Corn yield estimation method based on multispectral remote sensing and three-dimensional convolutional neural network

Abstract

The invention discloses a corn yield estimation method based on multispectral remote sensing and a three-dimensional convolutional neural network, and the method comprises the following steps: obtaining a multi-temporal remote sensing image of the whole growth period of corn in a target region, and obtaining a multi-temporal remote sensing image of land parcels by land parcels; generating a yield grid map in the target area by adopting an inverse distance weight method based on ground actual measurement yield data; segmenting the yield grid map by taking the boundaries of the land parcels as masks, and taking the average value of all pixels in each land parcel as a yield value; pairing the yield value of each land parcel with the multi-temporal remote sensing images one by one, and constructing an image-yield pairing sample set; matching a sample set by using an image-yield, and training a three-dimensional convolutional neural network model; identifying a to-be-estimated multi-temporal multispectral remote sensing image by using the trained three-dimensional convolutional neural network model to realize plot-level corn prediction; according to the method, high-precision plot-level corn yield prediction is realized by fusing the multi-temporal multispectral remote sensing image and the three-dimensional convolutional neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crop yield estimation, and more particularly to a corn yield estimation method based on multi-spectral remote sensing and a three-dimensional convolutional neural network. BACKGROUND

[0002] Food security is an important foundation of national security, and corn, as one of the three major staple foods, its yield information is of great significance to national food reserves, market regulation and agricultural production management. In order to realize accurate prediction of crop yield, academia and industry have long carried out various technical explorations.

[0003] Early crop yield estimation mainly relies on field investigation and statistical regression model. For example, the prediction system of USDA-NASS of the United States combines farmer survey and field sample survey, and is supplemented by a weather regression model, which has good stability at the national scale. However, this method relies on a large amount of manual sampling and statistical experience, has high cost, long prediction period, and is difficult to meet the application requirements of plot level and rapidness.

[0004] With the development of remote sensing technology, researchers have gradually introduced remote sensing empirical regression methods based on vegetation index (NDVI, EVI, etc.). By using multi-temporal satellite or unmanned aerial vehicle images, combined with meteorological data and crop index, yield prediction can be made weeks to months in advance. This method has relatively low cost and strong migration, but due to its dependence on simple index or linear regression, its ability to describe complex time series information and nonlinear characteristics is limited, and it often has the problem of insufficient generalization in different regions or climate conditions.

[0005] In recent years, machine learning and deep learning methods have been applied to crop yield prediction. For example, a random forest model uses multi-temporal spectral features for fitting, which can improve the accuracy to some extent, but due to its shallow structure, it cannot effectively model spatial texture and temporal dependence, and its prediction ability is limited. Two-dimensional convolutional neural network (2D-CNN) combined with LSTM structure can integrate spatial and temporal features to some extent, but due to the separate extraction of spatial and temporal features, it is easy to cause insufficient temporal dependence information and poor model generalization performance. The closest to the present application is the three-dimensional convolutional neural network (3D-CNN). This method takes the time series remote sensing image as a space-time cube input, which can model spatial and temporal features at the same time, and significantly improves the prediction accuracy compared with random forest and 2D-CNN. However, the existing 3D-CNN model still has two shortcomings: 1. Its parameter size is large, and the calculation and storage overhead is high, which limits its application in large-scale agricultural monitoring scenarios. 2. Lack of differential modeling of key wavebands and key growth period time series information, which tends to treat all inputs equally, leading to interference of redundant information on the prediction result, and insufficient model efficiency and robustness.

[0006] Therefore, how to introduce an effective attention mechanism to enhance the model's use of key information while maintaining the advantages of 3D-CNN in modeling spatio-temporal features, and optimize the input construction scheme to reduce data collection costs and improve prediction timeliness, has become a technical problem to be solved in the field. SUMMARY

[0007] The application aims to provide a corn yield estimation method based on multi-spectral remote sensing and a three-dimensional convolutional neural network, which realizes high-precision plot-level corn yield prediction by fusing multi-temporal multi-spectral remote sensing images and a three-dimensional convolutional neural network.

