Intelligent fertilization management method and system based on machine learning

Abstract

The application relates to the technical field of farmland fertilization management, and discloses an intelligent fertilization management method and system based on machine learning. The method collects a soil parameter set and an environmental parameter set of a target farmland, the soil parameters include soil humidity, nitrogen, phosphorus and potassium contents and pH values, and the environmental parameters include illumination intensity, temperature and rainfall; a soil nutrient dynamic change model is constructed according to historical data, and soil nutrient consumption trends in a future preset period are predicted; an initial fertilization scheme is generated based on the nutrient consumption trends and crop growth stage characteristics; a multi-modal sensor is used to monitor a crop growth state in real time, growth state data sets are obtained by acquiring leaf color indexes, stem heights and fruit development progress, and the growth state data sets and the initial fertilization scheme are input into a fertilization decision model; and through comparison of growth state deviations, nutrient proportions are adjusted to generate an optimized fertilization scheme.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of farmland fertilization management, in particular to an intelligent fertilization management method and system based on machine learning. BACKGROUND

[0002] In agricultural production, reasonable fertilization is a key link to ensure crop yield and quality. Traditional fertilization management mainly relies on farmers' experience or regional general fertilization recommendations, lacking precise guidance for specific fields and crop growth stages. This extensive fertilization mode is easy to cause mismatch between nutrient supply and demand, which may affect crop growth due to insufficient fertilization, or cause resource waste and environmental pollution due to excessive fertilization. With the deepening of the concept of precision agriculture, soil nutrient monitoring and variable fertilization technology have been developed, but there are still many limitations in existing technologies. For example, some methods only develop fertilization schemes based on soil test results, ignoring the influence of environmental factors on nutrient transformation and crop absorption, and not fully considering the actual demand changes in crop growth process. The dynamic changes of soil nutrients are affected by many factors, and it is difficult to reflect the nutrient supply status in time by relying on regular soil detection.

[0003] Crop growth status is a direct manifestation of nutrient demand, but traditional methods are difficult to obtain crop growth information in time and quantitatively for fertilization decision-making. Although satellite remote sensing, unmanned aerial vehicle aerial photography and other technologies can obtain crop canopy information, the monitoring precision of leaf color, stem morphology and other detailed features is limited, and is easily affected by weather conditions. Fertilization decision-making is a complex process coupled with multiple factors, which needs to integrate soil basic fertility, environmental conditions, crop real-time growth and other information. Existing technologies often consider these factors in isolation, lacking effective integrated analysis means. At the same time, the implementation effect of fertilization scheme needs to be fed back and verified, but traditional methods are difficult to evaluate the fertilization effect in time and used for optimizing subsequent decision-making. Therefore, an intelligent fertilization management method is needed, which can integrate multi-source information, dynamically predict nutrient demand, real-time monitor crop response and continuously optimize itself, to realize the precise management of farmland nutrients. SUMMARY

[0004] The present application aims to provide an intelligent fertilization management method and system based on machine learning to solve the problems raised in the background.

[0005] To achieve the above-mentioned purpose, the present application provides an intelligent fertilization management method based on machine learning, which comprises:

[0006] Collecting a set of soil parameters and a set of environmental parameters of the target farmland, the set of soil parameters including soil humidity, nitrogen, phosphorus and potassium content and pH value, and the set of environmental parameters including light intensity, temperature and rainfall;

[0007] Based on historical data from soil parameter sets and environmental parameter sets, a dynamic change model of soil nutrients is constructed, and the trend of soil nutrient consumption in the future within a preset period is predicted through time series analysis.

[0008] Based on soil nutrient consumption trends and combined with the characteristics of crop growth stages, an initial fertilization plan is generated.

[0009] Multimodal sensors are used to monitor crop growth status in real time, and leaf color index, stem height and fruit development progress are obtained to form a growth status dataset.

[0010] The growth status dataset and the initial fertilization plan are input into the fertilization decision model. By comparing the deviations in growth status, the nutrient ratio in the initial fertilization plan is adjusted to generate an optimized fertilization plan.

[0011] The system controls intelligent fertilization equipment to perform precise fertilization in designated zones according to the optimized fertilization plan, and records the actual fertilization data.

[0012] The actual fertilization data is associated and stored with soil parameter sets, environmental parameter sets, and growth status datasets to update the training data of the soil nutrient dynamic change model and the fertilization decision model.

[0013] Preferably, the specific steps for constructing the soil nutrient dynamic change model include:

[0014] Extract sampling data from historical soil parameter sets and environmental parameter sets for consecutive time periods, and divide the training samples according to the crop growth cycle;

[0015] Long Short-Term Memory (LSTM) networks were used to perform time-series modeling on the training samples to learn the decay pattern of soil nutrients as environmental parameters changed;

[0016] By optimizing the weight parameters of the long short-term memory network using the backpropagation algorithm, the error between the predicted soil nutrient consumption trend and the actual detected value is less than a preset threshold.

[0017] Preferably, the specific steps for generating the initial fertilization plan include:

[0018] Based on the soil nutrient consumption trend, calculate the expected deficit of nitrogen, phosphorus, and potassium in the future preset period;

[0019] Based on the standard fertilizer requirements of crops at their current growth stage, allocate basic fertilizer application based on the expected deficit.

[0020] By incorporating rainfall and temperature data from the environmental parameter set, the ratio of fast-acting fertilizer to slow-release fertilizer in the basic fertilizer application can be dynamically adjusted.

[0021] Preferably, the specific steps for adjusting the nutrient ratio in the initial fertilization plan by comparing growth state deviations include:

[0022] Comparing the leaf color index with the standard health threshold, determine the chlorophyll synthesis deficiency or excessive area;

[0023] According to the stem height deviation, correct the nitrogen fertilizer amount, and according to the fruit development progress deviation, correct the potassium fertilizer amount;

[0024] Adopting reinforcement learning algorithm to iteratively optimize nutrient allocation ratio until the comprehensive deviation rate of growth state data set is reduced to the target range.

[0025] Preferably, the training step of the fertilization decision model comprises:

[0026] Collecting historical fertilization schemes and corresponding growth state data of different crop varieties in multiple environments to construct a labeled data set;

[0027] Using graph convolution network to process the nonlinear relationship between fertilization scheme and growth state in the labeled data set to generate topological feature vector;

[0028] Improve the generalization ability of the graph convolution network through adversarial training, so that the output error of the fertilization scheme for unseen environment parameter combination is lower than the allowed value.

[0029] Preferably, the specific steps of controlling the intelligent fertilization equipment to perform zoned precision fertilization operation according to the optimized fertilization scheme comprise:

[0030] According to the fertility difference area divided by the geographic information system of farmland, match the zoned nutrient allocation ratio in the optimized fertilization scheme;

[0031] Send differential control instructions to each zoned intelligent fertilization equipment through the Internet of Things gateway, and start the fertilization operation synchronously;

[0032] Use Beidou positioning system to real-time check the consistency of the travel trajectory of the fertilization equipment and the nutrient spraying coverage range.

[0033] Preferably, the specific steps of updating the training data of the soil nutrient dynamic change model and the fertilization decision model comprise:

[0034] Add the deviation value of the actual fertilization data and the predicted soil nutrient consumption trend to the anomaly detection module of the labeled data set;

[0035] When the deviation value exceeds the warning threshold, trigger the incremental learning process of the soil nutrient dynamic change model, and recalibrate the time series prediction parameters;

[0036] Using the federal learning framework to aggregate the fertilization effect data of multiple plots, globally updating the weight matrix of the fertilization decision model.

