Agricultural production full-process digital management method and system

Abstract

The invention relates to the technical field of production management, and discloses an agricultural production whole-process digital management method and system, and the method comprises the steps: collecting multi-source heterogeneous data before production, during production and after production, constructing a space-time aligned production management unit, building a directed dependency graph based on a farm activity time sequence chain and an agricultural rule, and carrying out the construction of a digital management unit. And a decision scheme of a next operation link is generated through multi-task collaborative prediction by using the graph attention network coding process state characteristics, and finally a scheduling instruction is issued to realize closed-loop intelligent management and control. According to the invention, linkage of data fusion, process modeling and intelligent decision-making in the whole period of agricultural production can be realized.

Description

Technical Field

[0001] This invention relates to the field of production management technology, and more specifically, to a digital management method and system for the entire agricultural production process. Background Technology

[0002] With the rapid development of smart agriculture, digital management of the entire agricultural production process has become a key direction for improving agricultural efficiency, resource utilization, and sustainable development capabilities. While some existing systems attempt to optimize specific stages of agricultural production through data fusion and intelligent algorithms, they still have significant limitations in terms of coverage, system integration, and end-to-end collaborative capabilities. Although traditional digital agricultural management methods have made some progress in data collection for single production stages, local model building, and decision support in specific scenarios, they have not yet systematically solved core problems such as seamless data integration across multiple stages (pre-production planning, production execution, and post-production distribution), efficient collaboration between heterogeneous systems, flexible adaptation to a generalized platform architecture, and cross-business domain information fusion and intelligent linkage. Therefore, they cannot fully meet the integrated management needs of different crop types, production models, and regional environments.

[0003] Therefore, how to integrate multi-source data throughout the entire agricultural production cycle, connect the business logic of each stage before, during, and after production, achieve efficient collaboration among multiple systems, and possess good scalability and adaptability for full-process digital management methods and systems has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a digital management method and system for the entire agricultural production process, which solves the technical problems in the prior art of not being able to systematically achieve seamless data connection, efficient collaboration of heterogeneous systems, and integrated intelligent control of the entire process across multiple stages such as pre-production planning, production execution, and post-production circulation.

[0005] This invention provides a method and system for digital management of the entire agricultural production process, comprising: Firstly, a digital management method for the entire agricultural production process includes the following steps: Collect multi-source heterogeneous agricultural data covering all stages of production, including pre-production preparation, field management during production, and post-production harvesting and processing, to generate an agricultural production dataset. The agricultural production dataset includes plot boundary vectors, real-time meteorological monitoring data, soil nutrient detection indicators, crop sowing operation records, precision irrigation execution logs, field survey images of pests and diseases, measured yield data by region, and agricultural product circulation traceability information. Upon receiving crop planting information for the target agricultural production area, the agricultural production dataset is spatiotemporally correlated and matched according to production time sequence and plot spatial location. Based on crop variety characteristics, planting system type, and agronomic management model, the matched data is divided into several production management units to determine the process management baseline unit. Specifically, this includes: upon receiving crop planting information for the target agricultural production area, the agricultural production dataset is spatiotemporally correlated and matched; wherein the agricultural production dataset includes plot boundary vectors, real-time meteorological monitoring data, soil nutrient testing indicators, crop sowing operation records, precision irrigation execution logs, field survey images of pests and diseases, measured yield data for different zones, and agricultural product circulation traceability information. Based on the timestamp and spatial coordinates of each data record in the agricultural production dataset, meteorological monitoring, soil testing, operation records and plot boundary vectors are associated and mapped. The matched data are divided into units according to crop variety characteristics, planting system type and agronomic management mode. Based on the crop variety and planting scale of each plot in the target agricultural production area, several production management units of the target agricultural production area are determined as process management benchmark units. Within the production management unit, key nodes of agricultural operations and their execution parameters are extracted. Based on the temporal sequence and operation type of the agricultural operations, a time-series chain of agricultural activities is constructed, forming a standardized sequence of event parameter tuples. Specifically, this includes: extracting key nodes of agricultural operations from crop sowing operation records, precision irrigation execution logs, and agricultural product circulation traceability information within the production management unit; constructing a time-series chain of agricultural activities based on the execution time and parameters of each key node, and according to the position and dependencies of the operation in the agricultural production process; and forming a standardized sequence of event parameter tuples based on the operation type, timestamp, plot range, equipment number, and parameter set of each agricultural operation node in the time-series chain. Based on the event parameter tuple sequence and the agronomic dependency rule knowledge base, a directed dependency graph of the agricultural production process is constructed using a graph structure modeling method. Specifically, this includes: extracting each agricultural operation event from the event parameter tuple sequence and the operation dependency relationship from the agronomic dependency rule knowledge base; using each agricultural operation event as a graph node and the agronomic dependency relationship between operations as graph edges; and constructing a directed dependency graph of the agricultural production process using a graph structure modeling method based on the set of constraints for preceding and following processes determined by the agronomic dependency rules and the set of edge weight functions determined by the operation time window constraints. Using the node features and edge weights in the directed dependency graph as the input space, and historical production experience and current execution progress as the state representation targets, a deep learning model based on graph attention mechanism is established. The optimal network parameters are iterated based on the feature encoding quality. The graph structure deep learning model is used to encode the features of the directed dependency graph to generate process state feature vectors. Based on the process state feature vector and the multi-task decision network, a collaborative prediction algorithm is used to solve the initial decision scheme for the next operation stage. Specifically, this includes: extracting production process state information and historical operation execution modes from the process state feature vector; using the process state feature vector as input, and based on the multi-task objective function set determined by agricultural operation type classification constraints, input application rate regression constraints, and risk warning level identification constraints, the collaborative prediction algorithm in the intelligent decision model is used to solve the recommended agricultural operation type, recommended agricultural input application rate, and production risk warning level for the next operation stage as the initial decision scheme. Based on the recommended agricultural operation type, the recommended agricultural input application rate, the production risk warning level, and the land space information of the production management unit, an operation scheduling instruction containing operation parameters and execution instructions is generated. The operation scheduling instruction is then sent to the corresponding execution equipment through the agricultural Internet of Things platform to achieve intelligent management and control of the agricultural production process.

[0006] Furthermore, the agricultural production dataset is subjected to spatiotemporal correlation matching based on production time sequence and plot spatial location. The matched agricultural production dataset is then divided into several production management units, including: Collect multi-source heterogeneous agricultural data covering all stages of production, including pre-production preparation, field management during production, and post-production harvesting and processing, and generate an agricultural production dataset. Farmland plot boundary coordinates are extracted from the plot boundary vectors in the agricultural production dataset. The coordinate accuracy is corrected by a positioning terminal, and standardized farmland plot vector data is generated by combining the intelligent interpretation results of satellite remote sensing images. The coordinates of real-time meteorological monitoring data, soil nutrient detection indicators, field survey images of pests and diseases, and crop sowing operation records in the agricultural production dataset are unified. Field meteorological monitoring stations, soil moisture sensor networks, UAV agricultural inspection images, and agricultural technician operation records are uniformly mapped to the geographic coordinate reference system. Spatial units are aggregated according to grid management requirements to generate spatiotemporally aligned agricultural data. Based on the spatiotemporally aligned agricultural data, production management unit division criteria are set according to crop growth and development patterns, field rotation and cropping systems, and cultivation density standards to ensure that crop growth characteristics, environmental response patterns, and agronomic measures are consistent within the same production management unit. Based on the crop varieties and planting scale of each plot in the target agricultural production area, a globally unique identifier is assigned to each production management unit, the production management unit is identified as the baseline unit for process management, and the globally unique identifier is linked to a multi-year production record database.

