Crop yield prediction method based on LSTM and GRU

Abstract

The invention relates to the technical field of crop yield prediction, and discloses a crop yield prediction method based on LSTM and GRU, and the method comprises the steps: obtaining crop planting data; the crop planting data comprises climate parameters and soil parameters; constructing a crop yield prediction model based on LSTM and GRU; extracting a first time sequence feature from the crop planting data by using an LSTM network; extracting a second time sequence feature from the crop planting data by using the first GRU network; performing feature fusion on the first time sequence feature and the second time sequence feature by using a feature fusion layer in a weighted average mode; using a second GRU network to extract tense features from the fusion features; and performing linear transformation on the tense characteristics by using a full connection layer to obtain a final crop yield prediction result. According to the method, the multi-stage time features can be extracted and processed more accurately by adopting a double-circulation structure, so that the accuracy of crop yield prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of crop yield prediction, and in particular to a crop yield prediction method based on LSTM and GRU. Background Art

[0002] In the agricultural production process, environmental factors such as climate conditions, soil quality and water supply are key variables that affect crop growth and yield. These factors not only directly affect the growth and development of crops, but also cause large fluctuations in yield under extreme climate conditions. In addition, the continued growth of the population and the increasing impact of global climate change have made the need for accurate prediction of crop yields increasingly urgent. By providing timely and accurate forecasts, agricultural producers can take early action to mitigate the impact of adverse environmental conditions on crop yield and quality, thereby maximizing the stability of food supply and ensuring food security.

[0003] Predicting crop yields based on environmental factors has become a focus of current research. Traditional statistical models usually rely on historical crop yield data and related environmental factors (such as meteorological conditions and soil properties) to establish prediction models. Although these models can provide valuable predictions to a certain extent, they are often limited by data limitations and simplified model assumptions, and are not sufficient to achieve high-precision predictions in complex environments. With the rapid development of big data, machine learning, and artificial intelligence technologies, more and more research is using these cutting-edge methods to process large amounts of complex nonlinear data, thereby improving the accuracy and efficiency of predictions and providing more reliable decision support for agricultural producers.

[0004] Machine learning algorithms such as support vector machines (SVM), forests, and neural networks are widely used for crop yield prediction. These algorithms can learn complex patterns and relationships from large amounts of data, thereby improving prediction accuracy. Shah et al. designed a smart agriculture system using regression techniques to predict crop yields by using SVM, random forests, and multivariate polynomial regression. Monga combined neural networks to estimate grape yields in vineyards based on images. Wang et al. used deep transfer learning to predict crop yields by integrating remote sensing data and neural networks. Ranjan and Parida used linear regression algorithms to map rice-growing areas and predict yields in Sahibgani, Jharkhand, India, using Sentinel-based optical and synthetic aperture radar data. Morales et al. analyzed farm simulation yield datasets using different algorithms, including regularized linear models, random forests, and neural networks.

