A method and system for crop growth function prediction based on deep operator networks

Abstract

The application discloses a crop growth function prediction method and system based on a deep operator network, and comprises the following steps: collecting environment parameter vectors aligned according to timestamps; inputting a trained prediction model to obtain a crop growth function and output a prediction result. The prediction model comprises a recurrent neural network module and an improved deep operator network module. In the recurrent neural network module, the environment parameter vectors are taken as inputs to obtain time sequence feature vectors. The improved deep operator network module comprises a branch network and a trunk network; the branch network takes the time sequence feature vectors as inputs; the trunk network takes evaluation points as inputs to generate trunk vectors. A growth state vector is obtained from the outputs of the crop growth operators at the evaluation points. Thus, the deep operator network and the recurrent neural network are fused to form a continuous dynamic crop growth function, the problems of low prediction accuracy and weak generalization ability of existing models are overcome, continuous spatiotemporal data can be processed with less required computing power, and the method can be deployed in most planting scenes.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a method and system for predicting crop growth functions based on deep operator networks. Background Technology

[0002] Crop growth forecasting plays a crucial guiding role in agricultural production. A relatively simple and common method involves measuring ambient temperature and calculating the effective accumulated temperature required for plant growth to predict the critical growth stages of crops from sowing to maturity. With the development of intelligent technologies, the parameters used in crop growth forecasting are often increased, coupled models are established to improve forecast reliability, or more reliable forecasting models such as long short-term memory neural networks are adopted.

[0003] In the aforementioned existing prediction methods, various parameters are processed by discretizing them into time points or intervals. Among them, coupled models can integrate multivariate and LSTM time series models to process time series. However, as exemplified above, the various existing prediction models essentially approximate the crop growth process by learning and fitting discrete time points or fixed interval sequence data. They only make predictions based on historical data, resulting in poor prediction results and weak generalization ability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a crop growth function prediction method and system based on deep operator networks to achieve reliable prediction of crop growth.

[0005] In a first aspect, a crop growth function prediction method based on a deep operator network according to an embodiment of this application includes the following steps: Collect environment parameter vectors aligned to timestamps ; The obtained environmental parameter vector The environmental parameter vector is input into the trained prediction model to obtain the crop growth function, and outputs the crop growth prediction result; Includes d-dimensional environmental parameter components; The crop growth function is obtained through a trained prediction model, which includes a recurrent neural network module and the environmental parameter vector. As input, for each time step The network unit receives the current input environmental parameter vector. In addition to the hidden state and unit state of the previous time step, the state is updated through an internal control mechanism; finally, global average pooling is performed on the hidden states of all time steps along the time dimension to obtain a fixed-dimensional temporal feature vector. The improved deep operator network module includes a branch network and a backbone network; the branch network uses temporal feature vectors. As the input function; the time series feature vector Each one-dimensional feature is configured with a light quantum network to output a A dimensional vector; the outputs of all h subnetworks are concatenated through the first fusion layer to generate the vector. The branch vector b in dimension; the backbone network with evaluation points As input, t For the current time step, the evaluation point will be... Decomposed into d+1 Each component is configured with a light quantum network to receive a scalar input and output a p-dimensional vector, and then the second fusion layer is used to... d+1 The outputs of each subnetwork are fused to generate a p-dimensional backbone vector. C The crop growth operator G, as the crop growth function, is expressed as the output at the evaluation point y as follows: ; This is the predicted growth state vector.

[0006] Its effect is as follows: This application presents a lightweight prediction model that integrates deep operator networks and recurrent neural networks to construct a continuous and dynamic crop growth function. This overcomes the problems of low prediction accuracy and weak generalization ability in existing prediction models based on discrete time points or fixed-interval sequence data. It can process continuous spatiotemporal data, integrate environmental parameters and crop growth status in real time, and generate a high-precision, highly generalized crop growth function for crop growth prediction. Furthermore, it requires less computing power and can be deployed in most planting scenarios.

[0007] Furthermore, the current crop growth state vector is obtained through a visual neural network. This includes: identifying and segmenting the pixel-level mask of each crop in the acquired images; for each crop, reconstructing a 3D point cloud model based on multiple images taken at different times or from different perspectives, and converting the point cloud to a world coordinate system with physical scale by combining camera calibration parameters; defining the growth reference plane and direction, analyzing the bottom point set at the interface between the plant point cloud and the soil, determining the growth reference plane representing the ground plane through a plane fitting algorithm, and defining its normal direction as the main growth direction of the plant, in order to extract the growth state vector. .

