Agricultural precision fertilization method

Abstract

The application relates to the technical field of intelligent fertilization, and discloses an agricultural precise fertilization method, which comprises the following steps: step 1, synchronously collecting multi-modal real-time data of single wheat and automatically identifying the current growth stage of the single wheat; step 2, searching historical fertility data of the position where the single wheat is located; step 3, fusing the real-time data, the historical fertility data and the growth stage label to generate an enhanced fusion feature vector; step 4, according to the current growth stage, selecting a corresponding causal diagram model, and estimating the average treatment effect of different fertilization schemes based on the enhanced fusion feature vector; and step 5, inputting the enhanced fusion feature vector, the average treatment effect estimation value and the growth stage label into a prediction model and an adjustment model. The application not only improves the precision of fertilization and the crop response effect, but also realizes the continuous optimization of the system through a closed-loop intelligent decision framework, and provides a high-efficiency fertilization solution which can be implemented and evolved for smart agriculture.

Description

Technical Field

[0001] This invention relates to the field of intelligent fertilization technology, and in particular to a method for precision fertilization in agriculture. Background Technology

[0002] Precision fertilization is one of the key technologies for increasing yields and reducing environmental pollution in modern agriculture. Traditional fertilization methods mainly rely on regional soil testing and empirical models, lacking precise perception of the real-time growth status and historical fertility of individual crops, leading to widespread problems of "over-application" or "under-fertilization" in fertilization programs. Although some technologies have attempted to introduce remote sensing monitoring or sensor data for fertilization guidance, the following significant shortcomings still exist:

[0003] The data dimensions are limited and lack multimodal fusion: Existing methods mostly rely on single types of data (such as spectral or soil data), failing to integrate multi-source information such as canopy, root zone, stem and leaf, making it difficult to fully capture crop status.

[0004] Poor stage adaptability: The nutrient requirements of crops vary significantly at different growth stages, while traditional methods lack the ability to dynamically identify growth stages and make stage-specific decisions.

[0005] The causal mechanism is unclear: most methods make decisions based on correlation rather than causation, which makes it impossible to accurately assess the actual effect of fertilization intervention and is easily affected by confounding factors.

[0006] Model update lag: The system lacks closed-loop feedback and online learning mechanisms, making it unable to continuously optimize the model based on actual fertilization effects, resulting in a decline in decision accuracy over time.

[0007] Therefore, there is an urgent need for a precision fertilization method that can integrate multimodal data, dynamically identify growth stages, combine causal inference and machine learning, and has online learning capabilities, in order to achieve truly intelligent fertilization with a "one plant, one policy" approach. Summary of the Invention

[0008] To address the technical problems mentioned in the background section, this invention provides a method for precision fertilization in agriculture.

[0009] This invention is achieved using the following technical solution: a precision fertilization method for agriculture, comprising the following steps:

[0010] Step 1: Synchronously collect multimodal real-time data of individual wheat plants and automatically identify their current growth stage;

[0011] Step 2: Retrieve historical fertility data for the location of the wheat plant;

[0012] Step 3: Integrate real-time data, historical fertility data, and growth stage labels to generate an enhanced fusion feature vector;

[0013] Step 4: Based on the current growth stage, select the corresponding cause-effect graph model and estimate the average treatment effect of different fertilization schemes based on the enhanced fusion feature vector.

[0014] Step 5: Input the enhanced fusion feature vector, average treatment effect estimate, and growth stage label into the prediction model and adjustment model to generate a stage-specific optimal fertilization plan;

[0015] Step 6: Perform fertilization and monitor feedback. Use the feedback data to update the prediction model, adjust the model, and the causal graph model for the corresponding growth stage.

[0016] Furthermore, the synchronous acquisition of multimodal real-time data from a single wheat plant includes:

[0017] Establish a unified spatiotemporal benchmark to ensure that different modal data collected from the same single plant are aligned in time and space; and verify and calibrate the validity of the collected data.

[0018] Based on a unified spatiotemporal benchmark, canopy hyperspectral / thermal infrared data, root zone soil data, and stem / leaf characteristic data of the same wheat plant are collected, and all data are bound with plant ID, timestamp, and spatial coordinates to generate a structured data package;

[0019] Based on the time of data collection from the drone canopy, the timestamps of data from other devices are finely adjusted to ensure that the timestamps of data from the same single plant are completely consistent.

[0020] Set a data validity threshold, first check whether the collected data contains all preset modal data to complete the integrity verification, and then filter out outliers that exceed the normal range of agronomy to complete the rationality verification.

[0021] Cross-validation methods, such as comparing chlorophyll values ​​retrieved by hyperspectral inversion with SPAD instrument measurements, are used to correct data deviations and ensure the accuracy of the collected data.

[0022] Furthermore, the automatic identification of its current growth stage includes:

[0023] Preprocess the multimodal real-time data;

[0024] Input preprocessed multimodal real-time data, start a multi-dimensional growth stage recognition model that includes a convolutional neural network recognition model, a phenological stage recognition model, and a hyperspectral image recognition model in parallel, and output three recognition results;

[0025] The three identification results mentioned above are combined and the final growth stage label is output.

[0026] A weighted fusion algorithm is used to integrate the results of the three recognition models. Weight coefficients are set to reflect the accuracy of each model at different growth stages. If the consistency of the fusion results reaches a preset threshold, a clear growth stage label is output. If the consistency does not reach the threshold, the current growth stage of a single wheat plant is corrected by combining the recognition results of the phenological stage model with the recognition results of the convolutional neural network model and the hyperspectral image recognition results.

[0027] Furthermore, the three recognition models include

[0028] A convolutional neural network recognition model is used to analyze preprocessed visible light images and preliminarily determine the growth stage by extracting visual features such as plant height, number of tillers, and ear morphology.

[0029] The phenological stage identification model predicts the growth stage of a plant by inputting accumulated temperature data and the current calendar date.

[0030] The hyperspectral image recognition model extracts features from hyperspectral image data and matches them with the unique spectral response patterns of each growth stage to obtain the spectral dimension of growth stage recognition results.

[0031] Furthermore, the retrieval of historical fertility data for the location of the wheat plant includes:

[0032] Based on the established unique spatial identifier of a single wheat plant and the bound single plant ID, the specific geographical coordinates of the single wheat plant are accurately located from the system database, and then the historical fertility data of the coordinates point in the past at least one complete growing season are obtained. The historical fertility data includes soil organic matter content records, basic soil N / P / K nutrient level data, corresponding season yield map and historical fertilization measures records.

[0033] Then, by encoding and converting the historical fertility data, a historical fertility feature vector is generated.

[0034] Furthermore, the process of fusing real-time data, historical fertility data, and growth stage labels to generate an enhanced fusion feature vector includes:

[0035] Based on the calibrated and verified structured real-time data package of single wheat plants, historical fertility feature vectors, and growth stage labels, the three types of data are standardized and aligned in time and space.

[0036] Multimodal feature extraction and interactive fusion are performed on the spatiotemporally aligned real-time data features, historical fertility features, and growth stage coding features to generate a preliminary fused feature vector; then, redundant features are removed by feature dimension compression to optimize the preliminary fused feature vector, and the final fused feature vector is output.

[0037] The final fused feature vector is then bound again to the individual plant ID and spatial coordinates to generate an enhanced fused feature vector.