[0008] To achieve the above-mentioned purpose, the application provides the following technical scheme. A corn yield estimation method based on multi-spectral remote sensing and a three-dimensional convolutional neural network comprises the following steps: S1. A multi-spectral camera is carried by a UAV to obtain multi-temporal remote sensing images of corn in the whole growth period in a target area, and the images are preprocessed to obtain multi-temporal remote sensing images of each plot; S2. Based on ground-truth yield data, a yield grid map in the target area is generated by using an inverse distance weighting method; The yield grid map is segmented with a plot boundary as a mask, and the average value of all pixels inside each plot is taken as a yield value; the yield value of each plot is paired with the multi-temporal remote sensing images of each plot to construct an image-yield paired sample set; S3. The image-yield paired sample set is used to train a three-dimensional convolutional neural network model to obtain a trained three-dimensional convolutional neural network model; S4. The trained three-dimensional convolutional neural network model is used to identify multi-temporal multi-spectral remote sensing images to be estimated to obtain standardized yield values, which are then restored to true yield values through inverse standardization, thereby realizing plot-level corn prediction.

[0009] Further, the bands of the multi-temporal multi-spectral remote sensing images include blue, green, red, infrared, and near-infrared.

[0010] Further, in S1, the preprocessing of the multi-temporal remote sensing images of corn in the whole growth period in the target area comprises the following steps: aerial triangulation, splicing, orthorectification, and cropping are performed by using Pix4Dmapper and ENVI software to obtain multi-temporal remote sensing images of each plot.

[0011] Further, in S3, the three-dimensional convolutional neural network model comprises: A three-dimensional convolutional module comprises three layers of Conv3D structures, and the sizes of the convolution kernels are 1x3x3, 3x3x3, and 3x3x3 in sequence, and BatchNorm3D and LeakyReLU activation functions are combined to extract spatio-temporal features. The channel attention module is used to obtain a statistical description of each spectral channel through global average pooling, and then learn the weight coefficients of each channel through two fully connected structures, so as to perform weighted adjustment on different spectral channels. The time sequence attention module is used to receive the feature sequence after channel weighting, take the feature vectors at different time steps as input sequences, and calculate the inter-temporal dependency in multiple subspaces in parallel through a multi-head attention mechanism. The fully connected regression module is used to flatten and reduce the high-dimensional spatio-temporal features output by the time sequence attention module, and realize the mapping from the deep feature space to the yield value through a series of fully connected layers and nonlinear activation functions.

[0012] Further, the time sequence attention module MHA adopts 4 attention heads, and the calculation formula is: MHA(Q,K,V)=Concat(head1,head2,...,head h )×W (1) (2) where i=1,2,...,h; , , Q, K and V are the query, key and value of the input feature respectively; , , W i is the mapping parameter of the i th attention head; is the feature dimension of each head.

[0013] Further, in the S4, the training process of the three-dimensional convolutional neural network model includes: The Adam optimizer is used, and the batch size is 16; The mean square error is used as the loss function, and the fitting effect of the model is measured by calculating the square difference between the predicted value and the true value; The Dropout random inactivation mechanism is introduced in the convolutional layer and the fully connected layer, the proportion is 0.3-0.5, and the batch normalization operation is added, so as to alleviate the problems of gradient disappearance and distribution drift; The early termination mechanism is enabled, and the training is terminated when the validation loss does not decrease for 20 consecutive rounds.

[0014] Further, in the S4, the true yield is restored by inverse normalization; wherein the calculation formula of inverse normalization is specifically: (3) (4) where T is the time sequence length, C is the number of spectral channels, H and W are the spatial dimensions, indicates the parameter the determined deep learning model mapping function, a yield mean value for the training set, a standard deviation.