[0007] Furthermore, key nodes and execution parameters of agricultural operations are extracted within the production management unit. Based on the temporal sequence and operation type of the agricultural operations, a time-series chain of agricultural activities is constructed, forming a standardized sequence of event parameter tuples, including: Extract key agricultural operation nodes from crop sowing operation records, precision irrigation execution logs, and agricultural product circulation traceability information within the production management unit; Based on the execution time and execution parameters of each key node of the agricultural operation, identify the operation time, the personnel or machinery performing the operation, the target plot of the operation, and the quantitative parameter information of the operation. Based on the position and dependencies of the agricultural operation in the agricultural production process, each key node of the agricultural operation is converted into a standardized agricultural operation node data structure. Arrange all the agricultural operation nodes within the same production management unit according to the agricultural calendar order to construct an agricultural activity time sequence chain, forming a sequence of event parameter tuples that reflects the complete production process.

[0008] Furthermore, based on the event parameter tuple sequence and the agronomic dependency rule knowledge base, a directed dependency graph of the agricultural production process is constructed, including: Extract each agricultural operation event from the event parameter tuple sequence and the operation dependency relationship from the agronomic dependency rule knowledge base; Using each agricultural operation event as a graph node, according to the preceding and following process constraints defined in the agronomic dependency rule knowledge base, if agricultural operation event B must be executed within the appropriate agricultural time window after agricultural operation event A is completed, then a directed dependency edge is established between the graph node corresponding to agricultural operation event A and the graph node corresponding to agricultural operation event B. Using the agronomic dependencies between operations as graph edges, and based on the edge weight function determined by the operation time window constraint, the edge weight value of the directed dependent edge is set to the allowable interval days. A virtual node for the start of a production cycle and a virtual node for the end of a production cycle are introduced. The virtual node for the start of a production cycle is connected to all starting operation diagram nodes without prior agricultural operations, and the virtual node for the end of a production cycle is connected to all finishing operation diagram nodes without subsequent agricultural operations. A graph structure modeling method is used to construct a complete directed dependency graph of agricultural production process, and the topological orderliness of the directed dependency graph structure is checked to identify and eliminate cyclic dependency paths that violate crop growth patterns or agronomic technical regulations.

[0009] Furthermore, using the graph node features and edge weights of the directed dependency graph as the input space, and historical production experience and current execution progress as state representation objectives, a graph attention model is constructed. Using feature encoding quality as the iteratively optimal network parameter, the graph attention model is used to encode features in the directed dependency graph, generating a process state feature vector that integrates historical production experience and current execution progress, including: A graph attention mechanism network is used as a feature extractor to construct initial node features for each graph node in the directed dependency graph. The initial node features are composed of the operation type encoding of the corresponding agricultural operation event, the standardized vector of the parameter set, and the combination of the timestamp and the spatiotemporal coordinate position encoding of the plot range. The initial features of the nodes are input into the graph attention mechanism network. The association information of adjacent graph nodes is aggregated along the directed dependency edges through a multi-layer graph attention propagation mechanism, and the hidden state representation of each graph node is updated layer by layer. The hidden state representations of each graph node in the directed dependency graph are aggregated using a time-weighted aggregation method. The aggregation weights are calculated and determined by the time decay function of the current decision time and the timestamps of each agricultural operation event. Based on the quality of feature encoding, the optimal network parameters are iterated, and the timeliness-weighted aggregation result is output as a process status feature vector that integrates historical production experience and current execution progress.

[0010] Furthermore, the process state feature vector is input into an intelligent decision-making model containing a multi-task decision network to output an initial decision scheme for the next operation, including: Extract the production process status information and historical job execution patterns from the process status feature vector; Using the process state feature vector as input, and based on the multi-task objective function set determined by agricultural operation type classification constraints, input application amount regression constraints, and risk warning level identification constraints, an intelligent decision-making model is constructed, which includes an operation recommendation sub-network, an input application amount prediction sub-network, and a production risk warning sub-network. Each sub-network shares the underlying feature extraction layer to achieve production knowledge collaboration. The process state feature vector is input into the job recommendation subnetwork, and the job recommendation subnetwork outputs the probability distribution of the next suitable agricultural operation type, using a multi-classification decision mechanism. The process state feature vector is input into the input application rate prediction subnetwork. The input application rate prediction subnetwork regressively calculates the application rate of agricultural inputs, including irrigation water volume, fertilizer application rate, and pesticide dosage. Non-negative constraint regression is used and an agronomic safety upper limit is set. The process status feature vector is input into the production risk early warning sub-network, and the production risk early warning sub-network outputs the probability of occurrence of crop pest and disease risk level, drought stress risk level, and lodging disaster risk level, using a multi-label classification mechanism; The collaborative prediction algorithm in the intelligent decision-making model is used to obtain an initial decision scheme for the next operation, which includes the recommended agricultural operation type, the recommended amount of agricultural inputs, and the production risk warning level.

[0011] Furthermore, based on the initial decision-making scheme, a work scheduling instruction containing work parameters and execution instructions is generated. This instruction is then distributed to the corresponding execution devices via the agricultural IoT platform to perform intelligent management and control operations of the agricultural production process, including: The recommended agricultural operation type, the recommended agricultural input application rate, and the plot spatial information of the production management unit are combined to generate an agricultural operation plan containing operation parameters. The agricultural operation plan is converted into an operation scheduling instruction containing execution instructions, using an electronic task sheet format that conforms to agricultural machinery operation standards; The operation scheduling instructions are pushed to field edge computing nodes or agricultural machinery control terminals through the IoT communication protocol of the agricultural IoT platform. After parsing the operation scheduling instruction, the field edge computing node determines the corresponding execution equipment type based on the recommended agricultural operation type, and sends precise operation control commands to the intelligent tractor and plant protection machinery through the agricultural machinery bus interface, or adjusts the opening of the water and fertilizer integration irrigation system's proportion valve and the start / stop status of the irrigation pump station according to the recommended agricultural input application amount through the industrial communication protocol. The system collects real-time operation execution feedback data from the intelligent agricultural machinery, the integrated water and fertilizer irrigation system, or the post-harvest processing equipment, converts the operation execution feedback data into new agricultural operation events, and synchronously updates the event parameter tuple sequence to perform intelligent control operations on the agricultural production process.

[0012] Furthermore, the intelligent decision-making model based on historical production data continuously optimizes agronomic-dependent rules, including: Collect the event parameter tuple sequence of each process management benchmark unit in multiple production cycles, the operation execution feedback data, crop growth monitoring data and final yield and quality indicators to construct a data sample set of historical production; The data sample set is effectively labeled with agronomic techniques. By comparing and analyzing the deviation between the actual yield and quality indicators and the expected targets in the data sample set, efficient and inefficient production modes are identified. Key influencing factors are extracted from the event parameter tuple sequences corresponding to the efficient and inefficient production modes. The intelligent decision-making model is trained using a contrastive learning strategy, which widens the distance between the process state feature vectors corresponding to the efficient production mode and the process state feature vectors corresponding to the inefficient production mode in the feature space, thereby enhancing the intelligent decision-making model's ability to identify high-quality agricultural decisions and generating a trained and optimized intelligent decision-making model. High-weight directed dependency edges and their corresponding edge weight values ​​are extracted from the trained and optimized intelligent decision model. The graph node pairs connected by the high-weight directed dependency edges and the key time window parameters recorded by the edge weight values ​​are used to form agronomic dependency rules. The incremental update of the trained and optimized intelligent decision-making model is triggered based on the newly added task execution feedback data. The agronomic dependency rules are retained by the transfer learning mechanism, while adapting to local climate change, variety updates and cultivation technology improvements. The agronomic dependency rules obtained from mining are integrated with the experience of traditional agricultural experts to update the agronomic dependency rule knowledge base, establish a human-machine collaborative agricultural production decision-making knowledge graph, and support agricultural technicians in reviewing and revising the initial decision-making scheme. The agronomic dependency rules obtained from mining are integrated with the experience of traditional agricultural experts to update the agronomic dependency rule knowledge base, establish a human-machine collaborative agricultural production decision-making knowledge graph, and support agricultural technicians in reviewing and revising the initial decision-making scheme.