[0038] Furthermore, the step of selecting the corresponding causal graph model based on the current growth stage and estimating the average treatment effect of different fertilization schemes based on the enhanced fusion feature vector includes:

[0039] Pre-stored cause-effect graph models that correspond one-to-one with each major growth stage of wheat;

[0040] Based on the current growth stage label, the corresponding target causal graph model is invoked;

[0041] The enhanced fusion feature vector is preprocessed for feature adaptation to obtain feature data that conforms to the model input standard;

[0042] The preprocessed feature data is input into the target causal graph model, and the conditional average treatment effect (average treatment effect) corresponding to different pre-fertilization schemes is calculated by the causal inference algorithm.

[0043] The calculated average treatment effect estimate is validated for reasonableness, and the validated average treatment effect estimate is stored.

[0044] The feature adaptation preprocessing includes adjusting the feature dimensions to match the input requirements of the target causal graph model, and removing redundant features that are irrelevant to the estimation of the current fertilization scheme.

[0045] The causal reasoning algorithm includes the causal forest algorithm or the propensity score matching algorithm;

[0046] The rationality verification is based on determining the validity threshold within the reasonable range of the historical average treatment effect of fertilization over the past three growing seasons of the plot, and eliminating abnormal estimates that exceed the threshold.

[0047] Furthermore, the step of jointly inputting the enhanced fusion feature vector, the average treatment effect estimate, and the growth stage label into the prediction model and the adjustment model includes:

[0048] The prediction model, based on the enhanced fusion feature vector, simulates the growth response of a single wheat plant under different candidate fertilization schemes and outputs multiple sets of prediction results.

[0049] The adjustment model receives multiple sets of prediction results and average treatment effect estimates output by the prediction model, and optimizes and selects the optimal fertilization scheme from the candidate fertilization schemes based on a reward function that matches the current growth stage.

[0050] The optimal fertilization plan includes fertilizer type, fertilizer dosage, and application timing;

[0051] The prediction model is a time-series prediction model based on long short-term memory network or Transformer architecture, used to predict the nutrient requirements and growth indicators of plants in future time periods.

[0052] The adjustment model is implemented using a reinforcement learning algorithm, and its reward function dynamically configures priorities according to different growth stages:

[0053] During the vegetative growth stage, the reward function prioritizes biomass growth and tiller number.

[0054] During the reproductive growth stage, the reward function prioritizes grain yield, thousand-grain weight, and quality indicators.

[0055] The adjustment model uses the average treatment effect estimate as prior knowledge to guide the reinforcement learning agent to converge more quickly in the action space to the fertilization scheme that produces the maximum expected causal reward.

[0056] Furthermore, the step of performing fertilization and monitoring feedback, and using the feedback data to simultaneously update the prediction model, adjust the model, and the causal graph model for the corresponding growth stage includes:

[0057] The control and execution equipment precisely fertilizes the target individual plant according to the optimal fertilization plan and records the actual fertilization data;

[0058] Within a preset time window after fertilization, multimodal feedback data for the single plant was collected again.

[0059] Based on the feedback data, the prediction results of the prediction model, and the average treatment effect estimate, a quantitative feedback signal is generated;

[0060] Using the feedback signal and the current decision-execution data, the parameters of the prediction model, the adjustment model, and the causal graph model of the corresponding growth stage are updated simultaneously.

[0061] The multimodal feedback data includes changes in soil nutrient content in the root zone after fertilization, changes in plant chlorophyll content, changes in plant morphological growth indicators, and actual environmental data.

[0062] Furthermore, the updated prediction model includes:

[0063] The input features, fertilization plan, and feedback data of this decision are used to form new training samples, and the model parameters are updated through an online learning mechanism.

[0064] The complete interactive experience of this decision is stored in the reinforcement learning experience replay buffer and used to update the policy network;

[0065] The actual fertilization effect was used as a new observation sample to correct the parameters in the causal graph model of the corresponding growth stage, so as to optimize the subsequent causal effect estimation.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] This invention achieves true single-plant-level precision management: by establishing a unified spatiotemporal benchmark and a multimodal data synchronous acquisition mechanism, it ensures that the data of each wheat plant can be independently traced and analyzed, overcoming the extensive problem of traditional regional uniform fertilization.

[0068] This invention integrates multi-source information to enhance the comprehensiveness of state perception: it integrates multimodal data such as canopy hyperspectral, thermal infrared, root zone soil, and stem morphology, and combines them with historical fertility background to generate an enhanced fusion feature vector, providing richer and more accurate input information for decision-making.

[0069] This invention dynamically identifies growth stages and enables stage-specific decision-making: it automatically identifies the current growth stage through a parallel multi-model fusion mechanism (CNN + phenological stage + hyperspectral) and selects the corresponding causal graph model and reward function based on the stage, making the fertilization strategy more in line with the physiological needs of crops.

[0070] This invention introduces causal inference to enhance the scientific nature of decision-making: by using causal graph models and average treatment effect estimation methods, the expected effects of different fertilization schemes are quantified, avoiding spurious correlation interference, and making the recommended schemes more interpretable and reliable.

[0071] This invention integrates prediction and reinforcement learning to optimize decision adaptability: by simulating the growth response under different schemes through the LSTM / Transformer prediction model, and then combining the stage-specific reward function with the reinforcement learning algorithm, the optimal scheme search in dynamic environment is achieved.

[0072] This invention constructs a closed-loop learning system to achieve continuous self-optimization: through an execution-monitoring-feedback-update mechanism, it uses real feedback data to simultaneously update the prediction model, adjust the model and the causal graph model, enabling the system to have online learning and long-term evolution capabilities, becoming more and more accurate with use.

[0073] In summary, this invention not only improves the accuracy of fertilization and crop response, but also achieves continuous system optimization through a closed-loop intelligent decision-making framework, providing a practical and evolving efficient fertilization solution for smart agriculture. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the workflow of the precision fertilization method for agriculture proposed in this invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0076] Example:

[0077] The 'causal graph model' mentioned in this scheme refers to a structured causal model based on a directed acyclic graph (DAG) used to represent the causal relationship between variables; the average treatment effect (ATE) refers to the expected causal effect of a treatment (such as fertilization) on the outcome variable (such as yield).

[0078] Please combine Figure 1 This embodiment proposes a precision fertilization method for agriculture, which includes:

[0079] Step 1: Synchronously collect multimodal real-time data of individual wheat plants and automatically identify their current growth stage;

[0080] In this embodiment, the synchronous acquisition of multimodal real-time data of a single wheat plant includes:

[0081] Establish a unified spatiotemporal benchmark to ensure that different modal data collected from the same single plant are aligned in time and space; and verify and calibrate the validity of the collected data.

[0082] Based on a unified spatiotemporal benchmark, canopy hyperspectral / thermal infrared data, root zone soil data, and stem / leaf characteristic data of the same wheat plant are collected, and all data are bound with plant ID, timestamp, and spatial coordinates to generate a structured data package.

[0083] Based on the data collection time of the drone canopy, the timestamps of other equipment are finely adjusted to ensure that the timestamps of the data for the same single plant are completely consistent.

[0084] Set a data validity threshold, first check whether the collected data contains all preset modal data to complete the integrity verification, and then filter out outliers that exceed the normal range of agronomy to complete the rationality verification.

[0085] Cross-validation methods, such as comparing chlorophyll values ​​retrieved via hyperspectral inversion with SPAD measurements, are used to correct data bias and ensure the accuracy of the collected data.

