A crop growth prediction method and system based on digital twins

Abstract

This invention discloses a crop growth prediction method and system based on digital twins, belonging to the interdisciplinary field of smart agriculture and information technology. The method includes: acquiring multi-source heterogeneous time-series data; constructing an agronomic causal knowledge graph and quantifying the strength of conditional causal effects; dynamically adjusting spatiotemporal search weights based on the strength of causal effects when data is missing; constructing a digital twin agent comprising a crop growth agent, an environmental simulation agent, and at least one management decision-making agent, with each agent configured with a stress response memory function to record stress history and dynamically correct response functions; driving multi-agent adversarial inference based on reconstructed data to achieve cross-scenario stress response parameter transfer and calibration; collecting expert correction operations, and using inverse reinforcement learning to infer implicit reward functions to represent decision preferences, generating management suggestions matching expert intentions. This invention solves the problems of data missingness, limitations of static inference, and difficulty in inheriting implicit experience, thereby improving prediction accuracy and the level of decision-making intelligence.

Description

Technical Field

[0001] This invention relates to the intersection of smart agriculture and information technology, and in particular to a crop growth prediction method and system based on digital twins. Background Technology

[0002] With the rapid development of the Internet of Things, remote sensing technology, and artificial intelligence, precision agriculture has become an important direction for the development of modern agriculture. Crop growth prediction and intelligent decision-making systems integrate multi-source data to provide scientific guidance for agricultural operations, which is of great significance for improving crop yields and reducing resource consumption.

[0003] Existing technologies lack the capacity to process heterogeneous data from multiple sources across space, air, and ground. Currently, the core focus of these technologies is on the construction and coordination of data acquisition networks. However, when sensor nodes malfunction or communication is interrupted, leading to data loss, these technologies can only rely on the network's self-organizing function for supplementary acquisition. This supplementary acquisition itself still depends on the normal operation of hardware devices and cannot solve the problem of historical data gaps. Existing decision-making systems generally use static models trained on historical data. However, this decision-making approach ignores the real-time physiological state of crops and the cumulative effects of drought stress. The experience gained by senior agronomists through long-term practice is particularly important, but this experience is often difficult to fully express using explicit rules. Existing technologies allow experts to manually adjust parameters through human-computer interaction interfaces, but the system only records the final result of the adjustment and cannot understand the deep logic behind the expert's adjustment. This makes it impossible for the system to continue the expert's decision-making style in new situations and to explain the decision-making logic.

[0004] Furthermore, in existing technologies, data preprocessing, model inference, and human-computer interaction are usually independent modules, lacking information feedback and collaborative optimization mechanisms. This results in the overall system performance not being able to continuously improve with use, thus requiring a more suitable prediction method for crop growth. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a crop growth prediction method and system based on digital twins.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a crop growth prediction method based on digital twins, comprising:

[0007] S1: Obtain multi-source heterogeneous time-series data of the target plot;

[0008] S2: Construct an agronomic causal knowledge graph to quantify the strength of the conditional causal effects of various environmental factors on crop yield indicators;

[0009] S3: When missing data time series is detected, data is reconstructed based on the agronomic causal knowledge graph to generate reconstructed data;

[0010] S4: Construct a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent has the function of recording the history of environmental stress and dynamically modifying the stress response function.

[0011] S5: Based on the reconstructed data, drive the digital twin to perform multi-agent adversarial simulation and transmit stress response parameters between agents in different decision-making scenarios to calibrate environmental compensation effect parameters;

[0012] S6: Collect the correction operations of expert users, use inverse reinforcement learning to infer the implicit reward function that represents their decision preferences, and generate management suggestions based on the implicit reward function.

[0013] In a preferred embodiment of the present invention, the data reconstruction includes dynamically adjusting the weighting weights, wherein the dynamically adjusting the weighting weights includes:

[0014] Identify the set of key environmental factors whose causal effects exceed a threshold during the current reproductive period;

[0015] In spatial similarity searches, higher weights are given to parcel attributes related to key environmental factors;

[0016] In time similarity searches, historical periods that are similar to the change curves of key environmental factors are prioritized.

[0017] In a preferred embodiment of the present invention, the data reconstruction includes causal consistency verification, which includes:

[0018] Verify whether the substitute values ​​meet the numerical constraints in the causal relationship, and lower the fusion confidence of those that do not meet the constraints.

[0019] The weights of the weighted fusion are determined by both the confidence level and the strength of the causal effect.

[0020] In a preferred embodiment of the present invention, the transmission of stress response parameters between agents in different decision-making scenarios includes:

[0021] The stomatal conductance change curve, photosynthetic inhibition coefficient, and rehydration recovery time after stress are encoded into a stress memory vector;

[0022] By sharing the memory space with crop growth agents in other decision-making scenarios, the model of rehydration compensation effect is modified.

[0023] In a preferred embodiment of the present invention, the data reconstruction includes spatial similarity search, the spatial similarity search including:

[0024] Construct a plot similarity map, wherein the plot similarity map is characterized by a graph neural network to represent the similarity of plots in terms of soil properties, planting varieties and sowing time;

[0025] Based on graph neural network identification, a set of neighboring plots whose similarity to the target plot exceeds a preset threshold;

[0026] The spatial substitution value is extracted from real-time data of the neighboring plot set.

[0027] In a preferred embodiment of the present invention, the implicit reward function is a multi-objective weighted function; the inverse reinforcement learning algorithm learns the weight coefficients of each objective from expert demonstration data, and under the same environmental conditions, the implicit reward function scores the expert operation higher than the other candidate operation.

[0028] In a preferred embodiment of the present invention, after generating management suggestions, the method further includes:

[0029] Compare the simulation results of the first and second decision scenarios, and calculate the differences in expected output and resource consumption under different management strategies;

[0030] Based on the differences in expected yield and resource consumption, and in conjunction with the current crop growth stage, the optimal management strategy and its execution parameters are selected from the management recommendations.

[0031] The optimal management strategy and its execution parameters will be used as the final output.

[0032] A crop growth prediction system based on digital twins includes,

[0033] The data acquisition interface is used to acquire multi-source heterogeneous time-series data of the target plot;

[0034] The knowledge graph construction unit is used to construct an agronomic causal knowledge graph and quantify the strength of the conditional causal effects of various environmental factors on crop yield indicators.

[0035] The data reconstruction engine is used to dynamically adjust the weighting of spatial similarity search and temporal similarity search according to the strength of the conditional causal effect of the current reproductive period when data missing is detected. It extracts spatial substitute values ​​and historical substitute values ​​respectively, performs causal consistency verification with the strength of the conditional causal effect as a constraint, and then weights and fuses them to generate reconstructed data.

[0036] A digital twin construction unit is used to construct a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent includes a stress response memory module for recording environmental stress history and dynamically correcting the stress response function.

[0037] The adversarial simulation engine is used to drive the digital twin to perform multi-agent adversarial simulation based on the reconstructed data: at least two decision scenarios are run in parallel, wherein the crop growth agent in the first decision scenario transmits the stress response parameters to the crop growth agent in the second decision scenario through the stress response memory module after experiencing stress, and calibrates its subsequent environmental compensation effect parameters.

[0038] The expert intent capture unit is used to collect the management strategy correction operations of expert users and infer the implicit reward function of expert users through an inverse reinforcement learning algorithm. The implicit reward function represents the decision preferences of expert users.

[0039] The decision generation unit is used to input the implicit reward function into the management decision-making agent to generate management suggestions that match the expert's decision-making intent;

[0040] Output interface, used to output prediction results or management suggestions.

