Agricultural information management system and method based on big data platform

Abstract

The invention relates to the technical field of agricultural information management, and particularly discloses an agricultural information management system and method based on a big data platform, and the method comprises the steps: firstly deploying a multi-source data collection module at an edge calculation node, and obtaining and standardizing the soil moisture content, meteorological environment and equipment operation data in real time; secondly, constructing a local dynamic irrigation strategy model, and realizing multi-objective optimization through a reinforcement learning algorithm; establishing a federated learning framework at the cloud, dynamically distributing node weights by adopting an attention mechanism, and realizing model aggregation of privacy protection in combination with secure multi-party computing; an optimal irrigation instruction is generated through a multi-source data fusion engine, and a three-level response exception handling mechanism is established; and finally, a closed-loop feedback system containing short-term incremental learning and long-term architecture optimization is formed. The corresponding management system comprises six functional modules, namely a data acquisition module, a local modeling module, a federated learning module, a real-time decision-making module, an abnormal monitoring module and a closed-loop optimization module.

Description

Technical Field

[0001] This invention relates to the field of agricultural information management technology, and specifically to an agricultural information management system and method based on a big data platform. Background Technology

[0002] With the advancement of agricultural modernization, precision irrigation technology has become an important means to improve water resource utilization efficiency and ensure food security. Traditional agricultural irrigation management mainly relies on manual experience and judgment, which suffers from problems such as strong subjectivity and slow response. In recent years, the application of new-generation information technologies such as the Internet of Things and big data in the agricultural field has provided technical support for building intelligent irrigation systems. By deploying various environmental sensors and irrigation equipment, agricultural producers can obtain real-time field data and achieve remote control. However, how to effectively integrate these scattered data resources and transform them into scientific irrigation decisions remains a significant challenge facing agricultural informatization.

[0003] The existing technology has the following shortcomings:

[0004] In agricultural water and fertilizer irrigation resource optimization systems based on genetic algorithms, the algorithms rely solely on static soil parameters for global optimization, failing to integrate real-time data on dynamic soil salinity changes (such as time-series information on ion concentration from conductivity sensors). This leads to local optima, specifically generating excessive irrigation commands in sandy soil regions. When automated irrigation robots execute these commands, continuous high-load operation causes overheating and damage to the pump motors. Furthermore, long-term excessive irrigation accelerates soil salinization, resulting in irreversible farmland degradation. Existing systems lack real-time hardware status monitoring modules (such as Hall effect-based current sensors) and closed-loop feedback mechanisms for dynamic optimization algorithms, failing to trigger emergency shutdown protection under abnormal conditions. Ultimately, this leads to irrigation system paralysis and significant economic losses in agricultural production. Summary of the Invention

[0005] The purpose of this invention is to provide an agricultural information management system and method based on a big data platform to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] The agricultural information management method based on a big data platform includes the following steps:

[0008] S1: The data processing platform deploys multiple edge computing nodes under a distributed computing architecture. Each edge computing node collects dynamic soil parameters, environmental monitoring data, and irrigation equipment operation status data of the target area in real time, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data.

[0009] S2: Build a dynamic irrigation strategy model locally on each edge computing node, generate an initial irrigation baseline based on historical data, and optimize the strategy through reinforcement learning. The optimization objectives include irrigation effect, resource consumption and soil health constraints.

[0010] S3: Establish a global federated learning framework on the cloud server, periodically aggregate the model parameters of each edge computing node, dynamically allocate node weights using an attention mechanism, and achieve privacy-preserving model fusion through secure multi-party computation.

[0011] S4: Distribute the updated global model to each edge node, combine it with real-time sensor data to generate the optimal irrigation instructions, and dynamically adjust the execution strategy;

[0012] S5: During irrigation, monitor the equipment status and soil parameter changes in real time. If an abnormality is detected, trigger the conservative irrigation mode and feed the abnormal data back to the cloud optimization model.

[0013] S6: Establish a closed-loop feedback mechanism to update the local strategy evaluation model in the short term and optimize the global model architecture in the long term, so as to achieve dynamic and precise management of agricultural irrigation.

[0014] As a further aspect of the present invention: the reinforcement learning optimization strategy includes the following steps:

[0015] A dual-channel strategy evaluation network was constructed. The first channel processes the spatiotemporal sequence data of soil moisture and uses a deep temporal convolutional network to extract soil moisture transport features. The second channel processes crop growth image data and captures the phenotypic change patterns of crops through multi-scale feature fusion.

[0016] The design incorporates a composite reward function, quantifying irrigation effect as the deviation coefficient between crop growth rate and theoretical value, converting resource consumption into a normalized index of hydropower energy consumption, and reflecting soil health constraints as the reciprocal of the salt accumulation rate.

[0017] A hierarchical experience replay mechanism is used to store training samples, prioritizing the retention of decision sequences with significant policy improvements, and weighted sampling is performed according to sample importance during model updates;

[0018] By introducing strategy distillation technology, the knowledge of complex decision-making networks is transferred to lightweight execution networks, ensuring the real-time response capability of edge devices.

[0019] As a further aspect of the present invention, the specific implementation of the optimization objective includes:

[0020] A dynamic prediction model for soil health was established, and multidimensional parameters such as soil dielectric constant and redox potential were collected through an embedded sensor network to construct an early warning index system for soil quality degradation.

[0021] Design a spatiotemporal coupled assessment method for resource consumption, deploy pressure fluctuation sensors at key nodes of the irrigation network, and establish a heat map of network energy consumption distribution by combining water flow velocity measurement;

[0022] Develop a multimodal verification system for irrigation effectiveness, integrating near-ground remote sensing spectral data, leaf temperature distribution data, and root growth monitoring data to construct a three-dimensional evaluation model for crop water use efficiency;

[0023] Implement a strategy to resolve conflicting objectives. When different optimization objectives conflict, dynamically adjust the weight allocation scheme based on the current crop growth stage and the meteorological warning level.

[0024] As a further aspect of the present invention: the dynamic allocation of node weights using an attention mechanism includes the following steps:

[0025] A multi-dimensional node contribution evaluation system is constructed, which comprehensively considers three dimensions: data quality indicators of each edge node, model update stability, and differences in regional environmental characteristics.

[0026] The design of the spatiotemporal attention calculation assigns higher temporal attention to nodes with drastic changes in soil parameters and higher spatial attention weight to nodes with crop planting area exceeding a preset threshold.

[0027] A dynamic weighting adjustment strategy is implemented to automatically increase the weighting coefficient of soil moisture data during the dry season and increase the weighting ratio of drainage capacity assessment during the rainy season.

[0028] Establish a weighted verification mechanism and use cross-validation to ensure that the attention allocation results match the actual irrigation needs of each region.

[0029] As a further aspect of the present invention: the privacy-preserving model fusion achieved through secure multi-party computation includes the following steps:

[0030] Design a layered encrypted transmission protocol to perform gradient-level segmented encryption processing on model parameters;

[0031] Construct a distributed computing verification network, in which multiple relay nodes jointly complete the compliance verification of model parameters;

[0032] A differential privacy protection mechanism is implemented, adding adaptive noise conforming to a Gaussian distribution during the model aggregation stage. The noise intensity is negatively correlated with data sensitivity.

[0033] Establish a model watermarking traceability system to assign a unique digital fingerprint to each node participating in federated learning.

[0034] As a further aspect of the present invention: the step of generating optimal irrigation instructions and dynamically adjusting the execution strategy by combining real-time sensor data includes the following steps:

[0035] A multi-source data fusion decision engine is established to align the irrigation baseline value output by the global model with real-time collected data on soil moisture content, crop stem micro-changes, and weather station forecasts in a spatiotemporal manner and weight the data with confidence levels to generate a preliminary irrigation plan.

[0036] The design incorporates a dynamic correction mechanism for irrigation strategies. When a sudden change in soil parameters or equipment malfunction is detected, an emergency strategy library based on case-based reasoning is activated to match the optimal adjustment scheme under similar historical conditions.

[0037] Implement a real-time feedback system for irrigation effects, using a network of miniature sensors buried in the crop root zone to monitor water infiltration rate and distribution uniformity, and dynamically adjust the intensity and duration of sprinkler irrigation.

[0038] A command security verification system is built, which simulates the soil parameter change trend after execution before the command is issued, and automatically triggers the manual review process when a risk of salinization is predicted.

[0039] As a further aspect of the present invention: the real-time monitoring of the device status and soil parameter changes includes the following steps:

[0040] A multi-dimensional equipment health assessment system was constructed. The time-domain characteristic waveform of the water pump bearing was collected by vibration sensor, and the frequency domain analysis results of pipeline pressure sensor were combined to establish a predictive model for equipment deterioration trend.

[0041] Design a soil anomaly detection network and deploy a capacitive-resistive dual-mode sensor array in the irrigation area. By measuring the synergistic change characteristics of soil dielectric constant and conductivity, identify areas of abnormal salt accumulation.

[0042] Implement a dynamic monitoring strategy adjustment mechanism. When the weather forecast shows that the probability of heavy precipitation exceeds the threshold, automatically increase the sampling frequency of soil moisture content and expand the density of monitoring points.

[0043] Establish an equipment-soil correlation analysis model. When a decline in irrigation equipment efficiency is detected, simultaneously analyze the historical change curves of soil parameters in the corresponding area to identify potential blockage or salinization problems.

[0044] As a further aspect of the present invention: the triggering of the conservative irrigation mode includes the following steps:

[0045] The design of a tiered emergency response strategy divides the conservative mode into three levels according to the severity of the anomaly: Level 1 adjusts irrigation duration, Level 2 reduces irrigation intensity, and Level 3 switches to preset safety parameters.