[0005] In recent years, with the development of deep learning theory and the improvement of GPU performance, CNN and LSTM have achieved great success in the field of time series prediction with their powerful spatial feature capture capabilities. Khaki et al. designed a deep neural network (DNN) method using state-of-the-art modeling and solving techniques. The performance of this model is significantly better than other popular methods such as Lasso shallow neural network (SNN) and regression tree (RT). Nevavuori et al. developed and trained CNN-LSTM, convolutional LSTM and 3D-CNN architectures using time series drone RGB and weather data collected from nine fields in Pori, Finland. They used a complete 15-week image sequence from the 2018 growing season to evaluate the feasibility of spatiotemporal deep learning architectures in modeling and predicting time series crop yields using RGB time series data. Gavahi et al. proposed a hybrid structure DeepYield that combines ConLSTM layers with three-dimensional CNN (3DCNN) to achieve more accurate and reliable spatiotemporal feature extraction. The researchers compared the predictive performance of the developed model with competing methods such as decision trees, CNN with Gaussian processors, and CNN-LSTM, and the results showed that DeepYield significantly outperformed these techniques. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a crop yield prediction method based on LSTM and GRU.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A crop yield prediction method based on LSTM and GRU includes the following steps: Acquiring crop planting data; the crop planting data includes climate parameters and soil parameters; Constructing a crop yield prediction model based on LSTM and GRU; the crop yield prediction model includes a parallel LSTM network and a first GRU network, as well as a feature fusion layer, a second GRU network and a fully connected layer; The LSTM network is used to extract the first time series characteristics of the temporal dependence between long-term seasonal changes and climate fluctuations and crop yields from crop planting data; The first GRU network is used to extract the second time series features of the short-term weather changes and the temporal dependence between fertilizer application and crop yield from the crop planting data; The feature fusion layer is used to fuse the first time series features and the second time series features by weighted average method; The second GRU network is used to extract the temporal characteristics of the impact of agricultural activities and environmental factors on the yield of crops at different stages of the growth cycle from the fusion features; The fully connected layer is used to perform linear transformation on the temporal features to obtain the final crop yield prediction result.

[0008] Optionally, the climate parameters and soil parameters include: Crop type, crop year, cropping season, state, cropping area, yield, annual rainfall, fertilizer use, and pesticide use.

[0009] Optionally, the LSTM network includes: At least two LSTM units are cascaded in sequence, and each LSTM unit inputs a time series consisting of crop planting data.

[0010] Optionally, the first GRU network includes: At least two GRU units are cascaded in sequence, and each GRU unit inputs a time series consisting of crop planting data.

[0011] Optionally, when the feature fusion layer is used to fuse the first time series feature and the second time series feature in a weighted average manner: Two learnable weight vector parameters are used to represent the relative importance of LSTM and GRU outputs, and the sum of the two weight vector parameters is set to 1; The first time series feature output by the LSTM network and the second time series feature output by the first GRU network are weighted averaged according to the two weight vector parameters, expressed as:

[0012] in, represents the fusion feature, represents the weight vector parameter of the first time series feature, represents the first time series feature, represents the weight vector parameter of the second time series feature, Represents the second time series feature.

[0013] Alternatively, in the crop yield prediction model training process, the weight vector parameter of the first time series feature is adjusted by back propagation algorithm. and the weight vector parameter of the second time series feature Perform iterative optimization to obtain the optimal weight vector parameters and .

[0014] Optionally, the second GRU network includes: At least two GRU units are cascaded in sequence, and each GRU unit inputs a time series consisting of fused features.

[0015] The present invention has the following beneficial effects: The present invention uses LSTM and GRU layers in parallel to extract temporal features of climate and soil data, and further refines these features through additional GRU layers to enhance time series dependencies. This dual-loop structure can more accurately extract and process multi-level temporal features, thereby improving the accuracy of crop yield prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of a crop yield prediction method based on LSTM and GRU; Figure 2 This is a schematic diagram of the crop yield prediction model based on LSTM and GRU; Figure 3 This is a schematic diagram of the data volume distribution in different years in the data set; Figure 4 This is a schematic diagram of the data volume distribution of annual rainfall in the data set; Figure 5 This is a schematic diagram of the distribution of production data in the data set; Figure 6 This is a schematic diagram of the loss curve during model training; Figure 7 Compare bar graphs for different models; Figure 8 A scatter plot comparing actual and predicted values. DETAILED DESCRIPTION

[0017] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0018] This paper designs a new crop yield prediction model - DualSpinNet model, which integrates the long short-term memory network (LSTM) and gated recurrent unit (GRU) architecture, uses LSTM and GRU layers in parallel to extract the temporal features of climate and soil data, and further refines these features through additional GRU layers to enhance time series dependencies. This dual-loop structure can more accurately extract and process multi-level temporal features, thereby improving the accuracy of crop yield prediction. The model provides agricultural producers and decision makers with a highly accurate tool that helps improve crop yield and quality, promote food security and sustainable agricultural development.