[0041] In a preferred embodiment of the present invention, the stress response memory module maintains a temporal memory matrix to record the stress type, intensity, duration and recovery trajectory, which is used to dynamically adjust the sensitivity parameter and recovery rate parameter of the current stress response function when similar stress events occur.

[0042] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0043] (1) This invention constructs an agronomic causal knowledge graph and quantifies the intensity of conditional causal effects to achieve causal-guided reconstruction in scenarios with missing data, thus solving the problem of prediction interruption caused by sensor failure in existing technologies; by endowing the digital twin with stress response memory function and introducing cross-scenario adversarial inference, it simulates the physiological memory effect and rehydration compensation law of crops, breaking through the limitations of static comparison of traditional models; by inverse reinforcement learning, it reverses the implicit reward function from the expert correction operation, transforming the implicit experience of senior agronomists into computable decision preferences.

[0044] (2) Based on the quantified conditional causal effect intensity in the agronomic causal knowledge graph, the weighting of spatial similarity search and temporal similarity search is dynamically adjusted. When the sensor data of the target plot is missing in time series, the system no longer uses traditional linear interpolation or single spatial interpolation. Instead, it actively retrieves real-time data of neighboring plots with highly similar soil properties and historical data with highly matched curves of key factors based on the key environmental factors of the current growth period. By dynamically allocating weights through the conditional causal effect intensity, the system ensures that the key factors obtain higher confidence in the reconstruction, thereby ensuring the reliability of subsequent decisions.

[0045] (3) The crop growth agent is equipped with a stress response memory function. It records the type, intensity, duration and recovery trajectory of each stress event through a time-series memory matrix, and dynamically adjusts the sensitivity parameter and recovery rate parameter when similar stress occurs. At the same time, in the multi-agent adversarial simulation, the first decision scenario that experiences stress encodes the stress response parameter into a stress memory vector, which is passed to the second decision scenario through a shared memory space to calibrate its rehydration compensation effect model, improve the consistency with the biological characteristics of real crops, improve the utilization efficiency of simulation resources, and enable the system to provide effective simulation based on historical experience when facing extreme unknown situations.

[0046] (4) Collect manual modification operations of management strategies by expert users in specific environmental conditions, construct expert demonstration data, and infer the implicit reward function of experts through inverse reinforcement learning algorithm. After the learned implicit reward function is input into the management decision-making agent in a multi-objective weighted form, the system can generate management suggestions that match the decision intention of experts in new environmental conditions. When encountering new situations that have not appeared in historical data, the system can deduce on its own based on the learned reward function and select the action that best matches the values ​​of experts, thereby continuing the decision-making style of experts and enhancing the transparency of decision suggestions and user trust.

[0047] (5) The reconstructed data generated in step S3 ensures that the environmental state responded to by the crop growth agent in step S5 is accurate and reliable, so that the accumulation of stress memory and the transmission of parameters across scenarios are based on reality and avoid memory pollution caused by data noise. The diversified evolutionary trajectory generated by multi-scenario deduction in step S5 is used as additional expert demonstration data or negative samples for the inverse reinforcement learning algorithm in step S6, expanding the coverage of the training set and improving the accuracy of the reward function back-inference. The implicit reward function learned in step S6 can be input into the management decision agent in step S5, so that the decision suggestions generated are closer to real experts. At the same time, the reward function can also be used to guide the weight adjustment of the weighted fusion in step S3, forming a knowledge-driven positive iteration, forming a closed-loop collaboration, and providing intelligent decision-making solutions for smart agriculture. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a preferred embodiment of the present invention;

[0050] Figure 2This is a system block diagram of a preferred embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of the structure of an electronic device according to a preferred embodiment of the present invention. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0053] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0054] Application Overview:

[0055] Existing technologies rely excessively on the stable operation of hardware devices. When sensor failures or communication interruptions lead to data loss, the system's predictive capabilities drop sharply or even fail completely. Crop growth decisions are inherently highly complex: the physiological responses of crops exhibit lag and memory effects, making it difficult to intuitively assess the long-term impact of different decision paths. Although existing technologies can perform scenario simulations, they are essentially static comparisons and cannot simulate the physiological changes of crops after experiencing stress, let alone establish knowledge transfer between different decision paths.

[0056] The implicit experience gained by senior agronomists in practice is difficult to fully express using explicit rules. Existing technologies for manual instruction can only record the explicit operations of experts and cannot understand the deeper intentions behind the experts' adjustments.

[0057] This invention constructs an agronomic causal knowledge graph to quantify the conditional causal effects of various environmental factors on yield at different growth stages. When sensor data is missing, the system dynamically adjusts the weights of spatial and temporal similarity searches based on the current causal effect intensity. This invention overcomes the limitations of static scenario deduction by constructing a digital twin containing a stress response memory module. In adversarial deductions across different decision-making scenarios, the agent experiencing stress transmits its physiological response parameters to agents in other scenarios, achieving cross-scenario knowledge transfer. This invention also overcomes the limitations of manually recording explicit rules by using an inverse reinforcement learning algorithm to infer the implicit reward function from the expert's corrective actions. The learned reward function enables the system to continue the expert's decision-making style in new situations and is transformed into a readable decision rule tree through knowledge distillation.

[0058] Example 1:

[0059] A crop growth prediction method based on digital twins includes:

[0060] S1: Obtain multi-source heterogeneous time-series data of the target plot;

[0061] S2: Construct an agronomic causal knowledge graph to quantify the strength of the conditional causal effects of various environmental factors on crop yield indicators;

[0062] S3: When missing data time series is detected, data is reconstructed based on the agronomic causal knowledge graph to generate reconstructed data;

[0063] S4: Construct a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent has the function of recording the history of environmental stress and dynamically modifying the stress response function.

[0064] S5: Based on the reconstructed data, drive the digital twin to perform multi-agent adversarial simulation and transmit stress response parameters between agents in different decision-making scenarios to calibrate environmental compensation effect parameters;

[0065] S6: Collect the correction operations of expert users, use inverse reinforcement learning to infer the implicit reward function that represents their decision preferences, and generate management suggestions based on the implicit reward function.

[0066] Core challenges include: at the causal-guided data reconstruction level, how to separate universal causal effect coefficients from plot-specific correction factors in historical data, and establish a structural equation model to infer theoretical values ​​based on causal maps when measured data is completely missing; at the multi-agent adversarial inference level with stress memory, the recovery of stomatal conductance and the rehydration compensation effect after drought stress exhibit lag, threshold, and cumulative characteristics, as well as the problem of standardized vector protocols for lossless transmission of physiological parameters in decision-making scenarios; at the level of capturing expert intent through inverse reinforcement learning, the implicit preference weights of experts inferring from a finite number of correction operations are ill-conditioned, requiring the introduction of a hierarchical learning framework to identify the time-varying characteristics of expert intent, while the collaborative closed loop of the three modules also needs to unify the state space representation and control error propagation to avoid the system falling into positive feedback loop.

[0067] Specifically, in step S1, multi-source heterogeneous time-series data of the target plot are acquired. The multi-source heterogeneous time-series data includes remote sensing image data, meteorological data, soil sensor data, and crop phenotypic data.

[0068] Acquisition refers to the means of data collection, including but not limited to satellite remote sensing, drone aerial photography, ground-based IoT sensors, and manual measurement; receiving or actively collecting data from external data sources, and performing necessary format conversion, cleaning, and standardization processing to ensure that the data can be recognized and utilized by the system, thereby establishing a data channel between the physical and digital worlds, which is a prerequisite for subsequent analysis and decision-making.

