Corn optimal seeding rate decision method based on multi-source field information fusion

Abstract

The present application provides a corn optimal seeding rate decision-making method based on multi-source field information fusion, comprising the following steps: obtaining multi-source data of the target field, respectively constructing yield prediction model data set and optimal seeding rate true value data set, and carrying out data preprocessing; constructing an end-to-end optimal seeding rate decision network, the decision network comprising a feature coding module, a double-dimension weight learning module and a gated weighted fusion module; completing feature embedding and global context modeling through the feature coding module, completing adaptive contribution modeling of two types of dimension representation through the double-dimension weight learning module, and completing multi-dimensional information fusion and adaptive weighting of multi-path candidate decision through the gated weighted fusion module, and finally outputting the optimal seeding rate prediction value; using a yield-guided step-by-step training architecture to train the decision network, obtaining target multi-source field information of the to-be-seeded plot, inputting the trained optimal seeding rate decision model, and outputting the optimal corn seeding rate of the corresponding plot.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to the application of deep learning technology in agricultural sowing technology, specifically a method for determining the optimal sowing rate of maize based on the fusion of multi-source field information. Background Technology

[0002] The constant-rate planting strategy is widely used in current maize production. However, due to the significant spatial heterogeneity of the field environment, this strategy often fails to effectively match planting density with the supply of resources such as soil fertility, topography, and water and fertilizer conditions. This can easily lead to insufficient optimization of population structure, thereby reducing the utilization efficiency of key resources such as light, water, and fertilizer, and limiting the full release of yield potential.

[0003] Variable seeding technology uses differences in soil properties, topography, and water and fertilizer conditions within a plot as the control target, and spatially optimizes the seeding density according to local conditions. It is considered one of the effective ways to improve resource utilization efficiency and maize yield per unit area.

[0004] Existing research has shown that maize yield typically exhibits a parabolic response relationship of "first increasing and then decreasing" with increasing planting density, and there exists an optimal planting density to maximize yield; moreover, this optimal value is influenced by both field environmental characteristics and variety characteristics, and may vary significantly between different plots and even between different regions of the same plot.

[0005] In the field of variable seeding, machine learning methods, linear regression, and multiple linear fitting are commonly used to determine the optimal seeding rate under different field conditions. Currently, related existing technologies employ a two-stage modeling framework of yield prediction and optimal search. However, yield prediction errors can easily accumulate and amplify during subsequent optimization processes, thus affecting the stability and reliability of the optimal seeding rate decision. Summary of the Invention

[0006] The purpose of this invention is to provide a method for determining the optimal seeding rate for maize based on the fusion of multi-source field information, thereby addressing the technical problems raised in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution.

[0008] According to an embodiment of the present invention, a method for determining the optimal seeding rate for maize based on multi-source field information fusion is provided, comprising the following steps:

[0009] Acquire multi-source data from the target field, construct the optimal seeding rate true value dataset for the yield prediction model, and perform data preprocessing.

[0010] An end-to-end optimal seeding decision network is constructed, comprising a feature encoding module, a two-dimensional weight learning module, and a gated weighted fusion module. The input features of the decision network are divided into yield potential and crop growth optimization dimensions based on the mechanism of action. The feature encoding module performs feature embedding and global context modeling, the two-dimensional weight learning module performs adaptive contribution modeling of the two dimensions, and the gated weighted fusion module performs multi-dimensional information fusion and adaptive weighting of multi-candidate decisions, ultimately outputting the optimal seeding prediction value.

[0011] The decision network is trained using a yield-guided stepwise training architecture, which includes: first, training the yield prediction model based on the yield prediction model dataset, generating a pseudo-optimal broadcast dataset and pre-training the decision network to obtain prior weights; then, correcting and training the decision network loaded with prior weights based on the optimal broadcast ground truth dataset to obtain the trained optimal broadcast decision model.

[0012] Obtain target multi-source field information for the plots to be sown, input it into the trained optimal seeding rate decision model, and output the optimal seeding rate for maize for the corresponding plots.