[0013] Furthermore, coordinated scheduling and optimal allocation of agricultural resources include: Extract the initial decision schemes corresponding to each process management benchmark unit within the collaborative management plot set, obtain the recommended agricultural operation type, recommended agricultural input application amount, and production risk warning level from the initial decision schemes, align the agricultural operation requirements of each benchmark unit within the collaborative management plot set by time and prioritize them, and comprehensively consider the operation time window constraints in the directed dependency graph, the sensitivity of the critical period of crop growth, and the disaster warning information in the production risk warning level to generate a collaborative operation requirement sequence. Based on the collaborative operation demand sequence, a cross-plot resource sharing constraint model is established, setting an upper limit on the number of schedulable execution equipment, a constraint on the total water supply capacity of the integrated water and fertilizer irrigation system, a limit on manpower allocation capacity, and a requirement for logistics and transportation timeliness. Using the collaborative operation demand sequence and the cross-plot resource sharing constraint model as input, a multi-objective optimization algorithm is used to solve the global operation scheduling scheme. The optimization objectives include minimizing the overall operation completion time, maximizing the utilization rate of the execution equipment, minimizing the transportation cost of the recommended agricultural input application amount, and minimizing the number of conflicts in the critical agricultural time window. Based on the global operation scheduling scheme, the execution order of recommended agricultural operation types and the allocation scheme of recommended agricultural input application amount of each process management benchmark unit in the collaborative management plot set are dynamically adjusted to generate a cross-plot collaborative operation plan table containing operation parameters and execution instructions. According to the cross-plot collaborative operation plan, the operation execution feedback data and emergencies of each process management benchmark unit in the collaborative management plot set are monitored in real time. When there is a failure of the execution equipment, extreme weather, or an increase in the risk of pests and diseases in the production risk warning level, the dynamic rescheduling mechanism is triggered. The updated collaborative operation demand sequence and the cross-plot resource sharing constraint model are re-input into the multi-objective optimization algorithm to solve the updated global operation scheduling scheme, giving priority to ensuring the resource supply of key plots and key production links. The statistical analysis of the collaborative scheduling effect of the global operation scheduling scheme and the updated global operation scheduling scheme is used to evaluate the improvement in the utilization efficiency of the execution equipment, the reduction in transportation costs of the recommended agricultural input application amount, and the improvement in yield and quality, thereby generating a regional agricultural production collaborative management performance report.

[0014] Secondly, a digital management system for the entire agricultural production process includes: Data acquisition module: used to collect multi-source heterogeneous agricultural data during the agricultural production stage and generate agricultural production datasets; Production Management Module: Used to perform spatiotemporal correlation matching of the agricultural production dataset according to the production time sequence and the spatial location of the plots, and to divide the matched agricultural production dataset into several production management units; Model building module: This module extracts key nodes and execution parameters of agricultural operations within the production management unit, constructs a time-series chain of agricultural activities based on the temporal order and type of the operations, forming a standardized sequence of event parameter tuples. Based on the sequence of event parameter tuples and a knowledge base of agronomic dependency rules, it constructs a directed dependency graph of the agricultural production process. Nodes in the directed dependency graph represent key nodes of the agricultural operations, and edges represent agronomic dependencies and time window constraints between preceding and subsequent processes. Using the graph node features and edge weights of the directed dependency edges as the input space, and historical production experience and current execution progress as state representation targets, a graph attention model is constructed. Vector generation module: Used to iteratively optimize network parameters by using the feature encoding quality to encode features of the directed dependency graph through the graph attention model, and generate a process state feature vector that integrates historical production experience and current execution progress; Intelligent decision-making module: It is used to input the process state feature vector into an intelligent decision-making model containing a multi-task decision network to obtain an initial decision scheme for the next operation; generate an operation scheduling instruction containing operation parameters and execution instructions according to the initial decision scheme; and send the operation scheduling instruction to the corresponding execution device through the agricultural Internet of Things platform to perform intelligent management and control operations of the agricultural production process.

[0015] The beneficial effects of this invention are as follows: By collecting multi-source heterogeneous agricultural data covering all stages of agricultural production—pre-production preparation, field management during production, and post-production harvesting and processing—and performing spatiotemporal correlation matching, this invention achieves systematic integration of multi-source data across the entire agricultural production cycle and effective connection of business logic in each stage of pre-production, production, and post-production. By constructing a directed dependency graph of the agricultural production process based on an agronomic dependency rule knowledge base and using a graph structure deep learning model for feature encoding, this invention systematically represents the complex dependencies between agricultural operations, solving the problem that traditional methods suffer from static agricultural logic and cannot adapt to dynamic production scenarios. Furthermore, through a multi-task decision network… The network collaborative output recommends agricultural operation types, recommended agricultural input application rates, and production risk warning levels, and automatically generates operation scheduling instructions to be sent to the execution equipment to form a closed-loop management system, realizing efficient collaboration and intelligent linkage among multiple systems. Through continuous optimization of the intelligent decision-making model based on historical production data and the agronomic dependency rule mining mechanism, as well as cross-plot collaborative scheduling and agricultural resource optimization allocation functions, the system has good scalability and adaptability, and can fully meet the needs of integrated management of the entire process under different crop types, production modes, and regional environments, effectively improving agricultural production efficiency, resource utilization, and decision response speed. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the steps of a digital management method for the entire agricultural production process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a digital management system module for the entire agricultural production process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the data acquisition steps provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the digital management method for the entire agricultural production process provided in the embodiments of the present invention. Detailed Implementation

[0017] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses a method and system for digital management of the entire agricultural production process, comprising: like Figure 1 , Figure 3 and Figure 4 As shown, a digital management method for the entire agricultural production process includes the following steps: Step 1: Collect multi-source heterogeneous agricultural data from various stages of agricultural production to generate an agricultural production dataset; Step 2: Perform spatiotemporal correlation matching on the agricultural production dataset according to the production time sequence and the spatial location of the plots, and divide the matched agricultural production dataset into several production management units; Step 3: Extract key nodes and execution parameters of agricultural operations within the production management unit. Construct a time sequence chain of agricultural activities based on the time sequence and operation type of the agricultural operations, forming a standardized sequence of event parameter tuples. Based on the sequence of event parameter tuples and the agronomic dependency rule knowledge base, construct a directed dependency graph of the agricultural production process. Nodes in the directed dependency graph represent key nodes of the agricultural operations, and edges represent agronomic dependencies and operation time window constraints between preceding and subsequent processes. Using the graph node features and edge weights of the directed dependency edges in the directed dependency graph as the input space, and historical production experience and current execution progress as the state representation targets, construct a graph attention model. Step 4: Using the feature encoding quality as the iterative optimal network parameter, the directed dependency graph is feature encoded through the graph attention model to generate a process state feature vector that integrates historical production experience and current execution progress. Step 5: Input the process state feature vector into the intelligent decision model containing a multi-task decision network to obtain the initial decision scheme for the next operation; generate an operation scheduling instruction containing operation parameters and execution instructions according to the initial decision scheme, and send the operation scheduling instruction to the corresponding execution device through the agricultural Internet of Things platform to perform intelligent management and control of the agricultural production process.