[0086] Specifically, firstly, a unified spatiotemporal reference is established. Using a GPS time synchronization module or a Network Time Protocol (NTP) server, the system time of all data acquisition devices (including hyperspectral cameras and thermal infrared cameras mounted on UAVs, and soil sensor nodes deployed in the field) is synchronized to UTC time, with synchronization accuracy controlled within milliseconds. Simultaneously, real-time dynamic differential positioning technology provides centimeter-level spatial positioning for each mobile acquisition device, assigning a unique spatial identifier (i.e., individual plant ID) to each wheat plant, which is bound to its geographic coordinates. Based on this, the central control system presets the acquisition parameters of each modal sensor (such as the band range of the hyperspectral camera, the temperature measurement range of the thermal infrared camera, and the sampling frequency of the soil sensor), and plans the autonomous navigation paths of UAVs or ground mobile robots to achieve sequential "plant-by-plant" data collection, ensuring that the devices arrive at the predetermined collection location for each wheat plant in sequence.

[0087] Next, based on the aforementioned unified spatiotemporal benchmark, multimodal data of the same single wheat plant were simultaneously collected. Specifically, this included: using a drone hovering approximately 2-3 meters directly above the plant canopy to simultaneously trigger hyperspectral and thermal infrared cameras to acquire hyperspectral images (covering the 400-1000nm band) and thermal infrared images of the canopy; using wireless soil sensor nodes pre-embedded in the root zone of the wheat plant to collect real-time data on nitrogen, phosphorus, and potassium nutrient content, soil volumetric water content, temperature, and electrical conductivity (EC value) of the root zone soil; and, if necessary, using a LiDAR or visible light camera mounted on a robotic arm to collect morphological characteristic data such as stem diameter and leaf tilt angle. All collected raw data was immediately tagged with metadata, including a single plant ID, a high-precision timestamp (format YYYY-MM-DDHH:MM:SS:ms), spatial coordinates (latitude, longitude, and elevation), and packaged into a structured data package (in JSON or binary format).

[0088] Next, fine-tuning of the timestamps is performed. Using the time of data acquisition from the UAV canopy as the baseline (because its sampling time is the most accurate), software algorithms are used to fine-tune the timestamps of data recorded by other fixed or slower-moving devices (such as soil sensors). For example, the data stream from soil sensors may have communication delays; by calculating the transmission delay and applying time offset compensation, it is ensured that the timestamps of all modal data belonging to the same wheat plant are completely consistent, with a maximum time error not exceeding 100 milliseconds.

[0089] Then, the collected data undergoes rigorous validity verification. Data validity thresholds are set: First, a completeness check is performed, verifying that each data packet contains data for all preset modes (e.g., simultaneously including hyperspectral, thermal infrared, and soil NPK data). If any mode is missing, the data packet is marked as invalid. Second, a reasonableness check is performed, setting reasonable ranges for each parameter based on agronomical knowledge (e.g., soil nitrogen content is typically 0-200 mg / kg, and canopy temperature is typically -10℃ to 50℃). The program automatically filters out outliers exceeding these reasonable ranges (e.g., negative or extremely high soil nitrogen content).

[0090] Finally, data accuracy calibration is implemented. Cross-validation is used to correct systematic errors: for example, 10% of wheat plants in the field are randomly selected, and at the same time as the drone's hyperspectral image acquisition, the SPAD values ​​(characterizing relative chlorophyll content) of their leaves are measured in the field using a handheld SPAD meter (such as the Japanese Konica Minolta SPAD-502Plus). The chlorophyll index (such as the TCARI / OSAVI index) retrieved from the hyperspectral image is then subjected to linear regression analysis with the in-situ SPAD measurements to establish a calibration model. This model is then used to perform batch calibration of all hyperspectral retrieved data to ensure data accuracy and reliability. Similarly, soil sensor readings can also be periodically calibrated and corrected using soil sample laboratory test results.

[0091] Through the steps outlined above, this approach yields high-quality, highly synchronous, and traceable multimodal single-plant data, laying a solid foundation for subsequent fusion analysis and intelligent decision-making.

[0092] In this embodiment, the automatic identification of its current growth stage includes:

[0093] Preprocessing is performed on multimodal real-time data.

[0094] Input preprocessed multimodal real-time data, start a multi-dimensional growth stage recognition model that includes a convolutional neural network recognition model, a phenological stage recognition model, and a hyperspectral image recognition model in parallel, and output three recognition results.

[0095] The three identification results mentioned above are combined and the final growth stage label is output.

[0096] A weighted fusion algorithm is used to integrate the results of the three recognition models. Weight coefficients are set to reflect the accuracy of each model at different growth stages. If the consistency of the fusion results reaches a preset threshold, a clear growth stage label is output. If the consistency does not reach the threshold, the current growth stage of a single wheat plant is corrected by combining the recognition results of the phenological stage model with the recognition results of the convolutional neural network model and the hyperspectral image recognition results.

[0097] In this embodiment, the three recognition models include:

[0098] A convolutional neural network recognition model is used to analyze preprocessed visible light images. By extracting visual features such as plant height, number of tillers, and ear morphology, the growth stage can be preliminarily determined.

[0099] The phenological stage identification model predicts the growth stage of a plant by inputting accumulated temperature data and the current calendar date.

[0100] The hyperspectral image recognition model extracts features from hyperspectral image data and matches them with the unique spectral response patterns of each growth stage to obtain the spectral dimension of growth stage recognition results.

[0101] Specifically, in this embodiment, the detailed implementation of automatically identifying the current growth stage of a single wheat plant includes the following steps.

[0102] First, the acquired multimodal real-time data undergoes standardization preprocessing. For visible light image data, image cropping, grayscale normalization, and scaling to a fixed resolution (e.g., ...) are performed. (pixels), and a histogram equalization algorithm is used to enhance image contrast to highlight plant morphological features. For hyperspectral image data, radiometric calibration and atmospheric correction are performed, the average spectral curve of the region of interest (ROI) is extracted, and noise is removed using Savitzky-Golay convolution smoothing. For environmental data, including the accumulated effective temperature before the acquisition date and the current calendar date, formatting is performed. The formula for calculating the effective accumulated temperature is:

[0103]

[0104] in, and These are the highest and lowest temperatures of the day. The base temperature for wheat development (usually 0℃ or 4℃) is denoted as n, which is the number of days since emergence.

[0105] Subsequently, the preprocessed multimodal data is input in parallel into three independent recognition models.

[0106] Convolutional Neural Network Recognition Model: This model employs a pre-trained deep residual network (such as ResNet-18) as its backbone, with pre-processed visible light images as input. The model automatically extracts key visual features such as plant height, tiller number, leaf droop, and ear visibility through convolutional layers. A Softmax classifier is then applied after the fully connected layers, outputting the probability distribution vector of the wheat plant's stage of growth (emergence, tillering, jointing, booting, heading, flowering, grain filling, and maturity). .

[0107] Phenological Stage Identification Model: This model is a rule-based expert system. Its inputs are the calculated cumulative effective accumulated temperature (GDD) and the current calendar date. The system pre-stores a phenological stage-accumulated temperature lookup table based on local variety characteristics (e.g., 200-400°C·d accumulated temperature is required for tillering, and 400-600°C·d for jointing). The model compares the current GDD with the lookup table and outputs the most probable growth stage determination. .

[0108] Hyperspectral Image Recognition Model: This model first performs a first-order derivative transformation on the preprocessed hyperspectral curve to amplify the differences in spectral features, and then uses Principal Component Analysis (PCA) for dimensionality reduction. The dimensionality-reduced features are then fed into a Support Vector Machine (SVM) classifier. This SVM classifier has been pre-trained with a large number of labeled samples and is capable of recognizing unique spectral response patterns at each growth stage (such as specific reflectance changes at the red edge position during the jointing stage). The final output is a classification result for the growth stage in one spectral dimension. .