[0013] Furthermore, the multi-source data includes accumulated temperature, total precipitation, and total sunshine hours in the year before sowing, as well as nitrogen fertilizer application, phosphorus fertilizer application, potassium fertilizer application, soil organic matter content, corn sowing amount and corresponding yield data in the year of sowing.

[0014] The characteristics of the yield potential dimension include accumulated temperature, total precipitation, and total sunshine hours in the year prior to sowing;

[0015] The characteristics of the crop growth optimization dimension include the amount of nitrogen fertilizer applied, the amount of phosphorus fertilizer applied, the amount of potassium fertilizer applied, and the soil organic matter content in the year of sowing.

[0016] Furthermore, in the feature encoding module:

[0017] Each continuous input feature is uniformly encoded into a vector representation of a preset length T, and organized into a sequence form adapted to the attention mechanism. A classification notation cls is introduced to aggregate global information, and a group embedding is added to explicitly encode the yield potential dimension or crop growth optimization dimension to which the feature belongs. The encoded feature sequence is input into a multi-head self-attention encoder to complete the information interaction between features and obtain a feature representation containing global context information.

[0018] Furthermore, in the two-dimensional weight learning module:

[0019] The feature sequences output by the feature encoding module are then subjected to attention pooling for the features in the yield potential dimension and the crop growth optimization dimension, respectively, to obtain a yield potential dimension-level representation. Dimensional representation of crop growth optimization The two types of dimensional representations and their differences are fused to obtain the fused representation u, which is expressed as:

[0020]

[0021] The fused representation u is input into a linear mapping layer to generate adaptive weights of two dimensions. and The adaptive weights are used to characterize the relative contribution of the two types of dimensional information to the broadcast volume decision in the current sample.

[0022] Furthermore, the gated weighted fusion module specifically performs the following operations:

[0023] Characterizing production potential at the dimensional level Dimensional representation of crop growth optimization Dimensional differences Dimensional Interaction Items Weighted dimensional representation Weighted dimensional representation The global representation of the classification symbol cls is fused to construct a unified decision input representation s, which is represented as:

[0024]

[0025] Based on the unified decision input representation s, normalized weights are generated to produce n candidate predictions through weighted branch output. Simultaneously, the predicted values ​​of the corresponding n candidate broadcast volumes are output through the prediction branch. ;

[0026] The candidate weights and candidate broadcast predictions are weighted and aggregated to obtain the final optimal broadcast OSR, which is expressed as: In the formula, n is the preset number of candidate decision units, and k is the sequence number of the candidate decision units.

[0027] Furthermore, in the step of training the production prediction model based on the production prediction model dataset, an oversampling method is used to augment the sparse regions of the production prediction model dataset, as shown below:

[0028]

[0029] In the formula, Indicates minority class samples, express Nearest neighbor samples in feature space Represents a random number, and .

[0030] Furthermore, the process of constructing and training the yield prediction model includes:

[0031] A yield prediction model is constructed based on the gradient boosting algorithm. The model is trained using a data-augmented yield prediction model dataset. The training process optimizes the mean squared error (MSE) loss function, which is expressed as follows:

[0032]

[0033] In the formula, This represents the actual output value. This represents the predicted output value, and n represents the sample size.

[0034] Furthermore, the step of generating a pseudo-optimal seeding dataset and pre-training the decision network includes: using the trained yield prediction model, predicting yields for multiple gradient seeding schemes under different combinations of climatic conditions and field environment characteristics, selecting the seeding rate corresponding to the maximum predicted yield in each environmental scenario, and constructing a pseudo-optimal seeding dataset; using the pseudo-optimal seeding dataset to pre-train the end-to-end optimal seeding decision network to obtain the model's prior weights.

[0035] Furthermore, the step of correcting and training the decision network with loaded prior weights includes: using an oversampling method to augment the optimal broadcast ground truth dataset, using the pre-trained prior weights as the initial weights of the decision network, and using the augmented optimal broadcast ground truth dataset to correct and train the decision network to obtain the trained optimal broadcast decision model.

[0036] According to another embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, causes the processor to perform the steps of the maize optimal seeding rate decision method based on multi-source field information fusion as provided in the above embodiments.