[0019] In this embodiment, the target agricultural production area is a large grain production base located in the North China Plain. The main crop is winter wheat, and a single-crop-per-year rotation system is adopted. The total area of ​​the land is approximately 5,000 mu (about 333 hectares), divided into 200 standard farmland plots, each with an area of ​​approximately 25 mu (about 1 hectares). The base has deployed sensing equipment such as Beidou satellite navigation and positioning terminals, soil moisture sensor arrays, meteorological monitoring stations, UAV remote sensing platforms, and agricultural product traceability label readers and writers, and is connected to an agricultural Internet of Things (IoT) platform, possessing the basic conditions for executing intelligent operation scheduling.

[0020] First, the data acquisition module starts operating, acquiring the boundary coordinate sequence of each farmland plot through BeiDou satellite navigation and positioning terminals deployed in the farmland, forming the initial plot boundary vector. Simultaneously, a soil moisture sensor array collects data on soil moisture content, conductivity, pH, and nitrogen, phosphorus, and potassium nutrient content every 10 minutes. A meteorological monitoring station records real-time temperature, humidity, wind speed, precipitation, and solar radiation intensity. A UAV remote sensing platform performs a weekly flight mission, carrying a multispectral camera to acquire normalized vegetation index (NVR) images of the crop canopy and simultaneously capture high-resolution visible light images of suspected pest and disease areas. Agricultural product traceability label readers automatically read or write electronic tag information with timestamps and operation details during sowing, fertilization, spraying, harvesting, and warehousing. All the above raw data are timestamped at the time of acquisition, and obtained geographical coordinates in the World Geodetic Coordinate System (WGC) through the built-in satellite navigation module or post-processing differential correction. These coordinates are then uniformly converted to the National Geodetic Coordinate System 2000, completing coordinate correction and projection transformation to ensure all data records have a consistent spatial reference frame. For missing data items, such as when a weather station fails to upload data due to a power outage on a certain day, the data is filled in by interpolation between adjacent stations combined with the historical average for the same period. For outliers, such as soil sensors outputting values ​​outside the reasonable range due to malfunctions, they are removed using a sliding window standard deviation detection algorithm. All physical quantity units are uniformly converted to international standard units (such as kg, m³, ℃, etc.), ultimately forming a structured agricultural production dataset, which is stored in a distributed time-series database.

[0021] Upon receiving the winter wheat planting plan information for the target area (including the variety Jimai 22, the sowing date of October 10th, and the target yield of 600 kg / mu), the unit partitioning module begins its work. This module first extracts all plot boundary vectors from the agricultural production dataset and uses a high-precision BeiDou satellite navigation and positioning terminal to perform centimeter-level correction on the boundary coordinates. Then, combining this with a farmland mask map generated from Sentinel-2 satellite imagery interpreted by a deep learning semantic segmentation model, it performs topological correction and closure processing on the plot boundaries, generating standardized farmland plot vector data. Subsequently, real-time meteorological monitoring data, soil nutrient test indicators, pest and disease images, and sowing records were mapped according to the National 2000 Geodetic Coordinate System and spatially aggregated using a 100m × 100m regular grid as the unit. For example, the weighted average of multiple soil sensor readings falling within the same grid was taken, with the weight determined by the reciprocal of the distance between the sensor and the grid center. Meteorological data was directly assigned to the grid covering the meteorological station and its eight neighboring grids. After the location of lesions in pest and disease images was identified by a target detection deep learning model, their geographic coordinates were placed into the corresponding grid and the risk level was marked. Based on this, according to the growth and development stages of winter wheat (emergence, tillering, jointing, heading, grain filling, and maturity), the locally implemented two-year wheat-corn rotation system, and the planting density standard of 300,000 basic seedlings per mu, the criteria for dividing production management units were set: requiring that the crop varieties within the same unit be the same, the sowing time differ by no more than 3 days, the soil type be consistent (e.g., all are brown soil), and the irrigation conditions be the same (e.g., whether or not drip irrigation facilities are available). Accordingly, the 5,000-mu area was divided into 186 production management units, each assigned a globally unique identifier (such as PMU-2024-WH-001 to PMU-2024-WH-186). This identifier was then indexed and linked with the historical yield, fertilization records, disaster losses, and other data of the corresponding plots in the base's historical production archive database, serving as the benchmark unit for subsequent process modeling.

[0022] The process modeling module operates within each production management unit. Taking production management unit 045 as an example, this module extracts all agricultural operation records for that unit from October 1, 2024, from the agricultural production dataset. These records include: October 10th, sowing was completed using an imported large-scale seeder with the following parameters: row spacing of 20 cm, sowing depth of 4 cm, and seeding rate of 12.5 kg / mu; November 5th, the drip irrigation system was activated for winter irrigation with a water consumption of 45 cubic meters / mu; March 12th, 2025, a domestically produced plant protection drone was used to spray fertilizer for greening, with a urea application rate of 10 kg / mu; March 28th, a second spraying was conducted to control sheath blight with tebuconazole at a dosage of 30 g / mu. Each record includes the operation time, the machine number (e.g., seeder 073, plant protection drone 112), the target plot identifier (i.e., production management unit 045), and quantitative parameters. The module converts each record into a standardized event parameter tuple in the format (operation type, timestamp, plot range, equipment number, parameter set). The operation type uses a predefined encoding (e.g., sowing=1, irrigation=2, fertilization=3, spraying=4); the timestamp is a standard timestamp; the plot range is represented by the coordinates of the smallest bounding rectangle of that unit; and the parameter set is encapsulated in a structured data format, such as row spacing of 20 cm, sowing depth of 4 cm, and seeding rate of 12.5 kg per acre. Subsequently, all tuples are arranged chronologically to construct a time-series chain of agricultural activities, forming a complete sequence of event parameter tuples.

[0023] Based on this sequence and a pre-built agronomic dependency rule knowledge base (including structured rules such as the "Technical Regulations for High-Yield Cultivation of Winter Wheat" and the "Guidelines for Safe Use of Pesticides"), the process modeling module constructs a directed dependency graph. For example, the knowledge base stipulates that seedling inspection and replanting should be completed within 7 to 15 days after sowing, but in this example, there is no replanting record, so no dependency edge is generated; it also stipulates that the application of greening fertilizer must be completed within 5 days after the average daily temperature has stabilized above 3 degrees Celsius, and the local average daily temperature on March 12 has reached 5 degrees Celsius, which complies with the rule; it further stipulates that herbicides are prohibited during the jointing stage, and any violations are marked as abnormal. The specific graph construction process is as follows: each agricultural operation event is treated as a graph node. If event B must be executed within a specified number of days after event A is completed (this number of days is determined by agronomic rules), then a directed edge is added from node A to node B, with the edge weight set to the number of days. For example, an edge with a weight of 25 is added from the sowing node to the irrigation node (the upper limit is used because winter irrigation is usually applied 20 to 30 days after sowing); no edge is added from fertilization to pesticide spraying as there is no mandatory dependency. In addition, two virtual nodes are introduced: start and end. The start node is connected to all nodes without preceding events (such as sowing in this example), and the end node is pointed to by all nodes without subsequent events (such as the last pesticide spraying). After the graph is constructed, a topological sorting algorithm is executed. If a cycle is found (e.g., A points to B, B points to C, C points to A), it is checked whether the irreversibility of crop growth is violated (e.g., harvesting cannot occur before sowing). If so, the dependent edges causing the cycle are deleted, ensuring that the graph is a directed acyclic graph.

