Method and device for optimizing agricultural decision based on multi-modal perception and expert feedback

Abstract

The application provides an agricultural decision optimization method and device based on multi-modal perception and expert feedback, and relates to the technical field of agricultural data processing, comprising: acquiring environment time series data and crop image data in a facility agricultural environment; performing feature extraction and cross-modal fusion on the environment time series data and the crop image data to obtain a fusion feature representation; inputting the fusion feature representation into a preset decision model to generate a preliminary agricultural decision scheme through the decision model; collecting feedback data for the preliminary agricultural decision scheme through an expert interaction interface; wherein the feedback data comprises parameter correction data and operation behavior trajectory data; constructing an instant reward value based on the feedback data, and updating parameters of the decision model through a reinforcement learning algorithm using the instant reward value.

Description

Technical Field

[0001] This invention relates to the field of agricultural data processing technology, and in particular to an agricultural decision optimization method and apparatus based on multimodal perception and expert feedback. Background Technology

[0002] With the continuous development of smart agriculture technology, artificial intelligence has been widely applied to the monitoring and management of facility agriculture in order to reduce labor costs and improve agricultural production efficiency. Existing agricultural intelligent decision-making systems typically generate corresponding agricultural operation suggestions based on collected environmental sensor data or crop image data.

[0003] However, when faced with the complex and ever-changing actual production environment of facility agriculture, the decision-making schemes generated by existing agricultural intelligent decision-making systems often deviate from the actual physiological laws of plants and professional agricultural production logic, resulting in insufficient accuracy and practicality of the decision results, and may even give misleading decisions that cause irreversible damage to crops.

[0004] Therefore, how to improve the decision-making accuracy of agricultural intelligent decision-making models in complex and dynamic environments and avoid generating decision schemes that violate agricultural professional logic is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides an agricultural decision optimization method and apparatus based on multimodal perception and expert feedback, which addresses the challenge of improving the decision-making accuracy of agricultural intelligent decision-making models in complex dynamic environments and avoiding the generation of decision schemes that violate agricultural professional logic.

[0006] This invention provides an agricultural decision optimization method based on multimodal perception and expert feedback, comprising: Acquire environmental time-series data and crop image data in facility agriculture environments; Feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold; The fused feature representation is input into a preset decision model, and a preliminary agricultural decision-making scheme is generated through the decision model; Feedback data on the preliminary agricultural decision-making scheme is collected through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; An immediate reward value is constructed based on the feedback data, and the decision model is updated with parameters using a reinforcement learning algorithm based on the immediate reward value; wherein, the loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision scheme and the preset plant physiological safety boundary.

[0007] According to the present invention, an agricultural decision optimization method based on multimodal perception and expert feedback is provided, wherein the method for determining the dynamic gain coefficient includes: The environmental indicators are compared with preset crop disease thresholds to obtain positive deviation values ​​of the environmental indicators exceeding the crop disease thresholds. The positive deviation value is weighted and summed using preset environmental factor sensitivity weights, and the weighted summation result is superimposed on a preset benchmark value to obtain the dynamic gain coefficient.

[0008] According to the present invention, an agricultural decision optimization method based on multimodal perception and expert feedback is provided, which performs feature extraction and cross-modal fusion on the environmental time-series data and the crop image data to obtain a fused feature representation, including: The environmental time-series data is encoded to obtain an environmental time-series feature vector, and the crop image data is processed to obtain a visual feature tensor. A query vector is generated based on the environmental temporal feature vector, and a key vector and a value vector are generated based on the visual feature tensor, respectively. A relevance score is calculated based on the query vector and the key vector. The relevance score is adjusted according to the dynamic gain coefficient and then normalized to generate an environment-aware attention weight. The value vector is weighted and aggregated according to the environmental awareness attention weights to obtain the fused feature representation.

[0009] According to the present invention, an agricultural decision optimization method based on multimodal perception and expert feedback generates preliminary agricultural decision schemes through the decision model, including: A comprehensive situation vector is generated based on the fused feature representation; Using the comprehensive situation vector as a retrieval seed, semantic matching is performed in a pre-constructed vectorized agricultural knowledge base to obtain expert guidance knowledge; The comprehensive situation vector and the expert guidance knowledge are encapsulated into structured prompt words, and the structured prompt words are input into the agricultural big language model to generate the preliminary agricultural decision-making scheme.

[0010] According to the present invention, an agricultural decision optimization method based on multimodal perception and expert feedback is provided, which constructs an instant reward value based on the feedback data, including: Extract the duration of the operation from the operation behavior trajectory data; Based on the preset duration range in which the operation stays and the historical decision accuracy of the experts providing the feedback data, the expert confidence weight is calculated. The immediate reward value is obtained by weighting the parameter correction data using the expert confidence weight.

[0011] According to the present invention, an agricultural decision optimization method based on multimodal perception and expert feedback is provided, which uses the instantaneous reward value to update the parameters of the decision model through a reinforcement learning algorithm, including: Based on the degree of deviation between the decision parameters and the plant physiological safety boundary, the deviation risk level is determined, and the penalty weight factor is dynamically determined based on the deviation risk level. The penalty weighting factor is used to weight the agricultural logic penalty items to obtain a comprehensive penalty value; The total loss function value is obtained by subtracting the comprehensive penalty value from the shearing objective function value constructed based on the immediate reward value; The decision model is updated based on the total loss function value.

[0012] The agricultural decision optimization method based on multimodal perception and expert feedback provided by the present invention further includes: The preliminary agricultural decision-making scheme is calibrated based on the feedback data to obtain a calibrated agricultural decision-making scheme. After the calibrated agricultural decision-making scheme is confirmed and executed, the calibrated agricultural decision-making scheme is transformed into a high-dimensional vector through an embedding model; The high-dimensional vector is updated in the vectorized agricultural knowledge base for semantic retrieval in subsequent decision-making processes.

[0013] This invention also provides an agricultural decision optimization based on multimodal perception and expert feedback, comprising: The data acquisition module is used to acquire environmental time-series data and crop image data in the facility agriculture environment; The feature fusion module is used to extract features and perform cross-modal fusion on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold. The decision generation module is used to input the fused feature representation into a preset decision model and generate a preliminary agricultural decision scheme through the decision model; The expert feedback module is used to collect feedback data on the preliminary agricultural decision-making plan through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; The model optimization module is used to construct an immediate reward value based on the feedback data, and use the immediate reward value to update the parameters of the decision model through a reinforcement learning algorithm to obtain a decision model with updated parameters; wherein, the loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision scheme and the preset plant physiological safety boundary.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the agricultural decision optimization method based on multimodal perception and expert feedback as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the agricultural decision optimization method based on multimodal perception and expert feedback as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the agricultural decision optimization method based on multimodal perception and expert feedback as described above.

[0017] The agricultural decision optimization method and apparatus based on multimodal perception and expert feedback provided by this invention effectively eliminates decision bias in complex agricultural environments by fusing environmental time-series data and crop image data across modalities and using dynamic gain coefficients based on adaptive adjustment of environmental indicators for feature weighting, thereby improving the accuracy of underlying perception. Simultaneously, it constructs an instant reward value using expert parameter corrections and operational behavior trajectories as feedback data, combined with a reinforcement learning algorithm incorporating agricultural logic penalties to continuously update the parameters of the decision model. This enables the decision model to autonomously learn and continuously evolve from the implicit experience of agricultural experts. Plant physiological safety boundary constraints ensure that the model update direction and generated decision schemes always conform to the actual agricultural production logic, thus significantly improving the overall decision accuracy, reliability, and absolute safety of the intelligent system in complex dynamic environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the agricultural decision optimization method based on multimodal perception and expert feedback provided by the present invention. Figure 2 A schematic diagram of the agricultural decision optimization device based on multimodal perception and expert feedback provided by the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] The agricultural decision optimization method based on multimodal perception and expert feedback provided by this invention can be executed by a computing device with data processing capabilities. The computing device includes, but is not limited to: an edge computing terminal equipped with a graphics processor, a cloud server, a local industrial control computer, or a distributed computing system composed of a combination of the above devices.