[0046] Build a case-based reasoning and decision-making database to store the handling solutions and effectiveness evaluations of historical abnormal events. When a new abnormality occurs, match the optimized handling solution of the most similar case.

[0047] Implement a dynamic parameter compensation mechanism to dynamically calculate the minimum required irrigation amount based on real-time meteorological data and crop growth stage under conservative mode;

[0048] Establish an anomaly feedback optimization loop, and mark the data collected during the execution of the conservative mode as high-value samples, which will be given priority for incremental training of the cloud model.

[0049] As a further aspect of the present invention, the establishment of the closed-loop feedback mechanism includes the following steps:

[0050] A dual-timescale model update system is constructed, with lightweight incremental learning deployed at edge nodes. The strategy is fine-tuned and the local parameters of the model are evaluated every hour based on the latest sensor data. A model structure evolution algorithm is deployed on a cloud server, and the irrigation effect of each region is comprehensively evaluated and the global model architecture is reconstructed every week.

[0051] The design of the feedback data hierarchical processing mechanism divides the irrigation execution results into two categories: routine operation data and abnormal event data. Routine data is stored using sliding window compression, while abnormal data triggers an immediate model correction process.

[0052] Implement cross-seasonal knowledge transfer technology, extract the feature patterns of the previous quarter as prior knowledge and inject them into the new model when the crop growth cycle changes, to accelerate training convergence;

[0053] Establish a model performance traceability system to record parameter changes and corresponding effect improvements with each update, forming an interpretable optimization path map to provide a basis for decision-making in subsequent adjustments.

[0054] Agricultural information management systems based on big data platforms include:

[0055] The distributed data acquisition and preprocessing module is used by the data processing platform to deploy multiple edge computing nodes under the distributed computing architecture. Each edge computing node collects soil dynamic parameters, environmental monitoring data and irrigation equipment operation status data of the target area in real time, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data.

[0056] The local strategy modeling and optimization module constructs a dynamic irrigation strategy model locally on each edge computing node, generates an initial irrigation baseline based on historical data, and optimizes the strategy through reinforcement learning. The optimization objectives include irrigation effect, resource consumption, and soil health constraints.

[0057] The federated learning and model aggregation module establishes a global federated learning framework on the cloud server, periodically aggregates the model parameters of each edge computing node, dynamically allocates node weights using an attention mechanism, and achieves privacy-preserving model fusion through secure multi-party computation.

[0058] The real-time decision-making and dynamic control module distributes the updated global model to each edge node, generates the optimal irrigation instruction by combining real-time sensor data, and dynamically adjusts the execution strategy.

[0059] An anomaly monitoring and emergency response module monitors the equipment status and soil parameter changes in real time during irrigation. If an anomaly is detected, a conservative irrigation mode is triggered, and the abnormal data is fed back to the cloud optimization model.

[0060] The closed-loop optimization and knowledge evolution module is used to establish a closed-loop feedback mechanism, update the local strategy evaluation model in the short term, and optimize the global model architecture in the long term to achieve dynamic and precise management of agricultural irrigation.

[0061] The beneficial effects of this invention are:

[0062] (1) This invention achieves localized real-time processing and decision-making of farmland environmental data through a distributed edge computing architecture. The dynamic irrigation strategy model, which operates independently on each edge node, can quickly respond to changes in the field environment, effectively solving the response lag problem of traditional centralized cloud processing methods. Local data processing significantly reduces network transmission requirements, ensuring stable system operation even in farmland areas with poor wireless network conditions. This architecture design is particularly suitable for precision irrigation management of large-scale farmland, optimizing system resource utilization efficiency while ensuring real-time decision-making.

[0063] (2) The federated learning framework and privacy protection mechanism adopted in this invention achieve secure sharing and collaborative optimization of agricultural data. Through encrypted transmission and privacy protection technologies, key data of each farm is kept locally, and only the processed model parameters are uploaded to participate in global optimization, thus ensuring data privacy and security from a technical perspective. At the same time, the dynamic weight allocation mechanism ensures differentiated contributions of different farmland regions to the global model. This method not only achieves effective sharing of agricultural knowledge across regions but also meets the data security requirements of agricultural production and operation, providing a feasible technical solution to the problem of agricultural data silos. Attached Figure Description

[0064] The invention will now be further described with reference to the accompanying drawings.

[0065] Figure 1 This is a flowchart of the agricultural information management method based on a big data platform according to the present invention;

[0066] Figure 2 This is a flowchart of the agricultural information management system based on a big data platform in this invention. Detailed Implementation

[0067] 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.

[0068] Please see Figure 1 As shown, this invention is an agricultural information management method based on a big data platform, comprising the following steps:

[0069] S1: The data processing platform deploys multiple edge computing nodes under a distributed computing architecture. Each edge computing node collects dynamic soil parameters, environmental monitoring data, and irrigation equipment operation status data of the target area in real time, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data.

[0070] S2: Build a dynamic irrigation strategy model locally on each edge computing node, generate an initial irrigation baseline based on historical data, and optimize the strategy through reinforcement learning. The optimization objectives include irrigation effect, resource consumption and soil health constraints.

[0071] S3: Establish a global federated learning framework on the cloud server, periodically aggregate the model parameters of each edge computing node, dynamically allocate node weights using an attention mechanism, and achieve privacy-preserving model fusion through secure multi-party computation.

[0072] S4: Distribute the updated global model to each edge node, combine it with real-time sensor data to generate the optimal irrigation instructions, and dynamically adjust the execution strategy;

[0073] S5: During irrigation, monitor the equipment status and soil parameter changes in real time. If an abnormality is detected, trigger the conservative irrigation mode and feed the abnormal data back to the cloud optimization model.

[0074] S6: Establish a closed-loop feedback mechanism to update the local strategy evaluation model in the short term and optimize the global model architecture in the long term, so as to achieve dynamic and precise management of agricultural irrigation.

[0075] In S1, the data processing platform deploys multiple edge computing nodes under a distributed computing architecture. Each edge computing node collects real-time soil dynamic parameters, environmental monitoring data, and irrigation equipment operating status data of the target area, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data, specifically including:

[0076] The data processing platform adopts a distributed computing architecture, consisting of a cloud control center and multiple edge computing nodes. These edge computing nodes are deployed in farmland, with each node responsible for managing agricultural data collection and processing tasks in a specific area. The node hardware uses industrial-grade embedded devices equipped with multi-protocol communication interfaces, including RS485, LoRa, and 4G / 5G, ensuring stable connections with various agricultural sensors. Each node has a built-in data buffer with a circular buffer design, providing storage capacity sufficient for 72 hours of continuous data acquisition.

[0077] The sensor network connected to the edge computing nodes includes three types of monitoring devices: soil parameter monitoring units, environmental monitoring units, and equipment status monitoring units. The soil parameter monitoring units consist of embedded sensors with measurement depths of 20cm, 40cm, and 60cm, measuring parameters including soil volumetric moisture content, temperature, and electrical conductivity. The sensors use the frequency domain reflectance principle to measure moisture content, achieving an accuracy of ±2%. The environmental monitoring unit comprises a weather station and a crop growth monitoring camera. The weather station collects data on air temperature and humidity, light intensity, wind speed, and rainfall, while the camera is equipped with a multispectral imaging module to acquire images of the crop canopy. The equipment status monitoring unit is installed on the irrigation system and includes pump current sensors, pipeline pressure sensors, and valve opening detectors.

[0078] During data acquisition, each sensor operates according to a preset sampling frequency. The soil parameter sensor collects data every 15 minutes, the environmental monitoring unit updates its readings every minute, and the equipment status monitoring unit records data synchronously with irrigation operations. Edge computing nodes read data from each sensor via polling, with the reading cycle dynamically adjusted based on the data type. For slowly changing soil parameters, data is collected at 5-minute intervals; for rapidly changing meteorological data, data is collected at 30-second intervals. This differentiated acquisition strategy ensures data timeliness while avoiding unnecessary energy consumption.

[0079] The collected raw data is first processed for format standardization. Since sensors from different manufacturers output different formats, the edge computing nodes have built-in protocol parsers to convert various types of data into a unified JSON format. During the conversion process, the units of measurement and precision information of the original data are preserved. For example, soil moisture content is uniformly converted to a volume percentage, temperature data is uniformly converted to degrees Celsius, and electrical conductivity data is converted to millisiemens per centimeter.

[0080] Spatiotemporal alignment is a crucial step in data preprocessing. Edge computing nodes attach precise timestamps and spatial coordinates to each data point. The timestamps use a GPS-synchronized network time protocol with millisecond-level accuracy. Spatial coordinates are obtained through the node's built-in BeiDou positioning module, with a positioning error of less than 2 meters. For data collected by mobile monitoring equipment, such as crop growth indicators obtained from UAV patrols, spatial interpolation algorithms are used to establish a correlation between the data and data from fixed monitoring points.

[0081] Data standardization involves three steps: outlier removal, data normalization, and feature extraction. Outlier removal employs a dynamic threshold method, setting a reasonable range based on the statistical characteristics of historical data; data exceeding this range are marked as outliers and removed. Data normalization transforms parameters of different dimensions to the [0,1] interval using a min-max normalization method. Feature extraction primarily targets time-series data, calculating statistical characteristics within a sliding window, including mean, variance, and trend. For example, for soil moisture content data, the average value and rate of change over the past 6 hours are calculated as additional features.

[0082] The processed data is stored in a local database on the edge computing nodes. The database employs a time-series database structure, optimizing the storage and query efficiency of time-series data. Data tables are partitioned according to monitoring parameter types, with each partition having both time and spatial indexes to support efficient multi-dimensional queries. Data storage utilizes compression algorithms to reduce storage space usage without sacrificing accuracy.