[0008] Furthermore, it also includes plant height extraction: calculating the projection values ​​of all point coordinates in the plant point cloud onto the growth direction, and taking the difference between the maximum and minimum projection values ​​as the accurate plant height h.

[0009] Furthermore, plant height extraction also includes: extracting the ordered center point sequence of the plant point cloud, and classifying the height range... Discretize into m There are 1, 2, 3, 4, 5, 6, 7, 8, 9, 1 and These represent the minimum and maximum height projections in the point cloud, respectively, for each interval. k Then calculate the average position of all points within the interval as the center point. This yields an ordered sequence of center points. Arranged in ascending order of height. Corresponding to the top; based on an ordered sequence of center points, bending length L The approximation is obtained by summing the Euclidean distances between adjacent points: .

[0010] Further steps include stem thickness extraction, identifying stem regions in plant point clouds, taking a cross-sectional point cloud slice at a preset height above the ground, fitting two-dimensional circles or ellipses to the points within the slice, and using the fitted equivalent diameter as the stem thickness d; and the crown width obtained by converting pixel area with physical scale.

[0011] Furthermore, it is characterized by including a prediction model based on the current growth state vector. The prediction model is then updated by freezing most of its parameters and performing incremental training.

[0012] Furthermore, the incremental training includes the following sub-steps: data alignment, aligning the environment parameter vector... Growth state vector Perform time alignment; perform real-time context encoding to build a lightweight real-time feature encoder. Its input is the alignment data pair at the current time. The encoder outputs a real-time context encoding vector, along with statistical features calculated from the data cache queue. An adaptive offset is generated based on the context encoding vector to construct the adjusted backbone vector and branch vector, thereby updating the deep operator network module.

[0013] Furthermore, generating adaptive offsets based on the context encoding vector to construct adjusted backbone and branch vectors, thereby updating the deep operator network module, includes: adaptive offset generation, constructing a lightweight adaptive offset generator. Adaptive offset generator Includes two sub-networks and During prediction, the frozen, pre-trained master prediction model is first run once, and the intermediate features, including backbone features, generated by its deep operator network during the prediction process are recorded. b base and branching features C base Subsequently, the real-time context encoding vector Respectively with the main characteristics b base and branching features C base Fusion, input two subnetworks and Generate the corresponding trunk offset vector and branch offset vector The crop growth function is dynamically adjusted through learnable gain parameters. and By combining the offset with the base vector, we obtain the adjusted main vector and branch vectors: .

[0014] Furthermore, the prediction model training process employs supervised learning, and based on the diagonal sparsity of the Jacobian matrix blocks formed by the independent parameters of each sub-network in the branch network and the backbone network, a forward-mode automatic differentiation algorithm is used for gradient calculation. During gradient calculation, the local Jacobian matrix corresponding to each sub-network is calculated independently, reducing the overall computational complexity from that of traditional backpropagation. Reduce to P represents the total number of parameters. Let be the number of parameters for the i-th subnetwork.

[0015] Secondly, this application discloses a crop growth function prediction system based on deep operator networks, characterized in that it includes a memory and a processor, wherein the memory stores a computer program to implement the crop growth function prediction method based on deep operator networks, and the processor performs read and write operations on the memory. Attached Figure Description

[0016] Figure 1 This is a simplified flowchart illustrating a crop growth function prediction method based on deep operator networks according to some embodiments of this application. Figure 2 This is a simplified network structure diagram of a prediction module according to some embodiments of this application. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] refer to Figure 1 The crop growth function prediction method based on deep operator networks, as described above, specifically includes the following steps: Collect environment parameter vectors aligned to timestamps ; The obtained environmental parameter vector The environmental parameter vector is input into the trained prediction model to obtain the crop growth function, and outputs the crop growth prediction result; This includes d-dimensional environmental parameter components, that is, the environmental parameter vector. It includes d environmental parameter components.

[0019] The crop growth function output is a crop growth function, which is used to calculate the crop growth function at any time point t and any environmental parameter vector. Predict the growth status of crops.

[0020] Next, the network structure of the prediction model will be explained, combined with... Figure 2 It is understood that the prediction model includes a recurrent neural network module and an improved deep operator network module.