[0024] The state encoding module receives the directed dependency graph as input. For each graph node, an initial feature vector is constructed: an operation type encoding (one-hot encoded vector, dimension being the total number of operation types), a parameter set normalization vector (normalizing continuous parameters such as seeding rate and pesticide dosage using standard scores, and mapping discrete parameters such as machine type to embedding vectors), and a spatiotemporal embedding of timestamps and plot coordinates (decomposing timestamps into periodic features of year, month, day, and day of the week, and encoding plot coordinates using sine and cosine positions). All initial node features are input into a three-layer graph attention network, each layer containing four attention heads, aggregating neighbor information along directed edges. For example, in the first layer, the hidden state of a node's irrigation will aggregate information from the seeding node, with weights determined by the attention coefficients; the second layer further aggregates more distant neighbors; the third layer outputs the final hidden state of the node. Subsequently, the difference (in days) between the current decision time (assumed to be April 5, 2025) and the timestamps of each node is calculated. The difference is then substituted into the exponential decay function to obtain the timeliness weight (decay coefficient is 0.1). The hidden states of all nodes are weighted and summed to generate a 128-dimensional process state feature vector, which integrates historical operation patterns and current execution progress.

[0025] The intelligent decision-making module inputs this feature vector into the multi-task decision network. The network's bottom layer consists of a shared 3-layer fully connected layer (512 neurons per layer, with modified linear unit activation). The upper branches are three sub-networks: The job recommendation sub-network is a 2-layer multilayer perceptron with a normalized exponential function, outputting the probability distribution of the next operation type (e.g., irrigation: 0.1, fertilization: 0.05, spraying: 0.75, harvesting: 0.1), selecting the spraying type with the highest probability as the recommendation; the input application rate prediction sub-network is a 2-layer multilayer perceptron with a linear output layer, outputting the regression values ​​of irrigation, fertilizer, and pesticide amounts (e.g., pesticide amount = 35.2 grams per acre), and applying non-negativity constraints (minimum value 0) and agronomic upper limits (e.g., maximum dosage of tebuconazole 50 grams per acre), truncated to 35.2; the production risk warning sub-network is a 2-layer multilayer perceptron with a logistic function, outputting the probabilities of three risk categories (pests and diseases: 0.82, drought: 0.15, lodging: 0.08), with a threshold of 0.5 determining a high risk of pests and diseases. Combining the outputs of the three factors, an initial decision-making scheme is formed: the recommended operation type is spraying, the recommended application rate is 35.2 grams per acre, and the pest and disease risk level is high.

[0026] Finally, the operation scheduling instruction generation module combines the above decision-making scheme with the plot spatial information (standard geometric format polygon coordinates) of production management unit 045 to generate an agricultural operation plan, and encapsulates it into an electronic task order in the Extensible Markup Language (EXPLAIN) format of task data conforming to the international agricultural machinery bus standard. This task order is pushed to the edge computing node (model: industrial-grade edge computing gateway) deployed on the edge gateway of this unit via the message queue telemetry transmission protocol. After parsing the task order, the edge node identifies the operation type as plant protection operation, and then sends control instructions to the standby domestic plant protection drone (equipment number 112) through the agricultural machinery bus virtual terminal protocol on the controller area network bus. The instructions include parameters such as flight path (automatically generated flight path based on plot coordinates), spraying rate (35.2 grams per acre corresponding to a flow valve opening of 65%), and operating height (2.5 meters). After the drone performs its operation, its flight control system automatically transmits feedback data such as the operation trajectory, actual spraying volume, and battery consumption. The edge node converts this data into a new agricultural operation event (operation type = spraying, timestamp = 14:30 on April 5, 2025), and appends it to the end of the event parameter tuple sequence of production management unit 045 to complete the closed loop.

[0027] The entire system, without accompanying diagrams, relies on the tight coupling of data and control flows between modules through the above steps: the data acquisition module outputs a structured dataset to the unit partitioning module; the unit partitioning module outputs uniquely identified production management units to the process modeling module; the process modeling module outputs a directed dependency graph to the state encoding module; the state encoding module outputs process state feature vectors to the intelligent decision-making module; the intelligent decision-making module outputs decision schemes to the job scheduling instruction generation logic, and finally interacts with the execution equipment through the agricultural IoT platform. Each module runs in a container orchestration cluster deployed in the cloud, achieving asynchronous communication through a descriptive state transfer application programming interface and message queue middleware, ensuring real-time response capabilities under high concurrency.

[0028] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0029] In a 5,000-mu winter wheat production base in the North China Plain, the system first obtains the boundary coordinate point sequence of each standard plot through Beidou satellite navigation and positioning terminals deployed in the field to form an initial vector profile; at the same time, the soil moisture sensor array continuously collects soil moisture content, conductivity, pH value and nitrogen, phosphorus and potassium content at 10-minute intervals, the meteorological station records temperature, humidity, wind speed, precipitation and solar radiation simultaneously, the drone flies weekly to obtain multispectral normalized vegetation index images and visible light images of pests and diseases, and the traceability label reader automatically writes time-stamped electronic operation vouchers at key agricultural nodes. All raw data are appended with a UTC (United States Coordinated Universal Time) timestamp and obtained from the World Geodetic Coordinate System (UGC) coordinates through satellite navigation differential correction. These coordinates are then uniformly projected onto the National Geodetic Coordinate System 2000 to ensure spatial consistency. For meteorological data missing due to equipment power outages, inverse distance weighted interpolation from nearby stations is used, combined with historical average values ​​for the same period to fill the gaps. Abnormal abrupt changes in soil sensor outputs are detected and removed using a threshold of 3 times the standard deviation within a sliding window. All physical quantities, such as seeding rate (kg / acre) and irrigation rate (m³ / acre), are converted to SI units and finally written into a distributed time-series database built on an open-source time-series database, forming a spatiotemporally aligned structured agricultural production dataset.

[0030] When the system receives a planting plan for the Jimai 22 variety, sown on October 10th, with a target yield of 600 kg per mu, the unit partitioning module extracts all 200 initial plot boundaries from the database and calls the measured coordinates after centimeter-level real-time dynamic differential positioning correction. Combined with the farmland mask map output by the Sentinel-2 satellite imagery through a convolutional neural network semantic segmentation model, the boundaries are topologically closed and gaps are repaired to generate standardized plot vectors without overlap or gaps. Subsequently, heterogeneous data such as soil nutrients, meteorology, and pest and disease identification results are mapped to a 100m × 100m regular grid according to the National Geodetic Coordinate System 2000: multiple soil sensor readings within the same grid are weighted by the inverse of distance, meteorological station data is diffused to its own grid and a Moore neighborhood of 9 grids, and after the pest and disease images are located by the target detection deep learning model, the geographic coordinates are placed into the corresponding grid and the risk level is marked. Based on this, and according to the six-stage growth period of winter wheat, the wheat-corn rotation system, and the requirement of 300,000 basic seedlings per mu, rigid constraints were set for unit division: the variety within the same unit must be the same, the sowing date difference must be ≤3 days, the soil type must be brown soil, and the irrigation method must be either drip irrigation or flood irrigation. Adjacent plots that meet all conditions were clustered and merged, ultimately generating 186 production management units. Each unit was assigned a globally unique identifier, for example: PMU-2024-WH-001, and this ID was used to link to historical records of yield, fertilizer application, and disaster losses in the base's historical archive, serving as the spatiotemporal reference entity for process modeling.