[0069] Multi-source heterogeneous time series data:

[0070] Multi-source means that the data comes from various types of sensors or platforms, such as satellites, drones, and weather stations. Data from different sources complement each other in terms of spatiotemporal resolution, measurement principles, and characterization dimensions, and together depict the real state of crop growth from multiple perspectives.

[0071] Heterogeneous data refers to data with different structures, formats, and semantics. It has the ability to process heterogeneous data, ensuring subsequent data fusion and knowledge graph construction.

[0072] Time series data are collected continuously in chronological order to form a time sequence. Environmental conditions and crop responses at different growth stages evolve over time. It is necessary to record the time attributes of the data in order to capture growth patterns and analyze lag effects.

[0073] Remote sensing imagery data consists of ground images acquired by sensors mounted on satellites or drones. These images are used to monitor crop growth, identify stressed areas, and estimate biomass. They have a wide coverage area, are highly timely, and can quickly detect abnormal patches in the field.

[0074] Meteorological data, including temperature, precipitation, sunshine duration, wind speed, relative humidity, and evapotranspiration, are provided by meteorological stations or grid-based meteorological forecasting services near the target site.

[0075] Soil sensor data is real-time data on the physical and chemical properties of soil collected by IoT sensors buried in the field, including soil moisture, temperature, and nutrient content at different depths.

[0076] Crop phenotypic data directly reflects the measurement data of the crop's morphology, physiology and biochemical characteristics. It is acquired through field observation, handheld devices or drone remote sensing. Phenotypic data can verify and correct model simulation results.

[0077] This step constructs a multi-dimensional, three-dimensional monitoring system. Single data sources often have limitations, while multi-source heterogeneous data can form redundancy and complementarity. At the same time, the time-series characteristics enable the system to track the continuous trajectory of crop growth, providing a data foundation for subsequent steps to explore causal patterns, construct digital twins with stress memory, and capture experts' decision-making preferences at different times.

[0078] In step S2, an agronomic causal knowledge graph is constructed to store the causal relationships between crops at different growth stages and environmental factors, and to quantify the conditional causal effect strength of each environmental factor on crop yield indicators.

[0079] Specifically, the project involves extracting, organizing, and structuring professional knowledge in the field of agronomy from multi-source data using knowledge engineering techniques and machine learning methods, forming a knowledge base that can be queried, reasoned, and computed by computers. The knowledge is stored and represented in the form of a graph structure, and the construction process is continuously iterated and updated as data accumulates and understanding deepens.

[0080] Causal relationships between different reproductive stages and environmental factors:

[0081] The growth period refers to the various stages of crop growth, such as the sowing period, seedling stage, jointing stage, heading stage, grain-filling stage, and maturity stage; environmental factors include soil moisture, air temperature, light, and nutrients.

[0082] Causal relationships include specific states of environmental factors, changes in crop physiology or yield, and which growth stage of the crop applies to them; enabling the system to reason, conduct counterfactual deductions, and make intervention decisions.

[0083] Quantification is the process of converting qualitative relationships into numerical representations, enabling the system to compare the importance of different factors under different conditions. For example, the causal effect strength of soil moisture during the jointing stage is 0.7, while that of light is 0.2, meaning that the influence of moisture on yield is 3.5 times that of light.

[0084] Yield indicators are the final or intermediate indicators for measuring crop output.

[0085] Conditional causal effect strength refers to the expected change in output indicators caused by a unit change in a certain environmental factor under specific conditions. Using algorithms such as structural equation modeling and instrumental variable method, confounding factors are removed from historical observation data to separate the pure causal effect, and the effect is calculated in stratification according to different conditions.

[0086] Structural equation modeling can handle multiple causes and multiple outcomes simultaneously, and allows variables to influence each other. It can also extract the impact of each factor on output from historical data, that is, how much output changes when the factor changes by one unit while keeping other factors constant.

[0087] Instrumental variables are used to address the confusion between causality and correlation. In agricultural data, causality is easily obscured by various confounding factors. For example, low yields when temperatures are high do not necessarily mean that temperature directly causes the yield reduction. It may be that high-temperature years are often accompanied by drought. Instrumental variables can be used to eliminate the interference of confounding factors, thereby more accurately quantifying the true impact of each environmental factor.

[0088] Step S2 systematizes and quantifies fragmented experiential knowledge in agronomy. Crop yield is the result of the combined effects of multiple environmental factors. By introducing a causal directed acyclic graph to describe the generation mechanism between variables, and then using statistical methods to identify and quantify the strength of each causal path, a causal knowledge graph containing both structure and parameters is finally formed. This provides attention guidance for reconstruction when data is missing, provides causal constraints for multi-scenario inference, ensures that the inference process does not violate basic agronomic physiological laws, and provides an interpretable basis for decision-making recommendations.

[0089] Among them, the directed acyclic graph (DAG) determines the direction and causal relationship of all factors affecting crop yield, enabling the system to think based on causal logic, ensuring that the strength of causal effects can be quantified and counterfactual inferences can be performed.

[0090] In a causal knowledge graph that includes both structure and parameters, the structure represents causal relationships, and the parameters represent the weight of influence.

[0091] Step S3: When a data time series is detected as missing, the weighting of spatial similarity search and temporal similarity search is dynamically adjusted according to the strength of the conditional causal effect of the current reproductive period. Spatial substitute values ​​and historical substitute values ​​are extracted respectively. After causal consistency verification is performed with the strength of the conditional causal effect as a constraint, the data is weighted and fused to generate reconstructed data.

[0092] Specifically, detecting missing data timing is the trigger condition for this step; in agricultural production environments, sensor failure, power outages, and unstable communication are very common occurrences, so the method is designed with these issues in mind and contingency plans prepared for unreliable situations.

[0093] Among them, the detection is the system's real-time monitoring of the data stream of devices such as soil sensors, and automatically identifies data interruptions or anomalies through preset heartbeat mechanisms, timestamp continuity checks or data validity verification algorithms.

[0094] Time-series missing data refers to the discontinuity of data in the time dimension, that is, the complete absence of data within a certain time period.

[0095] The strength of the causal effect of the current growth period determines the priority of subsequent operations. Unlike traditional data completion methods, the proportion of environmental factors affecting crops varies greatly at different stages of crop growth. Therefore, prioritizing is more suitable for crop growth.

[0096] The dynamic adjustment of weighting includes: identifying the set of key environmental factors whose causal effect intensity exceeds the threshold under the current reproductive period conditions; assigning higher weights to land parcel attributes related to key environmental factors in spatial similarity search; and prioritizing the retrieval of historical periods with similar change curves to key environmental factors in temporal similarity search.

[0097] Specifically, the method involves identifying and filtering environmental factors whose causal effect intensity exceeds a threshold by querying a causal knowledge graph, thus forming a clear list.

[0098] The strength of conditional causal effect is a quantitative value obtained from the causal knowledge graph in step S2. It represents the expected impact of a unit change in an environmental factor on the yield index under specific reproductive conditions. The higher the value, the more critical the factor is at the current stage.

[0099] The threshold is a preset critical value used to screen out important environmental factors. Only factors whose causal effect intensity exceeds the threshold are included in the set of key environmental factors.

[0100] The set of key environmental factors consists of one or more environmental factors whose causal effect intensity exceeds a threshold during the current growth period; for example, the set during the heading stage is soil moisture and light, while the set during the maturity stage is temperature and wind speed.

