Method and system for modeling digital twin of farmland based on data and physical dual-core driving

Abstract

This invention discloses a method and system for modeling farmland digital twins based on a dual-core data and physical model, relating to the field of smart agriculture technology. The method includes acquiring sparse spatiotemporal observation data collected at multiple time steps by multiple sensor nodes set up within a target farmland area; extracting the dominant modes representing the evolution of the farmland environment from the sparse spatiotemporal observation data based on a dynamic mode decomposition algorithm, and performing linear combination and interpolation on the dominant modes to reconstruct a continuous spatial distribution field covering the target farmland area, which serves as the current initial field; using the output of the data kernel as the input of the physical kernel; and using the error between the prediction of the physical kernel and real-time observation as a feedback signal to drive the update of the embedding vector. This enables sparse observation → full-field reconstruction → time prediction → online calibration to form an automatically optimized complete closed loop, allowing the digital twin to continuously track the real state of the physical entity.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to a method and system for digital twin modeling of farmland driven by both data and physical cores. Background Technology

[0002] In smart agriculture, building accurate digital twin models of farmland is crucial for improving resource efficiency, but existing technological approaches all face significant bottlenecks in practical applications.

[0003] Currently, two main approaches are relied upon: one is purely data-driven models, such as deep learning, which rely entirely on training with massive amounts of historical data. However, the sparse deployment of sensors in farmland and the high cost of acquiring dense data result in poor prediction accuracy for such models in data-scarce areas. Furthermore, their "black box" nature means that predictions may violate physical laws, leading to insufficient reliability and difficulty in quickly adapting them to different farmlands.

[0004] Second, there are purely physical mechanism models, such as numerical simulations based on partial differential equations. While these models are physically interpretable, they are computationally extremely expensive and difficult to use for real-time prediction. Their accuracy heavily depends on soil parameters and boundary conditions that are difficult to obtain precisely, and their structure is fixed after construction, making it impossible to effectively utilize real-time observation data for online optimization.

[0005] As a closer attempt, hybrid methods such as physical information neural networks try to combine the two, but their training is complex and convergence is difficult. They also fail to effectively solve the initial problem of reconstructing the entire field state from sparse observations, and the cost of personalized model adaptation remains high.

[0006] In summary, existing technologies struggle to achieve a balance between prediction accuracy, physical consistency, and efficient adaptation when facing the core challenges of sparse data, limited computing resources, and the need for rapid personalization in farmland scenarios. This invention aims to overcome this systemic deficiency. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for farmland digital twin modeling based on a dual-core approach of data and physics, and to build a next-generation farmland digital twin system with low data requirements, high computational efficiency, reliable prediction accuracy, and flexible adaptability to different farmlands, providing core technical support for intelligent agricultural decision-making such as precision irrigation and fertilization.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a farmland digital twin modeling method based on a dual-core data and physical model, comprising the following steps: S1. Acquire sparse spatiotemporal observation data collected at multiple time steps by multiple sensor nodes set up in the target farmland area; S2. Based on the dynamic mode decomposition algorithm, the dominant modes representing the evolution law of farmland environment in the sparse spatiotemporal observation data are extracted, and the dominant modes are linearly combined and interpolated to reconstruct a continuous spatial distribution field covering the target farmland area, which is used as the current initial field. S3. Input the current initial field into the pre-trained Fourier neural operator model, and use the embedding vector corresponding to the target farmland area as a conditional input to deduce the farmland environment state prediction field at least one time step in the future. S4. After at least one time step, acquire a new round of sparse spatiotemporal observation data at the corresponding time, freeze the backbone parameters of the Fourier neural operator model, and update only the embedding vector based on the error between the new round of sparse spatiotemporal observation data and the corresponding farmland environmental state prediction field. S5. Based on the Fourier neural operator model, the updated embedding vector is used to re-predict the farmland environmental state prediction field to obtain the farmland environmental state calibration prediction field. S6. Based on the farmland environmental state calibration prediction field, use a pre-configured rule engine or a pre-trained lightweight machine learning decision model to generate structured agricultural production control decision information, and encode the control instructions in the agricultural production control decision information and send them to the corresponding execution device for automatic execution or execution after manual confirmation via a wireless communication protocol.

[0009] Furthermore, the method also includes preprocessing the sparse spatiotemporal observation data before performing S2, including the following steps: The first outlier was obtained by using a sliding window test on the time series of a single sensor node based on the statistical 3σ criterion. By comparing the synchronous observations of adjacent sensor nodes through spatial consistency testing, the second anomaly point is obtained. For the first and second anomalies and missing values ​​detected, the spatiotemporal kriging interpolation method is used to estimate and fill in the missing values ​​using effective data from adjacent time points and adjacent sensor nodes, thus obtaining the spatiotemporal observation data matrix.