[0031] Taking production management unit 045 as an example, the process modeling module extracts all its operation logs from the database since October 1, 2024. This includes the sowing operation performed by an imported large-scale seeder on October 10 (row spacing 20 cm, depth 4 cm, seeding rate 12.5 kg / mu), the application of winter irrigation water (45 cubic meters / mu) by the drip irrigation system on November 5, the spraying of regrowth fertilizer (10 kg / mu of urea) by a domestic plant protection drone on March 12, 2025, and the spraying of pesticides to prevent sheath blight (30 g / mu of tebuconazole) on March 28. Each record is parsed into a standardized tuple (operation type code, standard timestamp, minimum bounding rectangle coordinates, equipment identifier, structured parameter set). The operation type is mapped to integers according to a predefined dictionary (e.g., sowing = 1), and the parameter set retains the original units of measurement but is structured and encapsulated. These tuples are arranged in ascending order of time, forming a time-series chain of agricultural activities reflecting the actual execution process.

[0032] The time-series chain is logically matched with the built-in agronomic rule knowledge base: the knowledge base stores structured rules such as the winter irrigation water should be applied within 20 to 30 days after sowing. Based on this, the system establishes directed edges between the sowing node and the irrigation node, with the edge weight set to 25 (using the upper limit of the window to enhance robustness). Since the application of fertilizer for greening requires a daily average temperature greater than or equal to 3 degrees Celsius for 5 consecutive days, and the meteorological data from March 7 to 11 shows daily average temperatures of 2.8 degrees Celsius, 3.1 degrees Celsius, 3.5 degrees Celsius, 4.0 degrees Celsius, and 4.2 degrees Celsius respectively, the condition is met, so the fertilization node is allowed to exist without triggering an anomaly. If a herbicide spraying record is detected after the jointing stage, it is marked as a violation and the generation of dependent edges is blocked. When building the graph, a virtual start node is introduced, pointing to the first operation of sowing, and a virtual end node is pointed to by the last operation of spraying. After the graph is built, a topological sorting algorithm is run to sort the topology. If a cycle is found (such as harvesting before sowing due to data errors), conflicting edges are deleted according to the principle of irreversible crop growth, and a directed acyclic graph is forcibly output to ensure that the process logic conforms to biological laws.

[0033] After receiving the directed acyclic graph, the state encoding module constructs a 128-dimensional initial feature for each node: the operation type uses 8-dimensional one-hot encoding (supporting 8 types of agricultural operations), continuous parameters (such as seeding rate and pesticide dosage) are normalized using standard scores, discrete device identifiers (such as plant protection drone No. 112) are mapped to a 16-dimensional vector through a learnable embedding layer, the timestamp is decomposed into 8-dimensional sine and cosine periodic features of year, month, day, and day of the week, and the plot center coordinates are generated into a 32-dimensional spatial embedding through sine and cosine position encoding, with the remaining dimensions padded with zeros for alignment. This feature is input into a three-layer graph attention network, with each layer containing 4 attention heads, aggregating predecessor node information along directed edges—for example, the irrigation node aggregates the hidden state of the sowing node in the first layer, and the attention weights are dynamically calculated by the node feature similarity; the second layer aggregates indirect predecessors (such as the path from start to sowing to irrigation); the third layer outputs the final node representation. Then, the time difference (in days) between the current decision time (April 5, 2025) and the timestamps of each node is calculated. The timeliness weight is obtained by substituting it into the exponential decay function. The hidden states of all nodes are weighted and summed to generate a 128-dimensional process state feature vector. This vector contains both historical operation patterns (such as the influence of early sowing depth on later growth) and the high timeliness of recent events (such as the higher weight of spraying at the end of March on disease prediction in early April).

[0034] The intelligent decision-making module inputs this feature vector into a multi-task network: the shared bottom layer is a 3-layer, 512-neuron fully connected layer (with modified linear unit activation); the upper branch's task recommendation sub-network (2-layer multilayer perceptron with normalized exponential function) outputs four types of operation probabilities. Because it is currently in the early heading stage and the normalized vegetation index shows excessively high canopy density, the spraying probability reaches 0.75; the input prediction sub-network (2-layer multilayer perceptron with linear output layer) regresses the pesticide dosage to 35.2 grams per acre, retaining the original value after being truncated by non-negative constraints (greater than or equal to 0) and the maximum safe dosage of tebuconazole at 50 grams per acre; the risk warning sub-network (2-layer multilayer perceptron with logistic function) outputs a pest and disease probability of 0.82 (greater than the 0.5 threshold), classifying it as high risk. The combination of these three factors forms the decision scheme: execute plant protection operations, apply pesticide at 35.2 grams per acre, and establish a high disease warning level.

[0035] The operation scheduling instruction generation module merges the above scheme with the standard geometric format polygon coordinates of production management unit 045 to generate an agricultural operation scheme, and encapsulates it into an electronic task sheet according to the international agricultural machinery bus standard task data extensible markup language architecture. This task sheet is pushed to the industrial-grade edge computing node deployed on the edge gateway of this unit via message queue middleware. After parsing the extensible markup language, the edge node identifies the operation type as plant protection, and then sends an instruction to plant protection drone No. 112 through the controller area network bus agricultural machinery bus virtual terminal protocol: automatically generate a reciprocating flight path based on the standard geometric format coordinates, calculate the flow rate valve opening as 65% based on 35.2 grams per acre (obtained from the pre-calibrated dose-opening mapping table), and set the operation height to 2.5 meters. After the drone executes the operation, it transmits back data such as the actual trajectory, spraying amount of 34.8 grams per acre, and remaining battery power. The edge node converts this data into a new event tuple (spraying, 14:30 on April 5, 2025) and adds it to the event sequence of this unit, completing the perception-decision-execution-feedback closed loop.

[0036] The entire methodology achieves full-process digital control through a chain of processes, including data acquisition, unit division, process modeling, state coding, intelligent decision-making, and instruction issuance. Each module is deployed on a cloud-based container orchestration cluster, providing service interfaces through expressive state transfer application programming interfaces (APIs). Message queue middleware ensures reliable task delivery under high concurrency, and edge nodes achieve millisecond-level instruction response. This solves core problems in traditional agricultural systems such as fragmented multi-source data, static agricultural logic, and delayed decision-making, achieving dynamic collaboration across the entire chain from pre-production planning to production execution and post-production traceability.

[0037] like Figure 2 As shown, a digital management system for the entire agricultural production process includes: Data acquisition module: used to collect multi-source heterogeneous agricultural data during the agricultural production stage and generate agricultural production datasets; Production Management Module: Used to perform spatiotemporal correlation matching of the agricultural production dataset according to the production time sequence and the spatial location of the plots, and to divide the matched agricultural production dataset into several production management units; Model building module: This module extracts key nodes and execution parameters of agricultural operations within the production management unit, constructs a time-series chain of agricultural activities based on the temporal order and type of the operations, forming a standardized sequence of event parameter tuples. Based on the sequence of event parameter tuples and a knowledge base of agronomic dependency rules, it constructs a directed dependency graph of the agricultural production process. Nodes in the directed dependency graph represent key nodes of the agricultural operations, and edges represent agronomic dependencies and time window constraints between preceding and subsequent processes. Using the graph node features and edge weights of the directed dependency edges as the input space, and historical production experience and current execution progress as state representation targets, a graph attention model is constructed. Vector generation module: Used to iteratively optimize network parameters by using the feature encoding quality to encode features of the directed dependency graph through the graph attention model, and generate a process state feature vector that integrates historical production experience and current execution progress; Intelligent decision-making module: It is used to input the process state feature vector into an intelligent decision-making model containing a multi-task decision network to obtain an initial decision scheme for the next operation; generate an operation scheduling instruction containing operation parameters and execution instructions according to the initial decision scheme; and send the operation scheduling instruction to the corresponding execution device through the agricultural Internet of Things platform to perform intelligent management and control operations of the agricultural production process.