[0015] The application further provides an electronic device, comprising: a memory storing a computer program; a processor configured to execute the computer program to implement the method for estimating corn yield based on multi-spectral remote sensing and a three-dimensional convolutional neural network.

[0016] According to the embodiments provided in the application, the application has the following technical effects compared with the prior art: The application uses a multi-spectral camera carried by a UAV to obtain images in the whole growth period, which can dynamically capture the spectral characteristics and growth changes of corn in different growth stages. Compared with traditional single time phase or manual investigation, the data dimension is richer and the timeliness is stronger, which lays a solid data foundation for accurate prediction. The yield grid map is generated by the inverse distance weighting method and combined with the land parcel boundary mask, so as to convert the discrete ground measured data into a continuous label with the same spatial resolution as the remote sensing image and capable of accurately reflecting the average yield inside the land parcel, effectively solving the spatial scale mismatch problem of sample pairing and improving the quality and representativeness of the sample set. The three-dimensional convolutional neural network model can automatically learn deep and abstract features related to yield from two dimensions of space (distribution of pixels inside the land parcel) and time (time sequence changes in the whole growth period), fully excavates the growth rules and yield formation mechanism contained in multi-time phase images, has stronger feature expression ability and nonlinear fitting ability compared with traditional machine learning methods, and thus significantly improves the prediction accuracy. The application also realizes rapid, objective and non-destructive estimation of land parcel level corn yield, which can not only provide fine field management decision support for agricultural producers, but also provide accurate grain yield monitoring data for government departments, and has important practical value and broad application prospect for guaranteeing food security, optimizing agricultural resource allocation and promoting the development of intelligent agriculture. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute the embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor based on the provided drawings.

[0018] The method for estimating corn yield based on multi-spectral remote sensing and a three-dimensional convolutional neural network of the application will be further described below with reference to the drawings. Figure 1is a schematic diagram of the overall process of the corn yield estimation method based on multispectral remote sensing and three-dimensional convolutional neural network in Example 1 of the present application. Figure 2 is a 3D-CNN structure diagram in Example 1 of the present application. Figure 3 is a LOTSO experimental result diagram of different time sequence combinations in Example 1 of the present application; wherein (a) is a LOTSO eight time sequence experimental result diagram, (b) is a LOTSO seven time sequence experimental result diagram, (c) is a LOTSO six time sequence experimental result diagram, and (d) is a LOTSO five time sequence experimental result diagram. Figure 4 is a box plot of the results of different time sequence input schemes in Example 1 of the present application. Figure 5 is a marginal distribution regression scatter plot of the 3DCNN model in Example 1 of the present application. DETAILED DESCRIPTION

[0019] The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0020] In order to better understand the purpose, structure and function of the present application, the present application will be further described in detail below in conjunction with the accompanying drawings.

[0021] Example 1 As shown in Figure 1 , the present application provides a corn yield estimation method based on multispectral remote sensing and three-dimensional convolutional neural network, comprising the following steps: S1, using a unmanned aerial vehicle to carry a multispectral camera to obtain multispectral remote sensing images of corn in the target area during the whole growth period, pre-processing to obtain multispectral remote sensing images of each plot; S2, based on the ground measured yield data, using the inverse distance weighting method to generate a yield grid map in the target area; Segmenting the yield grid map with the plot boundary as a mask, taking the average value of all pixels inside each plot as the yield value; pairing the yield value of each plot with the multispectral remote sensing image of each plot one by one to construct an image-yield paired sample set; The present embodiment is specifically: A multispectral camera is used to obtain multispectral remote sensing images of corn planting plots in the target area. The camera has five narrow wavebands of blue light, green light, red light, red edge and near infrared. Single flight at a height of 100-120m to obtain images with a resolution of about 8cm, covering 8 time sequences of the whole growth period of corn; The aerial triangulation, splicing, orthorectification and cutting are performed by using Pix4Dmapper and ENVI software to obtain multi-temporal images of each plot. Then, the continuous plot yield map is generated by spatial interpolation, and is cut and matched to ensure that the remote sensing data correspond to the measured yield of each plot. All images are unified to 88x88 pixels as input samples of the three-dimensional convolutional neural network model.