[0037] Preferably, the working steps of the anomaly detection module comprise:

[0038] Establish the time alignment relationship between the actual fertilization data and the growth state data set, and extract the nutrient absorption response curve within the preset days after fertilization;

[0039] Identify outliers in the response curve through the Isolation Forest algorithm, and mark them as high-risk events requiring human intervention;

[0040] Store the soil parameter set and the environmental parameter set associated with the high-risk events separately for model failure analysis.

[0041] Preferably, the method further comprises the following steps:

[0042] After the crops are harvested, collect the final yield data and soil residual nutrient data to evaluate the actual benefits of the fertilization scheme;

[0043] According to the benefit evaluation results, adjust the prediction weights of the soil nutrient dynamic change model and the decision threshold of the fertilization decision model in reverse.

[0044] Preferably, the present application further comprises an intelligent fertilization management system based on machine learning, including a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, realizes the steps of the intelligent fertilization management method based on machine learning.

[0045] Compared with the prior art, the present application has the following advantages:

[0046] The method of the present application realizes the precision and intelligence of farmland fertilization by constructing a data-driven closed-loop management process. Firstly, the soil nutrient dynamic change model is established by comprehensively considering the soil background parameters and environmental factors, which can predict the natural consumption trend of nutrients in the future period and provide a forward-looking basis for the development of fertilization scheme. The initial fertilization scheme is generated by combining the characteristics of crop growth stages, making the fertilization strategy more in line with the physiological demand law of crops. Multi-modal sensors are introduced to monitor the growth state of crops in real time, obtaining indicators such as leaf color, stem height and fruit development that directly reflect the nutritional status of crops, forming a live test of soil nutrient supply conditions. The real-time growth state and the initial fertilization scheme are jointly input into the decision model, and the deviation between the actual growth and the expected target is compared to dynamically adjust the nutrient allocation ratio, so that the fertilization scheme can be optimized according to the actual response of crops, realizing the precise matching from "soil fertilization" to "crop fertilization".

[0047] The intelligent device is controlled to perform partitioned precise fertilization based on the optimization scheme, which reduces the overall application amount and field variation of the fertilizer and improves the fertilizer utilization rate. Actual fertilization data are recorded and stored in association with soil, environment, and crop growth data to form a complete fertilization effect dataset. The nutrient dynamic model and the decision model are continuously updated using the dataset, enabling the system to continuously learn from field practices, adapt to the characteristics of different plots and interannual climate changes, and gradually improve the decision accuracy. This method changes the fertilization management from static experience-based decision making to dynamic data-driven decision making, realizes on-demand supply and precise regulation of farmland nutrients, and helps to reduce agricultural non-point source pollution while ensuring crop yield, thereby promoting the sustainable development of agricultural production. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a soil nutrient dynamic change graph;

[0049] Figure 2 is a flowchart for constructing a soil nutrient dynamic change model;

[0050] Figure 3 is a flowchart for generating an initial fertilization scheme;

[0051] Figure 4 is a column chart of nitrogen, phosphorus, and potassium application amounts at different growth stages. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present application.

[0053] Please refer to Figure 1 The present application provides an intelligent fertilization management method and system based on machine learning. The method comprises: collecting a set of soil parameters and a set of environmental parameters of a target farmland, wherein the set of soil parameters includes soil humidity, nitrogen, phosphorus, and potassium content, and pH value, and the set of environmental parameters includes light intensity, temperature, and rainfall. These parameters are obtained in real time by a multi-modal sensor network deployed in the farmland, and the data are transmitted to a central processing unit via an Internet of Things gateway. Historical data are stored in a cloud platform database for constructing a soil nutrient dynamic change model. The soil nutrient dynamic change model uses time series analysis technology to process the historical data to predict the soil nutrient consumption trend in a future preset period. The prediction result is combined with crop growth stage characteristics, such as seedling stage, jointing stage, or grain filling stage, to generate an initial fertilization scheme. The initial fertilization scheme includes the basic ratio of nitrogen, phosphorus, and potassium and a fertilization schedule.

[0054] The multi-modal sensors continuously monitor the growth status of the crops, the leaf color index is captured by a hyperspectral camera, the stem height is measured by a laser radar, and the fruit development progress is evaluated based on image recognition algorithms. The growth status dataset is input into the fertilization decision model together with the initial fertilization scheme. The fertilization decision model uses machine learning algorithms to compare the growth status deviation and dynamically adjusts the nutrient allocation ratio. After the optimized fertilization scheme is output, the intelligent fertilization equipment receives the control instructions and performs the zoned precision fertilization operation. The intelligent fertilization equipment includes a variable rate fertilizer applicator and a drone spraying system, and the actual fertilization data is recorded in the blockchain log to ensure traceability. The actual fertilization data is stored in association with the soil parameter set, the environmental parameter set, and the growth status dataset, which is used to update the soil nutrient dynamic change model and the training data of the fertilization decision model.

[0055] Example 1: refer to Figure 2 The sampling data is derived from the Internet of Things sensor network deployed in the target farmland, the soil parameter set includes the timing measurement values of soil humidity, nitrogen, phosphorus, and potassium content, and pH value, and the environmental parameter set includes the weather station records of light intensity, temperature, and rainfall. The data collection frequency is set according to the crop growth cycle, and the sampling density is increased during the key growth periods of the crops, such as the jointing stage or the flowering stage. The raw data is cleaned and formatted, and the abnormal sensor readings are removed, and the missing values are supplemented using the time series interpolation method to form a regular historical data set. The historical data set is divided into training samples according to the crop growth cycle, and each training sample corresponds to a complete growth season, including the soil nutrient sequence and the corresponding environmental parameter sequence during the sowing to harvesting period. The long short-term memory network is selected as the core algorithm for time series modeling, and the structure of the long short-term memory network includes an input layer, multiple hidden layers, and an output layer. The input layer receives the training samples arranged in time sequence, and each time step data point includes multiple features of the soil parameter set and multiple features of the environmental parameter set. The hidden layer uses memory cells with forget gates, input gates, and output gates, which can capture the long-term dependence and decay law of soil nutrients with environmental parameters. During the training process, the long short-term memory network learns the dynamic change pattern of soil nutrients, such as the volatilization characteristics of nitrogen elements in high temperature and high humidity environments and the fixation law of phosphorus elements in acidic soils. The training samples are divided into multiple batches input into the network, and the predicted values are calculated through forward propagation. The difference between the predicted value and the actual detection value constitutes the loss function, and the loss function adopts the form of root mean square error.

[0056] The backpropagation algorithm is responsible for optimizing the weight parameters of the long short-term memory network. The backpropagation algorithm calculates the gradient of the loss function with respect to the network weights. The gradient descent optimization algorithm updates the weight parameters based on the gradient values, with the learning rate set to an adaptive adjustment mode. Training iterations continue until the error between the predicted soil nutrient consumption trend and the actual detected value is less than the pre-set threshold. The pre-set threshold is set according to the accuracy requirements of agricultural management, reflecting the acceptable error range of soil nutrient prediction. The early stopping method is used during training to prevent overfitting, and the performance on the validation set determines the timing of training termination. The long short-term memory network after training has the ability to predict the changes in soil nutrients in the future pre-set period, and the prediction period can be set to seven days or thirty days according to actual needs. The data preprocessing section includes several key steps, and the original sampling data needs to be normalized. Normalization scales the features of the soil parameter set and the environmental parameter set with different dimensions to the same numerical range, avoiding the dominance of certain features in the training process. Time series alignment ensures that the measurement time points of the soil parameter set and the recording time points of the environmental parameter set are accurately matched, and the time window sliding mechanism generates continuous time series fragments. Feature engineering extracts meaningful derived features, such as the influence index of consecutive rainfall days on soil nitrogen loss. The division of training samples considers the crop rotation mode, and the soil background values of different previous crops are included as additional features in training.