[0019] Reference Figure 1 , a crop yield prediction method based on LSTM and GRU provided by an embodiment of the present invention includes the following steps S1 to S7: S1. Acquire crop planting data; the crop planting data includes climate parameters and soil parameters; In an optional embodiment of the present invention, step S1 obtains multidimensional data of crop planting, including crop type, crop year, planting season, state, planting area, yield, annual rainfall, fertilizer usage, and pesticide usage, and performs standardization. The purpose of standardization is to ensure that these agricultural factors have the same weight when input into the model, so that the model can effectively learn the relationship between different features, thereby accurately predicting crop yields.

[0020] S2. Constructing a crop yield prediction model based on LSTM and GRU; the crop yield prediction model includes a parallel LSTM network and a first GRU network, as well as a feature fusion layer, a second GRU network and a fully connected layer; In an optional embodiment of the present invention, step S2 constructs a crop yield prediction model based on LSTM and GRU, which integrates two sequence modeling architectures: LSTM and GRU. It aims to better capture complex dependencies in time series through a multi-layer neural network structure. The overall architecture consists of three main components: an input processing layer, an RNN layer (including LSTM and GRU), and a fully connected layer (FCL) that finally generates a prediction output.

[0021] The model first processes the input sequence through stacked layers of LSTM and GRU, extracting features of the time series layer by layer. The multi-layer architecture enables the model to learn and represent patterns at different levels of abstraction, allowing for deeper feature extraction, which helps capture complex dependencies in time series and thus improves prediction accuracy. The outputs of LSTM and GRU are fused using a weighted average method, combining the representations learned by the two models from the input sequence. This fusion mechanism takes advantage of the strengths of both networks, reduces the bias that may arise from using a single model, and improves the robustness and stability of predictions.

[0022] After fusing the outputs of LSTM and GRU, an additional GRU layer is introduced for further processing. This layer is designed to learn the fused time series features more deeply and further optimize the representation and prediction performance of the model. Finally, the final prediction result is output through the fully connected layer. Dropout is applied between each RNN layer to effectively reduce the risk of overfitting and enhance the generalization ability of the model in real-world applications.

[0023] Finally, the model structure is as follows Figure 2 As shown in Figure 3, the model structure combining LSTM and GRU layers is able to more accurately extract and process multi-level features from the data while capturing complex time series dependencies, thereby improving the accuracy of crop yield prediction.

[0024] S3, using LSTM network to extract the first time series features of the temporal dependence between long-term seasonal changes and climate fluctuations and crop yields from crop planting data; In an optional embodiment of the present invention, the LSTM network used in step S3 includes: At least two LSTM units are cascaded in sequence, and each LSTM unit inputs a time series consisting of crop planting data.

[0025] The workflow of the LSTM network can be divided into the following steps: input processing, calculation of the forget gate, calculation of the input gate and candidate storage units, updating the storage unit, calculation of the output gate, and calculation of the final output. The specific steps are as follows: LSTM receives input at the current time step , the hidden state of the previous time step and the memory cell state at the previous time step First, the forget gate outputs , which determines how much of the previous cell state should be retained Then, enter the gate Determines how much new information will be added to the storage unit. Candidate storage unit Represents new candidate information. Forget gate Control previous storage unit Degree of oblivion. Input gate and candidate storage units Controls the storage unit at the current time step The LSTM finally outputs the hidden state of the current time step through the output gate . This hidden state contains not only the current input The processing result also retains the information passed down from the previous time step. The specific description is shown in the following formula, where and are the weight matrices and bias tensors.

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] The example uses stacked LSTM layers to model the temporal dependency between crop yield and input agricultural factors. The LSTM layers capture the impact of long-term seasonal changes and climate fluctuations on crop yield.