[0010] Furthermore, in order to ensure the stability of model training and the accuracy of inference, this invention also requires the standardization of data from different sources and with different dimensions, such as using Z-Score standardization, Min-Max normalization, and one-hot encoding.

[0011] Furthermore, S2 includes the following steps: S2.1 Arrange the sparse spatiotemporal observation data according to the time order and the different sensor nodes to obtain an M×N dimension snapshot matrix, where M represents the number of sensor nodes, N represents the time step length, each column corresponds to the observation data of each sensor node in a time step, and each row corresponds to the time series of observation data of a sensor node. S2.2. Mark the first N-1 columns of the snapshot matrix as the first data matrix and the last N-1 columns as the second data matrix. Perform singular value decomposition on the first data matrix and retain the preset truncation rank according to the singular value energy distribution to obtain the dimension-reduced orthogonal basis, singular value matrix and right singular vector. S2.3. Based on the second data matrix, orthogonal basis, singular value matrix, and right singular vector, construct the best-fitting linear operator in the dimensionality reduction space, and perform eigenvalue decomposition on the linear operator to obtain eigenvalues ​​and eigenvectors. Based on the second data matrix, singular value matrix, right singular vector, and eigenvectors, perform dynamic mode calculation to obtain the corresponding dominant mode vector. The magnitude of the eigenvalue represents the rate of change (growth rate or decay rate) and oscillation frequency of the corresponding dominant mode, and the dominant mode vector represents the spatial distribution shape of the corresponding dominant mode. S2.4. Using the linear combination of all dominant mode vectors as spatial basis functions, interpolate and estimate the spatial locations of undeployed sensor nodes to obtain a continuous spatial distribution field covering the target farmland area, which is then marked as the current initial field.

[0012] Furthermore, S2 can also employ a multi-resolution dynamic mode decomposition algorithm to separate the dominant mode vector into multiple dominant mode sub-vectors at different time scales under different time windows and sampling rates, and construct a continuous spatial distribution field at the corresponding time scale based on the dominant mode sub-vectors to generate the corresponding continuous spatial distribution field at the time scale.

[0013] Furthermore, the training steps for the pre-trained Fourier neural operator model are as follows: Obtain the historical continuous spatial distribution field of the target farmland area, and generate a certain number of simulated continuous spatial distribution fields using a high-fidelity physical model; Collect the historical continuous spatial distribution field and the simulated continuous spatial distribution field to generate a training sample set; Obtain the farmland static attributes corresponding to each training sample in the training sample set, and generate a farmland static attribute vector. The farmland static data includes at least one of latitude and longitude, soil texture classification (encoded as one-hot or continuous vector), crop type, annual average temperature, and annual average precipitation. The static attribute vector of the farmland is input into the encoder to obtain the corresponding embedding vector; Using training samples from multiple time steps as input, corresponding embedding vectors as conditional input, and training samples from at least one future time step relative to the input as output, with the GeLU function as activation function, the backbone parameters of the Fourier neural operator model and the parameters of the encoder are updated simultaneously through backpropagation, so as to minimize the relative L2 error between the predicted training samples output by the Fourier neural operator model and the corresponding actual training samples.

[0014] Furthermore, before the first execution of S3, the embedding vector corresponding to the target farmland area is initially obtained through the following steps: Select at least one of the following for the target farmland area: latitude and longitude, soil texture classification, crop type, annual average temperature, and annual average precipitation. Perform preprocessing to generate the initial static attribute vector of the target farmland. The initial static attribute vector of farmland is input into an encoder trained together with the Fourier neural operator model to obtain the corresponding embedding vector.

[0015] Furthermore, step S3 uses the embedding vector corresponding to the target farmland area as a conditional input, including the following steps: In at least one Fourier layer of the Fourier neural operator model, the embedding vector is transformed into scaling and translation parameters through a fully connected network, and the scaling and translation parameters are used to perform element-wise multiplication and addition modulation on the frequency domain feature map.

[0016] Furthermore, S4 includes the following steps: The cumulative number of steps is obtained by accumulating the time steps of the collected sparse spatiotemporal observation data; When the cumulative number of steps reaches the preset threshold, the farmland environmental state prediction field predicted by the Fourier neural operator model within the time window corresponding to the cumulative number of steps and the collected sparse spatiotemporal observation data are taken as a batch of samples, and the mean square error loss between the two is calculated. Freeze the backbone parameters of the Fourier neural operator model, calculate only the gradient of the mean squared error loss with respect to the embedding vector, optimize the embedding vector by stochastic gradient descent or Adam optimizer to obtain a new embedding vector that minimizes the mean squared error loss, update the current embedding vector with the new embedding vector, and reset the cumulative step count to zero.