[0038] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for digital management of the entire agricultural production process, characterized in that, include: Collect multi-source heterogeneous agricultural data from various stages of agricultural production to generate an agricultural production dataset. The agricultural production dataset is matched spatiotemporally according to the production time sequence and the spatial location of the plots, and the matched agricultural production dataset is divided into several production management units. Key nodes and execution parameters of agricultural operations are extracted within the production management unit. Based on the time sequence and operation type of agricultural operations, a time sequence chain of agricultural activities is constructed to form a standardized sequence of event parameter tuples. Based on the event parameter tuple sequence and the agronomic dependency rule knowledge base, a directed dependency graph of the agricultural production process is constructed; the nodes in the directed dependency graph represent the key nodes of the agricultural operation, and the edges represent the agronomic dependency relationship and operation time window constraints between the preceding and following processes. Using the graph node features and edge weights of the directed dependency graph as the input space, and historical production experience and current execution progress as the state representation targets, a graph attention model is constructed. Using the feature encoding quality as the iterative optimal network parameter, the directed dependency graph is feature encoded through the graph attention model to generate a process state feature vector that integrates historical production experience and current execution progress. The process state feature vector is input into an intelligent decision-making model containing a multi-task decision-making network, and the initial decision scheme for the next operation is output. Based on the initial decision-making scheme, an operation scheduling instruction containing operation parameters and execution instructions is generated. The operation scheduling instruction is then sent to the corresponding execution device through the agricultural Internet of Things platform to perform intelligent management and control operations of the agricultural production process.

2. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, The agricultural production dataset is subjected to spatiotemporal correlation matching based on production time sequence and plot spatial location. The matched agricultural production dataset is then divided into several production management units, including: Collect multi-source heterogeneous agricultural data covering all stages of production, including pre-production preparation, field management during production, and post-production harvesting and processing, and generate an agricultural production dataset. Farmland plot boundary coordinates are extracted from the plot boundary vectors in the agricultural production dataset. The coordinate accuracy is corrected by a positioning terminal, and standardized farmland plot vector data is generated by combining the intelligent interpretation results of satellite remote sensing images. The coordinates of real-time meteorological monitoring data, soil nutrient detection indicators, field survey images of pests and diseases, and crop sowing operation records in the agricultural production dataset are unified. Field meteorological monitoring stations, soil moisture sensor networks, UAV agricultural inspection images, and agricultural technician operation records are uniformly mapped to the geographic coordinate reference system. Spatial units are aggregated according to grid management requirements to generate spatiotemporally aligned agricultural data. Based on the spatiotemporally aligned agricultural data, production management unit division criteria are set according to crop growth and development patterns, field rotation and cropping systems, and cultivation density standards to ensure that crop growth characteristics, environmental response patterns, and agronomic measures are consistent within the same production management unit. Based on the crop varieties and planting scale of each plot in the target agricultural production area, a globally unique identifier is assigned to each production management unit, the production management unit is identified as the baseline unit for process management, and the globally unique identifier is linked to a multi-year production record database.

3. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, Key nodes and execution parameters of agricultural operations are extracted within the production management unit. Based on the time sequence and type of agricultural operations, a time-series chain of agricultural activities is constructed, forming a standardized sequence of event parameter tuples, including: Extract key agricultural operation nodes from crop sowing operation records, precision irrigation execution logs, and agricultural product circulation traceability information within the production management unit; Based on the execution time and execution parameters of each key node of the agricultural operation, identify the operation time, the personnel or machinery performing the operation, the target plot of the operation, and the quantitative parameter information of the operation. Based on the position and dependencies of the agricultural operation in the agricultural production process, each key node of the agricultural operation is converted into a standardized agricultural operation node data structure. Arrange all the agricultural operation nodes within the same production management unit according to the agricultural calendar order to construct an agricultural activity time sequence chain, forming a sequence of event parameter tuples that reflects the complete production process.

4. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, Based on the event parameter tuple sequence and the agronomic dependency rule knowledge base, a directed dependency graph of the agricultural production process is constructed, including: Extract each agricultural operation event from the event parameter tuple sequence and the operation dependency relationship from the agronomic dependency rule knowledge base; Using each agricultural operation event as a graph node, and based on the preceding and following process constraints defined in the agronomic dependency rule knowledge base, if agricultural operation event B must be executed within the appropriate agricultural time window after agricultural operation event A is completed, then a directed dependency edge is established between the graph node corresponding to agricultural operation event A and the graph node corresponding to agricultural operation event B. Using the agronomic dependencies between operations as graph edges, and based on the edge weight function determined by the operation time window constraint, the edge weight value of the directed dependent edge is set to the allowable interval days. A virtual node for the start of a production cycle and a virtual node for the end of a production cycle are introduced. The virtual node for the start of a production cycle is connected to all starting operation diagram nodes without prior agricultural operations, and the virtual node for the end of a production cycle is connected to all finishing operation diagram nodes without subsequent agricultural operations. A graph structure modeling method is used to construct a complete directed dependency graph of agricultural production process, and the topological orderliness of the directed dependency graph structure is checked to identify and eliminate cyclic dependency paths that violate crop growth patterns or agronomic technical regulations.

5. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, Using the graph node features and edge weights of the directed dependency graph as the input space, and historical production experience and current execution progress as state representation objectives, a graph attention model is constructed. Using feature encoding quality as the iteratively optimal network parameter, the graph attention model is used to encode features of the directed dependency graph, generating a process state feature vector that integrates historical production experience and current execution progress, including: A graph attention mechanism network is used as a feature extractor to construct initial node features for each graph node in the directed dependency graph. The initial node features are composed of the operation type encoding of the corresponding agricultural operation event, the standardized vector of the parameter set, and the combination of the timestamp and the spatiotemporal coordinate position encoding of the plot range. The initial features of the nodes are input into the graph attention mechanism network. The association information of adjacent graph nodes is aggregated along the directed dependency edges through a multi-layer graph attention propagation mechanism, and the hidden state representation of each graph node is updated layer by layer. The hidden state representations of each graph node in the directed dependency graph are aggregated using a time-weighted aggregation method. The aggregation weights are calculated and determined by the time decay function of the current decision time and the timestamps of each agricultural operation event. Based on the quality of feature encoding, the optimal network parameters are iterated, and the timeliness-weighted aggregation result is output as a process status feature vector that integrates historical production experience and current execution progress.

6. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, The process state feature vector is input into an intelligent decision-making model containing a multi-task decision network, and the initial decision scheme for the next operation is output, including: Extract the production process status information and historical job execution patterns from the process status feature vector; Using the process state feature vector as input, and based on the multi-task objective function set determined by agricultural operation type classification constraints, input application amount regression constraints, and risk warning level identification constraints, an intelligent decision-making model is constructed, which includes an operation recommendation sub-network, an input application amount prediction sub-network, and a production risk warning sub-network. Each sub-network shares the underlying feature extraction layer to achieve production knowledge collaboration. The process state feature vector is input into the job recommendation subnetwork, and the job recommendation subnetwork outputs the probability distribution of the next suitable agricultural operation type, using a multi-classification decision mechanism. The process state feature vector is input into the input application rate prediction subnetwork. The input application rate prediction subnetwork regressively calculates the application rate of agricultural inputs, including irrigation water volume, fertilizer application rate, and pesticide dosage. Non-negative constraint regression is used and an agronomic safety upper limit is set. The process status feature vector is input into the production risk early warning sub-network, and the production risk early warning sub-network outputs the probability of occurrence of crop pest and disease risk level, drought stress risk level, and lodging disaster risk level, using a multi-label classification mechanism; The collaborative prediction algorithm in the intelligent decision-making model is used to obtain an initial decision scheme for the next operation, which includes the recommended agricultural operation type, the recommended amount of agricultural inputs, and the production risk warning level.