[0109] Next, a weighted fusion algorithm is used to integrate the outputs of the three models to obtain the final growth stage labels. The weight coefficients of each model at different growth stages are set as shown in the table below (the weight coefficients are adjustable):

[0110] growth stage CNN model weights ( ) Phenological period model weights ( ) Hyperspectral model weights ( ) Dominant Model Vegetative growth (emergence - jointing) 0.4 0.4 0.2 CNN & Phenology Reproductive growth (heading to maturity) 0.3 0.3 0.4 Hyperspectral

[0111] Final fusion result Calculated by the following formula:

[0112]

[0113] in, The function converts the deterministic output of the phenological period and hyperspectral model into a form similar to... Probability distribution vectors of the same dimension (i.e., one-hot vectors where the value is 1 at the corresponding category position and 0 elsewhere). Final stage labels. Depend on The category corresponding to the highest probability value is determined.

[0114] Finally, consistency verification and conflict resolution are performed. A consistency index is calculated among the outputs of the three models. If at least two models output the same stage label, the consistency is considered to have reached a preset threshold (e.g., 80% confidence level), and the system directly outputs the final label. .

[0115] If the results of the three models are all different, and the consistency standard is not met, then the conflict resolution mechanism will be activated: the output of the phenological period model will be used as the benchmark. This is based on the assumption that accumulated temperature is the most stable indicator of developmental progress. Subsequently, the probability distribution output by the CNN model is compared. Results of the hyperspectral model .

[0116] If one of the results in the CNN or hyperspectral model is consistent with... If they are the same, then adopt the one. As the final output.

[0117] If the results of CNN and hyperspectral models are consistent, but different from those of other models... If they differ, check if the current GDD is near the critical range of the accumulated temperature for that phenological period (e.g., ±10%). If yes, adopt the consistent results of CNN and hyperspectral analysis; otherwise, still adopt the results. .

[0118] If all three are inconsistent, output a "cannot be determined" flag and prompt that manual intervention is required for judgment. At the same time, store the sample in the difficult sample library for subsequent model optimization.

[0119] Through the above-mentioned multi-model parallel processing and weighted decision fusion mechanism, the system can make full use of the advantages of different modal data, and significantly improve the accuracy and robustness of automatic identification of growth stages.

[0120] Step 2: Retrieve historical fertility data for the location of the wheat plant.

[0121] In this embodiment, retrieving historical fertility data for the location of the wheat plant includes:

[0122] Based on the established unique spatial identifier of a single wheat plant and the bound single plant ID, the specific geographical coordinates of the single wheat plant are accurately located from the system database, and then the historical fertility data of the coordinates point in the past at least one complete growing season are obtained. The historical fertility data includes soil organic matter content records, basic soil N / P / K nutrient level data, corresponding season yield map and historical fertilization measures records.

[0123] Then, by encoding and converting the historical fertility data, a historical fertility feature vector is generated.

[0124] Specifically, in this embodiment, the detailed implementation of retrieving historical fertility data of the location of the wheat plant and generating a feature vector includes the following steps:

[0125] First, based on the unique spatial identifier (i.e., its centimeter-precise geographic coordinates) and the associated plant ID for each wheat plant, the system performs a database retrieval operation. The system connects to a specially constructed relational database of farmland history information (which can use MySQL or PostgreSQL with PostGIS extensions). By executing parameterized SQL queries, the system precisely locates the coordinate point in the `plant_location` table of the database based on the input geographic coordinates of the plant (e.g., latitude and longitude: 116.405285, 39.904989). Considering the minor errors in field operations, the system sets a circular buffer zone with a radius of 10 centimeters for querying to ensure accurate matching of the corresponding plant record even with extremely slight positional offsets.

[0126] The system then retrieves historical fertility data for the geographic coordinate point from multiple related data tables for at least one past full growing season (e.g., the 2022-2023 season). This data includes:

[0127] Soil organic matter content record: Retrieve historical measurements of this location at different soil depths (e.g., 0-20cm, 20-40cm) from the `soil_organic_matter` table, including the measurement date and measurement value (unit: g / kg).

[0128] Soil basic N / P / K nutrient level data: retrieve the test data (unit: mg / kg) of available nitrogen (N), available phosphorus (P), and available potassium (K) at different stages (such as before sowing, jointing stage, and after harvest) at this location within the same growing season from the `soil_npk` table.

[0129] Yield map data for the corresponding season: The yield data (unit: kg / hm²) of the micro-region where this point is located was retrieved from the `yield_map` table for the previous growing season. This data comes from a high-precision yield map generated by the yield monitoring system on the combine harvester.

[0130] Historical fertilization records: Retrieve detailed records of all fertilization operations performed on the area near this location (e.g., within 1 square meter) during the previous growing season from the `fertilization_history` table. This includes the fertilization date, fertilizer type (e.g., urea, superphosphate), fertilizer application rate (unit: kg / hm²), fertilization method (e.g., basal fertilizer, top dressing), and the actual effective content of N, P₂O₅, and K₂O in the fertilizer.

[0131] Next, the system cleans and encodes the retrieved multi-source, heterogeneous, and time-series historical fertility data to generate a standardized historical fertility feature vector that can be used as input to the model. The process is as follows:

[0132] Numerical data standardization: For numerical data such as soil organic matter, N, P, and K content, firstly, obvious outliers are removed (for example, using the Z-score method to remove values ​​with an absolute value greater than 3). Then, all values ​​are standardized to the [0,1] interval. The standardization formula is:

[0133]

[0134] in, and These are the historical maximum and minimum values ​​of this parameter for this plot of land.

[0135] Categorical data encoding: For categorical data such as fertilizer type and fertilization method, one-hot encoding is used to convert them into binary vectors. For example, fertilizer types ["urea", "superphosphate", "potassium chloride"] are encoded as [1,0,0],[0,1,0],[0,0,1].

[0136] Temporal Feature Construction: For data with time-series characteristics (such as soil nutrient content at different periods), additional temporal statistical features are constructed. This includes calculating the average, maximum, minimum, and standard deviation for the entire growing season, as well as the rate of change (slope) between two adjacent measurements. These statistics collectively characterize the dynamic changes in soil fertility at that location.

[0137] The relationship between yield and fertilization: The partial factor productivity (PFP) of fertilizer in the previous growing season is calculated using the following formula:

[0138]

[0139] in, For the output at this point, This represents the total amount of nitrogen fertilizer applied during the season. Similarly, the amount can be calculated... and This indicator reflects the efficiency of historical fertilization.

[0140] Finally, all the cleaned, standardized, encoded, and derived features are concatenated in a predetermined order into a one-dimensional, high-dimensional numerical vector, which is the final historical fertility feature vector. The vector has a fixed dimension (for example, a vector containing soil properties, yield, fertilization operations, and efficiency indicators may have 50 dimensions), facilitating subsequent fusion with real-time data. The system binds this vector to the current plant ID and caches it for use by the next feature fusion module.

[0141] Through the steps described above, this solution transforms unstructured, scattered historical data into structured, machine-readable numerical features, providing the model with crucial historical background information and enabling its decisions to be based on a deep understanding of the long-term fertility status of the land parcel.

[0142] Step 3: Integrate real-time data, historical fertility data, and growth stage labels to generate an enhanced fusion feature vector.

[0143] In this embodiment, the process of fusing real-time data, historical fertility data, and growth stage labels to generate an enhanced fusion feature vector includes:

[0144] Based on the calibrated and verified structured real-time data package of single wheat plants, historical fertility feature vectors, and growth stage labels, the three types of data are standardized and aligned in time and space.