[0022] In the following embodiments, the collaborative operation of an edge computing terminal deployed in a facility agriculture greenhouse and a cloud-based evolutionary control engine is used as the execution subject. The edge computing terminal is responsible for performing real-time computing tasks of data acquisition and feature fusion, while the cloud-based evolutionary control engine is responsible for performing tasks of decision generation, feedback processing, and model parameter updating.

[0023] Those skilled in the art will understand that the above division of the execution subject is merely illustrative and does not constitute a limitation on the execution subject of the method of the present invention.

[0024] To facilitate understanding of the technical solutions of the various embodiments of the present invention, the following description uses a tomato greenhouse cultivation scenario as a unified embodiment throughout the text. In the tomato greenhouse cultivation scenario, tomatoes face the threat of fungal diseases such as gray mold at different growth stages, requiring a comprehensive judgment of environmental change trends and crop visual characteristics, and relying on the practical experience of agricultural experts to calibrate the planting management suggestions generated by the system.

[0025] Figure 1 This is a flowchart illustrating the agricultural decision optimization method based on multimodal perception and expert feedback provided by this invention. Figure 1 As shown, the method includes the following steps: Step S110: Obtain environmental time-series data and crop image data in the facility agriculture environment.

[0026] Environmental time-series data refers to environmental parameter data continuously collected by sensors and organized in chronological order in a facility agriculture environment. The environmental indicators included in the environmental time-series data may cover one or more of the following: temperature, humidity, light intensity, carbon dioxide concentration, soil temperature, and soil moisture.

[0027] Environmental time series data typically covers continuous sampling records within a preset time window, such as environmental parameter sampling sequences over the past 24, 48, or 72 hours.

[0028] Crop image data refers to image data obtained by photographing crops using image acquisition devices deployed in facility agriculture environments. These image acquisition devices can be monocular cameras, multi-angle camera arrays, or imaging devices with specific spectral responses.

[0029] Crop image data contains visual representation information of parts of the crop, such as leaves, stems, and fruits, including texture features, color distribution, and spot distribution.

[0030] In the context of tomato greenhouse cultivation, environmental time-series data can include time-series records of environmental parameters obtained from temperature sensors, humidity sensors, light sensors, and carbon dioxide concentration sensors in the greenhouse over the past 72 hours at fixed sampling intervals, such as every 5 minutes.

[0031] Crop image data includes, for example, images of tomato leaves and stems taken at the current moment by cameras deployed in different areas of the greenhouse.

[0032] Step S120: Perform feature extraction and cross-modal fusion on the environmental time-series data and the crop image data to obtain a fused feature representation; In this application, feature extraction refers to performing data transformation and information compression on the environmental time-series data and the crop image data respectively, and extracting high-dimensional feature expressions that can characterize the inherent laws of each data.

[0033] For environmental time-series data, the result of feature extraction is a feature vector reflecting the trend and pattern of environmental parameter changes. For the crop image data, the result of feature extraction is a feature tensor reflecting the spatial distribution characteristics of the crop's visual appearance.

[0034] Cross-modal fusion refers to the association and integration of features from different modalities, namely environmental temporal modalities and visual image modalities, so that the fused feature representation can simultaneously reflect the intrinsic relationship between environmental background information and crop visual representation information.

[0035] The cross-modal fusion process uses a dynamic gain coefficient for feature weighting. The dynamic gain coefficient is a numerical parameter used to adjust the contribution of different modal features during the fusion process. The dynamic gain coefficient is a dynamically adjusted value based on the comparison results of environmental indicators contained in the environmental time-series data with preset thresholds.

[0036] Specifically, the preset threshold is a critical value of environmental parameters that is pre-set based on the physiological laws of crops, such as the pathogenic environmental threshold of a specific crop disease.

[0037] When the real-time value of the environmental indicator meets or exceeds the preset threshold, the dynamic gain coefficient is adjusted from the default value to a larger value, thereby increasing the weight of visual features related to specific risk scenarios during the cross-modal fusion process.

[0038] When the real-time value of the environmental indicator does not reach the preset threshold, the dynamic gain coefficient remains at the default value or is adjusted towards a smaller value.

[0039] In the tomato greenhouse planting scenario, tomato gray mold is prone to outbreak under continuous low temperature and high humidity conditions.

[0040] If judgment is made solely based on crop image data, the system may misidentify normal leaf shadows as lesions due to insufficient lighting in the greenhouse. Through adaptive adjustment of the dynamic gain coefficient, when the humidity index continuously exceeds a preset threshold, the dynamic gain coefficient automatically increases, giving higher weight to visual features related to fungal diseases during the cross-modal fusion process. This allows for the differentiation between shadow interference and actual lesions at the fusion level.

[0041] The fusion feature representation is the output of the cross-modal fusion process, and is a high-dimensional feature representation that simultaneously encodes the relationship between environmental context and crop visual representation.

[0042] Step S130: Input the fused feature representation into a preset decision model, and generate a preliminary agricultural decision scheme through the decision model.

[0043] In this application, the decision model is a pre-built and deployed model that can generate agricultural planting management suggestions based on input features.

[0044] The specific form of the decision model can be a large language model fine-tuned with agricultural data, a generative model combining retrieval enhancement generation technology, or a hybrid decision model based on the fusion of rule engines and machine learning models.

[0045] The decision model receives the fused feature representation as input and generates the preliminary agricultural decision scheme based on the model's own parameters and knowledge.

[0046] The preliminary agricultural decision-making scheme is a planting and management recommendation output by the decision-making model, based on the current crop status and environmental conditions.

[0047] Preliminary agricultural decision-making plans may include, but are not limited to, one or more of the following: disease and pest diagnosis conclusions, environmental control instructions, irrigation plans, fertilization plans, and pesticide ratio recommendations.

[0048] Step S140: Collect feedback data on the preliminary agricultural decision-making plan through the expert interactive interface.

[0049] In this application, the expert interactive interface is a human-computer interface for agricultural experts to view, evaluate, and revise the preliminary agricultural decision-making plan. The expert interactive interface can be deployed on devices such as tablets, touch screens, and mobile terminal applications.

[0050] Feedback data is structured data generated by agricultural experts responding to the preliminary agricultural decision-making plan through the expert interaction interface. Feedback data includes parameter correction data and operational behavior trajectory data.

[0051] Parameter correction data refers to the data generated after experts adjust the specific parameters in the preliminary agricultural decision-making scheme. For example, if the expert corrects the irrigation volume recommended by the system from 15 liters / mu to 10 liters / mu, the corrected value and the correction range are recorded as parameter correction data.

[0052] Operational behavior trajectory data refers to the behavioral characteristic data generated by experts during the execution of operations on the expert interaction interface, such as the duration of expert pauses on specific decision options, the speed and continuity of swipe operations, and the sequence of click actions. Operational behavior trajectory data reflects the degree of subjective confidence of experts when making decision corrections.