[0033] S4, using the first GRU network to extract the second time series features of the time dependency between short-term weather changes and fertilizer application and crop yield from the crop planting data; In an optional embodiment of the present invention, the first GRU network used in step S4 includes: At least two GRU units are cascaded in sequence, and each GRU unit inputs a time series consisting of crop planting data.

[0034] GRU is a simplified version of the Long Short-Term Memory (LSTM) network. By introducing a gating mechanism, GRU solves the long-term dependency problem encountered by traditional RNNs when processing long sequence data. The main feature of GRU is that it uses two gates (reset gate and update gate) to control the flow of information, allowing it to capture dependencies in the sequence more effectively.

[0035] The update equation of GRU is as follows: Reset Gate:

[0036] Update Gate:

[0037] New Status:

[0038] Final state:

[0039] in, represents the sigmoid activation function, represents multi-element multiplication, is the weight matrix, is the current state, is the previous state, is the current input.

[0040] This embodiment uses stacked GRU layers to model the time dependency between crop yield and input agricultural factors. The GRU layer further enhances the impact of short-term weather changes, fertilizer application, etc. on yield. Through the synergy of the parallel LSTM network and the first GRU network, the impact of various factors on crop yield can be comprehensively considered to improve the accuracy of crop yield prediction.

[0041] S5, using a feature fusion layer to fuse the first time series features and the second time series features by using a weighted average method; In an optional embodiment of the present invention, when step S5 uses the feature fusion layer to fuse the first time series feature and the second time series feature in a weighted average manner: Two learnable weight vector parameters are used to represent the relative importance of LSTM and GRU outputs, and the sum of the two weight vector parameters is set to 1; The first time series feature output by the LSTM network and the second time series feature output by the first GRU network are weighted averaged according to the two weight vector parameters, expressed as:

[0042] in, represents the fusion feature, represents the weight vector parameter of the first time series feature, represents the first time series feature, represents the weight vector parameter of the second time series feature, Represents the second time series feature.

[0043] In this embodiment, when the outputs of the LSTM and GRU layers are merged, a weighted average method is used to weight the outputs of the two layers. The specific weighting method is as follows: 1. Calculate the weighting coefficient: Use a learnable weight vector parameter to represent the relative importance of the LSTM and GRU outputs. Define the weights as follows: Set the LSTM output weights to , set the GRU output weight to , so that the sum of these two weights is 1, that is, , ensuring that the weighted output is a valid combination.

[0044] 2. Weighted calculation method: Assumptions and are the outputs of the LSTM and GRU layers respectively. The fused output can be calculated by weighted average:

[0045] here, and are the parameters learned during the training process.

[0046] 3. Weight initialization and training: The optimal weights can be learned through the back propagation algorithm in the neural network training process and . In the early stages of training, these weights can be initialized to be uniformly distributed (e.g.: =0.5, = 0.5), but as training progresses, the model adjusts these two weights based on the characteristics of the data.

[0047] Through this weighted averaging method, the model can automatically adjust the role of LSTM and GRU in time series modeling according to the training data, so as to better integrate the features of long-term dependency (LSTM) and short-term dependency (GRU) and improve the accuracy of prediction.

[0048] S6. Use the second GRU network to extract the temporal characteristics of the impact of agricultural activities and environmental factors on the yield at different stages of the crop growth cycle from the fusion features; In an optional embodiment of the present invention, the second GRU network used in step S6 includes: At least two GRU units are cascaded in sequence, and each GRU unit inputs a time series consisting of fused features.

[0049] This example introduces a second GRU layer based on the features after fusion of LSTM and GRU outputs to further extract the complex relationship between crop yield and agricultural factors. Through this step, the model can better understand the impact of agricultural activities and environmental factors on yield at different stages of the crop growth cycle.

[0050] S7. Use the fully connected layer to perform linear transformation on the temporal features to obtain the final crop yield prediction result.