[0145] Multimodal feature extraction and interactive fusion are performed on the spatiotemporally aligned real-time data features, historical fertility features, and growth stage coding features to generate a preliminary fused feature vector; then, redundant features are removed by feature dimension compression to optimize the preliminary fused feature vector, and the final fused feature vector is output.

[0146] The final fused feature vector is then bound again to the individual plant ID and spatial coordinates to generate an enhanced fused feature vector.

[0147] Specifically, firstly, the three types of input data are standardized and preprocessed to ensure that they can be effectively integrated on the same scale and dimension.

[0148] Real-time data standardization: The calibrated and verified structured real-time data packets (containing hyperspectral features, thermal infrared features, soil sensor data, etc.) obtained from the data acquisition module are Z-score standardized. For each feature parameter (such as nitrogen content value, canopy temperature value), its mean (μ) and standard deviation (σ) in the batch of data are calculated and then applied according to the formula. The transformation is performed to obtain the standardized real-time feature vector. .

[0149] Historical feature vector standardization: This involves retrieving and encoding historical fertility feature vectors from the database. (This may include indicators with different dimensions such as organic matter content, historical yield, and PFP.) Min-Max standardization is performed, linearly scaling all values ​​to the [0,1] interval. The formula is as follows: ,get .

[0150] Growth stage label encoding: The current growth stage label (e.g., "jointing stage") is converted into a high-dimensional numerical representation. One-hot encoding is used to construct an 8-dimensional vector from all possible growth stages (e.g., emergence, tillering, jointing, booting, heading, flowering, grain filling, and maturity). The current stage is assigned a 1, and the rest are 0, generating the growth stage encoding vector. .

[0151] Secondly, perform spatiotemporal alignment to ensure that all features point to the same target and time.

[0152] Spatial alignment: All feature vectors are associated with their bound individual plant IDs and centimeter-level spatial coordinates. The system ensures this through database join queries. They all correspond to the same wheat plant.

[0153] Time alignment: Real-time data packets themselves have high-precision timestamps. Historical data, although from the past, has its time attributes (such as which growing season it belongs to) encoded within... Among the features, the growth stage label represents the phenological state at the current moment. The three are aligned temporally through the context of the "current decision-making moment."

[0154] Next, multimodal feature extraction and interactive fusion are performed. This system uses a neural network module based on the cross-attention mechanism to achieve deep feature fusion.

[0155] Model Input: This network has three input branches, which receive:

[0156] Real-time features after projection through a fully connected layer .

[0157] Historical features after projection through another fully connected layer .

[0158] Growth stage encoding vector Use it directly as a condition input.

[0159] Cross-attention calculation: Using real-time features as the Query (Q) and historical features as the Key (K) and Value (V), calculate the attention weights of historical information on the current state.

[0160]

[0161] This approach aims to enhance the data by retrieving information from historical fertility data that is most relevant to the current real-time state. Similarly, attention can be calculated in reverse, with historical data as Q, and real-time data as K and V, to capture the correlation between the current state and historical patterns.

[0162] Feature concatenation and modulation: The features obtained through the cross-attention mechanism are combined with the original projected features. The vectors are then concatenated. Subsequently, the encoding vectors from the growth stage are used. The concatenated features are then modulated, for example through Conditional Batch Normalization or affine transformation, to adapt the fusion process to different growth stages. This generates a rich preliminary fused feature vector. .

[0163] Then, the initially fused feature vectors are subjected to dimensionality reduction and optimization. Because Due to the high dimensionality and potential redundancy, feature compression is performed using principal component analysis (PCA) or the encoder portion of an autoencoder.

[0164] If PCA is used, the principal components of all preliminary fused feature vectors in this batch are calculated, and the top N principal components with a variance of more than 95% are selected to compress the vector dimension from the original D dimensions to N dimensions.

[0165] If an autoencoder is used, then a pre-trained encoder model will be utilized to... It is compressed into a low-dimensional, information-rich latent representation.

[0166] This step outputs an optimized, final fused feature vector that has eliminated redundant information.

[0167] Finally, vector binding and output are performed. The final generated fused feature vector is then output. The vector is then further bound to the individual wheat plant ID and spatial coordinates to generate the final enhanced fusion feature vector. This vector, as a structured data object, is fed into the downstream prediction and decision-making modules.

[0168] Through the above process, this solution achieves deep fusion of multi-source heterogeneous data at the feature level, generating an enhanced feature representation that reflects the current real-time status, contains historical fertility background, and fits the current phenological period.

[0169] Step 4: Select the corresponding causal graph model based on the current growth stage, and estimate the average treatment effect of different fertilization schemes based on the enhanced fusion feature vector.

[0170] In general, it includes:

[0171] Pre-stored cause-effect graph models that correspond one-to-one with each major growth stage of wheat.

[0172] Based on the current growth stage label, the corresponding target causal graph model is invoked.

[0173] The enhanced fusion feature vector is preprocessed for feature adaptation to obtain feature data that conforms to the model input standard.

[0174] The preprocessed feature data is input into the target causal graph model, and the conditional average treatment effect (average treatment effect) corresponding to different pre-fertilization schemes is calculated by the causal inference algorithm.

[0175] The calculated average treatment effect estimate is validated for reasonableness, and the validated average treatment effect estimate is stored.

[0176] The feature adaptation preprocessing includes adjusting the feature dimensions to match the input requirements of the target causal graph model, and removing redundant features that are irrelevant to the estimation of the current fertilization scheme.

[0177] The causal reasoning algorithm includes the causal forest algorithm or the propensity score matching algorithm.

[0178] The rationality verification is based on determining the validity threshold within the reasonable range of the historical average treatment effect of fertilization over the past three growing seasons of the plot, and eliminating abnormal estimates that exceed the threshold.

[0179] In this scheme, the average treatment effect is used to quantify the average expected impact of a particular fertilization intervention (i.e., "treatment") on crop growth outcomes. For example, it is the average difference in a specific crop outcome variable (such as yield, biomass, grain weight) between applying a particular fertilization regimen (e.g., applying 100 kg / ha of nitrogen) and not applying the regimen (or applying a control regimen), after controlling for other confounding factors.

[0180] More specifically, firstly, the system pre-stores a causal graph model library. This library contains causal graph models corresponding one-to-one with each major growth stage of wheat. Each model is a structured causal model (SCM) defined in the form of a directed acyclic graph, clearly depicting the causal relationships between various variables (such as soil nutrients, environmental factors, fertilization intervention, and plant growth indicators) at a specific growth stage. For example, in the causal graph of the tillering stage, the edge weights of "nitrogen application rate" pointing to "tiller number" and "biomass" are relatively large, while in the graph of the grain-filling stage, the paths of "potassium application rate" pointing to "thousand-grain weight" and "grain protein content" are more significant. These models are learned using PC or FCI algorithms based on a large amount of historical observation data, field trial data, and domain expert knowledge, and are stored in the model library in standardized data formats (such as XML or JSON).

[0181] Based on the current growth stage label (such as "jointing stage") received from the feature fusion module, the system queries and calls the target causal graph model that perfectly matches the stage from the pre-stored model library.

[0182] Next, the received enhanced fused feature vectors undergo feature adaptation preprocessing to make them conform to the input requirements of the target causal graph model. This preprocessing includes:

[0183] Dimensional Adjustment: The target causal graph model predetermines the dimension of its input features (e.g., requiring a 50-dimensional input vector). If the dimension of the enhanced fusion feature vector does not match this, it is projected to the target dimension through a pre-trained fully connected neural network layer. This projection layer is optimized along with the causal graph model during the model training phase.