[0021] In the recurrent neural network module, the environmental parameter vector As input, feature representations containing temporal context information are extracted; specifically, this embodiment employs a Long Short-Term Memory (LSTM) network, with its hidden state dimension set to... h For each time step The network unit receives the current input. In addition to the hidden state and unit state of the previous time step, the state is updated through an internal control mechanism; finally, global average pooling is performed on the hidden states of all time steps along the time dimension to obtain a fixed-dimensional temporal feature vector. .

[0022] The improved deep operator network in this embodiment is constructed based on a conventional deep operator network (DeepONet), which also includes a branch network and a backbone network. In this embodiment, the branch network processes the temporal feature vectors obtained by the aforementioned recurrent application network. The main differences between the improved deep operator network module and the conventional deep operator network (DeepONet) are as follows.

[0023] Because crop growth conditions are constantly changing, the improved deep operator network needs to be updated frequently in order to meet the requirements of lightweight design. The specific update method will be described in subsequent embodiments. Here, we will first explain the lightweight design.

[0024] The branch network uses time-series feature vectors As an input function; each branch network outputs a branch vector. b The branch vector b for p dimensional vector, The backbone network is based on evaluation points. As input, the output backbone vector C Similarly p dimensional vector, Then, the output of the crop growth operator G at the evaluation point y is used as the crop growth function, which is specifically implemented by the dot product of the two vectors mentioned above:

[0025] Where G is the crop growth operator to be learned, which serves as the crop growth function and the input function. u Characterizing historical environment and growth context, assessment points Specify the prediction time point and environmental conditions, and output The predicted growth state vector .

[0026] In this way, the current real-time prediction results can be displayed on the terminal device to provide users with reference, such as the current predicted growth trend. Furthermore, it can be combined with knowledge graphs and large language models, and the corresponding models can be directly called through API interfaces to output corresponding planting suggestions based on the output growth prediction results.

[0027] In this embodiment, to reduce model complexity, the multidimensional input is decomposed into independent one-dimensional coordinates, which are then processed separately by a lightweight quantum network; specifically, the evaluation points... Include d +1 component, namely time t and d For each environmental parameter, a simple subnetwork is constructed, such as a single-layer or multi-layer fully connected network, which processes their respective scalar inputs in parallel and independently.

[0028] Therefore, the backbone network, which originally processed the high-dimensional vector y directly, is now changed to... d It consists of +1 subnetworks, each subnetwork receiving one component, such as t and Then output a p These are local feature vectors of dimension 1. Then, a fusion layer is used to combine these local features into the final vector. p dimensional main vectorC For example, a linear transformation after splicing can be used to obtain the spliced ​​backbone vector. C .

[0029] Similarly, the branch network uses temporal feature vectors As the input function; the time series feature vector Each one-dimensional feature is configured with a light quantum network for processing, and each light quantum network outputs a... A dimensional vector; the outputs of all h subnetworks are input to the fusion layer and concatenated to generate the vector. The branch vector b is of dimension 1. The specifics are similar to those of the backbone network and will not be elaborated upon here.

[0030] Furthermore, in order to reduce training costs and network complexity, the prediction model training process adopts supervised learning, and based on the diagonal sparsity of the Jacobian matrix block formed by the independent parameters of each sub-network in the branch network and the backbone network, a forward mode automatic differentiation algorithm is used to calculate the gradient.

[0031] During gradient calculation, the local Jacobian matrix for each sub-network is calculated independently, reducing the overall computational complexity compared to traditional backpropagation. Reduce to P represents the total number of parameters. Let be the number of parameters for the i-th subnetwork.

[0032] Therefore, since the subnetworks independently process one-dimensional coordinates, the Jacobian matrix (the derivative matrix of the loss function with respect to the parameters) of the overall network exhibits a block diagonal structure. The parameter block corresponding to each subnetwork is only related to the input and output of that subnetwork and is independent of other subnetworks. This sparsity means that when calculating the gradient, it is only necessary to calculate the local Jacobian matrix of each subnetwork separately, without having to deal with the dense multiplication of the full parameter matrix.

[0033] Furthermore, in this application, the output of the crop growth function is a growth state vector (dimension...). m Typically, the dimensionality is small (e.g., 2-3), while the input parameters include environmental parameters, time, and network weights, resulting in a high overall input dimension. Therefore, using the forward mode for automatic differentiation is more efficient.