[0017] A farmland digital twin modeling system driven by both data and physics, comprising a physical layer, a digital twin model layer, and an application layer; The physical layer includes a sensor network, a wireless communication module, an edge gateway, a cloud, and multiple execution devices; The digital twin model layer includes a data kernel module, a physical kernel module, and an adaptive update module; The application layer includes a decision output module and a display device; The sensor network consists of multiple sensor nodes deployed according to a preset topology. Each sensor node is used to collect observation data in the target farmland area at a preset frequency. The observation data includes at least one of soil moisture, temperature and electrical conductivity at the location of the sensor node. The wireless communication module is used to aggregate the observation data collected by the sensor nodes to the edge gateway or the cloud via a wireless communication protocol; it is also used to send the control commands output by the decision output module to the corresponding execution device via a wireless communication protocol. The execution device includes at least one of a smart water valve, a fertilizer applicator, an alarm, etc., and is used to automatically execute operations according to the control command after receiving the control command. The data core module is deployed in at least one of the edge gateway and the cloud. After receiving the observation data collected by the sensor node and performing preprocessing, it reconstructs the preprocessed observation data into a continuous spatial distribution field covering the target farmland area based on the dynamic mode decomposition algorithm and sends it to the physical core module. The physical kernel module is deployed in at least one of the edge gateway and the cloud. It is used to receive the continuous spatial distribution field of the data kernel module, and to perform inference based on the pre-trained Fourier neural operator model and the embedded vector constructed based on the static features of the target farmland area as the conditional input to obtain the farmland environmental state prediction field for at least one future time step. It is also used to receive the updated embedded vector fed back by the adaptive update module, and to perform inference again based on the updated embedded vector to obtain the farmland environmental state calibration prediction field. The adaptive update module is deployed in at least one of the edge gateway and the cloud. It is used to acquire a new round of sparse spatiotemporal observation data at the corresponding time after at least one time step, freeze the backbone parameters of the Fourier neural operator model, and update only the embedding vector based on the error between the new round of sparse spatiotemporal observation data and the corresponding farmland environmental state prediction field and feed it back to the physical kernel module. The decision output module is deployed in at least one of the edge gateway and the cloud. It is used by the physical core module to obtain the farmland environmental state prediction field or the farmland environmental state calibration prediction field, use a pre-configured rule engine or a pre-trained lightweight machine learning decision model to generate structured agricultural production control decision information, and encode the control instructions in the agricultural production control decision information and send them to the corresponding execution device for automatic execution or execution after manual confirmation through a wireless communication protocol. The display device is used to provide a human-computer interaction interface, and to visualize the continuous spatial distribution field, the farmland environmental state prediction field, and the farmland environmental state calibration prediction field in the form of 2D or 3D models, while also visualizing the agricultural production control decision information of the target farmland area. Beneficial effects

[0018] Compared with existing technologies, the farmland digital twin modeling method and system provided by this invention, driven by both data and physics, utilizes a mode decomposition algorithm to extract low-rank dominant spatiotemporal modes that characterize the intrinsic evolutionary laws of the farmland environmental system from limited spatiotemporal "snapshot" data. Spatial interpolation and reconstruction based on these physically meaningful modes fundamentally improve the data estimation accuracy in unobserved areas, providing reliable continuous input for subsequent predictions and overcoming the bottleneck of pure data methods in data-scarce regions.

[0019] Employing a Fourier neural operator model as an approximate solver for physical laws, this method utilizes global convolution in the frequency domain to learn and simulate complex partial differential equation processes such as soil water, heat, and salt transport at extremely low computational cost via forward propagation. This avoids the drawbacks of traditional numerical methods, such as high computational cost and slow speed, and ensures that the prediction results conform to physical laws through its training data (integrating physical simulations and historical data), thus resolving the issue that pure data models may violate physical laws.

[0020] To address the differences among various farmlands, a "feature embedding vector" with a dimension much smaller than the model's main parameters is introduced to represent the individual characteristics of each farmland. During online updates, this embedding vector is fine-tuned only using optimization methods such as gradient descent, while keeping the main parameters of the pre-trained neural operator model frozen. This "freeze the backbone, optimize the plug" approach allows the model to adapt within minutes using a small amount of new observation data, with minimal computational overhead, achieving a rapid transformation from a "general-purpose model" to a "dedicated twin."

[0021] By using the output of the data kernel as the input of the physical kernel, and the error between the prediction of the physical kernel and the real-time observation as the feedback signal driving the update of the embedding vector, this makes sparse observation → full-field reconstruction → time prediction → online calibration form an automatically optimized complete closed loop, enabling the digital twin to continuously track the real state of the physical entity.