[0083] Data transmission employs a tiered strategy. Regular monitoring data is uploaded to the cloud in batches every hour, with data packets compressed and encrypted. Emergency alarm data, such as equipment malfunctions or sudden changes in soil parameters, immediately triggers a real-time transmission mechanism. The transmission protocol uses MQTT over TLS to ensure the security and reliability of data transmission. When network conditions are poor, edge computing nodes automatically cache data and resume transmission once network conditions are restored.

[0084] Data quality control is implemented throughout the entire acquisition and processing workflow. Edge computing nodes periodically perform sensor calibration checks, identifying potential sensor malfunctions by comparing monitoring data from neighboring nodes. Data quality assessment metrics include three dimensions: completeness, accuracy, and timeliness. These metrics are uploaded to the cloud along with the data for subsequent analysis.

[0085] Edge computing nodes possess local computing capabilities, enabling them to perform simple data analysis tasks. For example, they can calculate real-time trends in soil moisture content and issue early warnings when abnormal drops are detected. This edge intelligence reduces the computing burden on the cloud and improves system response speed. Nodes are connected via a mesh network, allowing them to maintain basic functionality even if individual nodes lose cloud connectivity, thanks to neighboring nodes.

[0086] The data processing platform adopts a modular design, allowing each functional module to be upgraded independently. Data acquisition supports hot-swapping, and new sensor types can be quickly integrated into the system via configuration files. Processing algorithms exist as plug-ins, which can be flexibly combined according to specific application scenarios. This design improves the system's adaptability and scalability.

[0087] Real-time monitoring of system operation status is a crucial measure to ensure data quality. Edge computing nodes periodically report their operational status, including CPU load, memory usage, and remaining storage space. The cloud control center comprehensively analyzes the status information of each node and dynamically adjusts data collection and processing strategies. For example, when a node is detected to be overloaded, its data collection frequency can be temporarily reduced.

[0088] Data security protection measures comprise three layers: transmission encryption, storage encryption, and access control. Data transmission employs the AES-256 encryption algorithm, data storage uses transparent encryption technology, and access control is based on role-based access control. Each edge computing node has a unique digital certificate, and all data operations are logged. These measures ensure the security of agricultural data throughout its entire lifecycle.

[0089] System maintenance supports a combination of remote operation and on-site maintenance. Routine software updates and parameter configurations can be completed remotely via the cloud. Hardware failures require on-site maintenance, and the system provides detailed fault diagnosis information to guide the maintenance work. Data collected during maintenance is not lost, and temporarily interrupted data collection will automatically re-collect key parameters upon resumption.

[0090] In S2, a dynamic irrigation strategy model is built locally on each edge computing node. An initial irrigation baseline is generated based on historical data, and the strategy is optimized through reinforcement learning. The optimization objectives include irrigation effect, resource consumption, and soil health constraints, specifically including:

[0091] The construction of the dynamic irrigation strategy model begins with the preprocessing stage of historical data. Edge computing nodes retrieve irrigation records and corresponding crop growth data from the local database over the past three years, which have already been cleaned and standardized. The historical data is organized according to the crop growth cycle, with each cycle divided into several growth stages. Data preprocessing includes filling in missing values, smoothing out anomalies, and feature engineering. Feature engineering mainly extracts three types of features: soil features (temporal variations in moisture content, temperature, and electrical conductivity), meteorological features (accumulated temperature, effective rainfall, and sunshine duration), and crop features (growth curves of plant height and leaf area index).

[0092] The initial irrigation baseline is generated using a rule-based reasoning method. The system has a built-in crop water requirement model library, containing water requirement parameters for major crops. The theoretical irrigation amount is calculated based on the currently planted crop type and its growth stage, combined with soil water holding capacity characteristics. Simultaneously, the baseline irrigation amount is adjusted by considering the three-day weather forecast provided by the local meteorological station. For example, irrigation is appropriately reduced when rainfall is forecast, and increased when high temperatures are forecast. The initial baseline includes not only the total irrigation amount but also details the duration, interval, and intensity of each irrigation session.

[0093] The reinforcement learning framework adopts the Actor-Critic architecture, which consists of a policy network and a value function network. The policy network is responsible for generating irrigation decisions, while the value function network evaluates the long-term benefits of the decisions. Network inputs include the current environmental state (soil moisture, weather conditions, crop growth) and equipment state (pump power, pipeline pressure), with outputs being irrigation control parameters (valve opening, irrigation duration). The network structure employs a design combining fully connected layers and an attention mechanism, enabling it to capture the complex relationships between different features.

[0094] The first channel of the dual-channel strategy evaluation network is dedicated to processing soil moisture data. This channel employs a deep temporal convolutional network structure, containing a stack of multiple convolutional and pooling layers. The convolutional kernels are designed as one-dimensional temporal convolutions, capable of extracting soil moisture transport characteristics at different depths. The network incorporates residual connections to prevent gradient vanishing and includes an attention mechanism to highlight changes at key time points. The input data consists of soil moisture content time-series data from the past 72 hours, with a sampling interval of 15 minutes, and the output is a dynamic feature vector of soil moisture.

[0095] The second channel processes crop growth image data, employing a multi-scale feature fusion architecture. Input data comes from visible light and multispectral images captured by field cameras, and after preprocessing, is divided into three scales. The large-scale scale captures the uniformity of growth across the entire field, the medium-scale analyzes the plant population structure, and the small-scale observes leaf morphological details. Feature maps from each scale are fused through a feature pyramid network, ultimately outputting a crop phenotypic feature vector. This channel pays particular attention to the visual characteristics of crop water stress, such as changes in leaf angle and color anomalies.

[0096] The composite reward function is designed to comprehensively consider three optimization objectives. The irrigation effect reward is calculated based on the deviation between the actual crop growth rate and the theoretical value, using an sigmoid function to map the deviation to the [0,1] interval. The resource consumption penalty is calculated based on hydropower consumption, weighted according to current energy prices and water supply constraints. The soil health constraint focuses on salinity accumulation, imposing an exponentially increasing penalty when an upward trend in salinity is detected. The three sub-rewards are combined into a total reward using learnable weighting coefficients, which are automatically adjusted according to the crop's critical growth stages.

[0097] The hierarchical experience replay mechanism improves upon traditional experience replay methods. The experience pool is divided into three priority regions: a high-value region stores decision sequences that significantly boost rewards, a median region stores ordinary decisions, and a baseline region stores random exploration samples. Sampling probabilities are assigned according to region, with samples in the high-value region having five times the probability of being sampled compared to those in the baseline region. Each sample is also assigned an importance weight, influencing gradient calculation during model updates. This mechanism accelerates the utilization of valuable experiences while maintaining sufficient exploratory nature.

[0098] Policy distillation is used to transfer trained complex models to lightweight execution networks. The teacher model is a high-capacity network trained in the cloud, while the student model is a streamlined network deployed on edge nodes. The knowledge transfer process includes two stages: output distillation and feature distillation. Output distillation allows the student model to mimic the decision distribution of the teacher model, while feature distillation allows the intermediate layers of the student model to learn the feature representations of the teacher model. The distillation loss function combines mean squared error and KL divergence to ensure the effective transfer of key knowledge. The resulting lightweight model is only one-quarter the size of the original model, and its inference speed is improved by three times.

[0099] The dynamic soil health prediction model is built upon data from an embedded sensor network. This network measures multidimensional parameters such as soil dielectric constant, redox potential, pH, and organic matter content. These parameters are collected every six hours, forming a time-series dataset. The prediction model employs a gated recurrent unit network, enabling it to capture long-term dependencies in soil parameters. Model outputs include a salinization risk index, compaction trend indicators, and biological activity levels, which constitute a soil quality early warning system. When any indicator exceeds a threshold, the system automatically adjusts irrigation strategies to mitigate soil degradation.

[0100] The resource consumption assessment method employs spatiotemporal coupled analysis. High-precision pressure sensors and flow meters are installed at key nodes of the irrigation network, with a sampling frequency of once per second. A network energy consumption model is established by calculating the spatiotemporal variations of the pressure gradient field and velocity field. An energy consumption distribution heatmap is updated at 15-minute intervals, visually displaying the system's energy loss hotspots. The assessment algorithm pays particular attention to the additional energy consumption caused by water hammer and increased local resistance; this information is used to optimize the network's operating parameters.

[0101] The irrigation effectiveness verification system integrates multi-source monitoring data. A near-ground remote sensing platform collects multispectral data of the fields weekly, calculating the normalized difference vegetation index (NDVI) and water stress index. An infrared thermal imager captures daily crop canopy temperature distribution maps to identify potential drought areas. A root system monitor measures root zone expansion dynamics and water uptake activity. These data are fused at the feature level to generate a three-dimensional assessment model of crop water use efficiency. This model quantifies spatial differences in irrigation water use efficiency, guiding the implementation of precision irrigation.

[0102] The target conflict resolution strategy employs a dynamic weight adjustment method. The system maintains a target weight matrix, recording the relative importance of each target at different growth stages. When a target conflict is detected (e.g., increasing irrigation promotes growth but exacerbates salinization), the system adjusts the weights based on the current crop growth stage and weather warning level. For example, during the crop flowering period, priority is given to ensuring growth needs, while before the rainy season, the focus is on controlling salt accumulation. The weight adjustment follows a pre-defined decision rule tree, ensuring the transparency and interpretability of the adjustment process.

[0103] The model training process employs a phased strategy. Initially, offline training is primarily conducted using historical data to establish basic decision-making capabilities. After the system goes live, online learning is implemented, with incremental training performed daily using newly collected data. A full training run is performed every weekend to recalibrate the model parameters. The training algorithm utilizes a proximal optimization strategy, which achieves a good balance between stability and sample efficiency. To prevent overfitting, the training data is divided into training, validation, and test sets in a 7:2:1 ratio.

