outputting a fixed-length image feature vector.

[0033] When interpolating target environmental data to form a fixed-length environmental time series, environmental parameters such as temperature and humidity can be resampled according to a preset sampling interval to form a fixed-length 72-dimensional environmental time series. When inputting this environmental time series into a recurrent neural network, a two-layer LSTM network can be used, with a hidden layer dimension of, for example, 128, to perform temporal modeling of the environmental sequence and output an environmental time series feature vector. In this way, both the image modality and the environmental modality are mapped into feature representations suitable for subsequent fusion calculations.

[0034] This embodiment clarifies the time alignment, image preprocessing, environment sequence formation, and CNN / LSTM feature extraction processes, providing a clear implementation path for the joint use of image data and environment data, thus avoiding insufficient disclosure due to only providing a general description of "feature extraction".

[0035] In this embodiment, the phrase "cross-modal fusion of the image features and the environmental temporal features to obtain fused features" is defined. When the image features and the environmental temporal features are input into the cross-modal attention mechanism, the image feature vector output by the convolutional neural network can be used as the query vector, and the environmental temporal feature vector output by the LSTM can be used as the key vector to calculate the correlation between the image features and the environmental temporal features. The correlation can be calculated using a dot product and then normalized to obtain the attention weights.

[0036] Based on the correlation between the image features and the environmental temporal features, when weighting the environmental temporal features, the calculated attention weights can be used to weight each dimension or time segment of the environmental temporal features, resulting in weighted environmental features that emphasize environmental information more relevant to the current disease image. For example, when the disease image shows lesion textures related to high humidity conditions, the attention mechanism will increase the weight of features from high humidity periods in the environmental temporal features. When fusing the weighted environmental temporal features with the image features, methods such as vector concatenation, element-wise addition, linear mapping followed by addition, or fully connected mapping can be used to obtain the final fused features.

[0037] This embodiment clarifies the correlation calculation, environmental feature weighting, and the fusion process of the weighted environmental features and image features, so that "cross-modal fusion" is no longer just an abstract concept, but has a technical path that can be directly implemented by those skilled in the art.

[0038] In this embodiment, the cross-modal fusion of the image features and environmental temporal features to obtain fused features includes: The image features and the environmental temporal features are input into a cross-modal attention mechanism, and the environmental temporal features are weighted based on the correlation between the image features and the environmental temporal features. The weighted environmental temporal features are fused with the image features to obtain the fused features.

[0039] In this embodiment, the phrase "determining the disease identification result based on the fusion features" is defined in detail. When the fusion features are input into the disease identification model, the fusion feature vector output by the cross-modal fusion module can be input into the deep learning identification module. The disease identification model can employ a neural network structure including a classification branch and an evaluation branch, wherein the classification branch is responsible for outputting the disease type, and the evaluation branch is responsible for outputting the severity level.

[0040] When the classification branch of the disease identification model outputs the disease type, it can use a fully connected layer with a Softmax function to output the predicted probabilities of multiple candidate disease categories, selecting the category with the highest probability as the disease type. When the evaluation branch of the disease identification model outputs the severity corresponding to the disease type, it can evaluate the proportion of infected area, lesion distribution density, or lesion coverage, mapping the evaluation results to severity levels such as mild, moderate, and severe. When determining the confidence level based on the predicted probability of each disease type by the disease identification model, the predicted probability value corresponding to the final disease type can be directly taken as the confidence level. When associating the disease type, severity, and confidence level to generate the disease identification result, a structured output can be formed, such as "Disease = Rice Blast, Severity = Moderate, Confidence = 0.87".

[0041] This embodiment clarifies the input, classification output, severity assessment, and confidence generation process of the disease identification model, making the "disease identification result" of this embodiment have a clear structure and source, and providing a directly callable structured result for subsequent candidate solution matching.

[0042] In this embodiment, determining the disease identification result based on the fused features includes: The fused features are input into the disease identification model; The disease type is output through the classification branch of the disease identification model; The severity of the disease is output by the evaluation branch of the disease identification model, corresponding to the disease type. The confidence level is determined based on the predicted probability of each disease type by the disease identification model. The disease type, severity, and confidence level are correlated to generate disease identification results.

[0043] In this embodiment, the step of establishing a cost model based on land area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control scheme includes: Obtain the current unit price of agricultural inputs involved in each candidate prevention and control plan; Obtain the land area and labor cost parameters of the target plot; Obtain the number of applications corresponding to each candidate control program; Based on the current unit price, the land area, the labor cost parameters, and the number of applications, a cost model is established for each candidate prevention and control scheme.