[0051] In an optional embodiment of the present invention, step S7 flattens the output of the last GRU layer and passes it to a fully connected layer (FCL). The fully connected layer generates the final crop yield prediction value by performing a linear transformation on these extracted features. This output will be combined with factors such as climate, fertilization and pesticide use to predict the final crop yield.

[0052] After obtaining the output of the fully connected layer, necessary post-processing steps are performed, such as comparing the predicted value with the actual yield, adjusting the dimension, etc. This step ensures that the output dimension is consistent with the predicted crop yield structure.

[0053] In order to evaluate the accuracy of the DualSpinNet crop yield prediction model proposed in this embodiment, a comparative experiment was used for comparative analysis.

[0054] The dataset used in this paper comes from Kaggle, named Agricultural Crop Yield in Indian States Dataset. This dataset contains agricultural data of various crops grown in various states of India from 1997 to 2020. It provides important features related to crop yield prediction, including crop type, crop year, planting season, state, planting area, yield, annual rainfall, fertilizer use, pesticide use, and yield. The focus of this dataset is to predict crop yield based on agronomic factors such as weather conditions, fertilizer and pesticide use, and other related variables. The dataset is presented in a tabular form, with each row representing data for a specific crop and its corresponding features. It consists of 19,698 rows and 10 columns (9 features and 1 label). Figure 3 The distribution of data volume in different years in the dataset is shown. Figure 4 The data distribution of annual rainfall in the dataset is shown, with rainfall in milliliters. Figure 5 The distribution of production data in the dataset is shown, with production in metric tons.

[0055] The effectiveness of the yield prediction model was evaluated using mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination ( ) is evaluated. The calculation method is as follows:

[0056]

[0057]

[0058]

[0059] In the above equation, y represents the average statistical yield in units of 103 kg / hm2. y_i and y_i^p represent the actual yield and predicted yield of the ith unit, respectively. The higher the value of the indicator R^2, the better the model performance, and the lower the values ​​of the indicators MSE, RMSE and MAE, the better the model performance.

[0060] The experimental environment uses the Linux operating system, equipped with a 10 vCPU Intel (R) Xeon (R) Gold 6148 CPU @ 2.40GHz processor and an NVIDIA RTX 3080 with 20GB memory. The development language is Python. During the model training phase, the initial learning rate is set to 0.001, and the optimizer used is AdamW. The loss curve during the model training process is shown in Figure 6 It can be observed that when the number of iterations reaches 200, the curve tends to be stable, indicating that the model converges.

[0061] The present invention uses the following seven benchmark models for comparison on the dataset: 1.Linear Linear refers to the basic component of a neural network that performs linear transformation of input data.

[0062] 2. LSTM LSTM is an advanced RNN designed to handle long-term dependencies using a gating mechanism and solve the gradient vanishing problem.

[0063] 3. GRU Gated recurrent units simplify LSTMs by combining gates and reducing complexity, providing similar performance at a lower computational cost.

[0064] 4. BiGRU BiGRU (Bidirectional Gated Recurrent Unit) is an enhanced version of the Recurrent Neural Network (RNN) and a variant of the GRU that incorporates a bidirectional propagation mechanism. Unlike the traditional GRU, BiGRU processes sequence data in both forward and backward directions, enabling it to capture the full context of the sequence.

[0065] 5. Seq2Seq Sequence-to-sequence models transform one sequence into another and are commonly used for translation and text generation, often enhanced by attention mechanisms.

[0066] Table 1 summarizes the performance of various yield prediction models, evaluated using four metrics: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). Among the tested models, DualSpinNet showed excellent performance, achieving the lowest MSE (1.5078E-07), RMSE (3.8830E-04), and MAE (9.7398E-05), as well as the highest R² value of 9.9993, indicating a good fit with the data. In contrast, the linear model showed the most unfavorable results in all metrics.