[0053] In the tomato greenhouse planting scenario, the system pushes the preliminary agricultural decision-making plan to the mobile terminal of Expert A, who has 10 years of tomato planting experience.

[0054] After reviewing the plan, Expert A fine-tuned the concentration of iprodione application from the system's recommended value to a lower concentration using a slider. The system simultaneously recorded the duration of Expert A's pause during this fine-tuning operation (8 seconds) and the continuity of the slider operation. These records together constitute the feedback data.

[0055] In the tomato greenhouse planting scenario, the decision model, based on the fused feature representation, determines that the current greenhouse humidity is continuously high and there are visual representations in the leaf images that match the early stage of gray mold, and generates a preliminary agricultural decision scheme accordingly.

[0056] For example, if the initial risk of gray mold is detected, it is recommended to turn on the internal circulation fan to reduce humidity and to spray preventively with 50% iprodione wettable powder.

[0057] Step S150: Construct an instant reward value based on the feedback data, and use the instant reward value to update the parameters of the decision model through a reinforcement learning algorithm to obtain the updated decision model.

[0058] The instant reward value in this embodiment is calculated based on the feedback data; the instant reward value is a numerical signal used to evaluate the quality of the preliminary agricultural decision-making scheme.

[0059] The magnitude and sign of the immediate reward value reflect the degree of expert approval of the preliminary agricultural decision. When the expert makes a small correction or confirms adoption, the immediate reward value tends to be positive; when the expert makes a large correction or rejects the plan, the immediate reward value tends to be negative.

[0060] The loss function of the reinforcement learning algorithm includes an agricultural logic penalty term. This agricultural logic penalty term is a numerical term related to plant physiological safety constraints, used to suppress decision directions that violate plant physiological laws during model parameter updates.

[0061] The agricultural logic penalty term is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision-making scheme and the preset plant physiological safety boundary.

[0062] The decision parameters are specific operational parameters that can be quantified in the preliminary agricultural decision-making scheme, such as irrigation volume, fertilizer concentration, and pesticide dosage.

[0063] Plant physiological safety boundaries are safety threshold ranges that are pre-set based on crop physiological knowledge and stored in an agricultural expert knowledge base, such as the maximum single irrigation volume under specific temperature conditions and the safe range of fertilizer concentration for specific growth stages.

[0064] When the decision parameters exceed the plant physiological safety boundary, the agricultural logic penalty term generates a positive value, and the greater the deviation, the larger the value of the agricultural logic penalty term. The agricultural logic penalty term participates in the calculation of the loss function in a detrimental manner, thereby forcing the decision model to move away from the decision region that violates plant physiological laws during the parameter update process.

[0065] The updated decision model is the one optimized by the aforementioned reinforcement learning algorithm, with its decision parameters adjusted. When faced with similar agricultural scenarios, the updated decision model can generate preliminary agricultural decision-making schemes that are closer to expert decision-making habits and conform to plant physiological laws.

[0066] In the tomato greenhouse planting scenario, it is assumed that the decision model generates a decision parameter of "irrigating 2 liters / plant when the temperature is 38℃" during the exploration process.

[0067] Because the decision parameter significantly exceeds the safe upper limit for a single irrigation volume under high temperature conditions set in the agricultural expert knowledge base, such as 500 ml / plant, the agricultural logic penalty term generates a large positive value.

[0068] The agricultural logic penalty term, applied to the loss function in a detrimental manner, causes the total loss function value corresponding to this decision direction to become significantly negative. Through parameter updates, when the decision model encounters similar high-temperature scenarios in the future, it will tend to suggest first opening shade curtains and spraying mist to cool down, and then irrigating after the temperature drops.

[0069] In this embodiment, by converting expert feedback into immediate reward values ​​and including an agricultural logic penalty term based on plant physiological safety boundaries in the loss function of the reinforcement learning algorithm, the decision-making model can continuously evolve during daily management, and the direction of evolution is always constrained by plant physiological laws, thus avoiding the model from producing decision outputs that violate agricultural science due to absorbing noisy feedback or exploring randomness.

[0070] Optionally, the method for determining the dynamic gain coefficient includes: The environmental indicators are compared with preset crop disease thresholds to obtain positive deviation values ​​of the environmental indicators exceeding the crop disease thresholds. The positive deviation value is weighted and summed using preset environmental factor sensitivity weights, and the weighted summation result is superimposed on a preset benchmark value to obtain the dynamic gain coefficient.

[0071] In this application, the crop pathogenicity threshold is a critical value of environmental parameters that is pre-set based on agronomic knowledge for a specific crop variety and a specific disease type.

[0072] When environmental indicators reach or exceed the disease threshold of the crop, it indicates that the current environmental conditions have entered the high-incidence range of the disease.

[0073] The positive deviation value refers to the portion of the real-time value of the environmental indicator that exceeds the critical threshold for crop disease. When the environmental indicator does not exceed the critical threshold for crop disease, the positive deviation value is zero.

[0074] In the tomato greenhouse cultivation scenario, the pathogenic threshold for tomato gray mold is set at 90% humidity. When the sensor detects a current greenhouse humidity of 95%, the positive deviation value is 95% minus 90%, which equals 5 percentage points. When the humidity is 85%, the positive deviation value is zero.

[0075] Environmental indicators can include environmental parameters in multiple dimensions, such as humidity, temperature, and light intensity. For each dimension of environmental parameter, there is a corresponding critical threshold for crop pathogenicity, and the system calculates the positive deviation value for each dimension.

[0076] Environmental factor sensitivity weights are pre-defined coefficients used to measure the contribution of each environmental parameter dimension to the occurrence of a specific disease. Different environmental parameter dimensions have different degrees of influence on the same disease, and the environmental factor sensitivity weights reflect this difference.

[0077] For example, for fungal diseases, the sensitivity weight of the humidity factor is usually higher than that of the temperature factor.

[0078] Weighted summation refers to multiplying the positive deviation values ​​of each environmental parameter dimension by their respective environmental factor sensitivity weights and then summing them.

[0079] In the tomato greenhouse cultivation scenario, assuming only humidity is considered as an environmental parameter, the sensitivity weight of the humidity factor is set to 0.1, and the positive deviation of the current humidity is 5 percentage points. Therefore, the weighted summation result is 0.1 multiplied by 5, which equals 0.5.

[0080] If the temperature dimension is also considered, for example, if the nighttime temperature falls within the pathogenic range of 15°C to 20°C, with a positive deviation of 2°C and a sensitivity weight of 0.05 for the temperature factor, then the weighted summation result is 0.1 multiplied by 5 plus 0.05 multiplied by 2, which equals 0.6. The weighted summation result is then superimposed on a preset benchmark value to obtain the dynamic gain coefficient.

[0081] The preset baseline value is the default value of the dynamic gain coefficient when no environmental factors are triggered. It is usually set to 1.0, which means that the feature weights in the fusion process are not added or subtracted in the default state.

[0082] The calculation process of the above dynamic gain coefficient can be described by the following mathematical expression: =1.0+ ; Where α represents the dynamic gain coefficient; 1.0 represents the preset reference value; n Indicates the number of dimensions of environmental indicators; E i Indicates the first iReal-time values ​​of environmental indicators across multiple dimensions; Indicates the first i The crop pathogenicity threshold corresponding to environmental indicators in each dimension; Indicates the first i The sensitivity weights of environmental factors in each dimension; This indicates the positive deviation of the real-time value of the environmental indicator from the critical threshold for crop disease. When the real-time value of the environmental indicator does not exceed the critical threshold for crop disease, it is taken as zero.