[0044] In this embodiment, the step of "establishing a cost model based on land area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control schemes" is refined. When obtaining the current unit price of agricultural inputs involved in each candidate prevention and control scheme, the system first analyzes the types of pesticides, biological agents, or fertilizers included in the matched candidate prevention and control schemes, and then obtains the real-time prices of these agricultural inputs through an agricultural input price database or by connecting to the API of an agricultural input e-commerce platform. The real-time prices can be converted into the corresponding agricultural input price items based on the usage per acre.

[0045] When obtaining the plot area and labor cost parameters for the target plot, the plot area can be obtained from a GIS map, user input, or historical plot archives. The labor cost parameter can use the regional default value, such as 80 yuan / mu, or it can be manually set by the user. When obtaining the number of applications corresponding to each candidate control scheme, the recommended application frequency can be read from the control knowledge base based on the pesticide combination of the candidate scheme and the severity of the disease identification results. For example, moderate disease may correspond to two applications, and severe disease may correspond to three applications.

[0046] When establishing cost models for each candidate prevention and control scheme based on the current unit price, the land area, the labor cost parameters, and the number of applications, the aforementioned parameters can be input into the dynamic cost model module to generate a cost expression or cost function for each candidate scheme, and the total cost value of that candidate scheme can be calculated. In this way, each candidate prevention and control scheme can correspond to a comparable economic cost result.

[0047] This embodiment clarifies the methods for obtaining agricultural input prices, plot area, labor costs, and the number of pesticide applications, as well as the path to entering the cost model, thus providing a complete set of input sources and processing steps for "establishing a cost model".

[0048] In this embodiment, establishing a cost model for each candidate prevention and control scheme based on the current unit price, the land area, the labor cost parameter, and the number of applications includes: Establish a cost function C = a × Pppp + b × Plab + c × A × D, where C represents the total cost corresponding to the candidate control scheme, Ppppppppplab ... Obtain the agricultural input price, labor cost, plot area, and number of applications for each candidate prevention and control scheme; Substitute the agricultural input price, labor cost, plot area, and number of applications corresponding to each candidate control scheme into the cost function to determine the total cost corresponding to each candidate control scheme.

[0049] In this embodiment, when establishing the cost function C = a × Ppharmaceutical + b × Plabor + c × A × D, Ppharmaceutical can represent the price of agricultural inputs corresponding to the candidate control scheme, which can be calculated from the unit price and usage of all agricultural inputs in the candidate scheme; Plabor can represent the labor cost, which can be calculated from the labor cost per unit area and the number of operations required to implement the scheme; A can represent the actual area of ​​the target plot; D can represent the number of times pesticides are applied corresponding to the candidate control scheme; a, b, and c can be preset weighting coefficients used to adjust the weight of different cost factors in the total cost.

[0050] When obtaining the agricultural input price, labor cost, plot area, and number of applications for each candidate control scheme, the system can access the agricultural input price database, labor cost configuration table, GIS plot information, and control knowledge base to obtain the corresponding parameters. Substituting the agricultural input price, labor cost, plot area, and number of applications for each candidate control scheme into the cost function, the dynamic cost model module can automatically assign and calculate the parameters, outputting the total cost for each candidate control scheme. For example, combining disease type, severity, plot area, and labor cost, the base cost C_base = 120 yuan / mu can be obtained.

[0051] This embodiment provides a clear cost function form, parameter meanings, and parameter acquisition methods, enabling the cost model to move beyond a general description and instead be supported by directly calculable formulas. This facilitates the unified quantification of the total cost of different candidate prevention and control schemes.

[0052] In this embodiment, the construction of a multi-objective optimization model based on the total cost, drug resistance risk, and environmental toxicity of each candidate control scheme includes: Construct a multi-objective optimization function with the objectives of minimizing total cost, minimizing drug resistance risk, and minimizing environmental toxicity; The resistance risk is determined based on the pesticide rotation coefficient, and the environmental toxicity is determined based on the pesticide median lethal dose and residual period.

[0053] This embodiment defines the "construction of a multi-objective optimization model based on the total cost, drug resistance risk, and environmental toxicity of each candidate prevention and control scheme." When constructing the multi-objective optimization function with the objectives of minimizing total cost, drug resistance risk, and environmental toxicity, the system first reads the total cost value corresponding to each candidate prevention and control scheme and uses it as the first objective value. Then, for each candidate prevention and control scheme, the drug resistance risk value and environmental toxicity value are calculated and used as the second and third objective values, respectively. In this way, a three-dimensional objective vector can be formed for each candidate prevention and control scheme, which can then be used by subsequent optimization algorithms for non-dominated sorting and selection.

[0054] When the pesticide resistance risk is determined based on the pesticide rotation coefficient, the corresponding rotation coefficient can be obtained by looking up a table or calculating based on the mechanism of action, rotation frequency, and historical usage of the pesticides used in the candidate control scheme, and this coefficient can be used as a measure of pesticide resistance risk. When the environmental toxicity is determined based on the pesticide median lethal dose and residual period, the LD50 value and residual period of each pesticide can be read from the pesticide information database, and an environmental toxicity index can be formed through preset calculation rules; for example, the longer the residual period and the lower the LD50, the higher the corresponding environmental toxicity risk.