[0101] Dynamically adjusting weighted values ​​determines retrieval attention based on causal importance. In agricultural production, numerous environmental factors influence crop growth, with key environmental factors differing at different growth stages. By introducing the intensity of conditional causal effects from a causal knowledge graph, the system clearly identifies importance levels and then searches accordingly.

[0102] Assigning higher weight to spatial similarity search: Once the system determines that soil moisture is the current key factor, it increases the weight of soil attributes when calculating the similarity between the target plot and its candidate neighbors. Soil attributes directly determine the retention and migration patterns of moisture. Even if two plots are geographically far apart, as long as their soil texture is exactly the same, their moisture change trends during drought will be highly similar. Conversely, geographically adjacent plots with sandy soil or clay soil may have drastically different moisture changes. Therefore, spatial similarity search can avoid being misled by geographical proximity. When all neighbors around the target plot show similarity deviations due to regional climate uniformity, the system can still find similar plots across regions by using soil attribute weights.

[0103] In time similarity searches, priority is given to historical periods with similar curves to the changes in key environmental factors. Crops respond to the environment continuously, and their current physiological state is the result of past environmental accumulation. Based on periods with highly similar historical curves, subsequent environmental evolution patterns can be deduced. For example, if soil moisture is currently experiencing the fifth day of continuous drought, and the curve drops sharply before entering a plateau period, the system will prioritize searching for years that have also experienced a sharp drop followed by a plateau pattern, rather than simply looking for the average of the same period in previous years, thus improving the accuracy of subsequent trend predictions.

[0104] Spatial similarity search includes: constructing a plot similarity map, which uses a graph neural network to represent the similarity of plots in terms of soil properties, planting varieties, and sowing time; identifying a set of neighboring plots whose similarity to the target plot exceeds a preset threshold based on the graph neural network; and extracting spatial substitution values ​​from real-time data of the neighboring plot set.

[0105] The land parcel similarity map is a network of land parcel relationships stored in a graph structure. Each node in the map represents a land parcel, and the edges between nodes represent the degree of similarity between two land parcels. The weight of the edges is determined by the similarity of the land parcels in terms of soil properties, planting varieties, sowing time, etc.

[0106] Graph Neural Networks (GNNs) are deep learning models specifically designed for processing graph-structured data. GNNs can simultaneously learn the features of a node itself and the connections between nodes, updating the representation vector of each node through a neighbor information aggregation mechanism.

[0107] Furthermore, the core idea of ​​graph neural networks is that the final feature of each node is determined by itself and its neighbors. After multiple iterations, the final vector of each plot not only encodes its own soil properties, but also the collective features of the entire plot community. Two plots may not have completely identical direct attributes, but if they are in similar agronomic communities, the model will still determine that they are of reference value. This solves the problems of traditional methods failing to capture implicit associations across features, having low accuracy in searching for spatial similarity, and having fixed similarity measures.

[0108] Soil properties are a set of indicators that describe the physical and chemical characteristics of soil in a plot. They typically include soil texture, organic matter content, pH value, cation exchange capacity, field water holding capacity, wilting coefficient, etc., which determine the soil's water retention capacity, nutrient supply capacity, and root growth environment.

[0109] The planting varieties refer to the crop varieties currently planted in the target plot. Different varieties have significant differences in their water and fertilizer requirements, stress resistance, and growth period, which is an important basis for judging the compatibility between plots.

[0110] Sowing time refers to the specific date on which a crop is sown, usually accurate to the day. Sowing time determines the current growth stage of the crop. Even if the same variety is planted, plots with sowing times that differ by one week will have different growth processes and environmental requirements.

[0111] Based on graph neural network recognition, after the graph neural network is trained, for any given target plot, the model calculates its similarity score with all other plots in the graph, and selects plots with scores exceeding a threshold as neighbors.

[0112] The neighboring plot set is a set of one or more plots that are highly similar to the target plot in key agronomic attributes, as identified by a graph neural network. They are not necessarily the closest geographical neighbors, but may be tens of kilometers apart, but have the same soil texture, are planted with the same variety, and are sown at similar times.

[0113] The preset threshold is a pre-defined threshold value for similarity scores. Only plots with a similarity score exceeding this value are included in the set of neighboring plots. The higher the threshold is set, the more accurate the selected neighbors will be; the lower the threshold is set, the more neighbors there will be but the lower their quality.

[0114] Spatial substitute values ​​are real-time observation values ​​read directly from sensors on neighboring plots at the time points corresponding to the missing data of the target plot.

[0115] The principle behind constructing time similarity search is that, since crop growth responds to environmental changes in a repetitive and periodic manner, the soil moisture change trajectory of the same plot is highly similar under similar meteorological fluctuation patterns. By matching the change curves, the dynamic rhythm of environmental evolution can be captured, making the reconstructed data values ​​more reasonable. Even if the current year has abnormal climate, the system can find periods with similar abnormal patterns in history, avoiding the use of normal year data to apply abnormal year data.

[0116] Time similarity search includes: constructing a historical time series database, extracting the change curves of current key environmental factors, retrieving historical similar periods based on similarity measurement algorithms, and extracting historical substitute value sequences.

[0117] Among them, the construction of the historical time series database involves collecting historical sensor data from the target site and surrounding areas over many years, including continuous time series of key environmental factors such as soil moisture, temperature, precipitation, and light intensity, and storing them in segments according to the growth period to form a standardized historical time series dataset.

[0118] Extracting the change curves of current key environmental factors involves extracting continuous observations of specific time periods and factors from the data segments that are not yet missing within the current reproductive period, forming an ordered data sequence. The environmental variable that has the greatest impact on yield during the current reproductive period is defined by the causal knowledge graph in step S2; the change curves reflect the dynamic evolution pattern.

[0119] Similarity measurement algorithms are mathematical methods used to quantify the degree of similarity between two curves. A commonly used algorithm is dynamic time warping, which allows two curves to stretch and shrink non-linearly on the time axis to find the best matching path. It is suitable for comparing time series data that are slightly different in length but similar in shape.

[0120] Searching for historically similar periods involves performing a similarity query in a historical database. The query is for a time interval from a specific year in history that is highly consistent with the current period in terms of changes in key environmental factors, fluctuation ranges, and trend characteristics. The query returns the top few records with the highest similarity scores.

[0121] Extracting historical substitute value sequences involves extracting factor values ​​from the retrieved historical similar periods within a timeframe of the same length as the current missing data window. If multiple highly similar periods exist, a comprehensive sequence can be generated through weighted fusion, with the weights determined by the similarity scores.

[0122] After extracting spatial and historical substitute values ​​in step S3, the causal consistency is verified by using the strength of conditional causal effect as a constraint, and then the data is weighted and fused to generate reconstructed data. The causal consistency verification includes: verifying whether the substitute values ​​meet the numerical constraints in the causal relationship, and lowering the fusion confidence of those that do not meet the constraints. The weight of the weighted fusion is determined by both the confidence level and the strength of the causal effect.

[0123] Among them, causal consistency verification is a data verification mechanism that checks whether the substitute values ​​obtained from spatial and temporal searches conform to the agronomic laws stored in the causal knowledge graph and verifies whether the data is consistent with the physiological logic of crops.

[0124] The substitute values ​​include spatial substitute values ​​obtained from spatial similarity search and historical substitute values ​​obtained from temporal similarity search.

[0125] Causal relationships are the causal chains and rules stored in the knowledge graph, including numerical boundaries and logical conditions.