[0083] In a tomato greenhouse cultivation scenario, considering only the humidity dimension, the parameters are set as follows: E1=95%, τ1=90%. =0.1. Substituting into the above formula: α=1.0+0.1×max(0,95-90)=1.0+0.1×5=1.5.

[0084] If we also consider the temperature dimension, the pathogenic range of nighttime temperature is between 15℃ and 20℃, with a positive deviation of 2℃, and the sensitivity weight of the temperature factor is 0.05; then α = 1.0 + 0.1 × 5 + 0.05 × 2 = 1.6.

[0085] It should be noted that the change in the dynamic gain coefficient α is directly driven by environmental data collected by physical sensors, rather than being the result of arbitrary mathematical calculations. Crop pathogenicity threshold. The records of plant physiological patterns are derived from the knowledge base of agricultural experts. For example, gray mold is highly prevalent when the humidity is consistently above 90% and the nighttime temperature is between 15°C and 20°C.

[0086] Environmental factor sensitivity weight This model is designed based on the physiological response characteristics of specific crop varieties to various environmental factors. Without the environmentally driven adjustment of the dynamic gain coefficient α, the accuracy of the pure visual model would significantly decrease under the interference of strong light or high humidity and shadow in a tomato greenhouse. This real-time intervention of environmental characteristics on visual recognition weights is based on a deep coupling with the specific physiological laws of tomatoes, rather than the conventional operation in the general computer field.

[0087] In this embodiment, by establishing the adjustment logic of the dynamic gain coefficient on the basis of the crop pathogenicity threshold and the sensitivity weight of environmental factors, the feature weighting in the cross-modal fusion process is ensured to have a clear agronomic basis. It can accurately adjust the weight for specific disease conditions of specific crops and improve the representation sensitivity of fused feature representation in high-risk pathogenic environments.

[0088] Optionally, feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation, including: The environmental time-series data is encoded to obtain an environmental time-series feature vector, and the crop image data is processed to obtain a visual feature tensor. A query vector is generated based on the environmental temporal feature vector, and a key vector and a value vector are generated based on the visual feature tensor, respectively. A relevance score is calculated based on the query vector and the key vector. The relevance score is adjusted according to the dynamic gain coefficient and then normalized to generate an environment-aware attention weight. The value vector is weighted and aggregated according to the environmental awareness attention weights to obtain the fused feature representation.

[0089] In this application, the system encodes environmental time-series data to obtain environmental time-series feature vector Fe.

[0090] The encoding process utilizes models capable of processing time-series data to transform environmental time-series data, extracting high-dimensional feature representations that reflect the trends and patterns of environmental parameters changing over time. Applicable encoding models include temporal convolutional networks, recurrent neural networks, long short-term memory networks, gated recurrent unit networks, and temporal Transformer models.

[0091] The environmental temporal feature vector Fe is the output of the encoding process, which is a fixed-length numerical vector with dimension d.

[0092] In a tomato greenhouse cultivation scenario, the system utilizes a temporal convolutional network to encode the time-series records of greenhouse temperature, humidity, light intensity, and carbon dioxide concentration over the past 72 hours. The temporal convolutional network extracts environmental change trend features, such as the slope of humidity increase and the magnitude of diurnal temperature variation, through causal convolution operations along the time axis.

[0093] Taking actual sampled data at a certain moment as an example, assuming the humidity rise slope extracted by the temporal convolutional network is approximately 0.35, the system determines that the current situation is one of high humidity-induced disease development. The aforementioned trend features are compressed into an environmental temporal feature vector Fe of dimension d.

[0094] Simultaneously, the system processes crop image data to obtain a visual feature tensor Fv. The processing utilizes a model capable of extracting spatial features from the image to transform the crop image data, extracting a high-dimensional feature representation that reflects the spatial distribution information of the crop's visual appearance.

[0095] Available processing models include deep residual networks, visual Transformer models, and convolutional neural networks. The visual feature tensor Fv is the output of the processing, and is a multidimensional numerical tensor with three dimensions: spatial height H, spatial width W, and number of channels C.

[0096] In the tomato greenhouse planting scenario, the system uses a deep residual network to process the currently captured tomato leaf images, extracting the leaf texture information, color distribution information, and spot location distribution information, and outputting the visual feature tensor Fv.

[0097] Next, the system generates a query vector Q using the environmental temporal feature vector Fe, and generates a key vector K and a value vector V using the visual feature tensor Fv.

[0098] In the cross-attention mechanism, the query vector Q represents the retrieval intent of the environment regarding what information it wants to focus on in the image; the key vector K represents the feature index contained in each spatial location in the image; and the value vector V represents the actual feature content carried by each spatial location in the image.

[0099] ; Key (K) & Value (V): Generated from image features, representing the visual spatial distribution of crops.

[0100]

[0101]

[0102] Among them, W Q Let be the learnable weight matrix used to map temporal feature vectors of the environment to the attention space, with dimension . d represents the original dimension of the environmental temporal feature vector. The mapped query dimensions; W K Let be a learnable weight matrix used to map visual feature tensors to the attention space, with dimension . C is the number of channels in the visual feature tensor; W V Let be a learnable weight matrix used to map visual feature tensors to a value space, with dimension . , The dimensions are the mapped values. The three learnable weight matrices mentioned above are automatically learned and updated during model training using the backpropagation algorithm.

[0103] In this application, the design of generating queries based on environmental features and generating key values ​​based on image features enables environmental features to actively guide selective attention to image features, thereby realizing environment-aware driven visual recognition.

[0104] The system calculates a relevance score based on the query vector Q and the key vector K. The relevance score reflects the degree of matching between the retrieval intent and the spatial features of each location in the image. The calculation method involves matrix multiplication of the transposes of the query vector Q and the key vector K, followed by division by a scaling factor to stabilize the numerical distribution. The scaling factor is taken from the query dimension. The square root of.

[0105] Subsequently, the system adjusts the correlation score using a dynamic gain coefficient α and generates environment-aware attention weights after normalization. The dynamic gain coefficient α is applied as a scaling factor to the correlation score, further amplifying the correlation score of image spatial locations related to disease features in high-risk disease environments. Then, the α-adjusted correlation score is normalized using a softmax function, ensuring the sum of attention weights at each spatial location is 1, thus obtaining the environment-aware attention weights.

[0106] The system uses environmental awareness attention weights to weight and aggregate the value vector V to obtain a fused feature representation.

[0107] The complete attention calculation process described above can be described by the following unified formula: Attention(Q,K,V) = Softmax ; in, It is a scaling factor used to stabilize the gradient; This is the dynamic gain coefficient.

[0108] In the scenario of tomato greenhouse cultivation, assuming It is 1.5. Compared to When the default value is 1.0, the correlation score of the spatial location in the image that matches the morphological features of water-soaked lesions is amplified, and after normalization, it obtains a higher attention weight.

[0109] Assuming in When set to the default value, the visual model's attention weight for a dark area on the leaf is approximately 0.5, indicating a hesitant judgment state. However, After being adjusted to 1.5, the attention weight of this area was significantly increased, and the system accurately captured early lesions that were not yet visually obvious, achieving the effect of preventing disease before it occurs.

[0110] In this embodiment, a cross-attention mechanism is designed to generate query vectors from environmental temporal feature vectors and key and value vectors from visual feature tensors. A dynamic gain coefficient is applied during the relevance score calculation, enabling directional guidance and precise focusing of visual information by environmental information. Compared to simple feature splicing or additive fusion methods, the cross-attention mechanism selectively amplifies visual region features relevant to the current environmental risk in the spatial dimension while suppressing irrelevant interference features, thereby obtaining a more discriminative fused feature representation.