[0022] By clearly defining the entire chain from sensor deployment, data communication, cloud / edge computing to decision generation and execution equipment control, this end-to-end system design ensures that the technical solution is not limited to the algorithm level, but is a complete, deployable, and operational solution, thus enhancing its industrial applicability. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0024] Figure 1 This is a flowchart illustrating the method steps provided in an embodiment of the present invention; Figure 2 This is a system structure block diagram provided for an embodiment of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] Exemplary embodiments will be described more fully below with reference to the accompanying drawings; however, these exemplary embodiments may be embodied in different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will enable those skilled in the art to fully understand the scope of this disclosure.

[0027] Where there is no conflict, the various embodiments of this disclosure and the features thereof in the embodiments may be combined with each other.

[0028] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded.

[0030] The embodiments described herein can be described with reference to plan views and / or cross-sectional views using the ideal schematic diagrams of this disclosure. Therefore, the example illustrations can be modified according to manufacturing techniques and / or tolerances. Therefore, the embodiments are not limited to those shown in the drawings, but include modifications to configurations formed based on manufacturing processes. Therefore, the areas illustrated in the drawings are schematic in nature, and the shapes of the areas shown in the figures illustrate specific shapes of areas of an element, but are not intended to be limiting.

[0031] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and this disclosure, and will not be interpreted as having an idealized or overly formal meaning, unless expressly so defined herein.

[0032] Please see Figure 1 The farmland digital twin modeling method based on a dual-core approach of data and physics includes the following steps: S1. Acquire sparse spatiotemporal observation data collected at multiple time steps by multiple sensor nodes set up within the target farmland area. Specifically: Multiple sensor nodes are deployed in the target farmland area according to a certain topology (such as a grid or Zigbee pattern). Each node integrates one or more sensors for soil moisture, temperature, conductivity, etc., and collects multidimensional sparse spatiotemporal observation data at a set frequency. This data is aggregated to an edge gateway or the cloud via wireless communication modules (such as LoRa or NB-IoT).

[0033] The sparse spatiotemporal observation data is then preprocessed, including anomaly detection and correction, to eliminate erroneous values ​​introduced by sensor noise, communication packet loss, or sporadic failures. This includes the following steps: The first outlier was obtained by using a sliding window test on the time series of a single sensor node based on the statistical 3σ criterion. By comparing the synchronous observations of adjacent sensor nodes through spatial consistency testing, the second anomaly point is obtained. For the first and second anomalies and missing values ​​detected, the spatiotemporal kriging interpolation method is used to estimate and fill in the missing values ​​using effective data from adjacent time points and adjacent sensor nodes, thus obtaining the spatiotemporal observation data matrix.

[0034] Furthermore, in order to ensure the stability of model training and the accuracy of inference, this invention also requires the standardization of data from different sources and with different dimensions, such as using Z-Score standardization, Min-Max normalization, and one-hot encoding.

[0035] S2. Based on the dynamic mode decomposition algorithm, the dominant modes representing the evolution of farmland environment in sparse spatiotemporal observation data are extracted, and the dominant modes are linearly combined and interpolated to reconstruct a continuous spatial distribution field covering the target farmland area, which serves as the current initial field. This includes the following steps: S2.1 Arrange the sparse spatiotemporal observation data according to the time order and the different sensor nodes to obtain an M×N dimension snapshot matrix, where M represents the number of sensor nodes and N represents the time step length. Each column corresponds to the observation data of each sensor node in a time step, and each row corresponds to the time series of observation data of a sensor node. S2.2. Label the first N-1 columns of the snapshot matrix as the first data matrix and the last N-1 columns as the second data matrix. Perform singular value decomposition on the first data matrix. Based on the singular value energy distribution, retain the preset truncation rank (e.g., r=12, to capture more than 95% of the system energy) to obtain the dimension-reduced orthogonal basis, singular value matrix and right singular vector. S2.3. Based on the second data matrix, orthogonal basis, singular value matrix and right singular vector, construct the best-fit linear operator in the dimension reduction space, where the linear algorithm = orthogonal basis * second data matrix * right singular vector * inverse of singular value matrix; Eigenvalues ​​and eigenvectors are obtained by performing eigendecomposition on the linear operator. Dynamic mode calculation is then performed based on the second data matrix, singular value matrix, right singular vector, and eigenvectors to obtain the corresponding dominant mode vector. The dominant mode vector = second data matrix * right singular vector * inverse of singular value matrix * eigenvector, where the magnitude of the eigenvalue represents the rate of change (growth rate or decay rate) and oscillation frequency of the corresponding dominant mode, and the dominant mode vector represents the spatial distribution shape of the corresponding dominant mode. Furthermore, a multi-resolution dynamic mode decomposition algorithm can be used to repeatedly execute S2.2 to S2.3 under different time windows (such as daily variations and seasonal trends) and sampling rates, separating the dominant mode vector into multiple dominant mode sub-vectors at different time scales, and constructing a continuous spatial distribution field at the corresponding time scale based on the dominant mode sub-vectors, thereby generating a corresponding continuous spatial distribution field at the time scale to improve the reconstruction accuracy.