[0037] Compared with existing technologies, the beneficial effects of the maize optimal seeding rate decision-making method based on multi-source field information fusion of this invention are:

[0038] This invention divides input features into yield potential dimension and growth optimization dimension according to their mechanism of action. It performs unified embedding encoding on the two types of features and introduces group embedding to explicitly identify the dimension affiliation. Then, it realizes cross-feature information interaction and global context modeling through a multi-head self-attention encoder. On this basis, it performs adaptive contribution modeling on the two-dimensional representation and directly outputs the best yield result, thereby reducing the interference of heterogeneous information mixing and improving decision stability and generalization ability at the structural level.

[0039] In the training of the decision network of this invention, a yield prediction model is first trained as a prior generator. Under different combinations of climate and field environment characteristics, the yield response of multiple gradient seeding schemes is evaluated. A pseudo-optimal seeding dataset is constructed with the seeding corresponding to the maximum predicted yield. The optimal seeding decision network is pre-trained using this pseudo-dataset to obtain prior weights. Subsequently, the prior weights are loaded into the decision network and combined with the real optimal seeding dataset for correction training to correct pseudo-label bias and enhance the ability to characterize local optima, thereby achieving robust transfer from global rules to real scenarios.

[0040] This invention integrates the yield potential dimension representation and the growth optimization dimension representation, along with their differences and interactions, the weighted dimension representation, and the global representation to form a unified decision input. It also sets weights to generate normalized weights for multiple candidate predictions from branch outputs, and sets prediction branches to generate corresponding candidate seeding predictions. Finally, the optimal seeding output is obtained through the weighted convergence of candidate weights and candidate predictions. This structure, by adaptively combining multiple candidate outputs, reduces the overfitting risk of a single regression head under small sample conditions and avoids reliability bias caused by simple averaging assumptions, thereby improving the robustness and generalizability of the prediction results. Attached Figure Description

[0041] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0042] In the attached diagram:

[0043] Figure 1 This is a flowchart illustrating the implementation of the optimal seeding rate decision-making method for maize based on multi-source field information fusion, as described in this invention.

[0044] Figure 2 This is a schematic diagram of the architecture of the optimal broadcast decision model in an embodiment of the present invention;

[0045] Figure 3 The model training flowchart of the optimal broadcast decision model provided by this invention;

[0046] Figure 4 A structural block diagram of a computer device provided by the present invention. Detailed Implementation

[0047] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0048] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0049] Please refer to Figure 1 In one embodiment of the present invention, a method for determining the optimal seeding rate of maize based on multi-source field information fusion is provided, comprising the following steps:

[0050] Step S101: Obtain multi-source data from the target field, construct the optimal seeding rate true value dataset for the yield prediction model data set, and perform data preprocessing;

[0051] The multi-source data includes accumulated temperature (AT), total precipitation (P), and total sunshine hours (SD) in the year before sowing. The multi-source data also includes nitrogen fertilizer application (NR), phosphorus fertilizer application (PR), potassium fertilizer application (KR), soil organic matter (SOM) content, maize planting amount and corresponding yield data in the year of sowing.

[0052] Step S101 in this embodiment of the invention is the data acquisition step. Specifically, data analysis is performed on seeding rate (SR), AT, P, SD, NR, PR, KR, SOM and yield data collected from field trials over many years to obtain the optimal seeding rate under different years and different fertility plots as the optimal seeding rate decision model dataset. The obtained dataset is then cleaned, standardized, feature-engineered and partitioned.

[0053] In one implementation of the present invention, step S101 is divided into the acquisition and processing of the yield prediction model dataset and the optimal broadcasting dataset.

[0054] In the collection and processing of the yield prediction model dataset: multi-year field trials were conducted at different locations, with SOM content, seeding rate and fertilizer application rate as the main experimental factors. The experimental area design needed to cover different SOM levels and different fertilizer intensities, and gradient seeding rate treatment was set under each combination of conditions to provide a foundation for the subsequent construction of the optimal seeding rate dataset; multi-source data collected from the field over many years were sorted out, and after preprocessing such as missing and outlier handling, data cleaning and normalization, the yield prediction model dataset was formed.