[0034] Specifically, in some embodiments, the forward propagation mode accumulates gradients by calculating the directional derivative. For scalar loss functions... ( (as a parameter vector), the forward mode calculates each parameter sequentially. directional derivative Its calculation process is carried out simultaneously with the function evaluation, and there is no need to store the complete Jacobian matrix of intermediate variables.

[0035] For example, this can be achieved using a deep learning framework such as the forward automatic differentiation interface provided by PyTorch, or through dual-mode programming. The computational complexity of the forward mode is linearly related to the number of parameters and has low memory consumption, making it suitable for the lightweight decomposed network structure in this application.

[0036] Therefore, this application constructs a lightweight prediction model that integrates deep operator networks and recurrent neural networks to build a continuous and dynamic crop growth function. This overcomes the problems of low prediction accuracy and weak generalization ability in existing prediction models based on discrete time points or fixed-interval sequence data. It can process continuous spatiotemporal data, integrate environmental parameters and crop growth status in real time, and generate a high-precision, highly generalized crop growth function for crop growth prediction. Furthermore, it requires less computing power and can be deployed in most planting scenarios.

[0037] To better illustrate this, let's look at a specific example: Assume the network output is loss function Where F represents the prediction model, and N is the total number of samples in the training set. Gradient calculation in the forward mode. At that time, for each parameter This allows for the calculation of the directional derivative. During the calculation, the intermediate derivative is decomposed into multiple independent sub-networks as a parameter. The calculation of each sub-network only involves local paths, thus reducing the amount of computation.

[0038] The crop growth function described in the above embodiments is trained using historical data, which includes environmental parameter vectors observed during past growth processes. and crop growth state vectors in historical data The trained prediction model is obtained by performing time-aligned training.

[0039] In some preferred embodiments, considering the rapid changes in crop growth, the prediction model needs to be updated in real time to ensure continuous and accurate prediction of the current crop growth trend. Therefore, the method also uses a visual neural network to obtain the current crop growth state vector. include: The pixel-level mask of each crop in the acquired image is identified and segmented. For example, the image obtained by the camera is processed based on the segmentation model MaskR-CNN. The model identifies the outline of each crop in the image and separates it from the background, weeds and soil to generate a binary mask for each plant.

[0040] For each crop, a 3D point cloud model is reconstructed based on multiple images taken at different times or from different viewpoints. The point cloud is then transformed to a physically scaled world coordinate system using camera calibration parameters. For example, for a monocular RGB image sequence, multi-view stereo vision (MVS) technology is used to reconstruct a sparse or dense 3D point cloud from multiple images of the same plant taken at different times or from slightly different viewpoints at the same time. Combined with camera calibration parameters, the image pixel coordinates can be transformed to a physically scaled world coordinate system.

[0041] Furthermore, since crops may grow at an angle and bend, directly measuring the vertical height of the image is inaccurate. This method constructs a growth reference plane. By analyzing the point set at the bottom of the plant point cloud (where it meets the soil), a plane is fitted as the ground plane. The main growth direction of the plant is defined as the normal direction perpendicular to this ground plane, rather than the absolute vertical axis of the image; then the length of the curved portion is calculated.

[0042] In detail, the plant height extraction includes: calculating the projection values ​​of all point coordinates in the plant point cloud along the growth direction, and using the difference between the maximum and minimum projection values ​​as the accurate plant height h. Furthermore, in the identification process of some specific plant varieties, the plant may exhibit curved growth, making straight-line distance calculations inapplicable. Therefore, in some embodiments, plant height extraction also includes: Extract the ordered center point sequence from the plant point cloud and classify the height range. Discretize into m There are 1, 2, 3, 4, 5, 6, 7, 8, 9, 1 and These represent the minimum and maximum height projections in the point cloud, respectively. For each interval... k Then, calculate the average position of all points within the interval as the center point. This yields an ordered sequence of center points. Arranged in ascending order of height. Corresponding to the top. Then, based on the ordered sequence of center points, the bending length... L The approximation is obtained by summing the Euclidean distances between adjacent points:

[0043] This summation approximates the actual curve length of the plant from bottom to top, and the approximation is more accurate when the point cloud is dense and the number of intervals m is large enough.

[0044] Furthermore, stem thickness extraction is performed. The stem region is identified in the plant point cloud. A cross-sectional point cloud slice is taken at a preset height above the ground. The points in the slice are fitted with a two-dimensional circle or ellipse. The equivalent diameter of the fitted shape is taken as the stem thickness d. The equivalent diameter or the length of the major / minor axis is used as the measure of stem thickness d.