[0111] Optionally, the decision model generates preliminary agricultural decision-making schemes, including: A comprehensive situation vector is generated based on the fused feature representation; Using the comprehensive situation vector as a retrieval seed, semantic matching is performed in a pre-constructed vectorized agricultural knowledge base to obtain expert guidance knowledge; The comprehensive situation vector and the expert guidance knowledge are encapsulated into structured prompt words, and the structured prompt words are input into the agricultural big language model to generate the preliminary agricultural decision-making scheme.

[0112] In this application, the system generates a comprehensive situation vector Γ based on the fused feature representation.

[0113] The comprehensive situation vector Γ is a comprehensive state representation obtained by aggregating the fused feature representation with the environmental temporal feature vector Fe. Feature aggregation methods include feature concatenation, residual connection, or a combination of both, and can be normalized to stabilize the numerical distribution.

[0114] The generation process of the comprehensive situation vector Γ can be described by the following mathematical expression: Γ= LayerNorm(Fe⊕Attention(Q, K, V)); Wherein, Γ is the comprehensive situation vector; Fe is the environmental temporal feature vector; ⊕ is the feature concatenation operation; Attention(Q, K, V) is the fused feature representation obtained after weighted aggregation through a cross-attention mechanism; LayerNorm is a layer normalization process used to stabilize the numerical distribution of the concatenated features. The comprehensive situation vector Γ fully preserves the crop physiological state information under the current environmental background, serving as the core input for subsequent retrieval and decision generation.

[0115] Then, the system uses the comprehensive situation vector Γ as a retrieval seed to perform semantic matching in a pre-built vectorized agricultural knowledge base to obtain expert guidance knowledge.

[0116] A vectorized agricultural knowledge base is a pre-built knowledge base that stores agricultural textual knowledge as high-dimensional vectors in a vector database. The knowledge entries stored in the knowledge base include planting technique recommendations, pest and disease control plans, historical successful management cases, and pesticide formulation guidelines. Each knowledge entry undergoes block processing and embedding model encoding before being stored in the database, transforming it into a high-dimensional vector form.

[0117] The semantic matching process calculates the similarity between the comprehensive situation vector Γ and the vectors of each knowledge entry in the vectorized agricultural knowledge base, and returns the knowledge entries with the highest similarity ranking as expert guidance knowledge.

[0118] In the tomato greenhouse planting scenario, the comprehensive situation vector Γ encodes the comprehensive state where the humidity is consistently higher than 90% and the leaves have suspected mold characteristics. The system matches knowledge items such as the low light treatment scheme for tomato gray mold and the physical humidity control priority strategy for fungal diseases in the vectorized agricultural knowledge base.

[0119] The system encapsulates the comprehensive situation vector Γ and expert guidance knowledge into structured prompt words, and inputs these structured prompt words into the agricultural big language model to generate a preliminary agricultural decision-making scheme that includes cause analysis and environmental control instructions.

[0120] In this application, structured prompts are text inputs formed by organizing real-time perceived data, retrieved expert guidance knowledge, and decision requirements according to a preset format. This provides sufficient contextual information for the agricultural big data language model, guiding it to generate professional and targeted decision suggestions.

[0121] More specifically, the agricultural big language model is a big language model that has been fine-tuned or adapted with agricultural data. It has the ability to understand agricultural terminology, follow agricultural logic, and generate structured agricultural recommendations.

[0122] In a tomato greenhouse cultivation scenario, a structured prompt could be: "Current greenhouse humidity is 95%, which is in a high-risk range. Humidity has been rising continuously over the past 72 hours. Images have detected suspected early signs of gray mold on leaf edges. Referring to low-light treatment solutions for tomato gray mold in the expert database, please generate management recommendations including physical humidity control and chemical pesticide ratios." The preliminary agricultural decision-making plan generated by the agricultural big data model includes two parts: cause analysis and environmental control instructions. The cause analysis states that the combination of persistent high humidity and low light conditions significantly increases the risk of gray mold spore germination. Environmental control instructions, for example, could prioritize turning on the internal circulation fan to reduce humidity to below 85%, while simultaneously applying a wettable powder of iprodione for preventative foliar spraying. The specific pesticide name and concentration ratio are determined by the system based on control solutions in the knowledge base.

[0123] In this embodiment, by using the comprehensive situation vector as the expert guidance knowledge for seed retrieval and encapsulating the retrieval results and real-time situation information into structured prompt words and inputting them into the agricultural big language model, the preliminary agricultural decision-making scheme generated by the decision-making model is both targeted based on real-time data and professional based on knowledge base accumulation. This effectively reduces the output that the big language model may produce when it lacks domain knowledge and does not conform to the actual farming logic.

[0124] Optionally, constructing an instant reward value based on the feedback data includes: Extract the duration of the operation from the operation behavior trajectory data; Based on the preset duration range in which the operation stays and the historical decision accuracy of the experts providing the feedback data, the expert confidence weight is calculated. The immediate reward value is obtained by weighting the parameter correction data using the expert confidence weight.

[0125] In this application, the operation dwell time refers to the time spent by an expert operating on a specific decision option or parameter adjustment area of ​​a preliminary agricultural decision-making scheme on the expert interaction interface. A shorter operation dwell time usually reflects that the expert is more familiar with the content and has a clear judgment, while a longer operation dwell time may reflect that the expert has doubts or hesitations in judgment.

[0126] The expert confidence weight ω is calculated based on the preset time interval of the operation dwell time and the historical decision accuracy of the experts who provide feedback data.

[0127] The preset time interval is a pre-defined time range used to judge the decisiveness of an expert's actions. For example, the duration of the action can be divided into several intervals: when the duration of the action is within a shorter decisive interval, the expert is judged to be decisive and experienced; when the duration of the action exceeds a longer hesitation interval, the expert is judged to be hesitant. The specific interval division can be flexibly set according to the actual application scenario.

[0128] Historical decision accuracy refers to the proportion of times that the expert's recommendations in previous managed decisions ultimately led to positive planting results, retrieved by the system from historical records.

[0129] The expert confidence weight ω is a dynamic numerical weight obtained after combining the behavioral characteristics reflected by the duration of operation and the static governance reflected by the accuracy of historical decisions. The value ranges from 0 to 1.

[0130] The system first determines the initial benchmark score S based on the expert's years of professional experience and historical decision-making accuracy. base Then, the behavior bonus coefficient S is determined based on the duration interval of the operation stay. actio Finally, the two are multiplied together and truncated to the range of 0 to 1 to obtain the expert confidence weight ω.

[0131] The calculation process can be described by the following mathematical expression: ; Where ω is the expert confidence weight; S base S is the initial baseline score; action S is the behavior bonus coefficient, which is applied when the duration of the operation is within the decisive range. action Take a positive additive value greater than or equal to 1. When the operation dwell time exceeds the decisive interval, S actionReduce the weight to less than 1; the Clip function restricts the product result to the range of 0 to 1.

[0132] Let's take a tomato greenhouse cultivation scenario as an example. Assume expert A has 10 years of experience and a high historical success rate in managed services; the system assigns an initial benchmark score S. base =0.90. Expert A's pause time on the tomato fertilizer application fine-tuning interface was 8 seconds. The system's preset decisive interval is 5 to 15 seconds. 8 seconds falls within this interval, and the system assigns a behavior bonus coefficient S accordingly. action =1.05. Substituting into the above formula: ω=Clip(0.90×1.05,0,1)=Clip(0.945,0,1)=0.945, which is approximately 0.95 after rounding. The above decisive interval division and the value of the additive coefficient are only examples. In actual applications, they can be adjusted according to the statistical characteristics of the operational behavior of the expert group.