[0057] The hyperparameter tuning of the long short-term memory network is achieved through systematic search, and the number of network layers and neurons is determined based on the size of the training data. The dropout rate setting controls the model complexity and prevents overfitting on the training data. The batch size affects the training stability and convergence speed, which needs to be adjusted according to the computing resources. The evaluation of hyperparameter combinations uses the cross-validation method, dividing the historical data set into training and test sets. The performance indicators on the test set guide the selection of hyperparameters, with the selection criteria focusing on the generalization ability of the model. The final selected hyperparameter combination is applied to the model training in the production environment, and the trained model is saved in a deployable format. The output of the soil nutrient dynamic change model is the predicted curve of soil nutrient concentration in the future time period, and the prediction results are presented in numerical and visual forms. Farm managers can view the prediction results through the network interface, which includes the daily change trend of nitrogen, phosphorus and potassium elements. The model prediction results form a closed-loop verification mechanism with real-time monitoring data, and the detected values of newly collected soil samples are used to evaluate the prediction accuracy. When the prediction deviation is consistently large, the model retraining process is triggered, and the model parameters are updated with the latest data. The model updating strategy uses incremental learning, integrating new observation data while preserving existing knowledge.

[0058] The deployment environment of the long short-term memory network model considers the actual application scenario, and the edge computing device undertakes the model inference task. The cloud computing platform is responsible for model training and version management, and the trained model is distributed to the field edge server. The edge server is equipped with a dedicated hardware accelerator to improve the inference speed of the long short-term memory network. Lightweight protocols are used for data transmission to ensure real-time operation in a limited bandwidth environment. The model inference service is packaged as a standardized application programming interface for other modules of the farmland management system to call. The maintenance of the soil nutrient dynamic change model includes regular performance evaluation, and the evaluation indicators include prediction accuracy and timeliness. The model degradation detection mechanism monitors the change of prediction quality, and automatically triggers an alarm when the performance indicators are below the set standard. The model version control system records the content of each update, supporting quick rollback to a stable version. Farmland management decisions are based on model prediction results, and the timing and amount of fertilization are determined by referring to the soil nutrient consumption trend. The uncertainty of model prediction is represented by a confidence interval, helping managers assess the risk of decision-making. The quality of historical data directly affects the performance of the soil nutrient dynamic change model, and strict quality control is implemented in the data collection process. Regular calibration of sensors ensures measurement accuracy, and an anomaly detection algorithm filters unreliable data in real time. Distributed database architecture is used for data storage to ensure the integrity and accessibility of historical data. Data backup strategies prevent accidental data loss, and multiple copies of storage ensure data security. The data access interface provides standard query functions, supporting flexible retrieval by time range and geographic location.

[0059] The explainability of the long short-term memory network model is enhanced by auxiliary analysis tools, and feature importance analysis reveals the influence of each environmental parameter on soil nutrient changes. The attention mechanism visualizes the model's focus on time series, helping to understand the basis for model decision-making. Reliability analysis of model prediction results considers the quality of input data, and prediction results corresponding to low-quality input data are marked as low confidence. Model output results are verified by traditional agricultural expert experience, forming a human-machine collaborative decision-making mode. The extensibility design of the soil nutrient dynamic change model supports multi-plot application, and different farmland model instances share the basic framework. Transfer learning technology uses existing farmland training data to accelerate the development of new plot models, and domain adaptation methods adjust model parameters to adapt to local conditions. The model cluster management platform monitors the running status of multiple deployment instances, and the resource scheduling algorithm optimizes the utilization rate of computing resources. The variability of the farmland environment requires the soil nutrient dynamic change model to have adaptability, and the online learning function of the model handles gradual changes in environmental conditions. The mutation environment detection algorithm identifies sudden events such as climate anomalies, triggering the model emergency update process. The uncertainty quantification of model prediction results helps managers assess the risk of decision-making, and probabilistic output represents the likelihood of multiple future scenarios. The prediction range of the model supports multi-time scale configuration, short-term prediction guides daily farm operations, and long-term prediction serves strategic planning.

[0060] Integrating soil nutrient dynamic models with other agricultural models forms comprehensive decision support systems, with crop growth models providing nutrient demand information. Meteorological prediction models input future weather conditions, improving the forward-looking nature of soil nutrient predictions. Economic benefit models evaluate the input-output ratio of different fertilization strategies, optimizing resource allocation. System integration employs modular design, with standard data interfaces ensuring smooth communication between models. Workflow engines coordinate model execution order, and data processing pipelines ensure efficient data flow. The representativeness of model training data influences the applicability of soil nutrient dynamic models, with data collection networks covering different soil types and climate regions. Data augmentation techniques generate synthetic training samples, improving model prediction capabilities for rare weather patterns. Active learning strategies prioritize data points with high information content, improving training data quality. Long-term observation plans accumulate cross-year data, capturing the long-term effects of climate change on soil nutrient cycling. Data standardization facilitates the integration and utilization of data from different sources, with metadata management recording data collection conditions and processing methods. Model verification employs multi-angle evaluation methods, with historical data backtesting verifying model performance under past conditions. Field comparison tests compare model-guided fertilization schemes with traditional methods, with a multi-index evaluation system comprehensively measuring model value. Model uncertainty analysis identifies the main sources of prediction errors, with targeted improvements to data collection or algorithm design. Peer review mechanisms incorporate expert opinions from the field, improving the theoretical basis and practical performance of models.

[0061] Example 2: see Figure 3, generate an initial fertilization plan based on soil nutrient depletion trends output by a soil nutrient dynamics model, the soil nutrient depletion trends representing expected changes in soil nitrogen, phosphorus, and potassium levels over a future period of time in the form of numerical curves. Calculating expected deficits of nitrogen, phosphorus, and potassium over a future predetermined period requires defining a nutrient demand baseline for crop growth, the nutrient demand baseline derived from standard nutrient requirements of a specific crop variety at various growth stages in an agricultural database. An expected deficit is equal to the difference between the nutrient reduction predicted by the soil nutrient depletion trends and the standard nutrient requirements of the crop, the difference calculated on a daily nutrient balance. The current growth stage of the crop is determined through phenology observation, and the standard nutrient requirements are queried from pre-set parameters in a crop growth model. The expected deficit is used as a basis to allocate a base fertilization amount, the base fertilization amount calculated with consideration of soil nutrient use efficiency and crop absorption rate, and the allocation results form preliminary recommendations of nitrogen, phosphorus, and potassium usage. Rainfall and temperature data in the environmental parameter set are introduced into a dynamic adjustment mechanism, the rainfall data obtained from real-time monitoring records of a farmland weather station. Under high rainfall conditions, the proportion of slow-release fertilizer is increased to reduce nutrient loss, and under low rainfall conditions, the proportion of quick-acting fertilizer is increased to promote rapid absorption by the crop. Temperature data affect soil microbial activity and nutrient release speed, and when the temperature is high, the proportion of quick-acting fertilizer is appropriately increased, and when the temperature is low, slow-release fertilizer is emphasized for long-term supply. The adjusted base fertilization amount, combined with a fertilization time plan, constitutes an initial fertilization plan document, the document format adopting standardized agricultural data specifications. The initial fertilization plan is presented through a user interface of a farmland management system, supporting manual review and fine-tuning by agricultural technicians. A growth status dataset is formed by real-time monitoring and acquisition of leaf color index, stem height, and fruit development progress through multi-modal sensors. The leaf color index is collected by a multi-spectral camera mounted on a drone, and the normalized difference vegetation index is calculated to reflect chlorophyll content. The stem height is measured by three-dimensional reconstruction using laser radar point cloud data, and the fruit development progress is evaluated based on image recognition algorithms of high-definition cameras. The growth status dataset is compared with the growth expectations preset in the initial fertilization plan, and the growth status deviation is represented as the standard deviation between the measured value and the expected value. The leaf color index is compared with a standard healthy threshold value, the standard healthy threshold value derived from the ideal spectral reflectance range of the same crop variety at the same growth stage. Areas with insufficient chlorophyll synthesis are marked with highlighted blocks on an electronic farmland map, and areas with excessive chlorophyll are separately marked as potential fertilizer damage risk zones.