[0055] This embodiment clarifies the source of the three target values ​​and the calculation basis for drug resistance risk and environmental toxicity, making the construction of the multi-objective optimization model no longer abstract, but based on clearly obtainable indicators and rules.

[0056] In this embodiment, the step of using the user constraint data as constraints to optimize the candidate prevention and control schemes based on the multi-objective optimization model, and obtaining and outputting at least one recommended prevention and control scheme, includes: The constraints of ensuring that the total cost does not exceed the user's budget and that banned pesticides are not used are input into the multi-objective optimization model. The NSGA-II algorithm is used to iteratively solve the candidate prevention and control schemes to obtain the Pareto optimal solution set; At least one recommended prevention and control scheme is determined from the Pareto optimal solution set, and the recommended prevention and control scheme is output through the application interface or web interface.

[0057] This embodiment refines the step of "using the user constraint data as constraints, optimizing the candidate control schemes based on the multi-objective optimization model, and obtaining and outputting at least one recommended control scheme." When the total cost not exceeding the user budget and the prohibition of using banned pesticides are input into the multi-objective optimization model as constraints, the system first reads the budget limit and the list of banned pesticides from the user constraint database. Then, it filters out schemes whose total cost exceeds the budget limit or contains banned pesticides from the candidate control scheme set, or writes these constraints as hard constraints into the objective constraint set during the optimization process.

[0058] When using the NSGA-II algorithm to iteratively solve the candidate prevention and control schemes, the population of candidate prevention and control schemes can be initialized first, with each individual representing a set of candidate scheme parameters; then, non-dominated sorting, crowding calculation, selection, crossover, and mutation are performed to form the next generation population; in each generation, three objectives are considered simultaneously: total cost, drug resistance risk, and environmental toxicity, and the budget and drug prohibition constraints are checked; when the iteration reaches the preset number of generations or the population converges, the Pareto optimal solution set is obtained.

[0059] When determining at least one recommended control scheme from the Pareto optimal solution set and outputting the recommended control scheme through an application interface or web interface, the system can select several representative schemes from the Pareto optimal solution set according to a preset strategy, such as low-cost high-risk, high-cost low-toxicity, and comprehensive balanced schemes, and display the scheme name, agent combination, number of applications, and estimated cost to the user. The user can then select and execute the scheme.

[0060] This embodiment clarifies the input method of constraints, the NSGA-II solution process, and the output method of recommended solutions, so that "optimizing to obtain recommended solutions" has a clear computational flow and interaction path, thereby supporting the recommendation decision-making steps.

[0061] In this embodiment, the method further includes: Receive user feedback on the implementation effect of the recommended prevention and control plan; The disease image data, the environmental data, the user constraint data, and the execution effect feedback are used as new samples; Based on the new samples, the local model parameters used for disease identification and / or prevention decisions are updated through federated learning, and model gradients are generated. The model gradients are encrypted and uploaded to the cloud for aggregation.

[0062] This embodiment adds user feedback and federated learning update processes to the method. When receiving user feedback on the implementation effect of the recommended prevention and control plan, users can evaluate the prevention and control effect after implementation on the APP or Web. The evaluation format can be "good", "average", "poor", etc., and can also include new on-site photos, text descriptions, or re-test results. The system associates this feedback with the original input data to form a feedback record.

[0063] When the disease image data, environmental data, user constraint data, and execution effect feedback are used as new samples, the system can package the pre-diagnosis image, environment, and constraint inputs with the post-diagnosis actual execution effect into supervised samples, on a per-diagnosis-task basis, for updating the disease identification model and / or prevention and control decision model. Based on these new samples, when updating the local model parameters used for disease identification and / or prevention and control decisions through federated learning and generating model gradients, the model can be trained iteratively once or multiple times on local nodes to obtain parameter update results and corresponding gradients. When the model gradients are encrypted and uploaded to the cloud for aggregation, encrypted communication can be used to send the local gradients to the cloud aggregation server. The cloud aggregates gradients uploaded from multiple terminals to form a new global model.

[0064] This embodiment clarifies the processes of user feedback collection, new sample composition, local training, and cloud aggregation, enabling the federated learning update path to have a complete data flow and processing steps, thereby providing an implementable technical solution for continuous model updates.