[0036] S2.4. Using the linear combination of all dominant mode vectors as spatial basis functions, interpolation estimation is performed on the spatial locations of undeployed sensor nodes to obtain a continuous spatial distribution field covering the target farmland area, which is marked as the current initial field. This field can include the full-field distribution of variables such as soil moisture, temperature, and electrical conductivity in the target farmland area, thus solving the problem of missing spatial data under sparse sensing conditions and providing complete initial conditions for physical deduction.

[0037] S3. Input the current initial field into the pre-trained Fourier neural operator model, and use the embedding vector corresponding to the target farmland area as a conditional input to deduce the farmland environment state prediction field at least one time step in the future.

[0038] The training steps for the pre-trained Fourier neural operator model are as follows: (1) Obtain the historical continuous spatial distribution field of the target farmland area, and generate a certain number of simulated continuous spatial distribution fields using a high-fidelity physical model. The high-fidelity physical model can be HYDRUS, SWAP, AquaCrop, FEFLOW, etc. For example, the meteorological conditions (rainfall, evapotranspiration), different initial soil moisture / salt distributions, and different irrigation scheme combinations in the past 6-10 years of the target farmland area can be simulated using a high-fidelity physical model to generate 6000-8000 sets of high-fidelity spatiotemporal evolution sequence data. Each set of data includes the initial field and the subsequent 24-72 hours of three-dimensional (two-dimensional spatial + temporal) distribution of soil moisture and salinity with a temporal resolution of 0.5-2 hours.

[0039] (2) Collect historical continuous spatial distribution fields and simulated continuous spatial distribution fields to generate a training sample set; (3) Obtain the farmland static attributes corresponding to each training sample in the training sample set, and generate farmland static attribute vectors. Farmland static data includes at least one of latitude and longitude, soil texture classification (encoded as one-hot or continuous vector), crop type, annual average temperature and annual average precipitation. (4) Input the static attribute vector of farmland into the encoder to obtain the corresponding embedding vector; wherein the encoder is a small neural network, such as a multilayer perceptron (MLP) or a Transformer encoder based on attention mechanism; (5) Using training samples from multiple time steps as input, corresponding embedding vectors as conditional input, and training samples from at least one future time step relative to the input as output, with the GeLU function as activation function, the backbone parameters of the Fourier neural operator model and the encoder parameters are updated simultaneously through backpropagation, so as to minimize the relative L2 error between the predicted training samples and the corresponding actual training samples output by the Fourier neural operator model. Here, using the embedding vector as conditional input means that in at least one Fourier layer of the Fourier neural operator model, the embedding vector is transformed into scale parameters and translation parameters through a fully connected network, and the scale parameters and translation parameters are used to perform element-wise multiplication and addition modulation on the frequency domain feature map.

[0040] By jointly training the encoder with the embedded vector and the Fourier neural operator model, the encoder can learn how to map the static attributes of farmland to a suitable embedding space. This embedding vector can optimally regulate the physical deduction process of the Fourier neural operator model to minimize prediction error. In other words, the encoder learns to compress "farmland identity information" into a "context vector" that the Fourier neural operator model can understand and utilize.

[0041] Furthermore, the encoder obtained through joint training can be used as the base encoder. Then, the backbone parameters of the Fourier neural operator model can be frozen for different farmlands, and only the fine-tuned encoder or the embedding vector itself can be trained to obtain an encoder and initial embedding vector adapted to different farmlands.

[0042] Before the first execution of S3, the embedding vector corresponding to the target farmland area is initially obtained through the following steps: Select at least one of the following for the target farmland area: latitude and longitude, soil texture classification, crop type, annual average temperature, and annual average precipitation. Perform preprocessing to generate the initial static attribute vector of the target farmland. The initial static attribute vector of farmland is input into the encoder trained together with the Fourier neural operator model to obtain the corresponding embedding vector.