[0022] Each plot image is paired with the corresponding yield data. The ground yield data is obtained by field sample measurement, and the inverse distance weighting (IDW) method is used for spatial interpolation of scattered sample points to generate a yield grid map covering the study area. After interpolation, the yield grid map is segmented with the plot boundary as a mask, and the average value of all pixels inside each plot is taken as the plot yield value to realize the spatial one-to-one correspondence of the image and yield data. The yield value is added to the image folder name as label information, so that the model can obtain yield information when reading samples, and a training sample set is constructed. The yield value is normalized by Z-score standardization to improve the convergence and training stability of the three-dimensional convolutional neural network model.

[0023] S3, using the image-yield paired sample set, training the three-dimensional convolutional neural network model to obtain the trained three-dimensional convolutional neural network model; In this embodiment, the three-dimensional convolutional neural network model comprises: Three-dimensional convolution module: a three-layer Conv3D structure is adopted, and the convolution kernel sizes are 1x3x3, 3x3x3 and 3x3x3 in sequence, combined with BatchNorm3D and LeakyReLU to extract spatial and temporal features; Channel attention module (SE): different spectral channels are weighted to enhance the use of key bands; Time sequence attention module (MHA): the dependency between different time sequences is calculated in multiple subspaces to highlight the attention to key growth periods; fully connected regression module: map high-dimensional features to continuous yield output to finally generate a prediction value.

[0024] In the deep learning modeling of multi-temporal remote sensing images, the connection relationship of each part on the data flow is as follows: the input of the model is the multi-temporal remote sensing image sample stacked in time sequence. First, the three-dimensional convolution module takes time and space latitude as the convolution operation object, and extracts the spatial structure features and growth dynamic features of crops at different time phases by sliding the three-dimensional convolution kernel in space. After several layers of convolution and pooling, a preliminary spatio-temporal feature tensor is obtained, which contains spatial evolution information between different growth period images.

[0025] The channel attention module acts on the channel dimension, obtains a statistical description of each spectral channel through global average pooling, and then learns the weight coefficients of each channel through two fully connected structures, thereby performing weighted adjustment on different spectral channels. The features output after the SE module enhance the response ability of the model to key spectral channels (such as infrared or near-infrared) while maintaining the spatial structure, and suppress the noise interference caused by redundant channels.

[0026] The time attention (MHA) module receives the feature sequence after channel weighting, takes the feature vectors at different time steps as the input sequence, and calculates the inter-temporal dependency in multiple subspaces in parallel through the multi-head attention mechanism. Each attention head dynamically allocates weights according to the correlation of the temporal features, thereby focusing on the key growth period that contributes more to yield change. The temporal feature representation output by the time attention (MHA) module integrates the global dependency of the whole time sequence, enabling the model to have long-term memory and adaptive weight allocation capabilities in the time dimension.

[0027] The fully connected regression module flattens and reduces the dimension of the high-dimensional spatio-temporal features output by the MHA module, and realizes the mapping from the deep feature space to the yield value through a series of fully connected layers and nonlinear activation functions. The output layer uses a linear activation function to predict continuous yield values. In the training stage, the model uses mean square error (MSE) as the loss function, and optimizes all convolution kernels and attention parameters through back propagation, finally realizing end-to-end spatio-temporal feature learning and yield regression.

[0028] The calculation formula of the MHA module is as follows: MHA(Q,K,V)=Concat(head1,head2,...,head h )×W (1) (2) Where i=1,2,...,h; , , Q, K and V are the query, key and value of the input features, respectively; , , is the mapping parameter of the i-th attention head; is the feature dimension of each head.