[0126] Fusion confidence indicates the degree to which a substitute value is reliable; the higher the confidence, the greater the weight that value will have in subsequent fusion; it is dynamically determined by the verification results.

[0127] Weighted fusion is the process of combining alternative values ​​from multiple sources into a final value according to a certain weight ratio; it is not a simple averaging, but an intelligent combination with evidence and bias.

[0128] Performing causal consistency checks can improve the reasonableness of data. Numerical constraints in causal relationships exist in the form of threshold rules or interval rules. When a substitute value violates the constraints, the confidence level does not simply reset to zero, but is dynamically adjusted according to the degree of violation. The adjusted weight is jointly determined by the confidence level and the strength of the causal effect. The confidence level represents the reliability of the numerical value, and the strength of the causal effect represents the importance of the factor. If the confidence level of the substitute value is high, but it corresponds to an unimportant factor, that is, the strength of the causal effect is low, the overall weight is not high, avoiding wasting effort on minor factors. If the factor is important, but the confidence level of the substitute value is extremely low, the weight will also be lowered to avoid unreliable final decisions.

[0129] Step S4 constructs a digital twin comprising a crop growth agent, an environmental simulation agent, and at least one management decision agent.

[0130] Specifically, a digital twin is a high-fidelity digital mapping of a physical entity in virtual space. It includes dynamic behavior rules, can synchronize physical states in real time, and is a simulation system capable of inference. The digital twin in this application consists of multiple coupled intelligent agents that jointly simulate the complete process of crop growth.

[0131] The crop growth agent is a computational unit in a digital twin that specifically simulates the physiological processes of the crop itself. It encapsulates the growth model of the crop from germination to maturity and can calculate photosynthesis, transpiration, dry matter accumulation, growth period, etc. based on the input environmental data, and output the crop status.

[0132] The environmental simulation agent is the computing unit in the digital twin responsible for simulating the dynamics of the farmland environment. It receives weather forecasts and real-time observation data, and simulates soil moisture transport, soil temperature changes, field microclimates, etc., providing boundary conditions for the crop growth agent.

[0133] The management decision-making intelligent agent is a computing unit in the digital twin that simulates human management measures. Based on preset decision rules or algorithms, it outputs operation instructions such as irrigation, fertilization, and pesticide application, which are then applied to the environmental simulation intelligent agent and the crop growth intelligent agent. In addition, it will receive reward functions from expert intentions to achieve intelligent decision-making.

[0134] The crop growth agent is equipped with a stress response memory function, which records the history of environmental stress and dynamically modifies the stress response function. The stress response memory function enables the crop growth agent to remember past environmental stress events and utilize these memories in future responses.

[0135] Recording the history of environmental stress allows the crop growth agent to save key information about each stress event that occurs; the recorded content includes stress type, intensity, start time, duration, recovery process, etc.

[0136] Dynamic correction means that the correction is not a one-time event, but continues as new coercive events occur.

[0137] The stress response function is a mathematical function that describes the physiological response of crops when faced with environmental stress. Typically, the input is the current environmental conditions and the intensity of stress, and the output is the change in the physiological state of the crop.

[0138] S5: Multi-agent adversarial simulation based on reconstructed data-driven digital twins: At least two decision scenarios are run in parallel. In the first decision scenario, the crop growth agent experiences stress and then transmits the stress response parameters to the crop growth agent in the second decision scenario through the stress response memory module to calibrate its subsequent environmental compensation effect parameters.

[0139] The reconstructed data is generated through step S3. It is a complete data cube generated after causal consistency verification and weighted fusion, containing continuous and reliable environmental factor data of the target plot in the current time period.

[0140] Multi-agent adversarial simulation is a parallel simulation method that runs at least two digital twin copies under different settings simultaneously. Each copy constitutes an independent decision-making scenario. The processes evolve in parallel and are independent of each other, but information exchange is allowed under specific conditions. The adversarial scenario represents different decision-making strategies, and the optimal strategy is finally selected by comparing the results.

[0141] The first decision scenario is one of multiple scenarios running in parallel, usually representing a specific management strategy; it is set as a stress scenario, that is, the crop in the scenario encounters adverse environments such as drought and high temperature.

[0142] The second decision-making scenario is another scenario that runs parallel to the first decision-making scenario, representing another management strategy. It is an information-receiving scenario, that is, receiving the coercive experience from the first scenario to calibrate its own parameters.

[0143] The process of information transfer from the first decision-making scenario to the second decision-making scenario is not simply about copying data. Instead, it involves transferring the response patterns learned by crops under stress in the first scenario to the second scenario through the sharing mechanism of the stress response memory module.

[0144] After receiving the stress response parameters from the first scenario, the second decision scenario does not directly replace its own parameters, but uses them as a reference to correct and adjust its own environmental compensation effect parameters.

[0145] Traditional decision support systems typically use a single model to output a single recommendation, which is limited by the inability to predict the potential consequences of different decision paths. The adversarial simulation process of this application can simulate different choices, obtain quantitative evaluations of each strategy, and preview the results of different choices in virtual space, thereby making more scientific decisions.

[0146] This application introduces a cross-scenario parameter transfer mechanism. The physiological response of crops to environmental stress has inherent regularity. After experiencing stress, the stress response parameters (such as sensitivity change and recovery rate) of the crop growth agent in the first decision scenario are encoded into empirical vectors. :

[0147] ;

[0148] Where τ represents the stress type, γ represents the stress intensity, and λ represents the duration. The parameter state after coercion.

[0149] In the second decision-making scenario, the crop growth agent continuously monitors the shared storage space during operation. When it detects an experience vector related to its potential risk, it calls a calibration function. Correct its own environmental compensation effect parameters:

[0150] ;

[0151] Here, α is the learning rate, which controls the strength of experience transfer. In this way, the second scenario allows for a more accurate understanding of the physiological response after stress, even without actually experiencing stress.

[0152] The stress response parameter transmission includes: encoding the stomatal conductance change curve, photosynthetic inhibition coefficient and rehydration recovery time after stress into a stress memory vector; and transmitting it to crop growth agents in other decision-making scenarios through a shared memory space to correct their rehydration compensation effect model.

[0153] Among them, the stomatal conductance change curve describes the continuous function or discrete sequence of the stomatal aperture of crop leaves changing with time. After stress occurs, stomatal conductance usually shows a trajectory of decrease, minimum point, and recovery.

[0154] The photosynthetic inhibition coefficient is a dimensionless parameter that quantifies the degree of decline in photosynthesis caused by stress. It is expressed as the ratio of the actual photosynthetic rate during stress to the photosynthetic rate under normal conditions, or the percentage decrease; it reflects the degree of damage to the crop's carbon assimilation capacity caused by stress.

[0155] The rehydration recovery time is the length of time required for the crop's physiological functions to return to normal levels after the stress is relieved, measured in hours or days; it reflects the crop's physiological resilience and self-repair ability.

[0156] Stress memory vectors serve as digital carriers of stress experiences, facilitating efficient transmission and storage between agents.

[0157] The shared memory space is a public storage area independent of each decision-making scenario, used to temporarily store the stress memory vectors generated by each scenario; it supports concurrent read and write operations by multiple agents.

[0158] The correction involves the receiver using the acquired stress memory vector to adjust the parameters of its internal rehydration compensation effect model; updating the original parameters based on new information to make the model more accurately reflect the actual observed compensation law.