[0062] The stem height deviation is corrected for the nitrogen fertilizer rate. The stem height deviation refers to the percentage difference between the actual measured height and the predicted height from the growth model. A negative deviation indicates growth retardation, and the nitrogen fertilizer rate needs to be increased to promote vegetative growth. A positive deviation indicates the phenomenon of excessive stem elongation, and the nitrogen fertilizer rate needs to be reduced to avoid lodging. The fruit development progress deviation is corrected for the potassium fertilizer rate. The fruit development progress deviation is calculated by the deviation of the fruit diameter measurement value from the standard growth curve. When the fruit development is lagging, the potassium fertilizer rate is increased to promote material transport. When the fruit development is too fast, the potassium fertilizer application rate is adjusted to prevent early decline. The nutrient allocation adjustment uses a reinforcement learning algorithm for iterative optimization. The reinforcement learning algorithm constructs a Markov decision process with the farmland state as the environment and the fertilizer adjustment as the action. The state space of the reinforcement learning algorithm includes all the feature vectors of the growth state dataset, and the action space defines the continuous value range of the nitrogen, phosphorus and potassium rate adjustment. The reward function is designed based on the comprehensive deviation rate of the growth state dataset. The comprehensive deviation rate integrates the weighted sum of the leaf color index deviation, the stem height deviation and the fruit development progress deviation. After each fertilizer action, the new growth state is observed, the reward value is calculated, and the action value function is updated. The policy network outputs the optimal nutrient allocation adjustment scheme, and the value network evaluates the long-term benefits of the current state. The training process uses the experience replay mechanism to break the data correlation, and the target network provides a stable learning goal. Iterative optimization continues until the comprehensive deviation rate falls within the target range, which is set according to the crop yield and quality requirements. The update frequency of the growth state dataset is matched with the crop growth speed, and the data collection frequency is increased during the rapid growth period. The calibration and maintenance of multi-modal sensors ensure data accuracy, and periodic cross-validation with manual measurement results is performed. The decision records generated during the deviation adjustment process are stored in the knowledge base, forming case data to support the rapid response to similar scenarios in the future. The farmland geographic information system integrates the growth state space distribution map, which can visualize the deviation degree in different regions. The nutrient allocation adjustment scheme generates differentiated fertilizer prescription maps, and the grid data of the prescription maps are imported into the intelligent fertilization equipment control system. The training of the reinforcement learning algorithm is carried out in a simulation environment, which simulates the crop growth response under different soil and climate conditions. The transfer learning technology migrates the model parameters trained in the simulation environment to the actual farmland application, reducing the training time in the field. Algorithm hyperparameters such as learning rate and discount factor are determined by parameter scanning, and the combination with the best performance on the validation set is selected. The online learning mechanism allows the model to continue to optimize in actual deployment and adapt to changes in the farmland environment. The safety check module prevents the fertilizer adjustment from being too large, and sets upper and lower limits for the single adjustment amount.

[0063] The analysis of leaf color index considers the influence of light conditions, and the measurements on cloudy and sunny days are corrected by a light compensation algorithm. The stem height measurement eliminates the error caused by the terrain slope, and the digital elevation model provides the elevation reference. The fruit development progress evaluation distinguishes the influence of different fruit positions, and the outer and inner layer fruits of the canopy are evaluated with different standards. The spatial and temporal correlation analysis of growth state deviation reveals the unevenness inside the farmland, providing the basis for zoned precision fertilization. The transparency of deviation adjustment decision is enhanced by the decision tree interpretability method, showing the influence weight of each feature on the adjustment result. The comparison analysis of the initial fertilization scheme and the optimized fertilization scheme forms a feedback loop, and the quantitative evaluation of the optimization effect guides the algorithm improvement. The historical optimization record constructs a case base, and the case base supports the case-based reasoning method to handle new scenarios. The user interface of the farmland management system provides visual tracking of the adjustment process, showing the deviation rate change curve after each iteration. The alarm mechanism monitors abnormal patterns in the adjustment process, such as triggering manual intervention when the deviation rate does not decrease for a long time. The permission management module controls the modification range of different users on the adjustment parameters, ensuring the safety of system operation. The data fusion of multi-modal sensors uses Kalman filtering technology to improve the accuracy and reliability of growth state measurement. The synchronization of data time stamps ensures the time consistency of different sensor readings, and the spatial-temporal indexing of growth state data set supports fast query. Data transmission uses encryption protocols to prevent tampering, and data integrity verification ensures decision reliability. Edge computing nodes process sensor data locally, reducing the demand for cloud transmission bandwidth. The energy consumption management of the growth state monitoring system uses solar power and low-power design, suitable for field working environment.

[0064] Referring to Figure 4 The figure takes crop growth stages as the horizontal axis and fertilization amount as the vertical axis, and presents the application amount of nitrogen, phosphorus, and potassium fertilizers in the seedling stage, jointing stage, filling stage, and mature stage through red, green, and blue columns, respectively. This figure directly reflects the differentiated demand of crops for nitrogen, phosphorus, and potassium at different growth stages, providing quantitative basis for the development of fertilization scheme. It is the first step to precisely match the soil fertilization potential and crop fertilization rules, and also lays the foundation for the subsequent dynamic optimization of fertilization scheme based on growth state, fully embodying the precision concept of time-based fertilization in intelligent fertilization management.

[0065] Embodiment 3: The training of the fertilization decision model requires collecting historical fertilization schemes and corresponding growth state data of different crop varieties in multiple environments. The historical fertilization scheme includes structured records such as fertilizer type, application amount, fertilization time, and fertilization method. The growth state data covers time series measurements such as leaf color index, stem height, and fruit development progress. The multiple environments refer to the combination of climate conditions in different geographical regions of farmland in different years. The construction of the labeled data set uses the ETL process to extract raw data from the databases of multiple agricultural test stations. The data cleaning step handles outliers and missing values. Data conversion unifies data from different sources into a standard format. Data loading stores the processed data into a model training dedicated database. Each sample in the labeled data set contains input features and output labels. The input features include fertilization scheme parameters and environmental parameters. The output label is the growth state index. The graph convolution network processes the nonlinear relationship between the fertilization scheme and the growth state in the labeled data set. The graph convolution network represents the fertilization scheme elements as nodes in the graph structure. The node feature vector encodes attributes such as fertilization amount and fertilization time. The edge feature represents the relationship between different fertilization operations. The graph convolution layer aggregates node information through the adjacency matrix. Each layer of convolution operation updates the node representation. The multi-layer graph convolution stack extracts high-order topological features to generate a topological feature vector containing global information. The topological feature vector is input into the fully connected layer for regression prediction, outputting the estimated value of the growth state. The training of the graph convolution network uses supervised learning. The loss function calculates the difference between the predicted growth state and the actual measured value.

[0066] The adversarial training enhances the generalization ability of the graph convolutional network. The adversarial training introduces the generative adversarial network framework. The generator network learns to generate synthetic fertilization scheme data, and the discriminator network distinguishes between real labeled dataset samples and generated samples. The graph convolutional network participates in the adversarial training as part of the discriminator, and the network parameters are optimized through the minimax game. The training goal is to make the graph convolutional network output an error lower than the allowed value for the fertilization scheme of the unseen environment parameter combination, and the allowed value is set according to the accuracy requirement of the agricultural application. Gradient penalty technology is used in the adversarial training process to enhance the stability of the training and prevent pattern collapse. The deployment of the fertilization decision model adopts a microservice architecture, and the model is encapsulated as a RESTful API service. The API interface receives input parameters including the current soil nutrient status, environmental conditions and crop growth stage, and returns the optimized fertilization scheme. The microservice container deployment realizes resource isolation and elastic scaling, and the service mesh manages the communication between microservices. The load balancer distributes user requests to multiple model instances to ensure system high availability. The API gateway handles authentication and traffic control, and access logs are recorded for security auditing. The partitioned precision fertilization operation is based on the fertility difference regions divided by the geographic information system of the farmland, and the fertility difference regions are determined by soil conductivity mapping and historical yield map analysis. Each fertility difference region is assigned a unique identifier and associated with the corresponding soil property parameters. The partition nutrient allocation ratio in the optimized fertilization scheme is mapped to the fertility difference region to form a fertilization prescription map. The fertilization prescription map uses a grid data format, and each grid cell contains the recommended nitrogen, phosphorus and potassium application amount. The prescription map generation algorithm considers the shape and size of the fertility difference region to ensure the implementability of the fertilization machine.

[0067] The Internet of Things gateway sends differentiated control instructions to each partitioned intelligent fertilization device, and the Internet of Things gateway converts data formats using industrial-grade communication protocols. The control instructions include target fertilization amount, travel speed and start-stop signals, and the instruction transmission uses encryption to prevent tampering. Each partitioned intelligent fertilization device synchronously starts the fertilization operation, and the time synchronization protocol ensures operation consistency. The controller of the intelligent fertilization device parses the received instructions and drives the hydraulic system to adjust the fertilizer discharge amount. The variable rate fertilizer applicator adjusts the fertilizer discharge shaft speed in real time according to the ground speed to maintain the accuracy of the fertilizer application amount per unit area. The Beidou positioning system real-time verifies the consistency of the fertilization device's travel trajectory and nutrient spraying coverage range, and the Beidou positioning system provides centimeter-level precision location information. The travel trajectory is compared with the preset operation path, and the deviation is calculated using the Euclidean distance formula. When the lateral deviation exceeds the threshold value, the control system automatically adjusts the steering mechanism to correct the direction. The nutrient spraying coverage range is calculated by the flow sensor and the travel speed, and the actual spraying amount is compared with the required value of the prescription map. When the deviation exceeds the tolerance, the system triggers an alarm, and the operator can manually intervene or start automatic correction.

[0068] The optimization process of the fertilization decision model generates the fertilization scheme considering multiple constraints. The fertilization cost constraint limits the upper limit of the fertilization cost per unit area. The environmental constraint controls the risk of nutrient loss, and the soil nutrient balance constraint maintains the sustainability of land productivity. A multi-objective optimization algorithm is used to solve the Pareto optimal solution set, and the decision maker selects the final scheme according to the preference. The key parameters in the optimization model, such as the fertilizer utilization coefficient, are calibrated through field tests, and different parameter values are used for different soil textures. The online learning function of the fertilization decision model integrates real-time feedback data, and the crop growth response after each fertilization operation is used for model updating. The incremental learning algorithm integrates new observation data while preserving the original knowledge, and the smooth evolution of model parameters avoids drastic fluctuations. The model performance monitoring panel displays the prediction accuracy trend, and the performance degradation triggers the retraining process. The A / B testing framework compares the performance of different versions of the model, and data-driven decision making determines the upgrade time of the model.

[0069] The mechanical calibration and maintenance of the zoning precision fertilization system ensure the operation accuracy. The fertilization machine is calibrated regularly in the test field. The flow sensor calibration uses a standard container to collect the actual fertilizer discharge, and a relationship curve between the electrical signal and the fertilizer discharge is established. The travel speed sensor calibration measures the wheel circumference to compensate for the slip rate, and the GPS antenna installation position error is measured and compensated. Regular wear and tear checks are performed on mechanical components, and consumable parts are replaced according to the planned usage time. The fault diagnosis system of the intelligent fertilization equipment monitors the equipment status in real time, and the vibration sensor detects mechanical abnormalities. The hydraulic pressure sensor monitors the system load, and the temperature sensor monitors the temperature rise of the key components. The fault code library contains common fault modes and processing suggestions, and the remote diagnosis center supports on-site troubleshooting. The preventive maintenance plan is based on the number of equipment operating hours, and the digital management of maintenance records supports life cycle management. The data exchange between the fertilization decision model and other intelligent systems in the farmland uses agricultural data standards, and the data interface defines clear input and output specifications. Soil moisture data is shared with the irrigation system to optimize water and fertilizer management. The weather station is connected to obtain short-term weather forecasts, and the fertilization time is adjusted to avoid rainfall. The final yield data is obtained through communication with the harvesting machinery to form a closed-loop feedback. The system integration platform uses an event-driven architecture, and the message queue asynchronously handles inter-system communication. The quality of the fertilization operation is evaluated through multi-source data verification, and unmanned aerial vehicle multi-spectral images are used to evaluate the crop growth changes after fertilization. Soil sampling and analysis verify the uniformity of nutrient distribution, and leaf nutrient diagnosis checks the nutrient absorption status. The evaluation results quantify the fertilization effect, providing a basis for model improvement. The quality evaluation report is automatically generated, including operation statistical data and effect indicators. The control algorithm of the variable rate fertilizer machine considers the mechanical response delay, and the advance compensation ensures the accuracy of the fertilization position. The PID controller adjusts the fertilizer discharge mechanism, and the parameter setting adapts to different fertilizer physical properties. CAN bus network connects each control unit, and real-time data transmission ensures coordinated control. The safety interlock device prevents misoperation, and the emergency stop button is located in a convenient position.

[0070] The knowledge representation of the fertilization decision model adopts the ontology method to uniformly describe the concepts and attributes in the agricultural field. The knowledge graph stores the field knowledge of crop fertilizer requirements and soil nutrient cycling, supporting semantic reasoning. The natural language processing interface allows users to query fertilization recommendations through voice, reducing the use threshold. The mobile application provides field operation guidance, and the offline mode ensures normal use in areas without network.

[0071] ;

[0072] wherein: represents the adjusted fertilization amount, which is a dimensionless relative value; represents the base fertilization amount, which is consistent in dimension with ; represents the actual stem height measurement value, with units of centimeters; represents the expected stem height standard value, with units of centimeters; represents the actual leaf color index measurement value, which is a dimensionless index; represents the expected leaf color index standard value, which is a dimensionless index; and are adjustment coefficients, which are dimensionless weight parameters. Both sides of the formula are dimensionless relative values, and the dimension is consistent.

[0073] The verification of the fertilization decision model uses the cross-validation method to divide the labeled data set into a training set and a test set. Statistical analysis of the prediction error of the model on the test set, and the error distribution histogram shows the performance characteristics of the model. Sensitivity analysis tests the response of the model to changes in input parameters, identifying key influencing factors. Robustness testing injects random noise to simulate data errors, evaluating the model's anti-interference ability. The verification results guide the optimization of the model structure and parameter adjustment, improving the practical reliability. The spatial analysis function of the farmland geographic information system supports dynamic adjustment in areas with different fertilizers, and the new soil sampling data updates the boundaries of the fertilizer partition. The spatiotemporal interpolation algorithm generates a continuous fertilizer distribution surface, and the Kriging method considers spatial autocorrelation. Redivision in areas with different fertilizers triggers the regeneration of the fertilization prescription map, ensuring the timeliness of accurate partition fertilization.

[0074] The anomaly detection module adds the deviation value of the actual fertilization data and the predicted soil nutrient consumption trend to the labeled data set. The actual fertilization data comes from the operation log recorded by the intelligent fertilization device. The operation log includes fields such as fertilization time, geographic coordinates, fertilizer variety, and application amount. The predicted soil nutrient consumption trend is the numerical sequence output by the soil nutrient dynamic change model. The deviation value is calculated using the absolute difference between the nutrient content measured by soil sampling after actual fertilization and the model prediction value. When the deviation value exceeds the warning threshold, the incremental learning process of the soil nutrient dynamic change model is triggered. The incremental learning process recalibrates the time series prediction parameters and integrates new observation data without retraining the complete model. The federated learning framework aggregates the fertilization effect data of multiple plots. The federated learning framework coordinates multiple local clients to train a global model. The global model updates the weight matrix of the fertilization decision model. The weight matrix update is based on the distributed average algorithm. The anomaly detection module establishes a time alignment relationship between the actual fertilization data and the growth state data set. Time alignment uses high-precision clock synchronization technology. The nutrient absorption response curve within a preset number of days after fertilization is extracted. The nutrient absorption response curve is drawn by continuously monitoring the change of soil nutrient concentration. The Isolation Forest algorithm identifies outliers in the response curve. The Isolation Forest algorithm constructs a random tree structure to isolate abnormal samples. Outliers are marked as high-risk events that require human intervention. The soil parameter set and environmental parameter set associated with high-risk events are stored separately. The stored data is used for model failure analysis, and an analysis report is generated to diagnose the root cause.

[0075] The incremental learning process of the soil nutrient dynamic change model uses an online sequence learning algorithm. Newly arrived soil parameter sets and environmental parameter set data are processed in a streaming manner. Model parameter updates use the recursive least squares method to adapt to slow changes in soil characteristics. The forgetting factor in the incremental learning process controls the degree of influence of old data, and dynamically adjusts the factor to adapt to the change rate in different seasons. Model version management records the timestamp and change content of each update, supporting model behavior traceability. The computational complexity of the incremental learning process is optimized for edge device deployment, with constant memory usage. The architecture of the federated learning framework includes a central server and multiple client nodes. The central server coordinates global model training. Client nodes correspond to local systems of different farmlands, and local data is retained in the original location without uploading. The federated average algorithm aggregates the model gradients of each client to generate an improved global model. Differential privacy technology protects local data privacy by adding noise to prevent information leakage. The model update frequency is self-adaptive according to network conditions and data volume, balancing timeliness and resource consumption.

[0076] The Isolation Forest algorithm of the anomaly detection module adopts an unsupervised learning approach, without the need for pre-labeled anomaly samples. During the algorithm construction process, features and split points are randomly selected to create multiple isolation trees. The anomaly score is based on the path length required for a sample to be isolated, with shorter path lengths indicating a higher likelihood of being an anomaly. The threshold setting is determined using the percentile method based on the anomaly score distribution of historical data. The algorithm supports large-scale datasets through parallel processing, and the distributed computing framework accelerates the detection process. The analysis process for high-risk events includes multiple diagnostic steps, with the anomaly pattern recognition of soil parameter sets identifying specific nutrient imbalances. The extreme value detection of environmental parameter sets relates to the impact of climate events, and the mutation point positioning of growth status datasets locates the turning point of fertilization effects. The diagnostic results are classified and stored in the knowledge base, forming a case library to support rapid response to similar events in the future. The analysis report includes event descriptions, possible causes, and recommended measures, which are pushed to the farm managers. The communication protocol of the federated learning framework optimizes the reduction of bandwidth requirements, and the model compression technique reduces the amount of data transmitted. The secure multi-party computation technology ensures the security of the gradient aggregation process, preventing malicious node attacks. The client selection strategy prioritizes nodes with high data quality for training, improving the quality of the global model. The model performance evaluation uses the reserved test dataset, with fairness considering the differences in data distribution among different clients.

[0077] The real-time processing capability of the anomaly detection module is achieved through a streaming computation engine, which manages the ordered processing of data streams in event time windows. The state backend stores intermediate results of detection algorithms, supporting state reconstruction after failure recovery. Complex event processing techniques identify time series features of abnormal patterns, providing early warning of high-risk events. Visualization of detection results uses time series charts, highlighting abnormal intervals. A version control system for labeled data sets manages the evolution history of data, generating new data versions with each update. Data provenance tracking records the origin and processing of each data item, meeting audit requirements. Quality check rules verify the integrity and consistency of new data, rejecting data inputs that do not meet standards. Data archiving strategies migrate historical data to low-cost storage, while online data retains the most recent active data. Joint updating of the soil nutrient dynamic change model and the fertilization decision model considers the mutual influence between models, with the update order determined based on dependency relationships. Updating of the soil nutrient dynamic change model triggers recalibration of the fertilization decision model, maintaining system consistency. A rollback mechanism restores to the previous version when performance decreases after updating, ensuring system reliability. Automated testing of the update process verifies that model behavior meets expectations, with integration testing covering common scenarios. Emergency response plans for high-risk events define processing procedures for different risk levels, with low-level risks handled automatically and high-level risks requiring human intervention. Event processing time targets are set to ensure timely response, with an escalation mechanism notifying higher-level management personnel when timeouts occur. Processing records are complete, including event descriptions, processing measures, and result evaluations. Subsequent tracking verifies the effectiveness of processing measures, forming a closed-loop event management. Model security mechanisms in the federated learning framework defend against poisoning attacks, with abnormal client detection excluding malicious data contributions. Model validation uses an independent validation set to evaluate global model quality and detect performance degradation. Model interpretability techniques analyze the decision-making basis of global models, identifying potential biases. Fairness adjustments avoid model discrimination against certain clients, balancing the interests of all parties. Refer to Table 1 for classification standards and response relationships for high-risk events in the anomaly detection module:

[0078] Table 1: High-risk event classification and response measures

[0079] Event classification code Abnormal indicator feature description Soil parameter abnormal range Environmental parameter correlation condition System automatic response level Manual intervention trigger condition HRE-01 Nitrogen absorption response curve sharply drops Daily nitrate nitrogen decrease exceeds 5 mg / kg Rainfall greater than 50 mm / day and temperature higher than 30°C Level 1: Adjust model parameters Early warning for 3 consecutive days without relief HRE-02 Phosphorus element fixation phenomenon is abnormally significant Weekly available phosphorus decrease exceeds 3 mg / kg Soil pH value is lower than 5.5 and humidity is less than 40% Level 2: Suspend phosphorus fertilizer recommendation Soil pH value continues to deteriorate HRE-03 Potassium element leaching rate exceeds expectation Daily available potassium decrease exceeds 8 mg / kg Rainfall greater than 100 mm / week and sandy soil Level 3: Start slow-release program Similar events occur in multiple plots HRE-04 Microelement coupling antagonism is abnormal Zinc / iron ratio fluctuation exceeds 2 times of standard deviation Temperature changes by more than 10°C and organic fertilizer is applied Level 2: Adjust microelement ratio Visible deficiency symptoms appear on crops HRE-05 Fertilizer residue accumulation exceeds standard early warning Salinity concentration increases by more than 0.2 dS / m in 10 days Evaporation is greater than rainfall and irrigation is insufficient Level 3: Start leaching recommendation Conductivity continues to rise

[0080] Embodiment 5: The yield monitoring system comprises a grain flow sensor, a moisture sensor, and a GPS receiving module. The grain flow sensor measures the weight of grain passing through the elevator in real time. The moisture sensor corrects the grain weight to the standard moisture content. The GPS receiving module records the coordinates of the harvesting location. The yield data is recorded in seconds, and a yield distribution map containing geographical references is generated. Soil residual nutrient data is obtained by soil sampling after harvesting. The sampling points are arranged in a grid method, and soil samples are collected at each sampling point. The soil samples are sent to the laboratory for determination of nitrogen, phosphorus, potassium and other nutrient contents. The determination results are stored in association with the geographical position. A multi-index evaluation system is needed to evaluate the actual benefits of the fertilization scheme. Economic benefit indicators include input-output ratio and net income. The input-output ratio calculates the ratio of fertilization cost to yield increase. The net income calculates the difference between yield increase and fertilization cost. Environmental benefit indicators consider nutrient use efficiency and soil nutrient balance. Nutrient use efficiency calculates the ratio of nutrient absorption to fertilizer application. Soil nutrient balance calculates the difference between fertilizer input and crop removal. A positive value indicates nutrient accumulation, and a negative value indicates nutrient depletion. Social benefit indicators include product quality grade and food safety indicators. Quality grade is evaluated according to market grading standards. The benefit evaluation results adjust the prediction weight of the soil nutrient dynamic change model and the decision threshold of the fertilization decision model in reverse. The prediction weight adjustment uses the Bayesian updating method. The prediction weight of the soil nutrient dynamic change model represents the influence degree of different environmental factors on nutrient change. The newly collected final yield data and soil residual nutrient data are used as evidence to update the prior distribution. The decision threshold adjustment is based on the optimization of the benefit function. The decision threshold of the fertilization decision model controls the judgment boundary of fertilization timing and amount. Genetic algorithm searches for the optimal threshold combination to maximize long-term cumulative benefits.

[0081] The process is illustrated with a winter wheat-summer corn rotation system in the North China Plain. At the winter wheat harvest stage, the yield monitoring system of the combine harvester generates a yield distribution map showing a yield variation range of 400-600 kg / acre. Soil sampling points are laid out at a density of one point per 5 acres, and sampling is done before corn planting. Laboratory analysis reports show soil residual nitrate nitrogen of 5-15 kg / acre, available phosphorus of 10-20 mg / kg, and available potassium of 100-150 mg / kg. The benefit assessment calculation considers the cost of fertilizer application during the winter wheat growing season, including the cost of fertilizer and the cost of fertilizer application. The yield increase is calculated based on the market price of wheat and the yield increase compared with the non-fertilized control area. The nutrient use efficiency calculation shows that the nitrogen fertilizer partial productivity is 35 kg of wheat per kg of nitrogen fertilizer, and the phosphorus fertilizer partial productivity is 50 kg of wheat per kg of phosphorus pentoxide. Soil nutrient balance calculations show that nitrogen is slightly in surplus, phosphorus is basically balanced, and potassium is in deficit. The prediction weight adjustment of the soil nutrient dynamic change model focuses on the impact factors of environmental parameters, and the weight of the temperature parameter is adjusted upward based on actual observation data. The weight of the rainfall parameter is adjusted downward because irrigation supplements crop water requirements. The prediction model considers the soil texture factor, and different nutrient release coefficients are set for clay loam and sandy loam. The decision threshold adjustment of the fertilizer decision-making model reduces the upper limit of nitrogen fertilizer application to avoid excessive fertilizer residues. The threshold for phosphorus fertilizer application is advanced to adapt to the low temperature conditions in early spring. The threshold for potassium fertilizer topdressing is increased to cope with the high potassium demand peak in the late growth period.

[0082] Quality control of data collection ensures reliability, and calibration of yield monitoring sensors ensures measurement accuracy. Soil sampling tools are sterilized to avoid cross-contamination, and national standard methods are used for laboratory analysis. Data verification procedures identify outliers, such as mechanical failure records in yield data. Spatial interpolation methods generate continuous yield distribution surfaces, with Kriging interpolation considering spatial autocorrelation characteristics. Benefit evaluation reports include textual analysis and graphical displays, with textual analysis explaining benefit performance and improvement directions. Graphical displays include yield distribution maps, nutrient residual distribution maps, and benefit comparison bar charts. Reports are pushed to farm management systems to support decision-makers in developing next-season planting plans. Historical benefit data trend analysis reveals long-term change patterns, such as soil fertility evolution trends. Forecasting weights of soil nutrient dynamic change models are updated using a sliding window method, with recent three-year data weights higher than early data. Model parameters are smoothly transitioned to prevent drastic fluctuations, and a forgetting factor controls the decay rate of historical data. Forecasting uncertainty is quantified by confidence intervals, with wide intervals indicating the need for more data. Independent data sets are used for model validation, with validation results guiding model complexity selection. Decision threshold optimization of fertilization decision models considers risk preferences, with risk-averse decision-makers choosing conservative thresholds. Multi-objective optimization balances economic benefits and environmental sustainability, with a Pareto frontier showing trade-offs. Threshold sensitivity analysis tests the degree of model output response to threshold changes, with threshold changes within stable intervals not causing decision reversals.

[0083] Benefit evaluation under a crop rotation system requires the coordinated optimization of fertilization schemes for both winter wheat and summer corn. Soil residual nutrient data guide the amount of base fertilizer for the next crop, reducing nutrient waste. Crop nutrient demand patterns complement each other, with wheat's phosphorus demand characteristics combined with corn's nitrogen demand characteristics. Long-term positioning observation data reveal nutrient cycling characteristics of the crop rotation system, guiding the design of sustainable fertilization systems. Farmer participation in evaluation collects practical operation feedback, with field observation records supplementing instrument measurement data. Farmer experience and model prediction results are mutually verified, forming a human-machine collaborative optimization mechanism. Visualization of evaluation results facilitates understanding, with a mobile application pushing concise reports. Training materials are based on typical cases, promoting best management practices. Data security management protects commercial privacy, with farm data encrypted and stored with access control. Data sharing follows the principle of informed consent, with desensitized data used for regional analysis. Blockchain technology records data usage logs, enabling complete audit tracking. Data ownership is clearly defined, encouraging data contributors to participate in cooperation. Model adjustments are verified through practice in the next growing season, with new data evaluating adjustment effectiveness. The verification period includes a complete crop rotation cycle, observing long-term effects. Comparative trials set different adjustment strategies, scientifically evaluating strengths and weaknesses. Third-party institutions participate in the evaluation process, ensuring objective and fair results.

[0084] The continuous improvement of the benefit evaluation system absorbs new technologies and methods. Remote sensing technology provides more detailed yield monitoring. Sensor networks expand the data collection dimension, and microbial indicators are included in soil health evaluation. Machine learning algorithms optimize evaluation models to improve prediction accuracy. International standards are introduced to facilitate result comparison and learn from global best practices. Market price data for economic benefit evaluation uses the moving average method to smooth fluctuations and reduce the impact of market anomalies. Cost calculation includes direct and indirect costs, and mechanical depreciation is allocated to each operation link. Environmental benefits are monetized, trying to quantify the value of ecological services, and reflecting the value of agricultural systems comprehensively.

[0085] The adjustment records of soil nutrient dynamic change model and fertilization decision model are saved completely, and the version control system manages the iteration history. The adjustment reasons and expected effects are documented to support traceability analysis. Model performance benchmarking is conducted regularly to ensure that adjustments do not cause performance degradation. The rollback mechanism restores the original parameters when the adjustment effect is not good, ensuring system reliability. Cross-regional data aggregation reveals large-scale rules, and data merging of climate similar areas enhances statistical effectiveness. Meta-analysis method integrates the results of multiple independent studies to draw more robust conclusions. Data standardization solves the compatibility problem of data from different sources, and unifies index definition and measurement method. International cooperation platform shares knowledge and experience, accelerates technological innovation and dissemination. The farmer decision support tool converts evaluation results into operation recommendations, and the fertilization scheme generation considers local conditions. Mobile terminal application provides real-time guidance, offline function adapts to areas with insufficient network coverage. Multilingual interface meets the needs of different users, and graphical operation reduces the threshold of use. User feedback mechanism collects user experience to continuously improve the practicality of the tool.

[0086] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0087] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based intelligent fertilization management method, characterized in that, Includes the following steps: Collect soil parameter sets and environmental parameter sets of the target farmland. The soil parameter set includes soil moisture, nitrogen, phosphorus and potassium content and pH value. The environmental parameter set includes light intensity, temperature and rainfall. Based on historical data from soil parameter sets and environmental parameter sets, a dynamic change model of soil nutrients is constructed, and the trend of soil nutrient consumption in the future within a preset period is predicted through time series analysis. Based on soil nutrient consumption trends and combined with the characteristics of crop growth stages, an initial fertilization plan is generated. Multimodal sensors are used to monitor crop growth status in real time, and leaf color index, stem height and fruit development progress are obtained to form a growth status dataset. The growth status dataset and the initial fertilization plan are input into the fertilization decision model. By comparing the deviations in growth status, the nutrient ratio in the initial fertilization plan is adjusted to generate an optimized fertilization plan. The system controls intelligent fertilization equipment to perform precise fertilization in designated zones according to the optimized fertilization plan, and records the actual fertilization data. The actual fertilization data is associated and stored with soil parameter sets, environmental parameter sets and growth status datasets to update the training data of the soil nutrient dynamic change model and the fertilization decision model. The training steps of the fertilization decision model include: Collect historical fertilization schemes and corresponding growth status data of different crop varieties under diverse environments, and construct a labeled dataset; A graph convolutional network is used to process the nonlinear relationship between fertilization schemes and growth status in the labeled dataset, and to generate topological feature vectors. Adversarial training is used to improve the generalization ability of graph convolutional networks, so that their output error for fertilization schemes with unseen combinations of environmental parameters is lower than the allowable value. The specific steps for controlling the intelligent fertilization equipment to perform zoned precision fertilization according to the optimized fertilization plan include: Based on the fertility difference zones defined by the farmland geographic information system, the nutrient allocation ratios in the fertilization scheme are matched and optimized. Differentiated control commands are sent to the smart fertilization equipment in each zone through the Internet of Things gateway to start the fertilization operation synchronously. The BeiDou positioning system is used to verify the consistency between the trajectory of the fertilization equipment and the coverage area of ​​the nutrient spraying in real time.

2. The intelligent fertilization management method based on machine learning according to claim 1, characterized in that, The specific steps for constructing the dynamic change model of soil nutrients include: Extract sampling data from historical soil parameter sets and environmental parameter sets for consecutive time periods, and divide the training samples according to the crop growth cycle; Long Short-Term Memory (LSTM) networks were used to perform time-series modeling on the training samples to learn the decay pattern of soil nutrients as environmental parameters changed; By optimizing the weight parameters of the long short-term memory network using the backpropagation algorithm, the error between the predicted soil nutrient consumption trend and the actual detected value is less than a preset threshold.

3. The intelligent fertilization management method based on machine learning according to claim 2, characterized in that, The specific steps for generating the initial fertilization plan include: Based on the soil nutrient consumption trend, calculate the expected deficit of nitrogen, phosphorus, and potassium in the future preset period; Based on the standard fertilizer requirements of crops at their current growth stage, allocate basic fertilizer application based on the expected deficit. By incorporating rainfall and temperature data from the environmental parameter set, the ratio of fast-acting fertilizer to slow-release fertilizer in the basic fertilizer application can be dynamically adjusted.

4. The intelligent fertilization management method based on machine learning according to claim 3, characterized in that, The specific steps for adjusting the nutrient ratio in the initial fertilization plan by comparing deviations in growth status include: By comparing the leaf color index with the standard health threshold, areas with insufficient or excessive chlorophyll synthesis can be identified. Adjust the amount of nitrogen fertilizer according to the deviation in stem height, and adjust the amount of potassium fertilizer according to the deviation in fruit development progress; The nutrient ratio is iteratively optimized using a reinforcement learning algorithm until the overall deviation rate of the growth state dataset is reduced to the target range.

5. The intelligent fertilization management method based on machine learning according to claim 1, characterized in that, The specific steps for updating the training data of the soil nutrient dynamics model and the fertilization decision model include: The deviation between the actual fertilization data and the predicted soil nutrient consumption trend is added to the anomaly detection module of the labeled dataset; When the deviation value exceeds the warning threshold, the incremental learning process of the soil nutrient dynamic change model is triggered to recalibrate the time series prediction parameters. A federated learning framework is used to aggregate fertilization effect data from multiple plots and globally update the weight matrix of the fertilization decision model.

6. The intelligent fertilization management method based on machine learning according to claim 5, characterized in that, The operation steps of the anomaly detection module include: Establish a time alignment relationship between actual fertilization data and growth status dataset, and extract nutrient absorption response curves within a preset number of days after fertilization; Outliers in response curves are identified using the Isolation Forest algorithm and marked as high-risk events requiring human intervention. The sets of soil parameters and environmental parameters associated with high-risk events are stored separately for model failure analysis.

7. The intelligent fertilization management method based on machine learning according to claim 6, characterized in that, It also includes the following steps: After crop harvesting, collect final yield data and soil residual nutrient data to evaluate the actual effectiveness of fertilization programs; The prediction weights of the soil nutrient dynamic change model and the decision thresholds of the fertilization decision model are adjusted in reverse based on the benefit assessment results.

8. A machine learning-based intelligent fertilization management system, 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 steps of the intelligent fertilization management method based on machine learning as described in any one of claims 1 to 7.

Images