[0043] S4. After at least one time step, acquire the new round of sparse spatiotemporal observation data at the corresponding time, freeze the backbone parameters of the Fourier neural operator model, and update only the embedding vector based on the error between the new round of sparse spatiotemporal observation data and the corresponding farmland environmental state prediction field, including the following steps: The cumulative number of steps is obtained by accumulating the time steps of the collected sparse spatiotemporal observation data; When the cumulative number of steps reaches the preset threshold, the farmland environmental state prediction field predicted by the Fourier neural operator model within the time window corresponding to the cumulative number of steps and the collected sparse spatiotemporal observation data are taken as a batch of samples, and the mean square error loss between the two is calculated. Freeze the backbone parameters of the Fourier neural operator model, calculate only the gradient of the mean squared error loss with respect to the embedding vector, optimize the embedding vector by stochastic gradient descent or Adam optimizer to obtain a new embedding vector that minimizes the mean squared error loss, update the current embedding vector with the new embedding vector, and reset the cumulative step count to zero.

[0044] S5. Based on the Fourier neural operator model, the farmland environmental state prediction field is re-predicted using the updated embedding vector to obtain the farmland environmental state calibration prediction field.

[0045] S6. Based on the farmland environmental condition calibration prediction field, using a pre-configured rule engine or a pre-trained lightweight machine learning decision model, structured agricultural production control decision information is generated. The control instructions in this information are encoded and sent to the corresponding execution equipment via wireless communication protocols for automatic execution or execution after manual confirmation. Examples of agricultural production control decision information include: "Start irrigation in area A at 3 PM for 30 minutes, using 5 cubic meters of water," or "The risk of salt accumulation in area B has increased; leaching is recommended this week."

[0046] See Figure 2 The present invention also provides a farmland digital twin modeling system driven by both data and physical layers, including a physical layer, a digital twin model layer and an application layer; The physical layer includes sensor networks, wireless communication modules, edge gateways, the cloud, and multiple execution devices: Sensor network: It consists of multiple sensor nodes deployed according to a preset topology. Each sensor node is used to collect observation data in the target farmland area at a preset frequency. The observation data includes at least one of soil moisture, temperature and electrical conductivity at the location of the sensor node. Wireless communication module: It aggregates the observation data collected by the sensor nodes to the edge gateway or cloud via wireless communication protocol; it is also used to send the control commands output by the decision output module to the corresponding execution device via wireless communication protocol. Execution device: including at least one of intelligent water valve, fertilizer applicator, alarm, etc., used to automatically execute operation according to control command after receiving control command.

[0047] The digital twin model layer includes a data kernel module, a physical kernel module, and an adaptive update module. Data kernel module: Deployed in at least one of the edge gateway and the cloud, it is used to reconstruct the preprocessed observation data into a continuous spatial distribution field covering the target farmland area based on the dynamic mode decomposition algorithm after receiving the observation data collected by the sensor node and performing preprocessing, and then send it to the physical kernel module. Physical kernel module: Deployed in at least one of the edge gateway and the cloud, it is used to receive the continuous spatial distribution field of the data kernel module, and perform inference based on the pre-trained Fourier neural operator model and the embedded vector constructed based on the static features of the target farmland area as the conditional input to obtain the farmland environmental state prediction field for at least one future time step; it is also used to receive the updated embedded vector fed back by the adaptive update module, and perform inference again based on the updated embedded vector to obtain the farmland environmental state calibration prediction field; Adaptive update module: Deployed in at least one of the edge gateway and the cloud, it is used to acquire a new round of sparse spatiotemporal observation data at the corresponding time after at least one time step, freeze the backbone parameters of the Fourier neural operator model, and update only the embedding vector based on the error between the new round of sparse spatiotemporal observation data and the corresponding farmland environmental state prediction field and feed it back to the physical kernel module.

[0048] The application layer includes a decision output module, a wireless communication module, and a display device. Decision output module: Deployed in at least one of edge gateway and cloud, used by the physical core module to obtain farmland environmental state prediction field or farmland environmental state calibration prediction field, use a pre-configured rule engine or a pre-trained lightweight machine learning decision model to generate structured agricultural production control decision information, and encode the control instructions in the agricultural production control decision information and send them to the corresponding execution device for automatic execution or execution after manual confirmation via wireless communication protocol. Display devices: used to provide a human-computer interaction interface, visually presenting the continuous spatial distribution field, farmland environmental state prediction field, and farmland environmental state calibration prediction field in the form of 2D or 3D models, while also visually presenting agricultural production control decision information for the target farmland area.