[0133] It should be noted that the confidence weight ω is not a static preset value, but a dynamic weight generated by combining expert behavior data captured in real time by the hardware terminal with historical archive data. When an expert hesitates in their actions, S... action The reduction leads to a decrease in ω, and the impact of this expert feedback on model evolution is automatically suppressed, thereby ensuring the purity of the learning samples.

[0134] Finally, the system uses the expert confidence weight ω to perform a weighted calculation on the parameter correction data contained in the feedback data to obtain the instant reward value.

[0135] The system first converts the expert's physical actions into numerical values ​​using a preset reward mapping function Φ, where positive values ​​indicate approval and negative values ​​indicate rejection. Then, the numerical values ​​are weighted by ω, and simultaneously, a supplementary weighting coefficient (1-ω) is used to weight the logical verification deviation value Δ calculated based on parameter correction data. The sum of these two weightings yields the immediate reward value r. t .

[0136] The calculation process can be described by the following mathematical expression: r t =ω·Φ(expert_action)+(1-ω)·Δ(logic_check); Where, r t ω is the immediate reward value; Φ(expert_action) is the expert confidence weight; Φ(expert_action) is the output value of the reward mapping function, which maps the expert's physical operation to a numerical value; (1-ω) is the supplementary weight coefficient; Δ(logic_check) is the logic check deviation value, which is calculated based on the extent of the expert's modification of the decision parameters. The larger the modification, the more negative Δ tends to be.

[0137] In a tomato greenhouse cultivation scenario, suppose expert A slightly adjusts the urea application rate from 15 kg / mu to 10 kg / mu, a change of 5 kg / mu, which is considered a moderate adjustment. The reward mapping function maps this operation to a near-positive numerical value. After weighting with ω approximately 0.95, a high positive immediate reward value is obtained, indicating that the system should learn in the direction of this expert correction.

[0138] After multiple such calibrations, the system gradually learned to make comprehensive judgments based on stem diameter, color, and ambient temperature and humidity during the tomato transplanting period, thereby reducing the risk of over-fertilization.

[0139] This step effectively distinguishes the reliability of different expert feedback by incorporating expert confidence weights calculated based on operation dwell time and historical decision accuracy during the construction of immediate reward values. This ensures that the feedback from experienced and decisive experts has a greater weight in model evolution, while the influence of expert feedback from hesitant experts or those with low historical accuracy is automatically suppressed, thus guaranteeing the purity of the model's learning samples and the correctness of the evolutionary direction.

[0140] Optionally, the decision model is updated using the immediate reward value via a reinforcement learning algorithm, including: Based on the degree of deviation between the decision parameters and the plant physiological safety boundary, the deviation risk level is determined, and the penalty weight factor is dynamically determined based on the deviation risk level. The penalty weighting factor is used to weight the agricultural logic penalty item to obtain a comprehensive penalty value; The total loss function value is obtained by subtracting the comprehensive penalty value from the shearing objective function value constructed based on the immediate reward value; The decision model is updated based on the total loss function value.

[0141] In this application, the degree of deviation is a quantitative difference in the deviation of decision parameters from the physiological safety boundary of plants. The risk level of deviation is classified according to the magnitude of the deviation and the type of risk involved. When the decision parameters involve operations that may lead to crop death and the degree of deviation is large, it is determined to be a high-risk level; when the decision parameters only deviate slightly within the normal growth range, it is determined to be a low-risk level.

[0142] Agricultural logic penalty term P( The calculation process is as follows. The system retrieves the plant physiological safety boundary S corresponding to the current environmental situation from the agricultural expert knowledge base. Γ Calculate the positive bias of decision parameters exceeding the plant physiological safety boundary: P( =max(0,VS) Γ ); Wherein, P( ) represents the original value of the agricultural logic penalty term; V represents the value of the decision parameters in the preliminary agricultural decision-making scheme; S Γ This represents the plant physiological safety boundary value retrieved from the agricultural expert knowledge base under the current environmental conditions; max(0,·) indicates taking a positive bias, and is set to zero when the decision parameter does not exceed the plant physiological safety boundary.

[0143] P( The original values ​​of the deviations need to be normalized and mapped to the range of 0 to 1 before participating in subsequent calculations. Normalization can be achieved by dividing the original deviation value by a preset maximum deviation benchmark value, or by using other mapping functions to compress the original values ​​to the range of 0 to 1.

[0144] The following example illustrates the process of tomato greenhouse cultivation. Assume the decision parameter V = 2000 ml / plant is the irrigation amount output by the model, and the plant physiological safety boundary S... Γ =500 ml / plant is the safe upper limit for irrigation when the temperature exceeds 35℃. Substituting into the above formula: P(θ)=max(0,2000-500)=1500. Assuming the system's preset maximum deviation benchmark value is 2000 ml / plant, then after normalization, P(θ)=1500 / 2000=0.75, indicating that this decision contains a serious logical violation.

[0145] Because high-temperature, high-flow-rate irrigation carries the risk of root death due to physiological shock caused by cold water, the system classifies deviations from the risk level as high-risk. Then, the system dynamically determines the penalty weighting factor λ based on the deviation risk level.

[0146] The penalty weighting factor λ is used to control the strength of the agricultural logic penalty term in the loss function, and is dynamically adjusted according to the deviation risk level. When the deviation risk level is high, λ takes a larger value, which strongly inhibits decision-making directions that violate the safety bottom line.

[0147] When the deviation from the risk level is low, λ takes a smaller value, allowing the decision model to make small adaptive explorations within a safe range to seek better production efficiency. For example, in high-risk scenarios, λ can take a value in the range of 100 to 500, while in normal scenarios, λ can take a value in the range of 0.1 to 1.0. The specific value is determined according to the crop variety and risk scenario.

[0148] In the above high-temperature irrigation scenario, the deviation risk level is high-risk level, and it is assumed that the system sets λ to 150.

[0149] Next, the system uses a penalty weighting factor λ to weight the agricultural logic penalty term P(θ) to obtain a comprehensive penalty value. The comprehensive penalty value is the product of λ and the normalized P(θ). In the above scenario, it is 150 × 0.75 = 112.5.

[0150] The system then subtracts the comprehensive penalty value from the shearing objective function value constructed based on the immediate reward value to obtain the total loss function value.

[0151] The complete calculation process of the total loss function can be described by the following mathematical expression: L total =L clip -λP( ); That is: L total = -λP( ); in, : Expected value for time step t. Probability ratio. The ratio of the probability of the new strategy and the old strategy selecting the same action, used to measure the magnitude of the update. The dominance function represents how much better the expert-calibrated solution is than the average expected outcome under the current situation Γ. clip( ) Shear function. Limit the policy update magnitude to [ Between ], to prevent parameters from being updated too drastically. Hyperparameters (e.g., 0.2) define the step size limit for the model to approach expert habits at each iteration. P( ): This is the normalized agricultural logic penalty term. It's the penalty value generated when the parameters generated by the model exceed the safe threshold for plant growth, ensuring that the evolutionary direction conforms to biological laws. λ: Penalty weighting factor, used to control the agricultural logic penalty term P( In the total loss function L total The proportion of the total loss function. The system updates the parameters of the decision model based on the total loss function value.