[0029] The training and optimization process of the three-dimensional convolutional neural network model includes: Firstly, the constructed plot image-yield paired sample set is divided into training set and validation set according to the ratio of 4:1 at the sample data level, wherein the training set is used for iterative updating of model parameters, and the validation set is used for evaluating the generalization performance of the model on unseen samples to prevent overfitting. In each plot sample, the input five-dimensional tensor data (5-band multispectral image) and the label are the Z-score standardized plot average yield value.

[0030] In terms of training hyperparameter setting, the batch size of the model is 16, and the Adam optimizer is used for parameter updating. The Adam optimizer combines adaptive learning rate and momentum correction mechanism, which can achieve faster convergence speed in high-dimensional non-convex space. In order to prevent the model from falling into local extreme value in the later training period, the loss function is selected as Mean Squared Error (MSE), which measures the fitting effect of the model by calculating the squared difference between the predicted value and the true value.

[0031] In order to prevent overfitting of deep network, Dropout random inactivation mechanism (proportion 0.3-0.5) is introduced in convolutional layer and fully connected layer, and Batch Normalization operation is added between layers to alleviate the problem of gradient disappearance and distribution drift. During training, the loss change of the validation set is monitored, and when the validation loss does not decrease significantly for a certain number of rounds (set to 20 epochs), the Early Stopping mechanism is triggered to terminate early to prevent performance degradation caused by invalid iterations. The total number of training rounds is set to 150 epochs, and the current model state is saved at the end of each epoch, and the coefficient of determination R 2 , mean absolute error (MAE) and root mean square error (RMSE) and other indicators are calculated on the validation set to dynamically evaluate the model performance.

[0032] In terms of hardware and software environment, the training work is completed on NVIDIA RTX 4090 GPU (24GB video memory) platform, the operating system is Ubuntu 22.04, and the deep learning framework uses PyTorch 2.1.2. In terms of memory management, the Automatic Mixed Precision (AMP) mode is enabled, which effectively reduces the memory occupation and speeds up the forward calculation. To improve the reproducibility of the model, the random seed (seed=42) is fixed at the start of training, and the number of data loading threads is controlled to ensure consistent batch order and repeatable gradient results.

[0033] After training, the model weights with the highest R2 value on the validation set are selected as the final model, and independently evaluated on the test set. The standardized results of the predicted output are de-standardized (multiplied by the standard deviation and added to the mean) to restore them to the true yield units (kg / plot) for comparison with measured yields. The entire training scheme ensures efficient model convergence while taking into account numerical stability, computational efficiency, and generalization performance, providing reliable technical support for crop yield estimation driven by multi-time-series remote sensing images.

[0034] S4. The trained three-dimensional convolutional neural network model is used to identify the multi-temporal and multispectral remote sensing images to be estimated, obtain standardized yield values, and then restore them to the true yield through destandardization, so as to realize plot-level corn prediction.

[0035] The specific formula for calculating destandardization is as follows: (3) (4) Where T is the time series length, C is the number of spectral channels, and H and W are the spatial dimensions, respectively. Indicates by parameters The determined mapping function of the deep learning model, The average output of the training set. The standard deviation is denoted as .

[0036] Example 2 Building upon Example 1, this invention proposes different time-series input construction schemes to optimize data acquisition costs and prediction timeliness, as shown in Figure 3. By gradually reducing the number of time-series nodes, it is found that using only the first four time series (June 30, July 8, July 20, and August 8) can achieve prediction accuracy close to that of a complete 8-time-series (R² = 0.7749, compared to 0.7998 for 8-time-series), enabling high-precision prediction results to be obtained approximately one month earlier, thereby significantly reducing data acquisition frequency and processing costs.