[0184] Feature selection: Irrelevant features are removed based on the type of fertilization program being evaluated. For example, if only nitrogen application is being evaluated, features highly correlated with phosphorus and potassium responses (such as historical soil available phosphorus content and available potassium change rate) are removed from the feature vector. This is achieved through a predefined feature-output correlation mapping table, which indicates which features contribute less than a preset threshold (e.g., <0.1) to the effect estimation of which fertilization intervention, and thus filter them out. The final output is a refined feature data array that meets the model input criteria. `.

[0185] Then, the preprocessed feature data... Input the target causal graph model, and calculate the average treatment effect corresponding to different pre-treatment fertilization schemes (Treatment, T) using a causal inference algorithm. Specifically:

[0186] This embodiment preferably employs the causal forest algorithm. This algorithm is an application of generalized random forest in causal inference. The system will... All features are used as covariates, fertilization schemes (such as "0 kg / hm² nitrogen", "100 kg / hm² nitrogen", "150 kg / hm² nitrogen") are used as treatment variables (T), and outcome variables of interest (Y, such as biomass increment or yield in the future) are used as outputs.

[0187] The algorithm estimates a personalized treatment effect for each pre-fertilization treatment `T=i` relative to the control treatment (e.g., `T=0`) for the current wheat plant. `, which is the average treatment effect. Its mathematical expression is:

[0188]

[0189] in, Indicates the application of processing Potential consequences This indicates the potential outcome under control.

[0190] As an alternative, propensity score matching (PSM) can be used. The system will first base its matching on features... The propensity score of the wheat plant to receive each fertilization scheme was estimated using a logistic regression model. Then, neighbor samples with similar propensity scores but different treatments were found in historical data. The average treatment effect was estimated by calculating the difference in outcome variables among the neighbor samples.

[0191] Subsequently, the calculated average treatment effect estimates were performed on all treatment effects. Perform a rationality check. The system maintains a database of historical average fertilization treatment effects for the past three growing seasons for this plot. For the current growth stage and similar soil background, calculate the distribution of historical average treatment effect values ​​(e.g., mean μ and standard deviation σ). Set the validity threshold range to `[μ-3σ,μ+3σ]` (based on the three Sigma rule) or use quantiles (e.g., 1% and 99th percentiles). Then, use the calculated... Compare with the threshold range:

[0192] If ` If the value falls within the threshold range, the valuation is considered reasonable and will be retained.

[0193] If ` If the value exceeds the threshold, it is marked as an abnormal estimate. The system will then trigger a review process: first, the input features are checked. Are there any outliers? If not, replace the outlier with the median or mean of the average treatment effect of similar fertilization programs in the same historical period.

[0194] Finally, the average treatment effect estimate of all verified values ​​is... The fertilization scheme `i`, along with the individual tree ID and timestamp, is stored in the average treatment effect database, and a confidence score is generated for each scheme `i` (based on the variance of tree estimates in a causal forest or the quality calculation of matched samples in a PSM). These stored estimates will be directly used in the subsequent optimization decision-making module.

[0195] Through the steps described above, this scheme achieves the quantification of personalized fertilization effects based on causal inference, providing a scientific and reliable quantitative basis for generating optimal decisions.

[0196] Step 5: Input the enhanced fusion feature vector, average treatment effect estimate, and growth stage label into the prediction model and adjustment model to generate a stage-specific optimal fertilization plan.

[0197] In general, it includes the following sub-steps:

[0198] The prediction model, based on the enhanced fusion feature vector, simulates the growth response of a single wheat plant under different candidate fertilization schemes and outputs multiple sets of prediction results.

[0199] The adjustment model receives multiple sets of prediction results and average treatment effect estimates output by the prediction model, and optimizes and selects the optimal fertilization scheme from the candidate fertilization schemes based on a reward function that matches the current growth stage.

[0200] The optimal fertilization plan includes the type of fertilizer, the dosage, and the timing of application.

[0201] The prediction model is a time-series prediction model based on long short-term memory networks or Transformer architecture, used to predict the nutrient requirements and growth indicators of plants in future time periods.

[0202] The adjustment model is implemented using a reinforcement learning algorithm, and its reward function dynamically configures priorities according to different growth stages.

[0203] During the vegetative growth stage, the reward function prioritizes biomass growth and tillering number.

[0204] During the reproductive growth stage, the reward function prioritizes grain yield, thousand-grain weight, and quality indicators.

[0205] The adjustment model uses the average treatment effect estimate as prior knowledge to guide the reinforcement learning agent to converge more quickly in the action space to the fertilization scheme that produces the maximum expected causal reward.

[0206] In detail, firstly, the system initializes the model and prepares the input data. The prediction model is a time-series prediction model based on an encoder-decoder architecture, where the encoder part can employ stacked Long Short-Term Memory (LSTM) network layers or Transformer encoder layers. The model is pre-trained on a large amount of historical wheat growth time-series data. The system then enhances the fusion of feature vectors. The features are concatenated with the one-hot encoded vectors of the growth stage labels to form the final model input features. At the same time, the system predefines a set of candidate fertilization schemes. Each of the schemes It is a vector that explicitly specifies the type of fertilizer (e.g., [nitrogen, phosphorus, potassium]), dosage (e.g., [100, 50, 80] kg / ha), and planned application time (e.g., "immediately" or "3 days later").

[0207] Next, the prediction model performs multi-scenario simulations. For each candidate fertilization plan... ;

[0208] Will and Input them into the prediction model together. The model will... As the initial state code, This is an intervention applied at the starting point of the prediction time step.

[0209] The model recursively predicts key growth indicators of plants on a daily basis during a critical growth period (e.g., from the present to the start of the next growth stage, typically 7-15 days). These indicators include:

[0210] Vegetative growth indicators: plant height, number of tillers, leaf area index (LAI), and aboveground biomass.

[0211] Reproductive growth / physiological indicators: number of ears, potential number of grains, grain filling rate, and canopy nitrogen content.

[0212] Nutrient requirement indicators: cumulative absorption of nitrogen, phosphorus and potassium by plants.

[0213] The predictive model output is for the solution Multi-step prediction results .

[0214] The model is then adjusted to optimize decision-making. This model is implemented by a reinforcement learning agent, for example, using proximal policy optimization (PPO).

[0215] State definition: Current state That is It contains all the current status information of the plant.

[0216] Action space: The action space is discrete, that is, the set of candidate fertilization schemes. .

[0217] Reward function design: Reward function The weights of the tags are dynamically configured based on the growth stage to reflect stage-specific objectives:

[0218] Vegetative growth stages (such as tillering stage and jointing stage):

[0219]

[0220] in, and It is the increase in biomass and tiller number at the end of the period relative to the beginning of the period, as output by the prediction model. These are higher weighting coefficients (e.g., 0.5, 0.4). It's the cost of fertilizer. It is a small cost penalty coefficient (e.g., 0.1).

[0221] Reproductive growth stages (such as the booting stage and grain-filling stage):

[0222]

[0223] in, It is based on the predicted yield, thousand-grain weight, and protein content estimated from the end-stage indicators. These are higher weighting coefficients (such as 0.4, 0.3, 0.2).

[0224] Value assessment and decision-making: agents through their policy networks Assessment in status Select each action below The expected cumulative reward. Average treatment effect estimate. In this process, it is introduced as valuable prior knowledge:

[0225] One way is to As part of the baseline or advantage function, it directly adjusts the value assessment. :

[0226]

[0227] in, It is a hyperparameter that adjusts the intensity of the average treatment effect.