[0152] Continuing with the example of tomato greenhouse cultivation, assuming the objective function value of shearing is 0.8—indicating that the decision to irrigate 2000 ml / plant at 38℃ is mathematically predicted to increase yield slightly—and the overall penalty value is 112.5, then the total loss function value L total = 0.8 - 112.5 = -111.7. This huge negative value, through backpropagation, affects the parameters of the decision-making model, causing the model to completely deviate from the decision-making region of high-temperature, high-flow-rate irrigation in subsequent parameter updates. After evolution, the model, in similar scenarios, generates a scientific decision-making logic that first turns on the sunshade and misting system to cool the temperature, and then irrigates only after the temperature drops to a safe range.

[0153] It is important to emphasize that the threshold of the agricultural logic penalty term P(θ) is derived from the plant physiological safety boundary in the agricultural expert knowledge base, rather than being randomly set. The existence of the penalty weight factor λ is equivalent to installing a speed bump on the reinforcement learning algorithm, solving the illusion and safety overshoot problems that may occur in AI decision-making in complex agricultural environments.

[0154] This step achieves hierarchical and precise control over the evolutionary direction of the decision-making model by dynamically determining the penalty weight factor based on the deviation risk level and constructing the total loss function in conjunction with the shearing objective function. In high-risk scenarios involving crop mortality risk, a larger penalty weight factor can quickly force the model to abandon dangerous decision paths; in normal growth scenarios, a smaller penalty weight factor preserves the exploration space for the model. This hierarchical control mechanism transforms reinforcement learning algorithms from general mathematical optimization tools into specialized evolutionary algorithms with agricultural physical logic constraints.

[0155] Optionally, the method further includes: The preliminary agricultural decision-making scheme is calibrated based on the feedback data to obtain a calibrated agricultural decision-making scheme. After the calibrated agricultural decision-making scheme is confirmed and executed, the calibrated agricultural decision-making scheme is transformed into a high-dimensional vector through an embedding model; The high-dimensional vector is updated in the vectorized agricultural knowledge base for semantic retrieval in subsequent decision-making processes.

[0156] In this application, the calibration process involves applying parameter corrections and logical adjustments provided by experts through feedback data to the initial agricultural decision-making scheme, resulting in a version that has been confirmed and adjusted by experts. The calibrated agricultural decision-making scheme integrates the system's intelligent analysis capabilities with expert domain experience.

[0157] Then, after the calibrated agricultural decision-making scheme is confirmed and implemented, the system transforms the calibrated agricultural decision-making scheme into a high-dimensional vector through an embedded model.

[0158] Confirmation and implementation means that the calibrated agricultural decision-making plan, after expert review and approval, is actually applied to planting and management operations, and the final implementation effect is verified to have achieved the expected goals, such as effective control of diseases or restoration of normal crop growth. Only plans that have been verified to be effective are included in the knowledge base update process to ensure that the knowledge base contains only high-quality decision-making cases.

[0159] Embedding models are used to transform text or structured data into high-dimensional vector representations. By encoding calibrated agricultural decision-making schemes using embedding models, high-dimensional vectors that can represent the semantic information of the schemes in vector space are obtained.

[0160] Finally, the system updates the high-dimensional vector into the vectorized agricultural knowledge base for semantic retrieval in subsequent decision-making processes.

[0161] In this embodiment of the application, the update process involves adding the newly generated high-dimensional vector and the associated original text of the scheme to the vectorized agricultural knowledge base to expand the knowledge entries.

[0162] In subsequent decision-making processes, when the system encounters similar agricultural scenarios again, it can retrieve previously calibrated and verified high-quality cases as references by performing semantic matching of the comprehensive situation vector in the vectorized agricultural knowledge base.

[0163] With continued expert involvement and feedback, high-quality cases are constantly accumulating in the vectorized agricultural knowledge base, while the parameters of the decision-making model are continuously updated through reinforcement learning algorithms. This synchronous evolution leads to a continuous improvement in the quality of the initial agricultural decision-making schemes generated by the system, gradually approaching the decision-making level of seasoned experts in the field.

[0164] In the context of tomato greenhouse cultivation, the gray mold control plan, calibrated by expert A, has been proven effective in multiple practical implementations, indicating that the disease was controlled in its early stages and did not spread.

[0165] This solution was used as a gold standard case study, transformed into a high-dimensional vector through an embedding model, and updated to a vectorized agricultural knowledge base. In subsequent planting cycles, when similar high-humidity environments and suspected gray mold characteristics reappear, the decision-making model can directly call upon the previously validated solution as a reference through semantic retrieval, significantly improving response speed and solution quality.

[0166] In this embodiment, by continuously updating expert-calibrated and practically validated agricultural decision-making schemes to a vectorized agricultural knowledge base, the implicit experience of experts is transformed from temporary corrections into permanent knowledge. The vectorized agricultural knowledge base and the decision-making model form a closed loop of knowledge accumulation and model evolution, enabling the system to continuously improve its knowledge accumulation and decision-making capabilities during daily management without requiring large-scale re-labeling of data.

[0167] The agricultural decision optimization device based on multimodal perception and expert feedback provided by the present invention will be described below. The agricultural decision optimization device based on multimodal perception and expert feedback described below and the agricultural decision optimization method based on multimodal perception and expert feedback described above can be referred to in correspondence.

[0168] Figure 2 This is a schematic diagram of the agricultural decision optimization device based on multimodal perception and expert feedback provided by the present invention, as shown below. Figure 2 As shown, it includes: Data acquisition module 210 is used to acquire environmental time-series data and crop image data in the facility agriculture environment; The feature fusion module 220 is used to extract features and perform cross-modal fusion on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold. The decision generation module 230 is used to input the fused feature representation into a preset decision model, and generate a preliminary agricultural decision scheme through the decision model; The expert feedback module 240 is used to collect feedback data on the preliminary agricultural decision-making scheme through an expert interactive interface; wherein, the feedback data includes parameter correction data and operation behavior trajectory data; The model optimization module 250 is used to construct an immediate reward value based on the feedback data, and use the immediate reward value to update the parameters of the decision model through a reinforcement learning algorithm to obtain a decision model with updated parameters; wherein, the loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision scheme and the preset plant physiological safety boundary.

[0169] In this application, cross-modal fusion of environmental time-series data and crop image data is performed, and feature weighting is carried out using dynamic gain coefficients that are adaptively adjusted based on environmental indicators. This effectively eliminates decision-making biases caused by single-modal or static recognition in complex agricultural environments and improves the accuracy of underlying perception. At the same time, the parameter corrections and operational behavior trajectories of experts on the preliminary decision-making scheme are used as feedback data to construct an instant reward value. Combined with a reinforcement learning algorithm that includes agricultural logic penalty terms, the decision-making model is continuously updated with parameters. This enables the decision-making model to learn autonomously and continuously evolve from the implicit experience of agricultural experts. Plant physiological safety boundary constraints ensure that the direction of model updates and the generated decision schemes always conform to the actual agricultural production logic. As a result, the overall decision-making accuracy, reliability, and absolute safety of the intelligent system in complex dynamic environments are significantly improved.

[0170] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340. The processor 310, communications interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions from the memory 330 to execute an agricultural decision optimization method based on multimodal perception and expert feedback. This method includes acquiring environmental time-series data and crop image data in a facility agriculture environment. Feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold; The fused feature representation is input into a preset decision model, and a preliminary agricultural decision-making scheme is generated through the decision model; Feedback data on the preliminary agricultural decision-making scheme is collected through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; An immediate reward value is constructed based on the feedback data, and the decision model is updated with parameters using a reinforcement learning algorithm based on the immediate reward value; wherein, the loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision scheme and the preset plant physiological safety boundary.