[0067] Table 2 shows the percentage improvement of the DualSpinNet model compared to the other models. The results show that DualSpinNet has significant enhancements over linear models, with an 89.13% improvement in MSE, a 67.03% improvement in RMSE, and a 57.82% improvement in MAE. In addition, it shows significant advantages over the LSTM, GRU, BiGRU, and Seq2Seq models, especially in terms of MSE and MAE metrics. These results highlight the effectiveness of the DualSpinNet model in improving the accuracy of yield prediction, marking its significant contribution to the field.

[0068] Table 1. Comparative experiments of different models

[0069] Table 2. DualSpinNet model improvement percentage

[0070] Figure 7 The histograms in Figure 1 show the comparison between the proposed DualSpinNet model and the RNN, LSTM, GRU, and Seq2Seq models. The DualSpinNet model has many improvements over the RNN, LSTM, and GRU models, and slightly improved over the Seq2Seq model. The visualization results comparing the predicted yield and the actual yield are shown in Figure 1. Figure 8 As shown. Figure 8 As can be seen from the figure, most of the data points are located on the equality line. The overall fit between the model and the actual data is good, indicating that the model performs well in the task of grain yield prediction.

[0071] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0072] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0074] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0075] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A crop yield prediction method based on LSTM and GRU, characterized in that: The following steps are involved: Acquiring crop planting data; the crop planting data includes climate parameters and soil parameters; Constructing a crop yield prediction model based on LSTM and GRU; the crop yield prediction model includes a parallel LSTM network and a first GRU network, as well as a feature fusion layer, a second GRU network and a fully connected layer; The LSTM network is used to extract the first time series characteristics of the temporal dependence between long-term seasonal changes and climate fluctuations and crop yields from crop planting data; The first GRU network is used to extract the second time series features of the short-term weather changes and the temporal dependence between fertilizer application and crop yield from the crop planting data; The feature fusion layer is used to fuse the first time series features and the second time series features by weighted average method; The second GRU network is used to extract the temporal characteristics of the impact of agricultural activities and environmental factors on the yield of crops at different stages of the growth cycle from the fusion features; The fully connected layer is used to perform linear transformation on the temporal features to obtain the final crop yield prediction result.

2. The crop yield prediction method based on LSTM and GRU according to claim 1, characterized in that: The climate parameters and soil parameters include: Crop type, crop year, cropping season, state, cropping area, yield, annual rainfall, fertilizer use, and pesticide use.

3. The crop yield prediction method based on LSTM and GRU according to claim 1, characterized in that: The LSTM network includes: At least two LSTM units are cascaded in sequence, and each LSTM unit inputs a time series consisting of crop planting data.

4. The crop yield prediction method based on LSTM and GRU according to claim 1, characterized in that: The first GRU network includes: At least two GRU units are cascaded in sequence, and each GRU unit inputs a time series consisting of crop planting data.

5. The crop yield prediction method based on LSTM and GRU according to claim 1, characterized in that: When the feature fusion layer is used to fuse the first time series features and the second time series features by weighted average: Two learnable weight vector parameters are used to represent the relative importance of LSTM and GRU outputs, and the sum of the two weight vector parameters is set to 1; The first time series feature output by the LSTM network and the second time series feature output by the first GRU network are weighted averaged according to the two weight vector parameters, expressed as: in, represents the fusion feature, represents the weight vector parameter of the first time series feature, represents the first time series feature, represents the weight vector parameter of the second time series feature, Represents the second time series feature.

6. The crop yield prediction method based on LSTM and GRU according to claim 1, characterized in that: In the process of crop yield prediction model training, the back propagation algorithm is used to adjust the weight vector parameters of the first time series feature. and the weight vector parameter of the second time series feature Perform iterative optimization to obtain the optimal weight vector parameters and .

7. The crop yield prediction method based on LSTM and GRU according to claim 1, characterized in that: The second GRU network includes: At least two GRU units are cascaded in sequence, and each GRU unit inputs a time series consisting of fused features.

Images