[0159] The water replenishment compensation effect model is a mathematical model that describes the phenomenon of excess recovery of physiological functions in crops when they regain water supply after experiencing drought stress. It includes parameters such as compensation initiation threshold, compensation intensity coefficient, and compensation duration.

[0160] Encoding into a coerced memory vector compresses the original information into a fixed-dimensional, information-dense vector representation. The specific encoding process includes:

[0161] Let the original stress response information be... ,in This is a curve showing the change in porosity. This is the photosynthetic inhibition coefficient. This refers to the time required for rehydration and recovery.

[0162] Step 1: To Extract key features and construct feature vectors :

[0163] ;

[0164] in, The average slope during the decreasing phase of porosity conductance; The lowest value reached by porosity conductance; The duration of time the value remains near its lowest point; The average slope during the recovery phase; The area under the curve loss represents the total stress loss;

[0165] Step 2: Combine the extracted features with the photosynthetic inhibition coefficient Rehydration recovery time After normalization, the parts are stitched together.

[0166] .

[0167] By setting up step S5, the system quantifies the differences between different management strategies under the same environmental conditions. Cross-scenario parameter transfer further eliminates evaluation bias caused by inaccurate model parameters, making the evolution results of each scenario more comparable. Through the cross-scenario transfer mechanism, the stress response patterns obtained through extensive calculations in one scenario can be reused in other scenarios, avoiding redundant calculations and improving the utilization efficiency of simulation resources. At the same time, the sharing of experience among multiple scenarios allows the cognitive ability of the entire system to continuously accumulate with the increase of simulation times. When the system faces extreme environments that have never been actually observed, by activating multiple adversarial scenarios to simulate different coping strategies and using cross-scenario transfer to quickly obtain estimates of crop physiological responses, the system can still provide evidence-based decision-making references when facing low-probability, high-impact events, rather than simply outputting default values ​​or reporting errors.

[0168] S6: Collect management strategy correction operations from expert users, use inverse reinforcement learning algorithm to infer the implicit reward function of expert users, and use the implicit reward function to represent the decision preferences of expert users; input the implicit reward function into the management decision-making agent to generate management suggestions that match the decision intentions of experts.

[0169] Among them, expert users are senior agronomists or planting experts with rich experience in agricultural production, and their decision-making behavior incorporates tacit knowledge and practical experience.

[0170] The management strategy correction operation is a manual adjustment behavior performed by expert users on the management suggestions automatically generated by the management decision-making agent during system operation, reflecting the expert's opinions on the system suggestions under the current environmental conditions.

[0171] Inverse reinforcement learning is a machine learning method that inversely infers the reward function followed by an expert from the behavioral trajectory demonstrated by that expert. The algorithm assumes that the expert’s behavior is to maximize some unknown reward function, and by observing the expert’s choices in various states, it inversely derives the mathematical form of the reward function and its parameters.

[0172] The implicit reward function is embedded in the expert's decision-making behavior and determines the expert's choice preferences in different situations; it is the output target of inverse reinforcement learning.

[0173] Decision preferences are the multi-objective trade-off criteria that expert users follow when making decisions, such as the degree of pursuit of maximizing output, the tendency to minimize risk, and the degree of importance attached to resource conservation.

[0174] Expert decision intent is the underlying purpose of expert users when making corrective actions, and this intent is mathematically modeled through an implicit reward function.

[0175] By inverse reinforcement learning, the reward function is derived, enabling the system to find actions that maximize cumulative rewards in any new state through planning or optimization. When an expert makes a mistake, the cumulative reward caused by that action is lower than the optimal path, and the inverse reinforcement learning algorithm automatically reduces its impact. The implicit reward function summarizes the expert's decision preferences as a function, which is a set of value judgments. For example, for experts who prefer stability, their implicit reward function will penalize output fluctuations, so that the stable output path will receive a higher score; for experts who prefer high output, their reward function will assign a higher weight to the final output.

[0176] Before obtaining the implicit reward function, the optimization objective of the management decision-making agent is preset. However, a fixed objective cannot adapt to the personalized preferences of different experts. This step inputs the implicit reward function obtained through reverse engineering into the management decision-making agent, changing the evaluation criteria for the agent. Under the new environmental state, the agent enumerates possible candidate management actions, combines the digital twin constructed in step S4 and the reconstructed data generated in step S3, simulates the possible outcome trajectory of each candidate action, and scores each trajectory using the implicit reward function. Finally, the action with the highest score is selected as the management suggestion output, ensuring that the final suggestion is highly consistent with the true intention of the experts.

[0177] By collecting expert feedback during daily interactions and using inverse reinforcement learning to deduce their implicit reward function, the system achieves a quantitative expression of tacit knowledge. Newly hired agronomists can leverage the system to continue the decision-making style of their predecessors, avoiding the industry dilemma of inexperienced newcomers. The implicit reward function can be broken down into multiple interpretable dimensions. When the system outputs management suggestions, it can simultaneously show users the learning results based on expert decision preferences and explain the results. This transparent decision-making basis helps build user trust in the intelligent system.

[0178] The implicit reward function is a multi-objective weighted function, including the objectives of maximizing output, minimizing risk, and conserving resources. The inverse reinforcement learning algorithm learns the weight coefficients of each objective from expert demonstration data, so that under the same environmental conditions, the implicit reward function scores the expert operation higher than the other candidate operation.

[0179] Among them, the multi-objective weighted function is a function that combines multiple decision objectives into a single comprehensive score. Its basic structure is the weighted sum of the characteristic values ​​of each objective and the corresponding weight coefficients; it can express the trade-offs between experts in multiple dimensions such as output, risk, and resources.

[0180] The expert demonstration data is a data set consisting of the expert correction operations collected in step S6 and their corresponding environmental states.

[0181] This implicit reward function considers multiple conflicting objectives. High yield is a fundamental requirement for agricultural production, which faces risks such as weather uncertainty and market fluctuations; excessive pursuit of high yield may amplify these risks. Meanwhile, water scarcity and rising fertilizer costs make resource conservation an increasingly important consideration. Therefore, the system adjusts weights based on patent intent during decision-making and simultaneously analyzes and learns the underlying logic of expert decisions. When encountering new situations, even those never seen in historical data, the system can automatically calculate the action that best aligns with expert preferences in the current state based on the learned weight coefficients. By encoding decision preferences into weight coefficients, the system can simultaneously store the preference configurations of multiple experts. When users want the system to adjust its decision-making style, they do not need to retrain the model; they only need to manually adjust the weight coefficients. This allows the system to quickly respond to changes in user needs and provides intervention tools for continuous system optimization.

[0182] Example 2:

[0183] Based on Example 1, after generating management recommendations that match the expert's decision-making intent, the method further includes: comparing the deduction results of the first decision scenario and the second decision scenario, calculating the expected yield difference and resource consumption difference under different management strategies; based on the expected yield difference and resource consumption difference, combined with the current crop growth stage, selecting the optimal management strategy and its execution parameters from the management recommendations; and using the optimal management strategy and its execution parameters as the final output.

[0184] In step S5, the system obtains the evolution trajectory under different decision-making scenarios through multi-agent adversarial inference. The two most critical dimensions are yield and resource consumption. Simply comparing the yield and resource consumption is insufficient. Therefore, this step introduces a comparison of the two dimensions of yield difference and resource consumption difference to construct a two-dimensional evaluation space. Combined with the current growth stage of the crop, the reliability of the decision is analyzed from multiple perspectives.