[0104] The decision-making and execution process includes multiple security checks. Generated irrigation instructions are first tested using a simulator to predict their effects and assess potential changes in soil parameters. Then, they are compared with an expert rule base to ensure compliance with basic agronomic principles. Finally, hardware-in-the-loop testing verifies the compatibility of the instructions with the irrigation equipment. Only instructions that pass all checks are issued for execution; any problem discovered at any stage triggers a decision review process.

[0105] A comprehensive indicator system has been established for system performance monitoring. Model performance indicators include decision accuracy, reward improvement, and generalization ability test results. Implementation effectiveness indicators include irrigation uniformity, water and electricity savings, and crop yield improvement. Soil health indicators track salinity, organic matter changes, and microbial activity. These indicators generate weekly evaluation reports to guide continuous system optimization.

[0106] Edge node computing resource management employs a dynamic allocation strategy. When excessive computing load is detected, the system automatically reduces the priority of non-critical tasks and decreases the depth of historical data analysis. When memory usage exceeds a threshold, intermediate computation results are compressed and the cache is cleared. This resource management mechanism ensures that core decision-making functions remain available even with limited hardware resources.

[0107] The knowledge transfer mechanism supports experience sharing across crops. When a new crop variety is introduced, the system transfers some parameters from models of similar crops, significantly reducing the training data required for the new model. The transfer process employs feature space alignment techniques to ensure the comparability of features between different crops. This mechanism is particularly suitable for crop rotation fields, allowing for rapid adaptation to changes brought about by crop succession.

[0108] System maintenance includes regular calibration and anomaly handling. Field sensors are calibrated monthly to ensure data accuracy. When model performance monitoring detects degradation, a retraining process is automatically triggered. The system can identify and isolate abnormal data caused by hardware failures to prevent impact on overall decision-making quality. Maintenance logs record all operations in detail, supporting traceability and analysis of problems.

[0109] In S3, a global federated learning framework is established on the cloud server to periodically aggregate model parameters from various edge computing nodes. An attention mechanism is used to dynamically allocate node weights, and privacy-preserving model fusion is achieved through secure multi-party computation. Specifically, this includes:

[0110] The cloud server employs a distributed architecture to deploy the federated learning framework, which consists of three core components: a model aggregation center, a node manager, and a secure computing engine. The model aggregation center coordinates the participation of all edge nodes, maintains the global model version, and executes the model parameter aggregation algorithm. The node manager maintains metadata for all participating nodes, including geographical location, hardware configuration, and data feature statistics. The secure computing engine provides encrypted computing capabilities to ensure the security of model parameters during transmission and aggregation. The system adopts a microservice architecture, with components communicating through well-defined interfaces, supporting horizontal scaling to adapt to farm networks of different sizes.

[0111] The multi-dimensional node contribution evaluation system is constructed and quantified from three dimensions. The data quality dimension evaluates indicators including sensor calibration status, data integrity rate, and signal-to-noise ratio, normalizing each indicator to a score between 0 and 1 using a quality scoring function. The model update stability dimension tracks the convergence of recent model updates for each node, calculating the variance of parameter changes and the consistency of gradient directions. The regional environmental characteristic difference dimension analyzes the uniqueness of the node's region, including soil type distribution, microclimate characteristics, and crop variety composition. The evaluation results from the three dimensions are weighted and summed to obtain a comprehensive contribution score, with the weighting coefficients dynamically adjusted according to the current learning task objective.

[0112] The spatiotemporal attention calculation design takes into account both temporal and spatial characteristics. The temporal attention component analyzes the temporal rate of change of soil parameters at each node, assigning higher attention weights to nodes where soil moisture fluctuations exceed a threshold. The temporal attention component employs a sliding window calculation technique with a 24-hour window size and a 1-hour step size, updating the temporal importance score of each node in real time. The spatial attention component considers the geographical distribution of nodes and crop planting density, assigning higher base weights to nodes in core production areas. The calculation of spatial weights incorporates the Voronoi diagram algorithm to ensure a balanced geographical distribution. The spatiotemporal attention scores are combined using learnable fusion parameters to form the final node weights.

[0113] The dynamic weight adjustment strategy responds to environmental changes. The system integrates meteorological monitoring data; when the meteorological drought index exceeds the threshold for three consecutive days, it automatically increases the weight of nodes dependent on soil moisture data, with the adjustment magnitude proportional to the severity of drought. During the rainy season, when continuous rainfall is detected, the system pays more attention to drainage capacity data, and the weight of corresponding nodes increases incrementally according to a preset curve. The adjustment process employs a smooth transition algorithm to avoid model oscillations caused by sudden weight changes. In the event of special weather events such as typhoons or cold waves, the system temporarily freezes weight adjustments to maintain strategy stability.

[0114] A weight validation mechanism ensures the rationality of the weight allocation. The validation process employs a hold-out method, randomly selecting a subset of nodes as the validation set and the remainder as the training set. Different weight allocation schemes are tested on the validation set, and the scheme that minimizes validation loss is selected. Validation metrics include irrigation decision accuracy, resource consumption rate, and soil health change rate. The validation process is executed once per round of federated learning. For weight allocation schemes that consistently perform poorly, the system triggers an expert review process to avoid potential biases in automated decision-making.

[0115] The layered encrypted transmission protocol is designed specifically for the characteristics of the model parameters. The protocol divides the model parameters into three layers according to their sensitivity: the first layer contains the model architecture and hyperparameters, protected by an asymmetric encryption algorithm; the second layer contains the parameters of the feature extraction layer, using lightweight symmetric encryption; and the third layer contains the output layer parameters, which only undergo data obfuscation. The encryption process is completed at the edge nodes, and the cloud always receives ciphertext. The transmission channel employs double encryption: the outer layer is the standard TLS protocol, and the inner layer is a custom lightweight encryption, ensuring that even if the channel encryption is broken, the model parameters remain secure.

[0116] The distributed computing verification network consists of trusted relay nodes. These relay nodes are elected from the edge nodes participating in federated learning, with selection criteria including computing power, network stability, and security record. The verification network employs a Byzantine fault-tolerant consensus mechanism, requiring verification by more than two-thirds of the relay nodes for each model parameter update to be accepted. Verification includes checking parameter value ranges, evaluating the reasonableness of update magnitudes, and testing gradient direction consistency. Abnormal parameter updates are flagged and isolated, and the relevant nodes are added to a watchlist. Verification results are recorded in an immutable distributed ledger, supporting post-verification auditing.

[0117] The differential privacy protection mechanism achieves a balance between data availability and privacy. The noise addition algorithm employs an adaptive strategy, first analyzing the sensitivity of parameters at each layer, and then dynamically adjusting the noise intensity based on the current learning stage. Larger noise is added in the early stages of training to protect the original data, and the noise is gradually reduced as the model converges. A truncated Gaussian distribution is used for noise distribution to limit the impact of outliers. Privacy budget management uses a tree-based cumulative approach to ensure that the total privacy loss is controllable under multiple query scenarios. Each participating node has an independent privacy budget to prevent excessive contributions from individual nodes that could lead to privacy leaks.

[0118] The model watermarking tracing system assigns a digital fingerprint to each node. This fingerprint information is embedded in the low-order bits of the model parameters, without affecting model performance but still recognizable by specialized detection algorithms. The watermark contains information such as node identification, participation time, and contribution level. When malicious models or parameter leaks are detected, the responsible node can be identified through reverse tracing. The watermarking system also supports model copyright protection, preventing unauthorized use of trained models. The watermark strength is dynamically adjusted based on the node's trust level, using a more covert embedding method for high-trust nodes.

[0119] The federated learning scheduler manages the entire training process. The scheduler employs an event-driven architecture, responding to three types of events: timed trigger events, node readiness events, and anomaly alert events. The basic training cycle is set to 24 hours, which can be manually adjusted in special circumstances. Each training round consists of four phases: node selection, parameter collection, safe aggregation, and model distribution. The scheduler maintains a task state machine to ensure sequential execution of each phase, and allows for rollback or retry in case of anomalies. Training progress is visualized in real time for easy monitoring by administrators.

[0120] The node selection strategy balances efficiency and fairness. Each training round selects a subset of nodes for participation, based on criteria including: device online status, remaining computing resources, data freshness, and historical contribution. The system employs a constrained random sampling algorithm to ensure a fair opportunity for all types of nodes to participate, while prioritizing nodes that provide the greatest information gain. For selected nodes, the system estimates their training time and energy consumption to avoid selecting nodes that might drop out midway. Nodes are notified of the selection results in advance, allowing sufficient preparation time.

[0121] The parameter aggregation algorithm employs an improved weighted average method. The base weights are derived from the node importance calculated by the attention mechanism, and three adjustment factors are considered: a data quantity factor (nodes with more data have increased weight), a quality factor (nodes with higher data quality have increased weight), and a timeliness factor (newly added nodes have temporarily increased weight). The aggregation process is executed in two steps: first, local aggregation results are computed in parallel on multiple worker nodes, and then the global model is aggregated on the master node. This hierarchical aggregation reduces the computational burden on the central node and improves system scalability.

[0122] Model update distribution employs incremental transmission technology. The cloud compares the parameter differences between the old and new models, distributing only the changed parts in a package. Data packets use chunked encoding, allowing edge nodes to request missing data blocks as needed. The transmission process supports breakpoint resumption and error retransmission, adapting to unstable network environments. Key parameters are stored using multiple replicas, ensuring that updates can be obtained from neighboring nodes even if some edge nodes are temporarily offline. Detailed logs are recorded during the distribution process, including transmission rate, integrity, and timeliness metrics, to optimize subsequent distribution strategies.