7. The method for digital management of the entire agricultural production process according to claim 6, characterized in that, Based on the initial decision-making scheme, a work scheduling instruction containing work parameters and execution instructions is generated. This instruction is then distributed to the corresponding execution devices via an agricultural IoT platform to perform intelligent management and control operations on the agricultural production process, including: The recommended agricultural operation type, the recommended agricultural input application rate, and the plot spatial information of the production management unit are combined to generate an agricultural operation plan containing operation parameters. The agricultural operation plan is converted into an operation scheduling instruction containing execution instructions, using an electronic task sheet format that conforms to agricultural machinery operation standards; The operation scheduling instructions are pushed to field edge computing nodes or agricultural machinery control terminals through the IoT communication protocol of the agricultural IoT platform. After parsing the operation scheduling instruction, the field edge computing node determines the corresponding execution equipment type based on the recommended agricultural operation type, and sends precise operation control commands to the intelligent tractor and plant protection machinery through the agricultural machinery bus interface, or adjusts the opening of the water and fertilizer integration irrigation system's proportion valve and the start / stop status of the irrigation pump station according to the recommended agricultural input application amount through the industrial communication protocol. The system collects real-time operation execution feedback data from the intelligent agricultural machinery, the integrated water and fertilizer irrigation system, or the post-harvest processing equipment, converts the operation execution feedback data into new agricultural operation events, and synchronously updates the event parameter tuple sequence to perform intelligent control operations on the agricultural production process.

8. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, Intelligent decision-making models based on historical production data are continuously optimized, along with agronomic-dependent rules, including: Collect the event parameter tuple sequence of each process management benchmark unit in multiple production cycles, the operation execution feedback data, crop growth monitoring data and final yield and quality indicators to construct a data sample set of historical production; The data sample set is effectively labeled with agronomic techniques. By comparing and analyzing the deviation between the actual yield and quality indicators and the expected targets in the data sample set, efficient and inefficient production modes are identified. Key influencing factors are extracted from the event parameter tuple sequences corresponding to the efficient and inefficient production modes. The intelligent decision-making model is trained using a contrastive learning strategy, which widens the distance between the process state feature vectors corresponding to the efficient production mode and the process state feature vectors corresponding to the inefficient production mode in the feature space, thereby generating a trained and optimized intelligent decision-making model. High-weight directed dependency edges and their corresponding edge weight values ​​are extracted from the trained and optimized intelligent decision model. The graph node pairs connected by the high-weight directed dependency edges and the key time window parameters recorded by the edge weight values ​​are used to form agronomic dependency rules. The incremental update of the trained and optimized intelligent decision-making model is triggered based on the newly added task execution feedback data. The agronomic dependency rules are retained by the transfer learning mechanism, while adapting to local climate change, variety updates and cultivation technology improvements. The agronomic dependency rules obtained from mining are integrated with the experience of traditional agricultural experts to update the agronomic dependency rule knowledge base, establish a human-machine collaborative agricultural production decision-making knowledge graph, and support agricultural technicians in reviewing and revising the initial decision-making scheme. The agronomic dependency rules obtained from mining are integrated with the experience of traditional agricultural experts to update the agronomic dependency rule knowledge base, establish a human-machine collaborative agricultural production decision-making knowledge graph, and support agricultural technicians in reviewing and revising the initial decision-making scheme.

9. The method for digital management of the entire agricultural production process according to claim 1, characterized in that, Coordinated scheduling and optimal allocation of agricultural resources, including: Extract the initial decision schemes corresponding to each process management benchmark unit within the collaborative management plot set, obtain the recommended agricultural operation type, recommended agricultural input application amount, and production risk warning level from the initial decision schemes, align the agricultural operation requirements of each benchmark unit within the collaborative management plot set by time and prioritize them, and comprehensively consider the operation time window constraints in the directed dependency graph, the sensitivity of the critical period of crop growth, and the disaster warning information in the production risk warning level to generate a collaborative operation requirement sequence. Based on the collaborative operation demand sequence, a cross-plot resource sharing constraint model is established, setting an upper limit on the number of schedulable execution equipment, a constraint on the total water supply capacity of the integrated water and fertilizer irrigation system, a limit on manpower allocation capacity, and a requirement for logistics and transportation timeliness. Using the collaborative operation demand sequence and the cross-plot resource sharing constraint model as input, a multi-objective optimization algorithm is used to solve the global operation scheduling scheme. The optimization objectives include minimizing the overall operation completion time, maximizing the utilization rate of the execution equipment, minimizing the transportation cost of the recommended agricultural input application amount, and minimizing the number of conflicts in the critical agricultural time window. Based on the global operation scheduling scheme, the execution order of recommended agricultural operation types and the allocation scheme of recommended agricultural input application amount of each process management benchmark unit in the collaborative management plot set are dynamically adjusted to generate a cross-plot collaborative operation plan table containing operation parameters and execution instructions. According to the cross-plot collaborative operation plan, the operation execution feedback data and emergencies of each process management benchmark unit in the collaborative management plot set are monitored in real time. When there is a failure of the execution equipment, extreme weather, or an increase in the risk of pests and diseases in the production risk warning level, the dynamic rescheduling mechanism is triggered. The updated collaborative operation demand sequence and the cross-plot resource sharing constraint model are re-input into the multi-objective optimization algorithm to solve the updated global operation scheduling scheme, giving priority to ensuring the resource supply of key plots and key production links. The statistical analysis of the collaborative scheduling effect of the global operation scheduling scheme and the updated global operation scheduling scheme is used to evaluate the improvement in the utilization efficiency of the execution equipment, the reduction in transportation costs of the recommended agricultural input application amount, and the improvement in yield and quality, thereby generating a regional agricultural production collaborative management performance report.

10. A digital management system for the entire agricultural production process, used to execute the digital management method for the entire agricultural production process as described in any one of claims 1-9, characterized in that, include: Data acquisition module: used to collect multi-source heterogeneous agricultural data during the agricultural production stage and generate agricultural production datasets; Production Management Module: Used to perform spatiotemporal correlation matching of the agricultural production dataset according to the production time sequence and the spatial location of the plots, and to divide the matched agricultural production dataset into several production management units; Model building module: This module extracts key nodes and execution parameters of agricultural operations within the production management unit, constructs a time-series chain of agricultural activities based on the temporal order and type of the operations, forming a standardized sequence of event parameter tuples. Based on the sequence of event parameter tuples and a knowledge base of agronomic dependency rules, it constructs a directed dependency graph of the agricultural production process. Nodes in the directed dependency graph represent key nodes of the agricultural operations, and edges represent agronomic dependencies and time window constraints between preceding and subsequent processes. Using the graph node features and edge weights of the directed dependency edges as the input space, and historical production experience and current execution progress as state representation targets, a graph attention model is constructed. Vector generation module: Used to iteratively optimize network parameters by using the feature encoding quality to encode features of the directed dependency graph through the graph attention model, and generate a process state feature vector that integrates historical production experience and current execution progress; Intelligent decision-making module: It is used to input the process state feature vector into an intelligent decision-making model containing a multi-task decision network to obtain an initial decision scheme for the next operation; generate an operation scheduling instruction containing operation parameters and execution instructions according to the initial decision scheme; and send the operation scheduling instruction to the corresponding execution device through the agricultural Internet of Things platform to perform intelligent management and control operations of the agricultural production process.

Images