[0045] The crown diameter and projected leaf area are calculated by projecting the plant mask onto a 2D image and combining it with 3D information through pixel area and physical scale conversion. The specific algorithm involves mathematical knowledge and will not be elaborated here. Therefore, the extracted plant height h, stem diameter d, and crown width calculated by pixel area and physical scale are weighted and summed to form the growth state vector. .

[0046] Based on this, the current growth state vector is identified The prediction model can be updated to adapt to the current growth status of most plants, ensuring the accuracy of the model's predictions. However, considering that updating across the entire network could easily lead to overfitting to new data and result in too much data, making real-time updates unsuitable, the following detailed steps are included: First, data alignment is performed, aligning the environment parameter vectors. Growth state vector Perform time alignment.

[0047] Freeze the initial crop growth function The parameters are fixed and used as the basis for prediction and prior knowledge base; then, real-time context encoding is performed to build a lightweight real-time feature encoder. A real-time feature encoder is a 2- or 3-layer fully connected neural network; its input is the aligned data pairs at the current time step. The encoder outputs a real-time context encoding vector, along with statistical features calculated from the data cache queue. .

[0048] Perform adaptive offset generation and build a lightweight adaptive offset generator. It is actually a parameterized function for real-time context encoding. and the initial crop growth function at the evaluation point The intermediate layer features, such as the backbone features before the output of the improved deep operator branch network module. b base Branch features before the backbone network output C base As input.

[0049] Specifically, adaptive offset generator Includes two sub-networks and The adaptive offset generator learns an increment for each vector, where: ; ; In other words, during prediction, the frozen, pre-trained master prediction model is first run once, and the intermediate features generated by its deep operator network during the prediction process are recorded, including backbone features. b base and branching features C base Subsequently, the real-time context encoding vector Respectively with the main characteristics b base and branching features C base Fusion, input two subnetworks and Generate the corresponding trunk offset vector and branch offset vector ; Furthermore, through learnable gain parameters and By combining the offset with the base vector, we obtain the adjusted main vector and branch vectors: .

[0050] Therefore, the above process incrementally trains and fine-tunes the prediction model based on the current real-time crop growth status, maintaining its accuracy in predicting the crop at the current moment. Furthermore, the prediction model in this embodiment is a lightweight model, requiring less computational power and allowing for low-cost model updates. It is worth noting that the method in this embodiment is particularly suitable for greenhouse cultivation, where the environment for crops within a certain space is largely the same. For large-scale cultivation, the planting area needs to be divided into different blocks for processing. Moreover, because the training process of the prediction model is also lightweight, it can quickly adapt to different planting scenarios to set the prediction model according to requirements.

[0051] Furthermore, a crop growth function prediction system based on deep operator networks according to an embodiment of this application includes a memory and a processor. The memory stores a computer program to implement the crop growth function prediction method based on deep operator networks, and the processor performs read and write operations on the memory.

[0052] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting crop growth functions based on deep operator networks, characterized in that, Includes the following steps: Collect environment parameter vectors aligned to timestamps ; The obtained environmental parameter vector The environmental parameter vector is input into the trained prediction model to obtain the crop growth function, and outputs the crop growth prediction result; Includes d-dimensional environmental parameter components; The crop growth function is obtained through a trained prediction model, which includes: The recurrent neural network module, the environmental parameter vector As input, for each time step The network unit receives the current input environmental parameter vector. In addition to the hidden state and unit state of the previous time step, the state is updated through an internal control mechanism; finally, global average pooling is performed on the hidden states of all time steps along the time dimension to obtain a fixed-dimensional temporal feature vector. ; An improved deep operator network module, including branch networks and backbone networks; The branch network uses time-series feature vectors As the input function; the time series feature vector Each one-dimensional feature is configured with a light quantum network to output a A dimensional vector; the outputs of all h subnetworks are concatenated through the first fusion layer to generate the vector. The branch vector b in dimension; The backbone network is based on evaluation points. As input, t For the current time step, the evaluation point will be... Decomposed into d+1 Each component is configured with a light quantum network to receive a scalar input and output a p-dimensional vector, and then the second fusion layer is used to... d+1 The outputs of each subnetwork are fused to generate a p-dimensional backbone vector. C ; The output of the crop growth operator G as the crop growth function at the evaluation point y is expressed as: ； This is the predicted growth state vector.