[0123] An anomaly detection and handling mechanism ensures system robustness. The detector monitors three types of anomalies: abnormal node behavior (e.g., sudden changes in parameter updates), abnormal model performance (e.g., decreased validation set accuracy), and abnormal system operation (e.g., resource exhaustion). Upon detecting an anomaly, the system takes action according to predefined strategies, including: suspending the problematic node's participation, rolling back to the previous stable model version, or triggering manual intervention. Severe anomalies generate diagnostic reports containing anomaly characteristics, possible causes, and handling suggestions to accelerate troubleshooting.

[0124] The performance evaluation system comprehensively measures the effectiveness of federated learning. Evaluation metrics are divided into three categories: model quality metrics (such as test accuracy and generalization ability), system efficiency metrics (such as training time and communication overhead), and resource consumption metrics (such as computational load and energy usage). Evaluation is performed automatically after each training epoch, and the results are compared with historical baselines to identify potential performance degradation. Long-term evaluation reports are generated monthly, analyzing metric trends and guiding system optimization.

[0125] Security audits record all critical operations. Audit logs include events such as node registration and deregistration, model updates and rollbacks, and parameter access and modification. Log entries include the operator's identity, timestamp, operation content, and result status. Audit traces support multi-dimensional queries by time range, operation type, and node ID. Log data is stored encrypted using Write Once Read Many (WORM) technology to prevent tampering. Regular log analysis is performed to detect potential security threats and system abuse.

[0126] System maintenance includes regular upgrades and emergency handling. Software upgrades employ a rolling update strategy, first testing on a small number of nodes and then gradually rolling them out after stability is confirmed. Hardware maintenance plans consider the agricultural calendar to avoid operations that could impact production during critical farming seasons. Emergency fault handling plans cover scenarios such as network outages, equipment failures, and data corruption, ensuring the system maintains basic functionality even in unforeseen circumstances. Maintenance operations follow the principle of least privilege to minimize the risk of human error.

[0127] In S4, the updated global model is distributed to each edge node, and the optimal irrigation instructions are generated by combining real-time sensor data. The execution strategy is then dynamically adjusted, including:

[0128] The global model deployment process employs a differential update mechanism. The cloud server compares the differences in parameters between the old and new models, compresses and packages only the changed parts, and transmits them to the edge nodes via an encrypted channel. The transmission protocol uses reliable UDP to ensure efficient transmission even in wireless network environments. After receiving the update packet, the edge node first performs integrity verification in the verification area. Only after successful verification is the packet merged into the running model. To cope with network interruptions, nodes locally retain the three most recent model versions, allowing for automatic rollback in case of update failure. Detailed logs are recorded throughout the deployment process, including transmission time, data integrity, and merging results. These logs are periodically sent back to the cloud for optimizing the deployment strategy.

[0129] The multi-source data fusion decision engine is constructed using a layered architecture. The first layer handles spatiotemporal alignment, attaching a unified timestamp and spatial coordinates to data from different sensors. Temporal alignment is based on the local clock of the edge nodes and synchronized via the NTP protocol, achieving millisecond-level accuracy. Spatial alignment utilizes a pre-mapped field coordinate grid to interpolate discrete sensor readings into continuous field data. The second layer performs confidence assessment, assigning a confidence score between 0 and 1 to each data source based on factors such as sensor accuracy, maintenance records, and environmental interference. The third layer performs weighted fusion, employing an adaptive weighting algorithm to dynamically adjust the weights of each data source. The fusion result generates a preliminary irrigation plan, including suggested irrigation amount, duration, and execution period.

[0130] The core of the dynamic correction mechanism for irrigation strategies is a case-based reasoning system. The case library stores historical records of handled anomalies, with each case including a description of the anomaly characteristics, the measures taken, and an evaluation of the final effect. The case index uses a multi-level classification tree structure: the first level categorizes by anomaly type, the second level by severity, and the third level by environmental conditions. A matching algorithm combines nearest neighbor search and rule-based reasoning to find the several cases most similar to the current situation. When generating solutions, the timeliness of the cases is considered, with recent cases receiving higher reference weight. Newly handled anomalies are added to the case library after expert review, enabling knowledge accumulation.

[0131] The real-time irrigation effect feedback system deploys a dedicated monitoring sensor network. Moisture sensor arrays are buried at different depths in the crop root zone, with the spacing between measurement points adjusted according to soil texture; denser monitoring points are used in clay soils. The sensors employ the frequency domain reflectance principle, measuring volumetric moisture content every five minutes. Data is transmitted to edge nodes via a low-power wireless network, where the nodes calculate the vertical distribution and lateral diffusion rate of moisture. The system establishes a moisture movement model to predict the infiltration range and residence time of irrigation water; these predictions are used to adjust the intensity and location of subsequent irrigation. When uneven moisture distribution is detected, the on / off duration of each sprinkler head is automatically adjusted.

[0132] The instruction safety verification module includes a digital twin simulation environment. The simulation model integrates soil hydrodynamics, crop water uptake, and salt transport models, capable of predicting soil parameter changes within 48 hours after irrigation. The verification process employs the Monte Carlo method, performing multiple simulations to account for the uncertainties in the input data. Risk warning indicators include salt accumulation rate, water seepage depth, and root zone hypoxia probability. When any indicator exceeds a safety threshold, the system automatically generates a risk assessment report, which is submitted to an agronomist for manual review. Instructions that pass review are marked with a special tag for more intensive monitoring during execution.

[0133] The command issuance and execution employs a two-level confirmation mechanism. After the edge node generates the final irrigation command, it is first sent to the field controller for syntax and scope checks. Commands that pass the check are displayed on the controller interface, awaiting secondary confirmation from the operator. The confirmation process has a 60-second timeout; after the timeout, it automatically enters standby mode to prevent accidental execution. Particularly critical commands, such as large-volume irrigation, require an additional management password to execute. All issued commands, regardless of whether they are executed, are logged in detail, including generation time, verification results, and execution status.

[0134] A multi-level monitoring system is established to ensure effective execution. At the equipment level, vibration, temperature, and current sensors are installed on irrigation units to monitor their operational status in real time. At the field level, water flow observation points and soil sampling points are set up for manual spot checks of command execution effectiveness. At the system level, edge nodes continuously receive feedback data from each monitoring point and compare it with the expected results. If deviations exceed the allowable range, the system selects to continue observation, adjust parameters, or terminate execution according to preset strategies. Execution interruptions automatically generate fault reports to guide subsequent troubleshooting.

[0135] The dynamic adjustment strategy considers multiple time scales. Minute-level fine-tuning addresses instantaneous fluctuations in sensor data, such as sudden wind speed changes affecting sprinkler uniformity, with the system automatically compensating for pressure loss. Hourly-level adjustments respond to changes in environmental trends, such as increasing irrigation frequency due to sustained high temperatures leading to increased evapotranspiration. Daily-level optimization is based on crop growth model predictions, such as adjusting irrigation regimes after entering critical water demand periods. Adjustments at different time scales are coordinated through a priority arbitration mechanism to avoid strategy conflicts.

[0136] The error handling process distinguishes between recoverable and unrecoverable errors. For recoverable errors, such as temporary sensor malfunctions, the system switches to degraded mode, continuing operation using alternative data sources or preset experience values. For unrecoverable errors, such as a main irrigation pipe rupture, a system-level alarm is immediately triggered, all executing commands are stopped, and maintenance personnel are notified. During error handling, the system maintains core monitoring functions and records complete error context information for post-incident analysis.

[0137] The system maintenance interface provides a variety of debugging tools. The simulator tool can replay historical data and reproduce the decision-making process in specific scenarios. The parameter adjustment tool allows agronomists to fine-tune algorithm weights within safe limits to adapt to individual needs. The diagnostic tool analyzes system performance bottlenecks, such as computational latency and communication failures. All tool operations are logged in detail, and access controls are implemented to prevent accidental operation.

[0138] Performance evaluation employs a multi-dimensional indicator system. Decision quality indicators include irrigation timeliness, water quantity accuracy, and distribution uniformity. Resource efficiency indicators include energy intensity, labor input, and equipment utilization. Agronomic effectiveness indicators include crop growth, yield prediction, and soil health. Evaluation data is derived from a combination of automated monitoring systems and manual sampling, with weekly evaluation reports generated to guide continuous system optimization.

[0139] In S5, during irrigation, the system monitors equipment status and soil parameter changes in real time. If an anomaly is detected, a conservative irrigation mode is triggered, and the abnormal data is fed back to the cloud-based optimization model. Specifically, this includes:

[0140] The construction of a multi-dimensional equipment health assessment system begins with sensor deployment. A triaxial vibration sensor is installed in the water pump bearing housing, with a sampling frequency of 10kHz, to capture high-frequency vibration characteristics in the early stages of bearing wear. Dynamic pressure sensors are installed at key nodes in the water pipeline, with a sampling frequency of 100Hz, to monitor water hammer effects and pressure fluctuations. The motor is equipped with a current harmonic analyzer to determine the rotor's health status by analyzing the harmonic components in the current waveform. Sensor data is transmitted to edge computing nodes via an industrial bus, where built-in signal processing algorithms extract feature values. Vibration signal analysis employs envelope demodulation technology to separate the bearing fault characteristic frequencies. Pressure signals undergo wavelet transform to identify specific frequency band energy changes caused by pipeline anomalies. Current signals are transformed using FFT to calculate the amplitude ratio of each harmonic. These feature values ​​are input into the equipment degradation prediction model, which uses a deep residual network structure to fuse multi-source signals for comprehensive judgment.