[0228] Another way is to use To guide exploration, in the early stages of training or when encountering new situations, prioritize exploring actions with high positive average treatment effect estimates.

[0229] The agent's final choice can maximize action As the optimal fertilization plan.

[0230] Finally, the system outputs the decision result: Optimal fertilization plan. It is formatted as a structured decision instruction.

[0231] Through the above process, this solution enables the system to make precise decisions by combining data-driven and causal inference, and can generate a theoretically optimal and practically feasible fertilization plan for each wheat plant at its specific growth stage.

[0232] Step 6: Perform fertilization and monitor feedback. Use the feedback data to update the prediction model, adjust the model, and the causal graph model for the corresponding growth stage.

[0233] In general, it includes the following sub-steps:

[0234] The control and execution equipment performs precise fertilization on the target individual plants according to the optimal fertilization plan and records the actual fertilization data.

[0235] Within a preset time window after fertilization, multimodal feedback data for the single plant were collected again.

[0236] Based on the feedback data, the prediction results of the prediction model, and the estimated average treatment effect, a quantitative feedback signal is generated.

[0237] Using the feedback signal and the current decision-execution data, the parameters of the prediction model, adjustment model, and causal graph model of the corresponding growth stage are updated simultaneously.

[0238] The multimodal feedback data includes changes in soil nutrient content in the root zone after fertilization, changes in plant chlorophyll content, changes in plant morphological growth indicators, and actual environmental data.

[0239] In this embodiment, the updated prediction model includes:

[0240] The input features, fertilization plan and feedback data of this decision are used to form new training samples, and the model parameters are updated through an online learning mechanism.

[0241] The complete interactive experience of this decision is stored in the reinforcement learning experience replay buffer and used to update the policy network;

[0242] The actual fertilization effect was used as a new observation sample to correct the parameters in the causal graph model of the corresponding growth stage, so as to optimize the subsequent causal effect estimation.

[0243] More specifically, the detailed implementation of performing fertilization and monitoring feedback, and using feedback data to simultaneously update the prediction model, adjust the model, and the causal graph model of the corresponding growth stage includes the following steps:

[0244] Step 1: Precise Execution and Data Recording

[0245] The system transmits the generated optimal fertilization plan to an autonomous fertilization execution device in the field (such as an agricultural robot or drone equipped with a precision fertilization mechanism) via a wireless communication protocol. This device navigates to the precise coordinates of the target plant (e.g., the plant with the ID "FIELD_01_ROW_005_PLANT_012") using RTK-GPS. The fertilization mechanism (such as a metering pump and nozzle) accurately applies the specified type and dosage of fertilizer according to the plan instructions. The execution device's built-in flow and position sensors monitor the fertilization process in real time, recording and transmitting actual fertilization data (including actual application rate, actual execution time, and machine operating speed) back to the central system for comparison and verification with the corresponding decision instructions, ensuring accuracy.

[0246] Step 2: Multimodal feedback data acquisition.

[0247] Within a preset time window after fertilization (set according to the growth stage, e.g., 5-7 days for vegetative growth and 7-10 days for reproductive growth), the system restarts the data acquisition process to track and monitor the target plant, collecting multimodal feedback data. Specifically, this includes:

[0248] Soil response data: The changes in the content of available nitrogen, phosphorus, and potassium nutrients in the soil and the changes in soil volumetric water content were collected on days 1, 3, 5, and 7 after fertilization using wireless soil sensor nodes buried in the root zone of the plant.

[0249] Plant physiological response data: Canopy images were collected again by a drone-borne hyperspectral imager, and the changes in chlorophyll content and leaf area index were calculated by inversion algorithm.

[0250] Top and side views of the plants are captured by a visible light camera mounted on a ground robot, and image processing algorithms are used to measure the increase in plant height and the change in the number of tillers.

[0251] Actual environmental data: The actual rainfall (mm), average temperature (°C), and total sunshine hours (h) within the time window after fertilization were obtained from field weather stations. These data were used to assess the actual impact of environmental factors on the fertilization effect.

[0252] Step 3: Quantization feedback signal generation.

[0253] The system will compare and analyze the monitored feedback data with the prediction results of the fertilization program by the prediction model in step 5 and the average treatment effect estimated by the causal reasoning submodule in step 4.

[0254] For continuous variables (such as biomass), calculate the root mean square error (RMSE):

[0255]

[0256] For effect assessment, calculate the average treatment effect estimation error:

[0257]

[0258] in, The estimate is made by comparing the changes in indicators of the target plant with those of similar plants in an unfertilized control area.

[0259] Based on these errors, a quantized comprehensive feedback signal is generated. The signal can be a scalar value (e.g., a reward value between -1 and 1), where a positive value indicates that the effect meets or exceeds expectations, and a negative value indicates that the effect does not meet expectations.

[0260] Step 4: Multi-model collaborative online update.

[0261] Using the generated feedback signals and the complete data of this decision-making and execution process (denoted as a quadruple) The three models are updated online:

[0262] Update the prediction model:

[0263] The input features for this decision (enhanced fusion feature vector) Implementation of fertilization program and the actual growth results measured in the feedback data. To form a new training sample .

[0264] Online learning mechanisms, such as stochastic gradient descent, are employed to incrementally update the weights of the prediction model with only the new sample at a small learning rate (e.g., 0.001). This aims to fine-tune the model rather than retrain it, making its future predictions more realistic.

[0265] Update and adjust the model (reinforcement learning agent):

[0266] The complete interactive experience quadruple Store it in its experience replay buffer.

[0267] Periodically (e.g., whenever 100 new experiences accumulate in the buffer), a small batch of experiences is randomly sampled from the buffer.

[0268] Using these sampled experiences, the parameters of its Actor network (policy network) and Critic network (value network) are updated through the update rules of the PPO or SAC algorithm, thereby optimizing its decision-making strategy so that it can make better choices in similar future states.

[0269] Update the cause-effect graph model:

[0270] The actual "fertilization intervention-result" data was used as a new observation sample to revise the causal graph model for the corresponding growth stage.

[0271] Specifically, incremental learning algorithms are used to update the model parameters used to calculate the average treatment effect. For example, if a causal forest is used, new samples are added to the training set, and some branches of the affected decision trees in the forest are partially reconstructed.

[0272] If an SCM-based approach is used, a Bayesian update rule is employed to adjust the coefficients in the structural equations, making the model's output more consistent with the newly observed data.

[0273] Step 5: Data Archiving and Version Management

[0274] All data from this cycle, including initial state, decision instructions, execution records, feedback results, model parameters before and after the update, are compressed, timestamped, and archived in the historical database. The system establishes a management log for model versions to ensure that each update is traceable and that a rollback to a previous stable version is possible in the event of model performance degradation.

[0275] Through the aforementioned closed-loop process, this solution not only enables the system to perform precise fertilization but, more importantly, completes a full learning iteration. This allows the intelligent agent to make more adaptive and accurate decisions when facing the same or similar situations in the future, truly achieving the system's self-optimization and continuous evolution.