[0065] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A decision-making method for agricultural pest and disease control, characterized in that, The method includes: Acquire disease image data of the target crop, as well as the corresponding environmental data and user constraint data of the disease image data; The disease image data and the environmental data are aligned, and image features and environmental temporal features are extracted respectively. The image features and the environmental temporal features are fused across modally to obtain fused features; The disease identification result is determined based on the fusion features, and the disease identification result includes disease type and severity. At least one candidate control scheme is matched based on the disease identification results; For each of the candidate prevention and control measures, a cost model is established based on the plot area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control measures. A multi-objective optimization model is constructed based on the total cost, drug resistance risk, and environmental toxicity of each candidate control scheme. Using the user constraint data as constraints, the candidate prevention and control schemes are optimized based on the multi-objective optimization model to obtain and output at least one recommended prevention and control scheme.

2. The method according to claim 1, characterized in that, The acquisition of disease image data of the target crop, as well as the corresponding environmental data and user constraint data, includes: The disease image data of the target crop is collected by mobile terminal or agricultural drone, and the timestamp and location information corresponding to the disease image data are obtained. Environmental data of the target crop's location is collected using field environmental sensors. The environmental data includes at least one of temperature, humidity, light intensity, rainfall, and soil moisture. Obtain at least one of the user-defined budget limit, list of banned pesticides, and organic farming requirements as the user constraint data.

3. The method according to claim 1, characterized in that, The process of aligning the disease image data and the environmental data, and extracting image features and environmental temporal features respectively, includes: The environmental data is matched based on the timestamps corresponding to the disease image data to obtain the target environmental data corresponding to the disease image data. The disease image data is scaled and normalized to obtain the processed disease image data. The target environment data is interpolated to form a fixed-length environmental time series. The processed disease image data is input into a convolutional neural network to extract the image features; The environmental time series is input into a recurrent neural network to extract the environmental time series features.

4. The method according to claim 3, characterized in that, The cross-modal fusion of the image features and environmental temporal features to obtain fused features includes: The image features and the environmental temporal features are input into a cross-modal attention mechanism, and the environmental temporal features are weighted based on the correlation between the image features and the environmental temporal features. The weighted environmental temporal features are fused with the image features to obtain the fused features.

5. The method according to claim 1, characterized in that, The determination of disease identification results based on the fused features includes: The fused features are input into the disease identification model; The disease type is output through the classification branch of the disease identification model; The severity of the disease is output by the evaluation branch of the disease identification model, corresponding to the disease type. The confidence level is determined based on the predicted probability of each disease type by the disease identification model. The disease type, severity, and confidence level are correlated to generate disease identification results.

6. The method according to claim 1, characterized in that, The cost model established based on land area, agricultural input prices, and labor costs to determine the total cost corresponding to the candidate prevention and control schemes includes: Obtain the current unit price of agricultural inputs involved in each candidate prevention and control plan; Obtain the land area and labor cost parameters of the target plot; Obtain the number of applications corresponding to each candidate control program; Based on the current unit price, the land area, the labor cost parameters, and the number of applications, a cost model is established for each candidate prevention and control scheme.

7. The method according to claim 6, characterized in that, The step of establishing a cost model for each candidate prevention and control scheme based on the current unit price, the land area, the labor cost parameters, and the number of applications includes: Establish a cost function C = a × Pppp + b × Plab + c × A × D, where C represents the total cost corresponding to the candidate control scheme, Ppppppppplab ... Obtain the agricultural input price, labor cost, plot area, and number of applications for each candidate prevention and control scheme; Substitute the agricultural input price, labor cost, plot area, and number of applications corresponding to each candidate control scheme into the cost function to determine the total cost corresponding to each candidate control scheme.

8. The method according to claim 1, characterized in that, The multi-objective optimization model, constructed based on the total cost, drug resistance risk, and environmental toxicity of each candidate control scheme, includes: Construct a multi-objective optimization function with the objectives of minimizing total cost, minimizing drug resistance risk, and minimizing environmental toxicity; The resistance risk is determined based on the pesticide rotation coefficient, and the environmental toxicity is determined based on the pesticide median lethal dose and residual period.

9. The method according to claim 8, characterized in that, The step of using the user constraint data as constraints, optimizing the candidate prevention and control schemes based on the multi-objective optimization model, and obtaining and outputting at least one recommended prevention and control scheme includes: The constraints of ensuring that the total cost does not exceed the user's budget and that banned pesticides are not used are input into the multi-objective optimization model. The NSGA-II algorithm is used to iteratively solve the candidate prevention and control schemes to obtain the Pareto optimal solution set; At least one recommended prevention and control scheme is determined from the Pareto optimal solution set, and the recommended prevention and control scheme is output through the application interface or web interface.

10. The method according to claim 1, characterized in that, The method further includes: Receive user feedback on the implementation effect of the recommended prevention and control plan; The disease image data, the environmental data, the user constraint data, and the execution effect feedback are used as new samples; Based on the new samples, the local model parameters used for disease identification and / or prevention decisions are updated through federated learning, and model gradients are generated. The model gradients are encrypted and uploaded to the cloud for aggregation.

Images