[0037] Example 3 To verify the effectiveness of the multi-temporal remote sensing yield prediction model combining 3D-CNN with SE channel attention and MHA temporal attention proposed in Example 1, this invention designed a multi-model comparison experiment. The experiments were conducted in the same study area, with the same plot samples and data preprocessing conditions. The input was a standardized multi-temporal UAV multispectral image sequence. First, using the traditional machine learning model Random Forest (RF) as a benchmark, the vegetation index and band features extracted from each plot image were input into the model for training and regression, such as... Figure 5The 2D-CNN+LSTM hybrid model is constructed, the single-time two-dimensional convolution feature is modeled through a time sequence network, and then yield prediction results are output; finally, the 3D-CNN+SE+MHA model proposed in the present research is used to perform end-to-end training and reasoning on multi-time images. All the models are trained in the same training set and validation set division manner (8:2 ratio), and the same optimizer and learning rate setting is used to ensure the fairness of the comparison results.

[0038] The experimental results show that under consistent experimental conditions, the R2 of the random forest model is 0.52, the R2 of the 2D-CNN+LSTM model is 0.48, and the R2 of the 3D-CNN+SE+MHA model proposed in the present research is significantly improved to 0.7998, and the root mean square error (RMSE) and the mean absolute error (MAE) are reduced to 73.98 kg / hm 2 and 58.10 kg / hm 2 , respectively. The results show that the model has stronger learning ability in capturing spatio-temporal features, suppressing noise interference and depicting dynamic changes in the growth period of crops, and not only significantly outperforms traditional methods in overall prediction accuracy, but also has higher stability and robustness at different growth stages, and has good early prediction ability.

[0039] In summary, the technical solutions in the present research have the following technical effects: (1) The present research uses a three-dimensional convolutional neural network combined with channel attention and multi-head time sequence attention mechanism to realize differential modeling of spectral and time sequence key information, thereby reducing redundant calculation while maintaining high accuracy; and by optimizing the time sequence input scheme, early prediction is realized under the condition of reducing data acquisition frequency. Compared with existing methods relying on vegetation index, two-dimensional convolution or traditional three-dimensional convolution, the present research has significant improvement in prediction accuracy, model efficiency and early yield estimation ability.

[0040] (2) Compared with the prior art, the present research has significant advantages under the same unmanned aerial vehicle multi-time multi-spectral data and verification conditions. The prediction accuracy R 2 of the random forest model in the present research is only 0.52, the root mean square error (RMSE) is 114.81 kg / hm 2 , and the mean absolute error (MAE) is 91.85 kg / hm2; the two-dimensional convolutional neural network (2D-CNN) combined with LSTM method also fails to surpass traditional machine learning, R 2 is 0.48, RMSE is 119.41 kg / hm 2 , and MAE is 95.53 kg / hm 2, which indicates that it is insufficient for the use of time-dependent information. The scheme of the three-dimensional convolutional neural network combined with the multi-dimensional attention mechanism proposed in the present application achieves the best performance under the same experimental conditions, in which the R2 of the 3D+SE+MHA model is improved to 0.7998, and the RMSE is reduced to 73.98 kg / hm 2 , and the MAE is 58.10 kg / hm 2 . The prediction accuracy and stability are significantly better than the existing methods.

[0041] In addition, by optimizing the time input construction scheme, the present application finds that only 4 key time nodes (June 30, July 8, July 20, and August 8) can achieve a prediction effect close to that of 8 time inputs (R 2 0.7749, compared with the R 2 0.7998 of the complete 8 time inputs), which reduces the data acquisition amount by nearly half while ensuring high accuracy and advances the prediction time by about a month. This not only reduces the cost of data acquisition and processing, but also makes the yield prediction more efficient and timely.

[0042] Therefore, compared with the existing methods based on vegetation index, random forest or two-dimensional convolutional network, the present application can significantly improve the prediction accuracy and robustness, reduce the data acquisition frequency, realize early high-precision corn yield prediction, and has higher practical value and promotion potential.

[0043] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks, characterized in that, Includes the following steps: S1. