Crop field fertilization decision system based on big data analysis

Abstract

The application provides a crop field fertilization decision system based on big data analysis, relates to the technical field of crop field fertilization, and comprises a data acquisition module, a data fusion and modeling module, an intelligent decision module, an execution output module and a feedback optimization module.The data acquisition module acquires multi-source heterogeneous big data in real time, the data fusion and modeling module adopts a graph attention network to perform weighted fusion on the big data, forms a mixed nutrient dynamic model, the intelligent decision module generates a variable fertilization prescription scheme by combining a multi-objective evolutionary algorithm and a digital twin through a reinforcement learning agent, the execution output module converts the prescription scheme into an instruction and issues the instruction to a variable rate fertilization device, and the feedback optimization module adopts a federated learning mechanism to cross-regionally iterate model parameters.Meanwhile, the system adopts a cloud-edge-end collaborative architecture, realizes precise modeling of field spatial heterogeneity, multi-objective dynamic optimization and privacy protection and decision interpretability, and effectively improves fertilization precision, fertilizer utilization rate and reduces environmental impact.

Description

Technical Field

[0001] This invention relates to the field of crop field fertilization technology, specifically a crop field fertilization decision system based on big data analysis. Background Technology

[0002] As agriculture develops towards precision and intelligence, higher demands are placed on the real-time, accuracy, and sustainability of crop field fertilization decisions. Especially in complex scenarios such as saline-alkali land, dryland farming areas, and terraced fields in Ningxia Hui Autonomous Region, and against the backdrop of climate change, soil degradation, and resource constraints, how to achieve a balance between yield improvement, fertilizer conservation, and environmental friendliness has become a focus of modern agriculture. For example, major food crops such as wheat and rice need to dynamically adjust the amount of nitrogen, phosphorus, and potassium fertilizers based on micro-variations in the field. Traditional fertilization methods that rely on experience are no longer able to meet the comprehensive needs of high yield, fertilizer conservation, and emission reduction. Currently, IoT sensors, remote sensing images, and agricultural big data platforms have been applied in some areas, and some systems provide basic fertilization recommendations through simple statistical models or empirical rules.

[0003] However, existing crop field fertilization decision-making technologies still face the following key technical challenges in practical applications: 1. Existing fertilization decision-making systems mostly use single sensor monitoring or simple spatial statistical averaging methods, lacking the ability to deeply integrate spatiotemporally heterogeneous data such as remote sensing images, soil nutrient distribution, and meteorological environment. This results in the system being unable to accurately depict the complex nutrient variation characteristics within the field, making it difficult to identify spatial differences in crop growth, and causing the fertilization recommendations to not match the actual needs, thus failing to meet the precision requirements of refined variable fertilization.

[0004] 2. Existing data-driven models are mostly pure statistical models or shallow neural networks, which often treat the soil-crop system as a "black box" and ignore the inherent agricultural physical laws such as nutrient migration and crop absorption. This makes the models prone to overfitting when facing small sample scenarios or cross-regional applications. Moreover, when there are drastic environmental fluctuations, such as extreme precipitation and high temperatures, the prediction results often deviate from the actual physical logic, resulting in poor stability and reliability of decision-making schemes.

[0005] 3. Existing fertilization programs often focus on the single goal of increasing yield, lacking a comprehensive and dynamic consideration of fertilizer input costs and the environmental impact of fertilizer runoff. At the same time, cross-farm data collaboration often faces serious privacy risks, and complex algorithm models lack interpretability, making it difficult for farmers and regulatory agencies to understand the logic behind the decisions. This not only limits the collaborative evolution of the system but also leads to a lag in the closed-loop feedback adjustment mechanism. Summary of the Invention

[0006] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a crop field fertilization decision-making system based on big data analysis, which solves the problems mentioned in the background section above.

[0007] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a crop field fertilization decision-making system based on big data analysis, wherein the system includes a data acquisition module, a data fusion and modeling module, an intelligent decision-making module, an execution output module, and a feedback optimization module, wherein: The data acquisition module is used to collect multi-source heterogeneous big data in real time, including UAV remote sensing image data, field IoT sensor data, meteorological data, historical yield data and crop variety parameter data. The data fusion and modeling module is used to clean and fuse the multi-source heterogeneous big data, and construct a field spatial heterogeneity knowledge graph based on graph attention network, while embedding physical information neural network to form a mixed nutrient dynamic model. The intelligent decision-making module is used to generate variable fertilization prescription schemes based on the mixed nutrient dynamic model and reinforcement learning agent, combined with a multi-objective evolutionary algorithm. The schemes simultaneously optimize three major objectives: yield, cost and environmental impact, and use digital twins to simulate the scenario. The execution output module is used to convert the variable fertilization prescription scheme into executable instructions and send them to the variable rate fertilization equipment; The feedback optimization module is used to collect execution feedback data and iterate model parameters across regions through a federated learning mechanism, while using interpretable artificial intelligence to generate decision explanation reports.

[0008] Preferably, in the data fusion and modeling module, the graph attention network uses a multi-head attention mechanism to perform weighted fusion of remote sensing image features and sensor time-series data, and the physical information neural network embeds the nutrient migration physical partial differential equation as a loss function constraint term into the neural network training process.

[0009] Preferably, the reinforcement learning agent in the intelligent decision-making module adopts a proximal policy optimization algorithm, with the current state of the field as the state space, the combination of fertilizer amount, timing and location as the action space, and a multi-objective weighted reward function as the optimization objective.

[0010] Preferably, the multi-objective evolutionary algorithm is used to generate Pareto front solution sets, and a digital twin module is used to perform dynamic nutrient simulation for each solution set over the next 48 hours to achieve uncertainty quantification and optimal solution selection.

[0011] Preferably, the feedback optimization module adopts a federated learning framework, where each farm edge node only uploads model gradient parameters and does not share the original field data, thereby achieving model generalization and data privacy protection.

[0012] Preferably, the interpretable artificial intelligence in the feedback optimization module quantifies the contribution of each input feature of the hybrid model using the Shapley value, and outputs a decision basis report in the form of a visual heatmap to support farmer terminal queries and regulatory compliance reviews.

[0013] Preferably, the crop field fertilization decision-making method of the system includes the following steps: Sp1: Real-time acquisition of multi-source heterogeneous big data, including satellite or UAV remote sensing image data, field IoT sensor data, meteorological data, historical yield data, and crop variety parameter data, through the data acquisition module; Sp2: The data fusion and modeling module cleans, standardizes and integrates the collected multi-source heterogeneous big data, constructs a field spatial heterogeneity knowledge graph based on graph attention network, and embeds the nutrient migration physical partial differential equation into the physical information neural network to form a hybrid nutrient dynamic model, while updating the field digital twin model. Sp3: The intelligent decision-making module runs a reinforcement learning agent based on the mixed nutrient dynamic model, and generates variable fertilization prescription schemes by combining a multi-objective evolutionary algorithm to achieve multi-objective optimization of maximizing yield, minimizing fertilizer cost and minimizing environmental impact. The digital twin module performs nutrient dynamic simulation and uncertainty quantification for each candidate scheme over the next 48 hours. Sp4: The variable fertilization prescription scheme is converted into executable instructions by the execution output module and sent to the variable rate fertilization equipment to realize automated fertilization; Sp5: The feedback optimization module collects soil nutrient, crop growth and environmental feedback data in real time after execution, uses federated learning mechanism to aggregate only the model gradient parameters for cross-regional iterative optimization, and uses an interpretable artificial intelligence module to generate decision basis reports and complete long-term model updates.

[0014] Preferably, in Sp3, the multi-objective optimization function is defined as maximizing output minus the first dynamic weight multiplied by the cost minus the second dynamic weight multiplied by the environmental impact, wherein the first dynamic weight and the second dynamic weight are adaptively adjusted by the digital twin simulation results.

[0015] Preferably, the system is deployed in a cloud-edge-device collaborative architecture, with the cloud responsible for global model training, edge nodes responsible for real-time inference at the field level, and the terminal application providing a decision visualization interface.

[0016] Beneficial effects This invention provides a crop field fertilization decision-making system based on big data analysis. It has the following beneficial effects: 1. This invention utilizes convolutional neural networks to extract deep features from remote sensing images, capturing subtle variations in crop growth. It also combines graph attention networks to spatially weight and fuse heterogeneous data from multiple sources, such as soil, meteorology, and imagery, to construct a refined knowledge graph of spatial heterogeneity in fields. This approach effectively solves the information gaps caused by traditional statistical averaging methods, and can highly restore the true differences in nutrient distribution within fields, providing a highly reliable digital foundation for variable operations.

[0017] 2. This invention uses a physical information neural network to embed the partial differential equations of nutrient migration and diffusion as soft constraints into the loss function of the neural network, enabling the model to have both data-driven and physical logic. This mechanism ensures that the prediction results always follow the laws of conservation of matter and natural diffusion, significantly improving the prediction reliability of the model in data-sparse scenarios or extreme environmental fluctuations. At the same time, through automatic error judgment and backoff mechanisms, the generalization applicability of the digital twin model under different soil types and climatic conditions is ensured, solving the pain point of poor stability of traditional black box models.

[0018] 3. This invention employs reinforcement learning agents combined with multi-objective evolutionary algorithms. In a digital twin simulation environment, through dynamic weight adjustment, it achieves a Pareto optimal balance between yield benefits, input costs, and environmental runoff risks, solving the problems of resource waste and environmental pollution caused by single-objective decision-making. At the same time, the system utilizes a federated learning mechanism to achieve privacy-preserving iteration of model parameters and introduces interpretable artificial intelligence attribution analysis to transform complex decision-making logic into a visual explanation. This not only safeguards data sovereignty in cross-farm collaboration but also enhances the transparency and auditability of decisions, forming a virtuous cycle of feedback and trust mechanism. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the decision-making method of the present invention; Figure 3 This is a diagram showing the main interface of the system according to the present invention. Detailed Implementation

[0020] 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 skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1: Please see Figure 1 and Figure 2As shown, the crop field fertilization decision-making system based on big data analysis includes a data acquisition module, a data fusion and modeling module, an intelligent decision-making module, an execution output module, and a feedback optimization module. The cloud is responsible for global model training and federated learning aggregation, the edge nodes (field-level gateways) are responsible for real-time inference and data preprocessing, and the terminal application (farmer's mobile terminal or monitoring platform) is responsible for decision visualization and instruction confirmation. By deeply integrating the closed-loop process of "perception, fusion, decision-making, execution, and feedback", an intelligent fertilization system with high spatial heterogeneity perception capability and physical constraint logic is constructed.

[0022] The first step is real-time multi-source heterogeneous big data acquisition: To provide the system with a continuous spatiotemporal decision-making foundation, the system first acquires multi-source heterogeneous big data in real time and at high frequency through the data acquisition module, initiating comprehensive field sensing work and comprehensively acquiring environmental factors affecting crop growth. This module includes a remote sensing acquisition submodule, a sensor acquisition submodule, a meteorological interface submodule, and a database query submodule, where the data sources are: The UAV remote sensing image data is acquired every 30 minutes or triggered by an event, with multispectral images including blue, green, red, red edge, and near-infrared bands, and a resolution of ≤0.5m. Feature vectors are extracted through a convolutional neural network with a dimension of 512. Field IoT sensor data is sampled every 15 minutes, including soil nitrate nitrogen (mg / kg), available phosphorus (mg / kg), available potassium (mg / kg), water content (%), and pH value (accuracy ±0.1). Outliers are normalized using Z-score. Range elimination; Meteorological data is obtained in real time and forecasts for the next 48 hours through the China Meteorological Administration or third-party application interfaces, including temperature (°C), rainfall (mm), wind speed (m / s), humidity (%), and timestamps aligned to UTC+8. Historical yield data and crop variety parameters were extracted from the National Agricultural Big Data Platform or local databases, including yields of the first three seasons (kg / mu) and variety genotype parameters (growth cycle days, nitrogen uptake coefficient).

[0023] Furthermore, during operation, the operator first starts the system via a terminal app or edge node console. The system automatically or manually initiates data collection based on a preset timer (every 15 minutes) or event triggers (i.e., weather warnings, abnormal crop growth). The system first directs a drone equipped with a multispectral camera to perform a full-coverage scan of the field. It then uses a convolutional neural network (CNN) algorithm to extract deep features from the original images, identifying spectral features reflecting the crop's nutritional status. This is achieved through sliding calculations across multiple convolutional kernels in the spatial dimension. The core convolution formula is expressed as: ; In this formula, This represents the original spectral image pixel matrix from the input. This represents a convolution kernel with specific weight parameters. For bias terms, These are the generated feature mapping values. The specific operation process includes: first, adjusting the image to a standard size and normalizing it; then, capturing subtle changes in edges, textures, and crop color gradients in the image through multiple convolutional layers, and using a non-linear activation function. The system retains effective growth characteristics and suppresses background noise. Then, it performs downsampling operations through pooling layers to reduce data dimensionality and enhance feature invariance. Finally, it outputs a high-dimensional feature vector that can quantify the crop nitrogen distribution state through a fully connected layer. Meanwhile, IoT sensors buried in the soil collect real-time data on nitrate nitrogen, available phosphorus, available potassium, as well as moisture, pH, and temperature. Because the data from different sources differ significantly in units and magnitudes, the system then calls a standardization algorithm, namely the Z-score standardization method, using the formula: ; in This is the original data. The mean, To determine the standard deviation, all heterogeneous data are transformed into a unified standardized vector. Then, real-time and 48-hour forecast data are obtained through the meteorological API, and historical yield and crop variety parameters are extracted from the National Agricultural Big Data Platform or local database. Subsequently, the system automatically executes the Extract-Transform-Load pipeline (ETL pipeline) to perform data integrity verification, outlier removal, timestamp alignment, and format standardization. Finally, a standardized multi-source heterogeneous large data package is generated and pushed to the next stage of data fusion and modeling. If any sub-module data is interrupted for more than 30 minutes, it automatically switches to the most recent valid historical backup data and generates an alarm log to push to the terminal application. To ensure data real-time latency of less than 1 minute and spatial coverage integrity, high-precision heterogeneous basic data is provided for subsequent fusion.

[0024] Taking a single-season rice planting field as an example, a farm with a planting area of ​​150 mu (approximately 10 hectares) uses the "Huanghuazhan" single-season rice variety. The soil is clayey, and there was excessive rainfall in the early stage. The farm faced typical problems: high temperature and humidity led to severe nitrogen volatilization and runoff loss, and traditional fertilization was prone to lodging and eutrophication of water bodies. After the system was deployed, when meteorological monitoring showed that there was recent high temperature and humidity and a risk of flooding, the acquisition module was automatically triggered. The acquired remote sensing images, after being extracted by CNN, clearly showed the leaf yellowing characteristics in the southeast low-lying area due to root damage. Meanwhile, the sensor data simultaneously fed back the specific values ​​of nutrient loss in the low-lying area. These cleaned and characterized data provided standardized digital basis for the subsequent construction of an accurate spatial distribution model.

[0025] The second step, data fusion, hybrid model construction, and digital twin update: In order to transform the above-mentioned scattered data points into a digital model with logical connections and physical laws, the system must reconstruct spatial heterogeneity and integrate physical information through the data fusion and modeling module. Although the original data reflects the current situation in the field, it lacks the logic of spatial mutual influence and the constraints of soil chemical changes. Therefore, this module realizes the deep fusion of multi-source data and the construction of a hybrid dynamic model to form a field digital twin that can be used for decision-making. The internal structure includes a graph attention fusion submodule, a physical information neural network training submodule, and a knowledge graph update submodule. This module first divides the field into fine grid nodes and uses a graph attention network (GAT) to perform weighted fusion of remote sensing image features and sensor time series data. The physical information neural network embeds the nutrient migration physical partial differential equation as a loss function constraint term into the neural network training process. The data sources are remote sensing feature vectors extracted by convolutional neural networks and sensor time series preprocessed by long short-term memory networks. Furthermore, during operation, after receiving the Sp1 data packet, the system first performs further data cleaning and Min-Max standardization. To achieve deep fusion of multi-source heterogeneous data and generate a digital model, the system uses a Graph Attention Network (GAT) to discretize the fields into a graph structure composed of grid nodes. Each node carries the remote sensing features, sensor readings, and meteorological parameters of the area. The fusion principle lies in the algorithm automatically learning spatial correlations using a multi-head attention mechanism to divide the fields into... Grid, Constructing a knowledge graph, node features The edges are adjacency relationships with Euclidean distance < 10m. A multi-head attention mechanism is used, with 8 heads and 256 hidden dimensions. The attention coefficient is calculated using the following formula: ; in For learnable weight matrix, For attention vectors, This represents vector concatenation. Represents a node For nodes The attention weights are fused and output as a global feature vector with a dimension of 1024. In this way, the originally isolated multi-source data is mapped into a high-dimensional spatial feature vector, realizing the transformation from physical entity data to digital logic model, and constructing a knowledge graph that represents the spatial heterogeneity of the field. This is the static skeleton of the digital twin model, which can compress multi-dimensional environmental features and map them into a unified spatial coordinate system. Meanwhile, to endow the model with the ability to predict the future and to prevent the model's predictions from violating the natural laws of soil fertility change, the system further introduces a Physical Information Neural Network (PINN) to strengthen the aforementioned framework. The data source is GAT fused feature vectors. Its core lies in constructing a comprehensive neural network total loss function, which consists of two parts: ,in, This is the traditional data-driven loss term (i.e., mean squared error), used to fit the measured data and ensure that the model output is as close as possible to the sensor's measured values. The key physical loss term is the physical loss term, which serves to embed the physical partial differential equation (PDE) of nutrient migration as a "soft constraint" into the optimization process of the neural network. This relationship is manifested in the fact that the system does not force the neural network to solve the equation precisely, but rather uses the residual of the physical equation as a penalty. The specific formula is as follows: ; in, Nutrient concentration, The diffusion coefficient (typical value 0.01m) 2 / h), For the migration velocity vector, For the response terms (crop uptake and microbial transformation), and data loss Joint optimization, with Adam as the optimizer, means that when the nutrient distribution predicted by the neural network does not conform to the partial differential equation describing the laws of conservation and diffusion of matter, the physical loss term will increase significantly, forcing the model to converge in a direction that conforms to the laws of agronomic physics. In this way, the system finally generates a hybrid nutrient dynamic model that is both data-supported and physically logical. After the model is generated, the system automatically initiates an accuracy assessment program. This assessment process is implemented through a built-in model validation operator, and its automated assessment logic is mainly based on the following three quantitative indicators: Data consistency residual assessment: The system compares the model's predicted values ​​with the sensor data from the validation set (hold-out data) that was not used in training, and calculates the root mean square error (RMSE). , If a preset industry accuracy threshold, such as 10%, is set, it is considered that the data level does not meet the standards; physical constraint violation rate determination: the system calculates the physical loss item. In the global grid distribution, if more than a certain proportion of grid points fail to meet the minimum residual requirement of the nutrient migration equation (i.e., the prediction results "disappear out of thin air" or "surge abnormally" in terms of physical logic), it is judged as a physical logic failure; Confidence interval determination: The system uses the Monte Carlo sampling method to evaluate the uncertainty of the prediction output. If the model's prediction confidence interval in the key fertilizer-demanding region is too wide (i.e., the coefficient of variation CV exceeds the set safety boundary), it is considered a failure. If the error is not met, it means that the features provided by the current data source are insufficient to support reliable modeling. At this time, the system will automatically determine that the error is not up to standard and trigger the rollback mechanism, and the instruction will jump back to Sp1 to re-collect data.

[0026] In the aforementioned implementation case of a single-season rice paddy, due to excessive rainfall and a hot and humid environment in the early stages, traditional pure data models are prone to data noise interference and give chaotic predictions. However, this system, through soft constraints of physical equations, ensures that the nutrient changes predicted by the model always follow the physical logic of migration from high concentration to low concentration and with the direction of water flow. In this process, after the model is initially fused, the automatic judgment program finds that the prediction confidence of the southeast low-lying area is low. The judgment engine identifies that the red edge index of this area is abnormal due to water accumulation. The system then triggers the rollback mechanism to the first step, by increasing the frequency of overlapping scans of the UAV in the low-lying area and retrieving denser historical sensor sequences for data completion. After resampling and secondary modeling, the system uses graph attention network to map these multispectral images and sensor data to a high-dimensional feature space. The resulting digital twin model not only achieves high-dimensional fusion and restoration of the current nutrient field and water field of the field, but also maintains extremely high prediction robustness in complex nitrogen volatilization and runoff scenarios, providing a solid and reliable decision-making foundation for the subsequent generation of fertilization prescriptions that are both spatially accurate and physically realistic.

[0027] The third step is intelligent fertilization decision-making based on multi-objective optimization: After having an accurate digital twin model, it is far from enough for the system to only understand the current situation. It also needs to formulate the optimal action plan based on the crop's growth goals. Therefore, the process moves to the third step, which uses the intelligent decision-making module to generate a multi-objective optimized variable fertilization prescription plan based on a hybrid model. The internal structure includes a reinforcement learning agent submodule, a multi-objective evolutionary submodule, and a digital twin simulation submodule. Based on the hybrid nutrient dynamic model and reinforcement learning agent, combined with the multi-objective evolutionary algorithm, a variable fertilization prescription plan is generated. The plan simultaneously optimizes the three major goals of yield, cost, and environmental impact. The scenario is simulated through digital twin, and massive strategy trial and error is carried out in the digital twin simulation environment. Furthermore, during operation, the system loads the hybrid model and digital twin state output by Sp2, extracts the current field state vector (256 dimensions, covering key features such as nutrient fields, Normalized Difference Vegetation Index (NDVI), and weather forecasts), and uses this as the basis to drive the reinforcement learning agent submodule. The system adopts the Proximal Policy Optimization (PPO) algorithm as the core decision engine, and its objective function formula is: ; Among them, the strategy pruning parameters Set to 0.2, updated 20 times per batch, through an actuator-evaluator network in the continuous action space. (Including nitrogen / phosphorus / potassium fertilizer application rate [0-50 kg / mu], fertilization timing [0-24 h], and grid index position) the strategy is iterated. In order to balance yield, cost, and environmental impact, the system constructs a multi-objective reward function: ; Among them, weight , The initial values ​​were set to 0.3 and 0.4, and were adaptively adjusted according to the simulation results. The reward function balanced yield prediction, fertilizer cost and environmental impact. PPO avoided training instability through policy pruning and achieved optimal sequential decision-making. Building upon this, the Multi-Objective Evolutionary Algorithm (MOEA) submodule inherits the initial PPO strategy, generating a Pareto front solution set through 50 generations of iteration with a population of size 100. Its multi-objective optimization function is expressed as: ; This function uses three sub-functions: yield prediction revenue (which is transformed into a minimization problem by taking negative values), fertilizer application costs, and environmental impact indicators. The system performs a weighted summation to find a solution set that allows multiple conflicting objectives to achieve optimal balance simultaneously. Subsequently, the system uses a digital twin module to conduct a 48-hour Monte Carlo simulation on the candidate solutions. Uncertainty is quantified through 100 samplings to ensure a confidence interval of 95% and a standard deviation of less than 5%. Finally, the solution with the largest supervolume is selected as the optimal prescription. Otherwise, the number of iterations is increased and Sp3 is rerun to achieve dynamic balance of the three objectives. A precise prescription map that can be directly executed is generated. The output is a variable fertilization prescription plan containing location coordinates, fertilization amount, and timing fields, which is pushed to the execution output module.

[0028] In the aforementioned implementation case study targeting a single-season rice paddy, when environmental impact exceeding standards was detected, the system automatically increased the weighting coefficient. Furthermore, when the Pareto front convergence index was below 95%, the number of iterations was increased. Finally, the system selected the optimal solution with the largest supervolume and generated a rasterized prescription map containing location coordinates, fertilization amount, and timing fields. This accurately identified the nutrient deficiency in the southeast low-lying area and formulated a small-scale, multiple-time fertilization plan, effectively suppressing runoff pollution that may be caused by high temperature and high humidity.

[0029] The fourth step is prescription conversion and automated execution output: Scientific prescriptions must be converted into mechanical language that agricultural machinery can execute. To this end, the system enters the fourth step, realizing the automated conversion and accurate issuance of instructions through the execution output module. The internal structure includes an instruction conversion submodule, an edge computing issuance submodule, and a confirmation receipt submodule. Furthermore, during operation, the module activates the instruction conversion submodule to map the gridded fertilizer prescription map into control messages conforming to the ISOBUS or Controller Area Network (CAN) bus standard. Subsequently, the edge computing distribution submodule in the system uses the mobile network to push the instructions to the variable rate fertilizer applicator equipped with the Beidou positioning system. Its core technology is the variable rate control algorithm (VRT), which adjusts the fertilizer spraying flow rate by controlling the pulse width modulation (PWM) signal of the actuator in real time. The general flow control formula is: ; in For traffic, for The duty cycle is then determined, and finally, after the device executes the command, a confirmation receipt is generated and sent to the next module.

[0030] In the above implementation case of a single-season rice field, the variable-rate fertilizer applicator analyzes the prescription values ​​based on real-time coordinates. When the machine travels to a low-lying area with high nutrient demand, the PWM control system automatically increases the duty cycle and increases the spraying flow. When it travels to an area with sufficient nutrients, it automatically reduces the pressure and limits the flow. The execution output module receives confirmation receipts from the agricultural machine in real time, ensuring that every step of the operation strictly follows the prescription instructions, and realizing deep collaboration between digital decision-making and mechanized operation.

[0031] The fifth step, feedback aggregation, interpretability analysis, and model iteration: After a fertilization operation is completed, in order to verify the effectiveness of the decision and enable the system to continuously evolve, the process finally enters the fifth step, which is to conduct closed-loop iteration and decision interpretability analysis through the feedback optimization module. The internal structure includes a federated learning aggregation submodule, an interpretable AI explanation submodule, and a model update submodule. This module collects soil and crop growth feedback again within a certain period of time after the operation and uses the federated learning mechanism to achieve cross-regional model parameter iteration. Furthermore, during operation, after the equipment completes its task, the module immediately verifies the remote sensing data (such as runoff nitrogen) collected via sensors and drones. It then uses the FedAvg algorithm to generalize the model, with the core formula being: ; in For the updated global model parameters, For edge nodes Sample size The total sample size is... The formula, which is the model gradient parameter uploaded by each farm edge node, aims to optimize the global model without sharing the original field data by weighted averaging of the local gradients of each farm. Meanwhile, to address the issue of decision-making transparency, the system introduces the Explainable Artificial Intelligence (SHAP) attribution analysis algorithm, whose formula is: ; By calculating features In all possible feature subsets The system calculates the expected marginal contribution of input features such as temperature, pH, and soil moisture content to the Shapley value of fertilization decisions, generates a visual heat map report, and finally completes long-term model updates and feeds back the parameters to the Sp2 and Sp3 modules to form a closed loop. If environmental indicators exceed the limits, a new round of iteration is immediately triggered. Operators can view the report and perform manual confirmation on the terminal app.

[0032] In the aforementioned implementation case of a single-season rice paddy, feedback data showed that the yellowing phenomenon of the crop was significantly improved, and the expected yield steadily rebounded. The SHAP heat map clearly showed that "soil moisture content" and "precipitation probability" were the key factors affecting the adjustment of fertilizer application. This not only enhanced the transparency of decision-making, but also transformed the successful experience in dealing with high temperature and high humidity environment into a system-wide knowledge accumulation through federated learning. Through this complete implementation process, this invention successfully constructed a precision agriculture fertilization closed loop with self-evolution capabilities.

[0033] Please see Figure 3 As shown, through the continuous execution of the above five steps, this method forms an efficient, closed-loop, and intelligent field fertilization decision-making process in practical use, realizing fully automated operation from data perception to precision fertilization, significantly improving fertilizer utilization, reducing environmental impact and ensuring crop yield. The triggering frequency of this process can be flexibly adjusted according to the actual field size, and logs are automatically recorded after each cycle to support subsequent optimization.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A crop field fertilization decision system based on big data analysis, characterized in that: The system includes a data acquisition module, a data fusion and modeling module, an intelligent decision-making module, an execution output module, and a feedback optimization module, wherein: The data acquisition module is used to collect multi-source heterogeneous big data in real time, including UAV remote sensing image data, field IoT sensor data, meteorological data, historical yield data and crop variety parameter data. The data fusion and modeling module is used to clean and fuse the multi-source heterogeneous big data, and construct a field spatial heterogeneity knowledge graph based on graph attention network, while embedding physical information neural network to form a mixed nutrient dynamic model. The intelligent decision-making module is used to generate variable fertilization prescription schemes based on the mixed nutrient dynamic model and reinforcement learning agent, combined with a multi-objective evolutionary algorithm. The schemes simultaneously optimize three major objectives: yield, cost and environmental impact, and use digital twins to simulate the scenario. The execution output module is used to convert the variable fertilization prescription scheme into executable instructions and send them to the variable rate fertilization equipment; The feedback optimization module is used to collect execution feedback data and iterate model parameters across regions through a federated learning mechanism, while using interpretable artificial intelligence to generate decision explanation reports.