[0141] The soil anomaly detection network employs a heterogeneous sensor architecture. Ten monitoring points are deployed per hectare of farmland, each equipped with a capacitive moisture sensor and a resistive conductivity sensor. The capacitive sensors measure at a frequency of 100MHz, accurately reflecting changes in the soil's dielectric constant. The resistive sensors utilize a four-electrode method to eliminate the influence of contact resistance. The sensor nodes are networked via a LoRa wireless network, with data transmission intervals dynamically adjusted according to irrigation status: one hour during non-irrigation periods and shortened to 15 minutes during irrigation. Each node incorporates an anomaly detection algorithm, calculating the rate of change of the ratio of dielectric constant to conductivity. An anomaly flag is triggered when this ratio deviates from the normal range by more than 20%. To further verify anomalies, the system initiates collaborative measurements with neighboring nodes, using spatial consistency analysis to eliminate false alarms caused by sensor malfunctions.

[0142] The dynamic monitoring strategy adjustment mechanism relies on meteorological data fusion. The system accesses short-term forecast data from the meteorological bureau, especially grid forecasts of precipitation probability and intensity. When the forecast indicates a heavy precipitation probability exceeding 60% within the next 3 hours, the enhanced monitoring mode is automatically activated. The soil moisture sampling frequency is increased from 15 minutes to 5 minutes, and the monitoring point density is supplemented and increased using mobile sensors. In enhanced monitoring mode, the system allocates more computing resources for data analysis and updates soil moisture infiltration model parameters in real time. Simultaneously, the data transmission strategy is adjusted, prioritizing the uploading of critical monitoring data while temporarily storing ordinary data locally. After actual precipitation occurs, the system assesses the precipitation effect based on rain gauge data and decides whether to prematurely terminate the enhanced monitoring mode.

[0143] The equipment-soil correlation analysis model is constructed using a graph neural network. The model input includes time series of equipment operating parameters and corresponding regional soil monitoring data, representing the spatial relationship between equipment nodes and soil monitoring nodes through a graph structure. The model comprises two feature extraction branches: the equipment branch analyzes pattern recognition of efficiency decline, such as blockage manifested as increased pressure and decreased flow, and wear manifested as a slow decline in efficiency; the soil branch analyzes the spatiotemporal characteristics of parameter changes, such as salinization showing changes in electrical conductivity gradients, and compaction manifested as a decrease in water infiltration rate. The features from both branches undergo cross-attention calculation in the correlation layer to uncover the potential links between equipment status and soil conditions. The model output is a joint diagnostic result, including possible problem types, severity, and suggested validation methods.

[0144] The tiered emergency response strategy is designed based on risk assessment results. Level 1 response addresses minor anomalies, such as a single sensor reading briefly exceeding the limit; the response measure is to extend the irrigation interval by 10%–20% and increase the monitoring frequency of relevant parameters. Level 2 response addresses persistent anomalies, such as multiple related sensors alarming simultaneously; the response measure is to reduce irrigation intensity by 30%–50% and activate backup monitoring equipment for verification. Level 3 response addresses severe anomalies, such as major equipment failure or large-scale soil parameter anomalies; the response measure is to immediately stop the current irrigation plan and switch to preset safe parameters. These safe parameters are set based on the crop's minimum water requirement to ensure that crops do not suffer from drought due to sudden cessation of irrigation. The response level is determined using a fuzzy logic system, considering three dimensions: anomaly duration, spatial range, and trend of change.

[0145] The case-based reasoning decision-making database is constructed using a structured representation method. Each case comprises six parts: a problem description recording the abnormal phenomenon and equipment status; an environmental context storing the weather, soil, and crop data at the time; a treatment measure section detailing the action steps taken; an execution result section evaluating the effectiveness of the measures; an experience summary extracting success factors and areas for improvement; and metadata recording the case's source and timestamp. Case retrieval uses semantic similarity calculation, transforming the current abnormal situation into a feature vector and searching for the k most similar cases in the database. During solution generation, the treatment measures of similar cases are adaptively modified, taking into account current environmental differences and changes in equipment configuration. New cases undergo a validation period; only confirmed effective treatment solutions are officially added to the database.

[0146] The core of the dynamic parameter compensation mechanism is the minimum water requirement model. Model inputs include the baseline water requirement for the current crop growth stage, real-time meteorological correction coefficients, and soil water retention capacity parameters. The baseline water requirement is derived from the crop growth model and set according to different growth stages. Meteorological corrections consider factors such as temperature, humidity, wind speed, and sunshine duration, calculating reference evapotranspiration using the Penman formula. Soil water retention capacity is set according to soil texture, with higher values ​​for clay and lower values ​​for sandy soil. During model execution, the actual evapotranspiration demand of the crop is first calculated, then the effective rainfall (if any) is subtracted, and finally adjusted based on the soil water storage capacity to obtain the minimum irrigation amount. In conservative mode, the system strictly follows the calculation results of this model to avoid over-irrigation.

[0147] The implementation of anomaly feedback optimization loop requires a special data processing workflow. Data collected in conservative mode is specially labeled, including information such as anomaly type, handling measures, and execution effects. This data is prioritized for transmission during upload, and the cloud server allocates independent storage areas and analysis channels for it. During incremental training, the sampling weight of this data is increased by 2-3 times, ensuring that the model focuses on learning from experience in handling anomalies. Simultaneously, the system generates a dedicated test set to evaluate the model's performance in similar anomaly scenarios, allowing for targeted improvements to weak points. To protect data diversity, the system controls the proportion of anomaly data in the total training set to no more than 30%, preventing the model from over-adapting to special cases.

[0148] The quality control of real-time monitoring data employs a multi-level verification mechanism. Raw sensor readings first undergo a range check to eliminate outliers that significantly exceed physical limits. Then, a consistency check is performed, comparing the correlation between neighboring sensors and different parameters. Finally, a trend rationality analysis is executed to determine whether current changes conform to known patterns. Problematic data discovered during verification is marked but not immediately discarded, and is used for subsequent analysis. For critical parameters, the system initiates redundant measurements, and the final adopted value is determined through majority voting among multiple sensors. The quality control results are fed back to the sensor health assessment module to optimize subsequent monitoring strategies.

[0149] The equipment maintenance early warning system is based on the analysis of monitoring data. The system establishes a health record for each piece of equipment, recording the service life, maintenance history, and current status of major components. Predictive maintenance algorithms analyze long-term trends in parameters such as vibration, temperature, and current, issuing early warnings when accelerated deterioration characteristics are detected. Warnings are categorized into three levels: observation, recommended maintenance, and immediate maintenance, each corresponding to different response requirements. Maintenance plans take into account the agricultural production cycle, scheduling them as much as possible during agricultural lulls. After maintenance is completed, technicians must record the maintenance content and replaced parts in the system; this information is used to optimize the equipment deterioration model.

[0150] The soil improvement recommendation module activates when persistent anomalies are detected. This module integrates soil testing data, irrigation records, and crop growth information to analyze potential causes of soil degradation. For salinization, recommendations include adjusting irrigation water quality, increasing leaching frequency, and improving drainage. For compaction, recommendations include reducing mechanical compaction, increasing organic matter, and improving tillage practices. After recommendations are generated, the system simulates and predicts the effects of the improvement measures, providing cost-benefit analyses of different options. Agronomists can refer to these recommendations to develop targeted soil management plans.

[0151] Standardized procedures are established for system calibration and maintenance. Sensor calibration is divided into two levels: field calibration and laboratory calibration. Field calibration is performed monthly, using standard reagents and simple equipment to verify basic functions. Laboratory calibration is performed annually, sending the sensors to a professional institution for comprehensive testing and calibration. Equipment maintenance is planned based on operating hours, and critical components such as water pump bearings must be inspected after 2000 hours of operation. Maintenance records are managed electronically, including maintenance content, replaced parts, and test results, forming a complete equipment health record. Calibration and maintenance plans are automatically reminded to ensure timely execution.

[0152] Anomaly data analysis reports are automatically generated. The report includes basic information such as the time of occurrence, duration, scope of impact, and handling measures. The analysis section displays the change curves of key parameters, comparisons with historical data, and possible causal inferences. The evaluation section summarizes the handling effectiveness and proposes system improvement suggestions. The report uses a structured format for easy import into other management systems. Reports of major anomalies are automatically sent to relevant personnel and require confirmation. All reports are archived and can be searched by time, type, and severity.

[0153] The cloud-based model optimization process employs a special processing channel. Data from anomaly scenarios is cleaned and labeled before entering a dedicated data pool. During training, a transfer learning strategy is used, fine-tuning the model first on a small-scale anomaly dataset before gradually integrating it into the main model. Multiple validation steps are implemented during optimization to ensure that the model's performance under normal conditions is not compromised while its anomaly handling capabilities are improved. Regression testing is performed before each update to verify the model's stability across all key scenarios. Updates utilize a canary release strategy, initially testing on a small number of nodes to confirm effectiveness before full rollout.

[0154] The system provides multiple manual intervention interfaces. For complex and abnormal situations, operators can access an expert intervention interface to view detailed diagnostic data and system recommendations. The interface offers manual parameter adjustment, allowing for overriding automated decision-making within safe limits. All manual intervention operations require double confirmation and are logged in a complete log. The system analyzes the differences between manual decisions and automated recommendations to optimize the algorithm. For scenarios requiring recurring manual intervention, optimization requirements are automatically generated, prompting the development team to focus on key improvements.