[0049] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for digital twin modeling of farmland based on a dual-core approach of data and physics, characterized in that, Includes the following steps: S1. Acquire sparse spatiotemporal observation data collected at multiple time steps by multiple sensor nodes set up in the target farmland area; S2. Based on the dynamic mode decomposition algorithm, the dominant modes representing the evolution law of farmland environment in the sparse spatiotemporal observation data are extracted, and the dominant modes are linearly combined and interpolated to reconstruct a continuous spatial distribution field covering the target farmland area, which is used as the current initial field. S3. Input the current initial field into the pre-trained Fourier neural operator model, and use the embedding vector corresponding to the target farmland area as a conditional input to deduce the farmland environment state prediction field at least one time step in the future. S4. After at least one time step, acquire a new round of sparse spatiotemporal observation data at the corresponding time, freeze the backbone parameters of the Fourier neural operator model, and update only the embedding vector based on the error between the new round of sparse spatiotemporal observation data and the corresponding farmland environmental state prediction field. S5. Based on the Fourier neural operator model, the updated embedding vector is used to re-predict the farmland environmental state prediction field to obtain the farmland environmental state calibration prediction field. S6. Based on the farmland environmental state calibration prediction field, use a pre-configured rule engine or a pre-trained lightweight machine learning decision model to generate structured agricultural production control decision information, and encode the control instructions in the agricultural production control decision information and send them to the corresponding execution device for automatic execution or execution after manual confirmation via a wireless communication protocol.

2. The method for digital twin modeling of farmland based on dual-core data and physical processing as described in claim 1, characterized in that, The method further includes preprocessing the sparse spatiotemporal observation data before performing S2, including the following steps: The first outlier was obtained by using a sliding window test on the time series of a single sensor node based on the statistical 3σ criterion. By comparing the synchronous observations of adjacent sensor nodes through spatial consistency testing, the second anomaly point is obtained. For the first and second anomalies and missing values ​​detected, the spatiotemporal kriging interpolation method is used to estimate and fill in the missing values ​​using effective data from adjacent time points and adjacent sensor nodes, thus obtaining the spatiotemporal observation data matrix.

3. The method for digital twin modeling of farmland based on dual-core data and physical processing as described in claim 1, characterized in that, S2 includes the following steps: S2.1 Arrange the sparse spatiotemporal observation data according to the time order and the different sensor nodes to obtain an M×N dimension snapshot matrix, where M represents the number of sensor nodes, N represents the time step length, each column corresponds to the observation data of each sensor node in a time step, and each row corresponds to the time series of observation data of a sensor node. S2.

2. Mark the first N-1 columns of the snapshot matrix as the first data matrix and the last N-1 columns as the second data matrix. Perform singular value decomposition on the first data matrix and retain the preset truncation rank according to the singular value energy distribution to obtain the dimension-reduced orthogonal basis, singular value matrix and right singular vector. S2.

3. Based on the second data matrix, orthogonal basis, singular value matrix, and right singular vector, construct the best-fitting linear operator in the dimensionality reduction space, and perform eigenvalue decomposition on the linear operator to obtain eigenvalues ​​and eigenvectors. Based on the second data matrix, singular value matrix, right singular vector, and eigenvectors, perform dynamic mode calculation to obtain the corresponding dominant mode vector. The magnitude of the eigenvalue represents the rate of change and oscillation frequency of the corresponding dominant mode, and the dominant mode vector represents the spatial distribution shape of the corresponding dominant mode. S2.

4. Using the linear combination of all dominant mode vectors as spatial basis functions, interpolate and estimate the spatial locations of undeployed sensor nodes to obtain a continuous spatial distribution field covering the target farmland area, which is then marked as the current initial field.

4. The farmland digital twin modeling method based on data and physical dual-core driven according to claim 3, characterized in that, The S2 can also employ a multi-resolution dynamic mode decomposition algorithm to separate the dominant mode vector into multiple dominant mode sub-vectors at different time scales under different time windows and sampling rates, and construct a continuous spatial distribution field at the corresponding time scale based on the dominant mode sub-vectors to generate the corresponding continuous spatial distribution field at the time scale.

5. The method for digital twin modeling of farmland based on dual-core data and physical processing as described in claim 1, characterized in that, The training steps for the pre-trained Fourier neural operator model are as follows: Obtain the historical continuous spatial distribution field of the target farmland area, and generate a certain number of simulated continuous spatial distribution fields using a high-fidelity physical model; Collect the historical continuous spatial distribution field and the simulated continuous spatial distribution field to generate a training sample set; Obtain the farmland static attributes corresponding to each training sample in the training sample set, and generate a farmland static attribute vector. The farmland static data includes at least one of latitude and longitude, soil texture classification, crop type, annual average temperature and annual average precipitation. The static attribute vector of the farmland is input into the encoder to obtain the corresponding embedding vector; Using training samples from multiple time steps as input, corresponding embedding vectors as conditional input, and training samples from at least one future time step relative to the input as output, with the GeLU function as activation function, the backbone parameters of the Fourier neural operator model and the parameters of the encoder are updated simultaneously through backpropagation, so as to minimize the relative L2 error between the predicted training samples output by the Fourier neural operator model and the corresponding actual training samples.