2. The crop field fertilization decision-making system based on big data analysis according to claim 1, characterized in that: In the data fusion and modeling module, the graph attention network uses a multi-head attention mechanism to perform weighted fusion of remote sensing image features and sensor time-series data, and the physical information neural network embeds the nutrient migration physical partial differential equation as a loss function constraint term into the neural network training process.

3. The crop field fertilization decision-making system based on big data analysis according to claim 1, characterized in that: The reinforcement learning agent in the intelligent decision-making module adopts a proximal policy optimization algorithm, with the current state of the field as the state space, the combination of fertilizer amount, timing and location as the action space, and a multi-objective weighted reward function as the optimization objective.

4. The crop field fertilization decision-making system based on big data analysis according to claim 1, characterized in that: The multi-objective evolutionary algorithm is used to generate Pareto front solutions, and a digital twin module is used to perform dynamic nutrient simulation for each solution over the next 48 hours to achieve uncertainty quantification and optimal solution selection.

5. The crop field fertilization decision-making system based on big data analysis according to claim 1, characterized in that: The feedback optimization module adopts a federated learning framework, where each farm edge node only uploads model gradient parameters and does not share the original field data, thereby achieving model generalization and data privacy protection.

6. The crop field fertilization decision-making system based on big data analysis according to claim 1, characterized in that: The interpretable artificial intelligence in the feedback optimization module uses the Shapley value to quantify the contribution of each input feature of the hybrid model and outputs a decision basis report in the form of a visual heatmap, supporting farmer terminal queries and regulatory compliance reviews.

7. The crop field fertilization decision-making system based on big data analysis according to any one of claims 1-6, characterized in that: The crop field fertilization decision-making method of the system includes the following steps: Sp1: Real-time acquisition of multi-source heterogeneous big data, including satellite or UAV remote sensing image data, field IoT sensor data, meteorological data, historical yield data, and crop variety parameter data, through the data acquisition module; Sp2: The data fusion and modeling module cleans, standardizes and integrates the collected multi-source heterogeneous big data, constructs a field spatial heterogeneity knowledge graph based on graph attention network, and embeds the nutrient migration physical partial differential equation into the physical information neural network to form a hybrid nutrient dynamic model, while updating the field digital twin model. Sp3: The intelligent decision-making module runs a reinforcement learning agent based on the mixed nutrient dynamic model, and generates variable fertilization prescription schemes by combining a multi-objective evolutionary algorithm to achieve multi-objective optimization of maximizing yield, minimizing fertilizer cost and minimizing environmental impact. The digital twin module performs nutrient dynamic simulation and uncertainty quantification for each candidate scheme over the next 48 hours. Sp4: The variable fertilization prescription scheme is converted into executable instructions by the execution output module and sent to the variable rate fertilization equipment to realize automated fertilization; Sp5: The feedback optimization module collects soil nutrient, crop growth and environmental feedback data in real time after execution, uses federated learning mechanism to aggregate only the model gradient parameters for cross-regional iterative optimization, and uses an interpretable artificial intelligence module to generate decision basis reports and complete long-term model updates.

8. The crop field fertilization decision-making system based on big data analysis according to claim 7, characterized in that: In Sp3, the multi-objective optimization function is defined as maximizing output minus the first dynamic weight multiplied by the cost minus the second dynamic weight multiplied by the environmental impact, wherein the first dynamic weight and the second dynamic weight are adaptively adjusted by the digital twin simulation results.

9. The crop field fertilization decision-making system based on big data analysis according to claim 1, characterized in that: The system is deployed in a cloud-edge-device collaborative architecture, with the cloud responsible for global model training, edge nodes responsible for real-time inference at the field level, and the terminal application providing a decision visualization interface.

Images