[0155] In S6, a closed-loop feedback mechanism is established to update the local strategy evaluation model in the short term and optimize the global model architecture in the long term, thereby achieving dynamic and precise management of agricultural irrigation. Specifically, this includes:

[0156] The dual-timescale model update system relies on a hierarchical computing architecture. The incremental learning module deployed on edge nodes employs a sliding window training strategy, with the window size being the monitoring data from the most recent 72 hours, updated hourly. The update process focuses on the local parameters of the strategy evaluation model, particularly the feature extraction layer related to real-time environmental response. The learning rate is set to a low value to ensure that incremental updates do not destroy existing knowledge. The model structure evolution algorithm on the cloud server uses a population-based optimization method, maintaining multiple candidate architectures and performing selection, crossover, and mutation operations weekly based on the irrigation effect evaluation results for each region. Evaluation indicators include water use efficiency, crop responsiveness, and soil health, and the optimal architecture is determined through multi-objective optimization. Updates at the two timescales are kept compatible through a parameter interface protocol, ensuring that edge nodes can correctly load the architecture updates distributed from the cloud.

[0157] The lightweight incremental learning design takes into account the limitations of edge computing resources. Model updates employ a knowledge distillation framework, using the complete model as the teacher model to generate softened labels to guide the training of the lightweight student model. Training data undergoes rigorous screening, retaining only the samples with the most information; sample importance is determined by calculating gradient magnitude. The optimization process uses stochastic gradient descent with momentum, setting the momentum coefficient to 0.9 to balance convergence speed and stability. To prevent catastrophic forgetting, the loss function includes constraint terms on historically important parameters, with the constraint strength adaptively adjusted based on parameter importance. Updated models are tested on a validation set; only after performance improvement is confirmed are they deployed to the online system; otherwise, they are rolled back to the previous version.

[0158] The model architecture evolution algorithm employs a modular design. The global model is decomposed into three functional modules: feature extraction, decision generation, and performance evaluation. Each module defines multiple candidate implementation methods. The feature extraction module can use convolutional networks, attention mechanisms, or traditional feature engineering methods. The decision generation module supports rule-based, learning-based, or hybrid strategies. Evolutionary operations include module replacement, parameter tuning, and connection modification. The evaluation phase uses cross-validation, testing each candidate architecture on historical data across multiple regions to ensure generalization ability. The evolutionary process retains elite individuals to avoid performance regression while maintaining sufficient diversity to prevent premature convergence.

[0159] The core of the feedback data tiered processing mechanism is the intelligent classification algorithm. The classification model input includes multi-dimensional features such as irrigation deviation, crop response level, soil parameter change rate, and equipment operating status. The output is a probability score for either routine or anomalous data. Routine data refers to irrigation operations performed as expected and their results. This type of data is stored using sliding window compression, retaining complete data from the most recent three months and statistical characteristics from earlier periods. Anomalous data includes situations such as equipment failure, deviation from expected results, and responses to special weather conditions. This type of data triggers an immediate processing flow: first, root cause analysis is performed; then, targeted training samples are generated; and finally, the model's emergency update channel is activated. The classification threshold is dynamically adjusted according to the system's operational phase, with a more lenient standard set immediately after the model update to collect more feedback.

[0160] The sliding window compressed storage employs a hierarchical coding technique. Recent complete data retains its original time-series characteristics, including sensor readings, control commands, and execution results. Medium-term data (1-3 months) stores statistical characteristics such as daily averages, extreme values, and trends. Long-term data (over 3 months) is further compressed into seasonal pattern characteristics. The compression process ensures the preservation of key data information, especially features relevant to model evaluation. The storage system establishes efficient indexes, supporting multi-dimensional queries by time range, crop type, and regional location. A data aging strategy automatically discards low-value historical data based on its information value, prioritizing the retention of typical scenarios and special events.

[0161] The instant model correction process is designed as an emergency response channel. When anomalous data arrives, the system first builds a mini-training set in an isolated environment, containing anomalous cases and their correct handling examples. Then, a rapid training mode is initiated, updating only the parameters relevant to the anomaly, with training epochs limited to five rounds. The updated model is first tested in a sandbox environment to verify its improvement effect in anomalous scenarios and to confirm that it will not negatively impact other scenarios. The validated model is immediately deployed to the production environment and marked as a temporary version, to be integrated into the main model during the next regular update. The entire process is typically completed within one hour, ensuring timely correction of system deviations.

[0162] The implementation of cross-seasonal knowledge transfer technology consists of three stages. The first stage involves knowledge extraction at the end of the season, separating general feature representations and domain-specific parameters from the current seasonal model. The general features are validated for cross-seasonal stability through adversarial training, ensuring that they capture the essential laws of crop growth rather than seasonal noise. The second stage prepares the transfer model before the season, using the general features as a fixed layer and training only the top seasonal adaptation layer. The adaptation layer has a high learning rate to quickly adapt to the specific conditions of the new season. The third stage involves joint fine-tuning at the beginning of the season, gradually unfreezing the underlying parameters to achieve a smooth transition from general knowledge to specific knowledge. The transfer process retains multiple checkpoints, allowing for regression to a previous state if necessary.

[0163] The model performance traceability system employs a version-based management approach. Each model update generates a complete snapshot, including model parameters, training data features, and evaluation results. The system constructs a directed graph to represent the update path, with nodes representing model versions and edges representing update operations. Each node is associated with a performance change matrix, displaying the improvement or degradation of each evaluation metric relative to the parent version. Traceability queries support multiple analytical perspectives: a timeline perspective showcasing the model's evolution; a performance perspective focusing on the optimization path of specific metrics; and a decision-making perspective reconstructing the decision-making process for key improvements. The system automatically identifies performance bottlenecks and breakthroughs, generating optimization suggestions for subsequent updates.

[0164] The construction of the optimization path graph employs a feature importance backtracking technique. For important decisions output by the model, the system traces back the key input features influencing those decisions and their evolution history. Graph nodes represent features or combinations of features, and edges represent dependency or reinforcement relationships. Node size reflects current influence, and color indicates the direction of change (increase or decrease). Graph updates are incremental, adjusting the weights of relevant nodes and edges after each model update. Analysis tools support path filtering, such as displaying only optimization paths related to water conservation or adaptation paths for specific crop types. Graph data is periodically simplified into optimization principle documents to guide adjustments in agronomic practices.

[0165] The performance evaluation subsystem enables multi-dimensional performance monitoring. Short-term evaluations are performed hourly, primarily focusing on deviations between model predictions and actual measurements, such as irrigation errors and soil moisture content deviations. Mid-term evaluations are summarized daily, calculating resource use efficiency and crop response indicators. Long-term evaluations are conducted according to growth stages, analyzing yield formation and soil change trends. Evaluation results are standardized into an index of 0-100 points for easy comparison across indicators. The system establishes a dynamic baseline mechanism; baseline values ​​are automatically adjusted based on climatic conditions and crop varieties to ensure that evaluation results reflect genuine improvements rather than environmental changes.

[0166] A model rollback mechanism ensures system reliability. A complete backup is created before each update, including model parameters, configuration files, and dependency libraries. Rollback trigger conditions include: a significant drop in evaluation metrics, operational anomalies, or user requests. The rollback process retains diagnostic information of the current state, facilitating analysis of the cause of problems. Major updates adopt a canary release strategy, first testing on select nodes and then fully rolling them out after confirming stability. The system maintains a rollback decision tree to guide recovery strategies under different failure scenarios, minimizing the impact on irrigation operations.

[0167] Please see Figure 2 As shown, the agricultural information management system based on a big data platform includes:

[0168] The distributed data acquisition and preprocessing module is used by the data processing platform to deploy multiple edge computing nodes under the distributed computing architecture. Each edge computing node collects soil dynamic parameters, environmental monitoring data and irrigation equipment operation status data of the target area in real time, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data.

[0169] The local strategy modeling and optimization module constructs a dynamic irrigation strategy model locally on each edge computing node, generates an initial irrigation baseline based on historical data, and optimizes the strategy through reinforcement learning. The optimization objectives include irrigation effect, resource consumption, and soil health constraints.

[0170] The federated learning and model aggregation module establishes a global federated learning framework on the cloud server, periodically aggregates the model parameters of each edge computing node, dynamically allocates node weights using an attention mechanism, and achieves privacy-preserving model fusion through secure multi-party computation.

[0171] The real-time decision-making and dynamic control module distributes the updated global model to each edge node, generates the optimal irrigation instruction by combining real-time sensor data, and dynamically adjusts the execution strategy.

[0172] An anomaly monitoring and emergency response module monitors the equipment status and soil parameter changes in real time during irrigation. If an anomaly is detected, a conservative irrigation mode is triggered, and the abnormal data is fed back to the cloud optimization model.

[0173] The closed-loop optimization and knowledge evolution module is used to establish a closed-loop feedback mechanism, update the local strategy evaluation model in the short term, and optimize the global model architecture in the long term to achieve dynamic and precise management of agricultural irrigation.

[0174] The working principle of this invention is as follows: This invention acquires heterogeneous data such as soil moisture, meteorological environment, and equipment operating status in real time at edge computing nodes, and performs spatiotemporal alignment and standardization processing. Based on historical data, a local irrigation strategy model is constructed, employing a dual-channel neural network to process soil moisture and crop growth data, and achieving multi-objective optimization through a composite reward function. A federated learning framework is established in the cloud, dynamically allocating node weights using an attention mechanism, and combining secure multi-party computation to achieve privacy-preserving model aggregation. An optimal irrigation command is generated through a multi-source data fusion engine, and an anomaly handling system with a three-level response mechanism is established. Finally, a closed-loop optimization mechanism with dual time scales is formed, achieving collaborative optimization of short-term incremental learning and long-term architectural evolution. The corresponding management system includes six functional modules: data acquisition, local modeling, federated learning, real-time decision-making, anomaly monitoring, and closed-loop optimization. These modules are interconnected through standardized interfaces to form a complete intelligent irrigation management closed loop. By leveraging a distributed computing architecture to decentralize data processing capabilities, using federated learning technology to ensure data privacy, and utilizing reinforcement learning algorithms to optimize irrigation strategies, the invention ultimately aims to improve irrigation accuracy, reduce resource consumption, and protect soil health. The system supports dynamic adaptation to different crops, soils, and climate conditions, and has good generalization ability and scalability.