2. The crop growth function prediction method based on deep operator networks according to claim 1, characterized in that, The current crop growth state vector is obtained through a visual neural network. include: Identify and segment the pixel-level mask of each crop in the acquired image; For each crop, a three-dimensional point cloud model of the plant is reconstructed based on multiple images taken at different times or from different perspectives, and the point cloud is converted to a world coordinate system with physical scale by combining camera calibration parameters. Define the growth reference plane and direction, analyze the bottom point set at the interface between the plant point cloud and the soil, and determine the growth reference plane representing the ground plane through a plane fitting algorithm. The normal direction of the plane is defined as the main growth direction of the plant, so as to extract the growth state vector. .

3. The crop growth function prediction method based on deep operator networks according to claim 2, characterized in that, This also includes plant height extraction: Calculate the projection values ​​of all point coordinates in the plant point cloud onto the growth direction, and take the difference between the maximum and minimum projection values ​​as the accurate plant height h.

4. The crop growth function prediction method based on deep operator networks according to claim 3, characterized in that, Plant height extraction also includes: Extract the ordered center point sequence from the plant point cloud and classify the height range. Discretize into m There are 1, 2, 3, 4, 5, 6, 7, 8, 9, 1 and These represent the minimum and maximum height projections in the point cloud, respectively, for each interval. k Then calculate the average position of all points within the interval as the center point. This yields an ordered sequence of center points. Arranged in ascending order of height. Corresponding to the top; based on an ordered sequence of center points, bending length L The approximation is obtained by summing the Euclidean distances between adjacent points: 。 5. The crop growth function prediction method based on deep operator networks according to claim 2, characterized in that, This includes extracting stem thickness, identifying the stem region in the plant point cloud, taking a cross-sectional point cloud slice at a preset height above the ground, fitting the points in the slice with a two-dimensional circle or ellipse, and using the fitted equivalent diameter as the stem thickness d. And the crown width obtained by converting pixel area to physical scale.

6. The crop growth function prediction method based on deep operator networks according to any one of claims 2-5, characterized in that, It also includes the prediction model based on the current growth state vector The prediction model is then updated by freezing most of its parameters and performing incremental training.

7. The crop growth function prediction method based on deep operator networks according to claim 6, characterized in that, The incremental training includes the following sub-steps: Data alignment, aligning environmental parameter vectors Growth state vector Perform time alignment; Perform real-time context encoding to build a lightweight real-time feature encoder. Its input is the alignment data pair at the current time. The encoder outputs a real-time context encoding vector, along with statistical features calculated from the data cache queue. ; An adaptive offset is generated based on the context encoding vector to construct the adjusted backbone vector and branch vectors, thereby updating the deep operator network module.

8. The crop growth function prediction method based on deep operator networks according to claim 7, characterized in that, Generate adaptive offsets based on the context encoding vector to construct adjusted backbone and branch vectors, thereby updating the deep operator network module, including: Adaptive offset generation: Building a lightweight adaptive offset generator. Adaptive offset generator Includes two sub-networks and During prediction, the frozen, pre-trained master prediction model is first run once, and the intermediate features, including backbone features, generated by its deep operator network during the prediction process are recorded. b base and branching features C base Subsequently, the real-time context encoding vector Respectively with the main characteristics b base and branching features C base Fusion, input two subnetworks and Generate the corresponding trunk offset vector and branch offset vector ; The crop growth function is dynamically adjusted through learnable gain parameters. and By combining the offset with the base vector, we obtain the adjusted main vector and branch vectors: 。 9. The crop growth function prediction method based on deep operator networks according to claim 1, characterized in that, The prediction model training process adopts supervised learning, and based on the diagonal sparsity of the Jacobian matrix block formed by the independent parameters of each sub-network in the branch network and the backbone network, the gradient calculation is performed using the forward mode automatic differentiation algorithm. During gradient calculation, the local Jacobian matrix for each sub-network is calculated independently, reducing the overall computational complexity compared to traditional backpropagation. Reduce to P represents the total number of parameters. Let be the number of parameters for the i-th subnetwork.

10. A crop growth function prediction system based on deep operator networks, characterized in that, The device includes a storage unit and a processor, wherein the storage unit stores a computer program to implement the crop growth function prediction method based on deep operator networks as described in any one of claims 1-9, and the processor performs read and write operations on the storage unit.

Images