[0185] Following step S6 in Example 1, multiple management suggestions are generated for user judgment and selection. This step, by quantifying yield and resource consumption differences, presents the advantages and disadvantages of different strategies with specific numerical values, and automates the selection process. This represents a leap from providing options to providing decision-making support, truly leveraging the decision support function of the intelligent system. By incorporating the current crop growth stage as a selection criterion, the system ensures that the final output management strategy is adapted to the crop's stage-specific physiological needs. For example, during the critical period of yield formation, the system tends to select strategies that guarantee yield; during the water-required maturity stage, the system tends to select resource-saving strategies.

[0186] The selection rules in this step can be flexibly configured according to actual needs. In water-scarce areas, resource conservation priority rules can be set to increase the weight of resource consumption differences; in the construction of high-standard farmland, yield priority rules can be set, enabling the system to adapt to diverse application scenarios. After the final output management strategy is implemented, its actual effect can be compared and verified with the projection results. If there is a deviation between the actual yield and the expected yield, it is possible to trace back to which link in the data reconstruction, stress memory, expert intent learning, or selection rules caused the deviation, thereby guiding the continuous optimization of the system.

[0187] Example 3:

[0188] A crop growth prediction system based on digital twins includes,

[0189] The data acquisition interface is used to acquire multi-source heterogeneous time-series data of the target plot;

[0190] The knowledge graph construction unit is used to construct an agronomic causal knowledge graph and quantify the strength of the conditional causal effects of various environmental factors on crop yield indicators.

[0191] The data reconstruction engine is used to dynamically adjust the weighting of spatial similarity search and temporal similarity search according to the strength of the conditional causal effect of the current reproductive period when data missing is detected. It extracts spatial substitute values ​​and historical substitute values ​​respectively, performs causal consistency verification with the strength of the conditional causal effect as a constraint, and then weights and fuses them to generate reconstructed data.

[0192] A digital twin construction unit is used to construct a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent includes a stress response memory module for recording environmental stress history and dynamically correcting the stress response function.

[0193] The adversarial simulation engine is used to drive the digital twin to perform multi-agent adversarial simulation based on the reconstructed data: at least two decision scenarios are run in parallel, wherein the crop growth agent in the first decision scenario transmits the stress response parameters to the crop growth agent in the second decision scenario through the stress response memory module after experiencing stress, and calibrates its subsequent environmental compensation effect parameters.

[0194] The expert intent capture unit is used to collect the management strategy correction operations of expert users and infer the implicit reward function of expert users through an inverse reinforcement learning algorithm. The implicit reward function represents the decision preferences of expert users.

[0195] The decision generation unit is used to input the implicit reward function into the management decision-making agent to generate management suggestions that match the expert's decision-making intent;

[0196] Output interface, used to output prediction results or management suggestions.

[0197] Specifically, the system corresponds to the method and implements the aforementioned method steps in the form of functional modules in hardware or software. The various components of the system are connected sequentially according to data flow and functional logic to form a complete intelligent decision-making closed loop.

[0198] The data acquisition interface establishes a data channel between the physical world and the digital system, providing raw data input for the digital representation of the target plot. By integrating multi-source heterogeneous data, it constructs a multi-dimensional, three-dimensional monitoring system, providing a comprehensive, real-time, and continuous data foundation for subsequent analysis and avoiding information blind spots that may arise from a single data source. This interface connects to various sensors and data sources integrated across space, air, and ground, including satellite remote sensing platforms, meteorological monitoring networks, ground-based IoT sensors, and mobile field acquisition devices. It acquires remote sensing imagery, meteorological data, soil sensor data, and crop phenotypic data of the target plot in real time. The interface has multi-protocol adaptation and multi-format parsing capabilities, enabling it to unify heterogeneous data sources into a standardized format that the system can process.

[0199] The knowledge graph construction unit transforms qualitative experience and laws in the field of agronomy into formal knowledge that can be understood and computed by machines. This provides causal guidance for the system's data reconstruction and decision inference, forming a computable agronomic causal knowledge base. This enables the system to have causal reasoning capabilities, provides attention guidance for subsequent data reconstruction, provides agronomic constraints for adversarial inference, and provides explanatory basis for decision recommendations.

[0200] The knowledge graph construction unit extracts knowledge from agronomic literature, expert experience, and historical planting data to build a knowledge graph that stores the causal relationships between crops at different growth stages and environmental factors. Furthermore, it quantifies the conditional causal effect of each environmental factor on crop yield indicators through causal inference algorithms, that is, the expected impact of a unit change in a certain environmental factor on yield under specific growth stages and specific soil conditions.

[0201] The data reconstruction engine addresses data loss issues caused by sensor malfunctions and communication interruptions in agricultural scenarios, ensuring continuous and stable system operation. It enables intelligent data completion in the event of missing sensor data, filling temporal gaps and freeing the system from rigid dependence on hardware stability, thus improving robustness and adaptability. The engine monitors the data stream in real time and automatically initiates a reconstruction mechanism when missing temporal data is detected. It obtains the strength of conditional causal effects for the current growth period from knowledge graph construction units, dynamically adjusting the weighting of spatial and temporal similarity searches accordingly, assigning higher search priority to environmental factors with strong causal effects. Spatially, it extracts real-time data from neighboring plots with similar soil properties and planting varieties as spatial substitutes. Temporally, it extracts data from historical periods with similar curves of key environmental factor changes as historical substitutes. The causal consistency of the two types of substitutes is verified by the strength of conditional causal effects, eliminating outliers that violate agronomic principles. Finally, weighted fusion generates reconstructed data that conforms to crop physiological laws.

[0202] The digital twin building unit constructs a high-fidelity virtual copy of the crop-environment system of the target plot, providing a simulation platform for decision-making simulation. It constructs virtual crops with memory, making their behavior closer to the physiological memory effect of real crops, and providing a biologically realistic simulation subject for subsequent multi-agent adversarial simulation.

[0203] The digital twin building block creates a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent is given a stress response memory function, which can record its environmental stress history and dynamically adjust its stress response function when similar stress events occur. The environmental simulation agent is responsible for simulating soil moisture transport and field microclimate, while the management decision agent is responsible for generating management actions.

[0204] The stress response memory module maintains a temporal memory matrix, recording the stress type, intensity, duration, and recovery trajectory. This is used to dynamically adjust the sensitivity and recovery rate parameters of the current stress response function when similar stress events occur.

[0205] The adversarial simulation engine runs multiple decision scenarios in parallel, simulating the long-term consequences of different management strategies in a virtual space. This provides quantitative evidence for scientific decision-making, enabling a leap from static comparison to dynamic adversarial analysis. Cross-scenario parameter transfer allows stress experience gained in one scenario to be shared across all scenarios, improving the utilization efficiency of simulation resources. Simultaneously, the final simulation results for each scenario provide reliable quantitative data for subsequent strategy selection. Driven by reconstructed data output from the data reconstruction engine, the engine runs at least two independent decision scenarios in parallel. During the simulation, if the crop growth agent in the first decision scenario experiences environmental stress, its stress response parameters are encoded as a stress memory vector. This vector is then transferred to the crop growth agent in the second decision scenario through a shared memory space to calibrate the latter's rehydration compensation effect model. Through this cross-scenario knowledge transfer, the model parameters of each scenario are continuously optimized, and the accuracy of the simulation results is continuously improved.