[0276] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for precision fertilization in agriculture, characterized in that, Includes the following steps: Step 1: Synchronously collect multimodal real-time data of individual wheat plants and automatically identify their current growth stage; Step 2: Retrieve historical fertility data for the location of the wheat plant; Step 3: Integrate real-time data, historical fertility data, and growth stage labels to generate an enhanced fusion feature vector; Step 4: Based on the current growth stage, select the corresponding cause-effect graph model and estimate the average treatment effect of different fertilization schemes based on the enhanced fusion feature vector. Step 5: Input the enhanced fusion feature vector, average treatment effect estimate, and growth stage label into the prediction model and adjustment model to generate a stage-specific optimal fertilization plan; Step 6: Perform fertilization and monitor feedback. Use the feedback data to update the prediction model, adjust the model, and the causal graph model for the corresponding growth stage. The process of fusing real-time data, historical fertility data, and growth stage labels to generate an enhanced fusion feature vector includes: Based on the calibrated and verified structured real-time data package of single wheat plants, historical fertility feature vectors, and growth stage labels, the three types of data are standardized and aligned in time and space. Multimodal feature extraction and interactive fusion are performed on the spatiotemporally aligned real-time data features, historical fertility features, and growth stage coding features to generate a preliminary fused feature vector; then, redundant features are removed by feature dimension compression to optimize the preliminary fused feature vector, and the final fused feature vector is output. The final fused feature vector is then bound again to the individual plant ID and spatial coordinates to generate an enhanced fused feature vector. The step of selecting the corresponding causal graph model based on the current growth stage and estimating the average treatment effect of different fertilization schemes based on the enhanced fusion feature vector includes: Pre-stored cause-effect graph models that correspond one-to-one with each major growth stage of wheat; Based on the current growth stage label, the corresponding target causal graph model is invoked; The enhanced fusion feature vector is preprocessed for feature adaptation to obtain feature data that conforms to the model input standard; The preprocessed feature data is input into the target causal graph model, and the conditional average treatment effect corresponding to different pre-fertilization schemes is calculated by the causal inference algorithm. The calculated average treatment effect estimate is validated for reasonableness, and the validated average treatment effect estimate is stored. The feature adaptation preprocessing includes adjusting the feature dimensions to match the input requirements of the target causal graph model, and removing redundant features that are irrelevant to the estimation of the current fertilization scheme. The causal reasoning algorithm includes the causal forest algorithm or the propensity score matching algorithm; The rationality verification is based on the reasonable range of the average treatment effect of historical fertilization in the three growing seasons of the plot to determine the effectiveness threshold, and abnormal estimates that exceed the threshold are eliminated. The step of inputting the enhanced fusion feature vector, the average treatment effect estimate, and the growth stage label into the prediction model and the adjustment model includes: The prediction model, based on the enhanced fusion feature vector, simulates the growth response of a single wheat plant under different candidate fertilization schemes and outputs multiple sets of prediction results. The adjustment model receives multiple sets of prediction results and average treatment effect estimates output by the prediction model, and optimizes and selects the optimal fertilization scheme from the candidate fertilization schemes based on a reward function that matches the current growth stage. The optimal fertilization plan includes fertilizer type, fertilizer dosage, and application timing; The prediction model is a time-series prediction model based on long short-term memory network or Transformer architecture, used to predict the nutrient requirements and growth indicators of plants in future time periods. The adjustment model is implemented using a reinforcement learning algorithm, and its reward function dynamically configures priorities according to different growth stages: During the vegetative growth stage, the reward function prioritizes biomass growth and tiller number. During the reproductive growth stage, the reward function prioritizes grain yield, thousand-grain weight, and quality indicators. The adjustment model uses the average treatment effect estimate as prior knowledge to guide the reinforcement learning agent to converge more quickly in the action space to the fertilization scheme that produces the maximum expected causal reward.

2. The precision fertilization method for agriculture as described in claim 1, characterized in that, The synchronous acquisition of multimodal real-time data from a single wheat plant includes: Establish a unified spatiotemporal benchmark to ensure that different modal data collected from the same single plant are aligned in time and space; and verify and calibrate the validity of the collected data. Based on a unified spatiotemporal benchmark, canopy hyperspectral / thermal infrared data, root zone soil data, and stem / leaf characteristic data of the same wheat plant are collected, and all data are bound with plant ID, timestamp, and spatial coordinates to generate a structured data package; Based on the time of data collection from the drone canopy, the timestamps of data from other devices are finely adjusted to ensure that the timestamps of data from the same single plant are completely consistent. Set a data validity threshold, first check whether the collected data contains all preset modal data to complete the integrity verification, and then filter out outliers that exceed the normal range of agronomy to complete the rationality verification. Cross-validation methods, such as comparing chlorophyll values ​​retrieved via hyperspectral inversion with SPAD measurements, are used to correct data bias and ensure the accuracy of the collected data.

3. The precision fertilization method for agriculture as described in claim 1, characterized in that, The automatic identification of its current growth stage includes: Preprocess the multimodal real-time data; Input preprocessed multimodal real-time data, start a multi-dimensional growth stage recognition model that includes a convolutional neural network recognition model, a phenological stage recognition model, and a hyperspectral image recognition model in parallel, and output three recognition results; The three identification results mentioned above are combined and the final growth stage label is output. A weighted fusion algorithm is used to integrate the results of the three recognition models. Weight coefficients are set to reflect the accuracy of each model at different growth stages. If the consistency of the fusion results reaches a preset threshold, a clear growth stage label is output. If the consistency does not reach the threshold, the current growth stage of a single wheat plant is corrected by combining the recognition results of the phenological stage model with the recognition results of the convolutional neural network model and the hyperspectral image recognition results.

4. The precision fertilization method for agriculture as described in claim 3, characterized in that, The three recognition models include A convolutional neural network recognition model is used to analyze preprocessed visible light images. By extracting the following visual features: plant height, number of tillers, and ear morphology, the growth stage is preliminarily determined. The phenological stage identification model predicts the growth stage of a plant by inputting accumulated temperature data and the current calendar date. The hyperspectral image recognition model extracts features from hyperspectral image data and matches them with the unique spectral response patterns of each growth stage to obtain the spectral dimension of growth stage recognition results.

5. The precision fertilization method for agriculture as described in claim 1, characterized in that, The historical fertility data retrieved for the location of the wheat plant includes: Based on the established unique spatial identifier of a single wheat plant and the bound single plant ID, the specific geographical coordinates of the single wheat plant are accurately located from the system database, and then the historical fertility data of the coordinates point in the past at least one complete growing season are obtained. The historical fertility data includes soil organic matter content records, basic soil N / P / K nutrient level data, corresponding season yield map and historical fertilization measures records. Then, by encoding and converting the historical fertility data, a historical fertility feature vector is generated.

6. The precision fertilization method for agriculture as described in claim 1, characterized in that, The process of performing fertilization and monitoring feedback, and using the feedback data to simultaneously update the prediction model, adjust the model, and the causal graph model for the corresponding growth stage, includes: The control and execution equipment precisely fertilizes the target individual plant according to the optimal fertilization plan and records the actual fertilization data; Within a preset time window after fertilization, multimodal feedback data for the single plant was collected again. Based on the feedback data, the prediction results of the prediction model, and the average treatment effect estimate, a quantitative feedback signal is generated; Using the feedback signal and the current decision-execution data, the parameters of the prediction model, the adjustment model, and the causal graph model of the corresponding growth stage are updated simultaneously. The multimodal feedback data includes changes in soil nutrient content in the root zone after fertilization, changes in plant chlorophyll content, changes in plant morphological growth indicators, and actual environmental data.

7. The precision fertilization method for agriculture as described in claim 1, characterized in that, The updated prediction model includes: The input features, fertilization plan, and feedback data of this decision are used to form new training samples, and the model parameters are updated through an online learning mechanism. The complete interactive experience of this decision is stored in the reinforcement learning experience replay buffer and used to update the policy network; The actual fertilization effect was used as a new observation sample to correct the parameters in the causal graph model of the corresponding growth stage, so as to optimize the subsequent causal effect estimation.

Images