[0055] The normalized values ​​are expressed as follows:

[0056]

[0057] In the formula, x represents the original data. The minimum value under the characteristic, The maximum value under the characteristic, The value is the normalized value;

[0058] In the collection and processing of the optimal seeding rate prediction model dataset, since existing studies have determined that yield initially increases and then decreases with increasing seeding rate, a parabolic fitting function was used to construct the seeding rate-yield relationship in experimental areas with the same SOM content and fertilizer application rate. The seeding rate at the yield maximum value in the function with the highest fitting accuracy was selected as the optimal seeding rate under that field condition. Based on this method, optimal seeding rates were collected under various field conditions, and the collected optimal seeding rate datasets were normalized to form the optimal seeding rate dataset.

[0059] The relationship between seeding rate and yield is constructed using a parabolic fitting function, expressed as follows:

[0060]

[0061]

[0062]

[0063] In the formula, S is the seeding rate, Y is the yield, and a, b, c, A, B, M, k, ... These are the parameters to be estimated.

[0064] Please continue to refer to Figure 1 and Figure 2 The present invention, a method for determining the optimal seeding rate of maize based on multi-source field information fusion, further includes the following steps:

[0065] Step S102: Construct an end-to-end optimal seeding rate decision network. The decision network includes a feature encoding module, a two-dimensional weight learning module, and a gated weighted fusion module. Based on the mechanism of action, the input features of the decision network are divided into yield potential dimension and crop growth optimization dimension. The feature encoding module completes feature embedding and global context modeling. The two-dimensional weight learning module completes adaptive contribution modeling of the two-dimensional representations. The gated weighted fusion module completes multi-dimensional information fusion and adaptive weighting of multi-candidate decisions, and finally outputs the optimal seeding rate prediction value.

[0066] In this embodiment of the invention, considering the different pathways through which the two types of factors play a role in yield formation and seeding rate regulation, the invention divides them into a yield potential dimension and a crop growth optimization dimension to reduce the interference of heterogeneous information mixing on the training of the optimal seeding rate model at the structural level; wherein:

[0067] In one specific implementation of the present invention, the characteristics of the yield potential dimension include the accumulated temperature, total precipitation, and total sunshine hours of the year before sowing, which reflect the climatic background of the crop in a specific year and region and its constraints on the upper limit of potential yield. These are environmental driving factors that are difficult to change directly through field management.

[0068] In one specific implementation of the present invention, the characteristics of the crop growth optimization dimension include the application of nitrogen fertilizer, phosphorus fertilizer, and potassium fertilizer in the sowing year, and the soil organic matter (SOM) content, which directly affect nutrient supply and population growth status. These are controllable and optimizable management and soil conditions in production.

[0069] This invention achieves a unified characterization of the coupled effects of climate environment, soil conditions and management factors on the optimal seeding rate within the same model by adaptively weighting and fusing multi-source features, thereby improving the matching degree between the recommended seeding rate and the field resource supply capacity.

[0070] Furthermore, in the feature encoding module:

[0071] Each continuous input feature is uniformly encoded into a vector representation of a preset length T, and organized into a sequence form adapted to the attention mechanism. A classification notation cls is introduced to aggregate global information, and a group embedding is added to explicitly encode the yield potential dimension or crop growth optimization dimension to which the feature belongs. The encoded feature sequence is input into a multi-head self-attention encoder to complete the information interaction between features and obtain a feature representation containing global context information.

[0072] Therefore, in the feature encoding stage of the decision network, this invention encodes each continuous feature into a vector representation of length T and organizes them into a sequence to adapt to the attention mechanism. Simultaneously, embodiments of this invention introduce a classification label (cls) to aggregate global information;

[0073] In one specific implementation of the present invention, in order to explicitly encode the dimensional attribution of features, the model of the present invention incorporates group embedding, which enables the network to distinguish between the inputs of the yield potential dimension and the growth optimization dimension during the training process, thereby learning the differentiated action patterns of the two types of information in a more targeted manner.

[0074] Furthermore, in the dual-dimensional weight learning module, attention pooling is performed on the feature sequences output by the feature encoding module for both the yield potential dimension and the crop growth optimization dimension, respectively, to obtain a yield potential dimension-level representation. Dimensional representation of crop growth optimization To characterize the complementary and differential relationship between the two dimensions of information, the two types of dimensional representations and their differential information are fused to obtain the fused representation u, which is expressed as:

[0075]

[0076] The fused representation u is input into a linear mapping layer to generate adaptive weights of two dimensions. and The adaptive weights are used to characterize the relative contribution of the two types of dimensional information to the broadcast volume decision in the current sample.

[0077] In one specific implementation of the present invention, the input features are divided into yield potential dimension and growth optimization dimension according to their mechanism of action. The two types of features are uniformly embedded and encoded, and group embedding is introduced to explicitly identify the dimension affiliation. Then, a multi-head self-attention encoder is used to realize cross-feature information interaction and global context modeling. On this basis, adaptive contribution modeling is performed on the two-dimensional representations and the best yield result is directly output, thereby reducing the interference of heterogeneous information mixing and improving decision stability and generalization ability at the structural level.

[0078] Furthermore, the gated weighted fusion module specifically performs the following operations:

[0079] Characterizing production potential at the dimensional level Dimensional representation of crop growth optimization Dimensional differences Dimensional Interaction Items Weighted dimensional representation Weighted dimensional representation The global representation of the classification symbol cls is fused to construct a unified decision input representation s, which is represented as:

[0080]

[0081] Based on the unified decision input representation s, normalized weights are generated to produce n candidate predictions through weighted branch output. Simultaneously, the predicted values ​​of the corresponding n candidate broadcast volumes are output through the prediction branch. ;

[0082] The candidate weights and candidate broadcast predictions are weighted and aggregated to obtain the final optimal broadcast OSR, which is expressed as:

[0083]

[0084] In the formula, n is the preset number of candidate decision units, and k is the sequence number of the candidate decision units.

[0085] In one specific implementation of the present invention, the present invention integrates the yield potential dimension representation and the growth optimization dimension representation, along with their difference and interaction information, the weighted dimension representation, and the global representation to form a unified decision input. It also sets weights to generate normalized weights for multiple candidate predictions from branch outputs, and sets prediction branches to generate corresponding candidate seeding prediction values. Finally, the optimal seeding output is obtained by weighted convergence of candidate weights and candidate predictions. This structure, through adaptive combination of multiple candidate outputs, reduces the overfitting risk of a single regression head under small sample conditions and avoids reliability bias caused by simple averaging assumptions, thereby improving the robustness and generalizability of the prediction results.

[0086] The gated weighted aggregation mechanism provided by this invention combines multiple candidate predictions with adaptive weights. When the sample size is limited, it can avoid overfitting a single regression head to a small number of samples and reduce the assumption that each candidate output is equally reliable implied by simple averaging, thereby improving the robustness and generalizability of the optimal broadcast prediction results.

[0087] Furthermore, this invention proposes a yield-guided stepwise training architecture. This training architecture uses a yield prediction model to generate pseudo-optimal broadcast labels and pre-trains the end-to-end decision network to obtain transferable prior weights. Subsequently, fine-tuning training is carried out using a small number of data-augmented real optimal broadcast samples, thereby significantly enhancing the training stability of the model under conditions of small samples and uneven feature distribution, and improving the accuracy and generalization ability of optimal broadcast prediction.

[0088] Specifically, in one particular implementation of the present invention, please continue to refer to... Figure 1 and Figure 3 The present invention, a method for determining the optimal seeding rate of maize based on multi-source field information fusion, further includes the following steps:

[0089] Step S103: The decision network is trained using a yield-guided stepwise training architecture, including: first, training the yield prediction model based on the yield prediction model dataset, generating a pseudo-optimal broadcast dataset and pre-training the decision network to obtain prior weights; then, correcting and training the decision network loaded with prior weights based on the optimal broadcast ground truth dataset to obtain the trained optimal broadcast decision model.

[0090] Step S104: Obtain target multi-source field information of the plot to be sown, input the trained optimal sowing rate decision model, and output the optimal sowing rate of corn for the corresponding plot.

[0091] In step S103 of this invention, the decision network is trained by first pre-training the model using a pseudo-optimal seeding dataset, and then using the true values ​​of the optimal seeding dataset for calibration training. The reason for using this training method is that the long field trial period and uncontrollable conditions result in a limited amount of available data, and the uneven distribution of samples in the feature interval weakens the learning effect and prediction accuracy of the optimal seeding decision model, affecting its generalizability and reliability in real production scenarios. Specifically, if the true values ​​of the optimal seeding dataset are obtained directly, the long field trial period and insufficient amount of data will lead to overfitting and decreased prediction accuracy if the true values ​​are used to train the model directly.

[0092] Therefore, this invention first constructs a pseudo-optimal seeding dataset using a yield prediction model, enabling the optimal seeding decision model to achieve initial convergence (i.e., learn global patterns). Then, it uses a small amount of true optimal seeding data for further correction. This effectively reduces overfitting, improves prediction accuracy, and prevents the subsequent accumulation of errors in the yield prediction model.

[0093] Unlike the traditional two-stage framework of yield prediction and optimal search, this invention moves the indirect optimization process of yield prediction and seeding rate derivation / search forward and solidifies it in the pre-training stage, using it only as a means of injecting prior knowledge. In the inference stage, the end-to-end model directly outputs the optimal seeding rate, avoiding the accumulation and amplification of errors caused by the two-stage chain, and reducing the impact of yield prediction model errors on the final optimal seeding rate.

[0094] Specifically, in the step of training the yield prediction model based on the yield prediction model dataset described in this invention, since the dataset is obtained from field experiments, uncontrollable conditions in the field lead to intervals with fewer data values ​​(outlier intervals), resulting in an uneven distribution of the dataset and affecting the prediction accuracy of the model. To solve this problem, this invention uses an oversampling method to augment the data in the outlier intervals;

[0095] In this invention, an oversampling method is used to augment the sparse regions of the yield prediction model dataset, as shown below:

[0096]

[0097] In the formula, Indicates minority class samples, express Nearest neighbor samples in feature space Represents a random number, and .

[0098] Furthermore, in the process of constructing and training the yield prediction model, a yield prediction model based on the gradient boosting algorithm is constructed and trained. Specifically, this invention constructs a yield prediction model based on the gradient boosting algorithm, uses a data-augmented yield prediction model dataset to complete model training, and the training process uses the mean squared error loss function (MSE) as the optimization objective. The mean squared error loss function is expressed as:

[0099]

[0100] In the formula, This represents the actual output value. This represents the predicted output value, and n represents the sample size.

[0101] Furthermore, the step of generating a pseudo-optimal seeding dataset and pre-training the decision network includes: using the trained yield prediction model, predicting yields for multiple gradient seeding schemes under different combinations of climatic conditions and field environment characteristics, selecting the seeding rate corresponding to the maximum predicted yield in each environmental scenario, and constructing a pseudo-optimal seeding dataset; using the pseudo-optimal seeding dataset to pre-train the end-to-end optimal seeding decision network to obtain the model's prior weights.

[0102] In one implementation, a set of optimal seeding rate ground value samples typically requires setting up multiple seeding rate treatments and obtaining corresponding yield data under the same field conditions before the optimal seeding rate can be determined through fitting. Therefore, the ground value dataset can only be obtained point by point through a large number of field trials, which is costly and time-consuming, resulting in a limited sample size. On the other hand, although pseudo-optimal seeding rate data generated based on yield prediction models can significantly expand the sample size, there are still unavoidable prediction biases, making it unsuitable for directly expanding the ground value dataset.

[0103] Therefore, this invention employs a pre-training strategy to initially train the optimal seeding decision network, enabling the model to prioritize learning the main response patterns and global laws between multi-source information in the field and OSR.

[0104] Specifically, this invention utilizes a trained yield prediction model to predict yields for multiple gradient seeding schemes under different climatic conditions and corresponding combinations of field environmental features. In each environmental scenario, the seeding rate corresponding to the highest predicted yield is selected as the optimal seeding rate for that scenario. By repeatedly calculating and screening multiple environmental combinations, a larger pseudo-optimal seeding rate dataset is constructed.

[0105] Subsequently, this invention uses this pseudo-dataset to pre-train the optimal broadcast volume prediction model to obtain prior weights. The main loss function of the training process is expressed as follows:

[0106]

[0107] In the formula, Truth value Predicted value.

[0108] Furthermore, the step of correcting and training the decision network with loaded prior weights includes:

[0109] Oversampling is used to augment the ground truth dataset of optimal broadcast volume. The prior weights obtained from pre-training are used as the initial weights of the decision network. The decision network is then trained and corrected using the augmented ground truth dataset of optimal broadcast volume to obtain the trained optimal broadcast volume decision model.

[0110] In the training of the decision network of this invention, a yield prediction model is first trained as a prior generator. Under different combinations of climate and field environment characteristics, the yield response of multiple gradient seeding schemes is evaluated. A pseudo-optimal seeding dataset is constructed with the seeding corresponding to the maximum predicted yield. The optimal seeding decision network is pre-trained using this pseudo-dataset to obtain prior weights. Subsequently, the prior weights are loaded into the decision network and calibrated by combining it with the real optimal seeding dataset to correct pseudo-label bias and enhance the ability to characterize local optima, thereby achieving robust transfer from global patterns to real-world scenarios.

[0111] In summary, this invention loads the prior weights obtained in the pre-training stage into the optimal broadcast prediction model, which can improve the stability of the training process and reduce the risk of overfitting caused by the small size of the ground truth dataset. Given the limited number of ground truth samples for optimal broadcast, this paper uses a data augmentation method to perform preliminary data augmentation. Subsequently, the augmented dataset is input into the model with the loaded prior weights for calibration training to correct potential biases caused by pseudo-labels and improve the fitting ability to real-world scenarios.

[0112] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0113] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0114] Please refer to Figure 4 According to another embodiment of this application, a computer device is provided, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the maize optimal seeding rate decision method based on multi-source field information fusion as described in any of the above embodiments.

[0115] The computer equipment can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal equipment may include, but is not limited to, processors and memory.

[0116] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting various parts of the terminal device via various interfaces and lines.

[0117] The memory can be used to store the computer program. The processor implements various functions of the terminal device by running or executing the computer program stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc.

[0118] In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card, secure digital card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0119] Another preferred embodiment of the present invention provides a storage medium, which is a computer-readable storage medium, and a computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above embodiment of the maize optimal seeding rate decision method based on multi-source field information fusion.

[0120] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0121] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for determining the optimal seeding rate for maize based on multi-source field information fusion, characterized in that, Includes the following steps: Acquire multi-source data from the target field, construct yield prediction model dataset and optimal seeding rate true value dataset respectively, and perform data preprocessing; An end-to-end optimal seeding decision network is constructed, comprising a feature encoding module, a two-dimensional weight learning module, and a gated weighted fusion module. The input features of the decision network are divided into yield potential and crop growth optimization dimensions based on the mechanism of action. The feature encoding module performs feature embedding and global context modeling, the two-dimensional weight learning module performs adaptive contribution modeling of the two dimensions, and the gated weighted fusion module performs multi-dimensional information fusion and adaptive weighting of multi-candidate decisions, ultimately outputting the optimal seeding prediction value. The decision network is trained using a yield-guided stepwise training architecture, which includes: first, training the yield prediction model based on the yield prediction model dataset, generating a pseudo-optimal broadcast dataset and pre-training the decision network to obtain prior weights; then, correcting and training the decision network loaded with prior weights based on the optimal broadcast ground truth dataset to obtain the trained optimal broadcast decision model. Obtain target multi-source field information for the plots to be sown, input it into the trained optimal seeding rate decision model, and output the optimal seeding rate for maize for the corresponding plots.

2. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 1, characterized in that, The multi-source data includes accumulated temperature, total precipitation, total sunshine hours in the year before sowing, as well as nitrogen fertilizer application, phosphorus fertilizer application, potassium fertilizer application, soil organic matter content, corn sowing amount and corresponding yield data in the year of sowing. The characteristics of the yield potential dimension include accumulated temperature, total precipitation, and total sunshine hours in the year prior to sowing; The characteristics of the crop growth optimization dimension include the amount of nitrogen fertilizer applied, the amount of phosphorus fertilizer applied, the amount of potassium fertilizer applied, and the soil organic matter content in the year of sowing.

3. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 2, characterized in that, In the feature encoding module: Each continuous input feature is uniformly encoded into a vector representation of a preset length T, and organized into a sequence form adapted to the attention mechanism. A classification notation cls is introduced to aggregate global information, and a group embedding is added to explicitly encode the yield potential dimension or crop growth optimization dimension to which the feature belongs. The encoded feature sequence is input into a multi-head self-attention encoder to complete the information interaction between features and obtain a feature representation containing global context information.

4. The optimal seeding rate decision method for maize based on multi-source field information fusion according to claim 3, characterized in that, In the two-dimensional weight learning module: The feature sequences output by the feature encoding module are then subjected to attention pooling for the features in the yield potential dimension and the crop growth optimization dimension, respectively, to obtain a yield potential dimension-level representation. Dimensional representation of crop growth optimization The two types of dimensional representations and their differences are fused to obtain the fused representation u, which is expressed as: The fused representation u is input into a linear mapping layer to generate adaptive weights of two dimensions. and The adaptive weights are used to characterize the relative contribution of the two types of dimensional information to the broadcast volume decision under the current sample.

5. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 4, characterized in that, The gated weighted fusion module specifically performs the following operations: Characterizing production potential at the dimensional level Dimensional representation of crop growth optimization Dimensional differences Dimensional Interaction Items Weighted dimensional representation Weighted dimensional representation The global representation of the classification symbol cls is fused to construct a unified decision input representation s, which is represented as: Based on the unified decision input representation s, normalized weights are generated to produce n candidate predictions through weighted branch output. Simultaneously, the predicted values ​​of the corresponding n candidate broadcast volumes are output through the prediction branch. ; The candidate weights and candidate broadcast predictions are weighted and aggregated to obtain the final optimal broadcast OSR, which is expressed as: In the formula, n is the preset number of candidate decision units, and k is the sequence number of the candidate decision units.

6. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 5, characterized in that, In the step of training the production prediction model based on the production prediction model dataset, an oversampling method is used to augment the sparse regions of the production prediction model dataset, as shown below: In the formula, Indicates minority class samples, express Nearest neighbor samples in feature space Represents a random number, and .

7. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 6, characterized in that, The process of constructing and training the yield prediction model includes: A yield prediction model is constructed based on the gradient boosting algorithm. The model is trained using a data augmentation dataset, with the mean squared error (MSE) loss function as the optimization objective. The MSE loss function is expressed as follows: In the formula, This represents the actual output value. This represents the predicted output value, and n represents the sample size.

8. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 7, characterized in that, The steps of generating a pseudo-optimal broadcast dataset and pre-training the decision network include: Using the trained yield prediction model, yield predictions were made for multiple gradient seeding schemes under different combinations of climatic conditions and field environment characteristics. Select the seeding rate corresponding to the maximum predicted yield in each environmental scenario to construct a pseudo-optimal seeding rate dataset; The end-to-end optimal broadcast decision network was pre-trained using a pseudo-optimal broadcast dataset to obtain the model's prior weights.

9. The method for determining the optimal seeding rate of maize based on multi-source field information fusion according to claim 8, characterized in that, The step of correcting and training the decision network with loaded prior weights includes: Oversampling was used to augment the optimal broadcast volume true value dataset. The prior weights obtained from pre-training are used as the initial weights of the decision network; The decision network was trained and corrected using the data-augmented optimal broadcast ground truth dataset to obtain the trained optimal broadcast decision model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the maize optimal seeding rate decision method based on multi-source field information fusion as described in any one of claims 1 to 9.

Images