[0171] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0172] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute the agricultural decision optimization method based on multimodal perception and expert feedback provided by the above methods, the method including: acquiring environmental time-series data and crop image data in the facility agriculture environment; Feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold; The fused feature representation is input into a preset decision model, and a preliminary agricultural decision-making scheme is generated through the decision model; Feedback data on the preliminary agricultural decision-making scheme is collected through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; An immediate reward value is constructed based on the feedback data, and the decision model is updated with parameters using a reinforcement learning algorithm based on the immediate reward value; wherein, the loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision scheme and the preset plant physiological safety boundary.

[0173] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the agricultural decision optimization method based on multimodal perception and expert feedback provided by the above methods, the method comprising: acquiring environmental time-series data and crop image data in a facility agriculture environment; Feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold; The fused feature representation is input into a preset decision model, and a preliminary agricultural decision-making scheme is generated through the decision model; Feedback data on the preliminary agricultural decision-making scheme is collected through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; An immediate reward value is constructed based on the feedback data, and the decision model is updated with parameters using a reinforcement learning algorithm based on the immediate reward value; wherein, the loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the degree of deviation between the decision parameters in the preliminary agricultural decision scheme and the preset plant physiological safety boundary.

[0174] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0175] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An agricultural decision optimization method based on multimodal perception and expert feedback, characterized in that, include: Acquire environmental time-series data and crop image data in facility agriculture environments; Feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold; The fused feature representation is input into a preset decision model, and a preliminary agricultural decision-making scheme is generated through the decision model; Feedback data on the preliminary agricultural decision-making scheme is collected through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; An immediate reward value is constructed based on the feedback data, and the decision model is updated using the immediate reward value through a reinforcement learning algorithm. The loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the deviation of the decision parameters in the preliminary agricultural decision scheme from a preset plant physiological safety boundary. The plant physiological safety boundary is a value retrieved from an agricultural expert knowledge base under the current environmental conditions. The construction of the instant reward value based on the feedback data includes: Extract the duration of the operation from the operation behavior trajectory data; Based on the preset duration range in which the operation stays and the historical decision accuracy of the experts providing the feedback data, the expert confidence weight is calculated. The expert's actions are converted into numerical values ​​using a preset reward mapping function. These numerical values ​​are then weighted using the expert confidence weight, and a supplementary weight coefficient is used to weight the logical verification deviation value calculated based on the parameter correction data. The sum of these weighted results yields the immediate reward value. The supplementary weight coefficient is the difference between 1 and the expert confidence weight, and the logical verification deviation value is calculated based on the extent to which the expert modifies the decision parameters in the preliminary agricultural decision-making scheme.

2. The method according to claim 1, characterized in that, The method for determining the dynamic gain coefficient includes: The environmental indicators are compared with preset crop disease thresholds to obtain positive deviation values ​​of the environmental indicators exceeding the crop disease thresholds. The positive deviation value is weighted and summed using preset environmental factor sensitivity weights, and the weighted summation result is superimposed on a preset benchmark value to obtain the dynamic gain coefficient.

3. The method according to claim 1, characterized in that, Feature extraction and cross-modal fusion are performed on the environmental time-series data and the crop image data to obtain a fused feature representation, including: The environmental time-series data is encoded to obtain an environmental time-series feature vector, and the crop image data is processed to obtain a visual feature tensor. A query vector is generated based on the environmental temporal feature vector, and a key vector and a value vector are generated based on the visual feature tensor, respectively. A relevance score is calculated based on the query vector and the key vector. The relevance score is adjusted according to the dynamic gain coefficient and then normalized to generate an environment-aware attention weight. The value vector is weighted and aggregated according to the environmental awareness attention weights to obtain the fused feature representation.

4. The method according to claim 1, characterized in that, The decision model generates preliminary agricultural decision-making schemes, including: A comprehensive situation vector is generated based on the fused feature representation; Using the comprehensive situation vector as a retrieval seed, semantic matching is performed in a pre-constructed vectorized agricultural knowledge base to obtain expert guidance knowledge; The comprehensive situation vector and the expert guidance knowledge are encapsulated into structured prompt words, and the structured prompt words are input into the agricultural big language model to generate the preliminary agricultural decision-making scheme.

5. The method according to claim 1, characterized in that, The decision model is updated using the immediate reward value via a reinforcement learning algorithm, including: Based on the degree of deviation between the decision parameters and the plant physiological safety boundary, the deviation risk level is determined, and the penalty weight factor is dynamically determined based on the deviation risk level. The penalty weighting factor is used to weight the agricultural logic penalty item to obtain a comprehensive penalty value; The total loss function value is obtained by subtracting the comprehensive penalty value from the shearing objective function value constructed based on the immediate reward value; The decision model is updated based on the total loss function value.

6. The method according to claim 1, characterized in that, The method further includes: The preliminary agricultural decision-making scheme is calibrated based on the feedback data to obtain a calibrated agricultural decision-making scheme. After the calibrated agricultural decision-making scheme is confirmed and executed, the calibrated agricultural decision-making scheme is transformed into a high-dimensional vector through an embedding model; The high-dimensional vector is updated in the vectorized agricultural knowledge base for semantic retrieval in subsequent decision-making processes.

7. An agricultural decision optimization device based on multimodal perception and expert feedback, comprising: The data acquisition module is used to acquire environmental time-series data and crop image data in the facility agriculture environment; The feature fusion module is used to extract features and perform cross-modal fusion on the environmental time-series data and the crop image data to obtain a fused feature representation; wherein, the cross-modal fusion process is based on feature weighting using a dynamic gain coefficient, and the dynamic gain coefficient is adaptively adjusted according to the comparison results between the environmental indicators contained in the environmental time-series data and the preset crop disease threshold. The decision generation module is used to input the fused feature representation into a preset decision model and generate a preliminary agricultural decision scheme through the decision model; The expert feedback module is used to collect feedback data on the preliminary agricultural decision-making plan through an expert interactive interface; wherein, the feedback data includes parameter correction data and operational behavior trajectory data; The model optimization module is used to construct an immediate reward value based on the feedback data, and to update the parameters of the decision model using the immediate reward value through a reinforcement learning algorithm to obtain a parameter-updated decision model. The loss function of the reinforcement learning algorithm includes an agricultural logic penalty term, which is generated based on the deviation between the decision parameters in the preliminary agricultural decision-making scheme and a preset plant physiological safety boundary. The plant physiological safety boundary is a value retrieved from an agricultural expert knowledge base under the current environmental conditions. The construction of the instant reward value based on the feedback data includes: Extract the duration of the operation from the operation behavior trajectory data; Based on the preset duration range in which the operation stays and the historical decision accuracy of the experts providing the feedback data, the expert confidence weight is calculated. The expert's actions are converted into numerical values ​​using a preset reward mapping function. These numerical values ​​are then weighted using the expert confidence weight, and a supplementary weight coefficient is used to weight the logical verification deviation value calculated based on the parameter correction data. The sum of these weighted results yields the immediate reward value. The supplementary weight coefficient is the difference between 1 and the expert confidence weight, and the logical verification deviation value is calculated based on the extent to which the expert modifies the decision parameters in the preliminary agricultural decision-making scheme.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the agricultural decision optimization method based on multimodal perception and expert feedback as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the agricultural decision optimization method based on multimodal perception and expert feedback as described in any one of claims 1 to 6.

Images