[0175] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An agricultural information management method based on a big data platform, characterized in that: Includes the following steps: S1: The data processing platform deploys multiple edge computing nodes under a distributed computing architecture. Each edge computing node collects dynamic soil parameters, environmental monitoring data, and irrigation equipment operation status data of the target area in real time, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data. S2: Build a dynamic irrigation strategy model locally on each edge computing node, generate an initial irrigation baseline based on historical data, and optimize the strategy through reinforcement learning. The optimization objectives include irrigation effect, resource consumption and soil health constraints. S3: Establish a global federated learning framework on the cloud server, periodically aggregate the model parameters of each edge computing node, dynamically allocate node weights using an attention mechanism, and achieve privacy-preserving model fusion through secure multi-party computation. S4: Distribute the updated global model to each edge node, combine it with real-time sensor data to generate the optimal irrigation instructions, and dynamically adjust the execution strategy; S5: During irrigation, monitor the equipment status and soil parameter changes in real time. If an abnormality is detected, trigger the conservative irrigation mode and feed the abnormal data back to the cloud optimization model. S6: Establish a closed-loop feedback mechanism to update the local strategy evaluation model in the short term and optimize the global model architecture in the long term, so as to achieve dynamic and precise management of agricultural irrigation.

2. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The optimization strategy through reinforcement learning includes the following steps: A dual-channel strategy evaluation network was constructed. The first channel processes the spatiotemporal sequence data of soil moisture and uses a deep temporal convolutional network to extract soil moisture transport features. The second channel processes crop growth image data and captures the phenotypic change patterns of crops through multi-scale feature fusion. The design incorporates a composite reward function, quantifying irrigation effect as the deviation coefficient between crop growth rate and theoretical value, converting resource consumption into a normalized index of hydropower energy consumption, and reflecting soil health constraints as the reciprocal of the salt accumulation rate. A hierarchical experience replay mechanism is used to store training samples, prioritizing the retention of decision sequences with significant policy improvements, and weighted sampling is performed according to sample importance during model updates; By introducing strategy distillation technology, the knowledge of complex decision-making networks is transferred to lightweight execution networks, ensuring the real-time response capability of edge devices.

3. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The specific implementation of the optimization objective includes: Establish a dynamic prediction model for soil health, collect multidimensional parameters such as soil dielectric constant and redox potential through an embedded sensor network, and construct an early warning index system for soil quality degradation. Design a spatiotemporal coupled assessment method for resource consumption, deploy pressure fluctuation sensors at key nodes of the irrigation network, and establish a heat map of network energy consumption distribution by combining water flow velocity measurement; Develop a multimodal verification system for irrigation effectiveness, integrating near-ground remote sensing spectral data, leaf temperature distribution data, and root growth monitoring data to construct a three-dimensional evaluation model for crop water use efficiency; Implement a strategy to resolve conflicting objectives. When different optimization objectives conflict, dynamically adjust the weight allocation scheme based on the current crop growth stage and the meteorological warning level.

4. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The dynamic allocation of node weights using an attention mechanism includes the following steps: A multi-dimensional node contribution evaluation system is constructed, which comprehensively considers three dimensions: data quality indicators of each edge node, model update stability, and differences in regional environmental characteristics. The design of the spatiotemporal attention calculation assigns higher temporal attention to nodes with drastic changes in soil parameters and higher spatial attention weight to nodes with crop planting area exceeding a preset threshold. A dynamic weighting adjustment strategy is implemented to automatically increase the weighting coefficient of soil moisture data during the dry season and increase the weighting ratio of drainage capacity assessment during the rainy season. Establish a weighted verification mechanism and use cross-validation to ensure that the attention allocation results match the actual irrigation needs of each region.

5. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The privacy-preserving model fusion achieved through secure multi-party computation includes the following steps: Design a layered encrypted transmission protocol to perform gradient-level segmented encryption processing on model parameters; Construct a distributed computing verification network, in which multiple relay nodes jointly complete the compliance verification of model parameters; A differential privacy protection mechanism is implemented, adding adaptive noise conforming to a Gaussian distribution during the model aggregation stage. The noise intensity is negatively correlated with data sensitivity. Establish a model watermarking traceability system to assign a unique digital fingerprint to each node participating in federated learning.

6. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The process of generating optimal irrigation instructions and dynamically adjusting execution strategies by combining real-time sensor data includes the following steps: A multi-source data fusion decision engine is established to align the irrigation baseline value output by the global model with real-time collected data on soil moisture content, crop stem micro-changes, and weather station forecasts in a spatiotemporal manner and weight the data with confidence levels to generate a preliminary irrigation plan. The design incorporates a dynamic correction mechanism for irrigation strategies. When a sudden change in soil parameters or equipment malfunction is detected, an emergency strategy library based on case-based reasoning is activated to match the optimal adjustment scheme under similar historical conditions. Implement a real-time feedback system for irrigation effects, using a network of miniature sensors buried in the crop root zone to monitor water infiltration rate and distribution uniformity, and dynamically adjust the intensity and duration of sprinkler irrigation. A command security verification system is built, which simulates the soil parameter change trend after execution before the command is issued, and automatically triggers the manual review process when a risk of salinization is predicted.

7. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The real-time monitoring of equipment status and soil parameter changes includes the following steps: A multi-dimensional equipment health assessment system was constructed. The time-domain characteristic waveform of the water pump bearing was collected by vibration sensor, and the frequency domain analysis results of pipeline pressure sensor were combined to establish a predictive model for equipment deterioration trend. Design a soil anomaly detection network and deploy a capacitive-resistive dual-mode sensor array in the irrigation area. By measuring the synergistic change characteristics of soil dielectric constant and conductivity, identify areas of abnormal salt accumulation. Implement a dynamic monitoring strategy adjustment mechanism. When the weather forecast shows that the probability of heavy precipitation exceeds the threshold, automatically increase the sampling frequency of soil moisture content and expand the density of monitoring points. Establish an equipment-soil correlation analysis model. When a decline in irrigation equipment efficiency is detected, simultaneously analyze the historical change curves of soil parameters in the corresponding area to identify potential blockage or salinization problems.

8. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The process of triggering the conservative irrigation mode includes the following steps: The design of a tiered emergency response strategy divides the conservative mode into three levels according to the severity of the anomaly: Level 1 adjusts irrigation duration, Level 2 reduces irrigation intensity, and Level 3 switches to preset safety parameters. Build a case-based reasoning and decision-making database to store the handling solutions and effectiveness evaluations of historical abnormal events. When a new abnormality occurs, match the optimized handling solution of the most similar case. Implement a dynamic parameter compensation mechanism to dynamically calculate the minimum required irrigation amount based on real-time meteorological data and crop growth stage under conservative mode; Establish an anomaly feedback optimization loop, and mark the data collected during the execution of the conservative mode as high-value samples, which will be given priority for incremental training of the cloud model.

9. The agricultural information management method based on a big data platform according to claim 1, characterized in that, The establishment of the closed-loop feedback mechanism includes the following steps: A dual-timescale model update system is constructed, with lightweight incremental learning deployed at edge nodes. The strategy is fine-tuned and the local parameters of the model are evaluated every hour based on the latest sensor data. A model structure evolution algorithm is deployed on a cloud server, and the irrigation effect of each region is comprehensively evaluated and the global model architecture is reconstructed every week. The design of the feedback data hierarchical processing mechanism divides the irrigation execution results into two categories: routine operation data and abnormal event data. Routine data is stored using sliding window compression, while abnormal data triggers an immediate model correction process. Implement cross-seasonal knowledge transfer technology, extract the feature patterns of the previous quarter as prior knowledge and inject them into the new model when the crop growth cycle changes, to accelerate training convergence; Establish a model performance traceability system to record parameter changes and corresponding effect improvements with each update, forming an interpretable optimization path map to provide a basis for decision-making in subsequent adjustments.

10. An agricultural information management system based on a big data platform, characterized in that: The agricultural information management method based on a big data platform as described in any one of claims 1-9 includes: The distributed data acquisition and preprocessing module is used by the data processing platform to deploy multiple edge computing nodes under the distributed computing architecture. Each edge computing node collects soil dynamic parameters, environmental monitoring data and irrigation equipment operation status data of the target area in real time, and performs spatiotemporal alignment and standardization processing on the collected heterogeneous data. The local strategy modeling and optimization module constructs a dynamic irrigation strategy model locally on each edge computing node, generates an initial irrigation baseline based on historical data, and optimizes the strategy through reinforcement learning. The optimization objectives include irrigation effect, resource consumption, and soil health constraints. The federated learning and model aggregation module establishes a global federated learning framework on the cloud server, periodically aggregates the model parameters of each edge computing node, dynamically allocates node weights using an attention mechanism, and achieves privacy-preserving model fusion through secure multi-party computation. The real-time decision-making and dynamic control module distributes the updated global model to each edge node, generates the optimal irrigation instruction by combining real-time sensor data, and dynamically adjusts the execution strategy. An anomaly monitoring and emergency response module monitors the equipment status and soil parameter changes in real time during irrigation. If an anomaly is detected, a conservative irrigation mode is triggered, and the abnormal data is fed back to the cloud optimization model. The closed-loop optimization and knowledge evolution module is used to establish a closed-loop feedback mechanism, update the local strategy evaluation model in the short term, and optimize the global model architecture in the long term to achieve dynamic and precise management of agricultural irrigation.

Images