6. The method for digital twin modeling of farmland based on dual-core data and physical processing as described in claim 1, characterized in that, Before the first execution of S3, the embedding vector corresponding to the target farmland area is initially obtained through the following steps: Select at least one of the following for the target farmland area: latitude and longitude, soil texture classification, crop type, annual average temperature, and annual average precipitation. Perform preprocessing to generate the initial static attribute vector of the target farmland. The initial static attribute vector of farmland is input into an encoder trained together with the Fourier neural operator model to obtain the corresponding embedding vector.

7. The method for digital twin modeling of farmland based on dual-core data and physical processing as described in claim 1, characterized in that, The step S3, which uses the embedding vector corresponding to the target farmland area as a conditional input, includes the following steps: In at least one Fourier layer of the Fourier neural operator model, the embedding vector is transformed into scaling and translation parameters through a fully connected network, and the scaling and translation parameters are used to perform element-wise multiplication and addition modulation on the frequency domain feature map.

8. The method for digital twin modeling of farmland based on dual-core data and physical processing as described in claim 1, characterized in that, S4 includes the following steps: The cumulative number of steps is obtained by accumulating the time steps of the collected sparse spatiotemporal observation data; When the cumulative number of steps reaches the preset threshold, the farmland environmental state prediction field predicted by the Fourier neural operator model within the time window corresponding to the cumulative number of steps and the collected sparse spatiotemporal observation data are taken as a batch of samples, and the mean square error loss between the two is calculated. Freeze the backbone parameters of the Fourier neural operator model, calculate only the gradient of the mean squared error loss with respect to the embedding vector, optimize the embedding vector by stochastic gradient descent or Adam optimizer to obtain a new embedding vector that minimizes the mean squared error loss, update the current embedding vector with the new embedding vector, and reset the cumulative step count to zero.

9. A farmland digital twin modeling system based on a dual-core data and physical modeling architecture, used to execute the farmland digital twin modeling method based on a dual-core data and physical modeling architecture as described in any one of claims 1-8, characterized in that, A farmland digital twin modeling system driven by both data and physics, comprising a physical layer, a digital twin model layer, and an application layer; The physical layer includes a sensor network, a wireless communication module, an edge gateway, a cloud, and multiple execution devices; The digital twin model layer includes a data kernel module, a physical kernel module, and an adaptive update module; The application layer includes a decision output module and a display device; The sensor network consists of multiple sensor nodes deployed according to a preset topology. Each sensor node is used to collect observation data in the target farmland area at a preset frequency. The observation data includes at least one of soil moisture, temperature and electrical conductivity at the location of the sensor node. The wireless communication module is used to aggregate the observation data collected by the sensor nodes to the edge gateway or the cloud via a wireless communication protocol; it is also used to send the control commands output by the decision output module to the corresponding execution device via a wireless communication protocol. The execution device includes at least one of a smart water valve, a fertilizer applicator, and an alarm, and is used to automatically execute operations according to the control command after receiving the control command. The data core module is deployed in at least one of the edge gateway and the cloud. After receiving the observation data collected by the sensor node and performing preprocessing, it reconstructs the preprocessed observation data into a continuous spatial distribution field covering the target farmland area based on the dynamic mode decomposition algorithm and sends it to the physical core module. The physical kernel module is deployed in at least one of the edge gateway and the cloud. It is used to receive the continuous spatial distribution field of the data kernel module, and to perform inference based on the pre-trained Fourier neural operator model and the embedded vector constructed based on the static features of the target farmland area as the conditional input to obtain the farmland environmental state prediction field for at least one future time step. It is also used to receive the updated embedded vector fed back by the adaptive update module, and to perform inference again based on the updated embedded vector to obtain the farmland environmental state calibration prediction field. The adaptive update module is deployed in at least one of the edge gateway and the cloud. It is used to acquire a new round of sparse spatiotemporal observation data at the corresponding time after at least one time step, freeze the backbone parameters of the Fourier neural operator model, and update only the embedding vector based on the error between the new round of sparse spatiotemporal observation data and the corresponding farmland environmental state prediction field and feed it back to the physical kernel module. The decision output module is deployed in at least one of the edge gateway and the cloud. It is used by the physical core module to obtain the farmland environmental state prediction field or the farmland environmental state calibration prediction field, use a pre-configured rule engine or a pre-trained lightweight machine learning decision model to generate structured agricultural production control decision information, and encode the control instructions in the agricultural production control decision information and send them to the corresponding execution device for automatic execution or execution after manual confirmation through a wireless communication protocol. The display device is used to provide a human-computer interaction interface, and to visualize the continuous spatial distribution field, the farmland environmental state prediction field, and the farmland environmental state calibration prediction field in the form of 2D or 3D models, while also visualizing the agricultural production control decision information of the target farmland area.

Images