[0206] The expert intent capture unit digitizes the tacit experience of senior agronomists. This unit collects manual adjustments made by expert users to management strategies under specific environmental conditions, and constructs expert demonstration data by associating these adjustments with the environmental conditions at the time of adjustment. An inverse reinforcement learning algorithm is then used to infer the expert's implicit reward function from this demonstration data. This function is expressed in a multi-objective weighted form, including yield maximization, risk minimization, and resource conservation objectives, with weight coefficients quantifying the expert's decision preferences. The expert's tacit knowledge is quantified into a computable, storable, and transferable reward function, allowing the system to continue its decision-making style even after the expert leaves.

[0207] The decision generation unit transforms the learned expert intent into practical management suggestions, completing a closed loop from knowledge to action. It inputs the implicit reward function, derived from the expert intent capture unit, into the management decision agent within the digital twin, enabling the agent to optimize around this reward function. In new environmental states, the agent combines the current reconstructed data with the digital twin's deductive capabilities to evaluate the expected cumulative rewards of multiple candidate management actions, selecting the action with the highest score as the management suggestion matching the expert's decision intent. Simultaneously, this unit also performs a strategy selection function: comparing the deductive results of multiple decision scenarios, calculating expected yield differences and resource consumption differences, and considering the current crop growth stage, selecting the optimal management strategy and its specific execution parameters from multiple candidate suggestions.

[0208] The output interface presents the prediction results and management suggestions generated by the system in a user-understandable form. This interface transmits the optimal management strategy and its execution parameters output by the decision generation unit, as well as the crop growth prediction results generated by the digital twin, to users or intelligent agricultural machinery through visual interfaces, mobile push notifications, voice broadcasts, or direct command issuance. This achieves seamless connection between the system output and the user, ensuring that the decision suggestions can be accurately understood and effectively implemented, and truly realizing the practical value of the intelligent decision-making system.

[0209] Example 4:

[0210] Figure 3 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0211] like Figure 3 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0212] The processor 11 can perform various appropriate actions and processes based on a computer program stored in the read-only memory (ROM) 12 or a computer program loaded from the storage unit 18 into the random access memory (RAM) 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0213] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0214] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as crop growth prediction methods based on digital twins.

[0215] In some embodiments, the digital twin-based crop growth prediction method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the digital twin-based crop growth prediction method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the digital twin-based crop growth prediction method by any other suitable means (e.g., by means of firmware).

[0216] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0217] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0218] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement the crop growth prediction method based on digital twins provided by this invention. The computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0219] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD)) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0220] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0221] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0222] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0223] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A crop growth prediction method based on digital twins, characterized in that, include: S1: Obtain multi-source heterogeneous time-series data of the target plot; S2: Construct an agronomic causal knowledge graph to quantify the strength of the conditional causal effects of various environmental factors on crop yield indicators; S3: When missing data time series is detected, data is reconstructed based on the agronomic causal knowledge graph to generate reconstructed data; S4: Construct a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent has the function of recording the history of environmental stress and dynamically modifying the stress response function. S5: Based on the reconstructed data, drive the digital twin to perform multi-agent adversarial simulation and transmit stress response parameters between agents in different decision-making scenarios to calibrate environmental compensation effect parameters; S6: Collect the correction operations of expert users, use inverse reinforcement learning to infer the implicit reward function that represents their decision preferences, and generate management suggestions based on the implicit reward function.

2. The crop growth prediction method based on digital twins according to claim 1, characterized in that: The data reconstruction includes dynamically adjusting the weighting weights, and the dynamic adjustment of the weighting weights includes: Identify the set of key environmental factors whose causal effects exceed a threshold during the current reproductive period; In spatial similarity searches, higher weights are given to parcel attributes related to key environmental factors; In time similarity searches, historical periods that are similar to the change curves of key environmental factors are prioritized.

3. The crop growth prediction method based on digital twins according to claim 1, characterized in that: The data reconstruction includes causal consistency verification, which includes: Verify whether the substitute values ​​meet the numerical constraints in the causal relationship, and lower the fusion confidence of those that do not meet the constraints. The weights of the weighted fusion are determined by both the confidence level and the strength of the causal effect.

4. The crop growth prediction method based on digital twins according to claim 1, characterized in that: The transmission of stress response parameters between agents in different decision-making scenarios includes: The stomatal conductance change curve, photosynthetic inhibition coefficient, and rehydration recovery time after stress are encoded into a stress memory vector; By sharing the memory space with crop growth agents in other decision-making scenarios, the model of rehydration compensation effect is modified.

5. The crop growth prediction method based on digital twins according to claim 1, characterized in that: The data reconstruction includes spatial similarity search, which includes: Construct a plot similarity map, wherein the plot similarity map is characterized by a graph neural network to represent the similarity of plots in terms of soil properties, planting varieties and sowing time; Based on graph neural network identification, a set of neighboring plots whose similarity to the target plot exceeds a preset threshold; The spatial substitution value is extracted from real-time data of the neighboring plot set.

6. The crop growth prediction method based on digital twins according to claim 1, characterized in that: The implicit reward function is a multi-objective weighted function; the inverse reinforcement learning algorithm learns the weight coefficients of each objective from expert demonstration data, and under the same environmental conditions, the implicit reward function scores the expert operation higher than the other candidate operation.

7. The crop growth prediction method based on digital twins according to claim 1, characterized in that: Following the generation of management recommendations, the following is also included: Compare the simulation results of the first and second decision scenarios, and calculate the differences in expected output and resource consumption under different management strategies; Based on the differences in expected yield and resource consumption, and in conjunction with the current crop growth stage, the optimal management strategy and its execution parameters are selected from the management recommendations. The optimal management strategy and its execution parameters will be used as the final output.

8. A crop growth prediction system based on digital twins, used to implement the crop growth prediction method based on digital twins as described in any one of claims 1-7, characterized in that: include, The data acquisition interface is used to acquire multi-source heterogeneous time-series data of the target plot; The knowledge graph construction unit is used to construct an agronomic causal knowledge graph and quantify the strength of the conditional causal effects of various environmental factors on crop yield indicators. The data reconstruction engine is used to dynamically adjust the weighting of spatial similarity search and temporal similarity search according to the strength of the conditional causal effect of the current reproductive period when data missing is detected. It extracts spatial substitute values ​​and historical substitute values ​​respectively, performs causal consistency verification with the strength of the conditional causal effect as a constraint, and then weights and fuses them to generate reconstructed data. A digital twin construction unit is used to construct a digital twin containing a crop growth agent, an environmental simulation agent, and at least one management decision agent. The crop growth agent includes a stress response memory module for recording environmental stress history and dynamically correcting the stress response function. The adversarial simulation engine is used to drive the digital twin to perform multi-agent adversarial simulation based on the reconstructed data: at least two decision scenarios are run in parallel, wherein the crop growth agent in the first decision scenario transmits the stress response parameters to the crop growth agent in the second decision scenario through the stress response memory module after experiencing stress, and calibrates its subsequent environmental compensation effect parameters. The expert intent capture unit is used to collect the management strategy correction operations of expert users and infer the implicit reward function of expert users through an inverse reinforcement learning algorithm. The implicit reward function represents the decision preferences of expert users. The decision generation unit is used to input the implicit reward function into the management decision-making agent to generate management suggestions that match the expert's decision-making intent; Output interface, used to output prediction results or management suggestions.

9. A crop growth prediction system based on digital twins according to claim 8, characterized in that: The stress response memory module maintains a temporal memory matrix, recording the stress type, intensity, duration, and recovery trajectory, which is used to dynamically adjust the sensitivity parameter and recovery rate parameter of the current stress response function when similar stress events occur.

Images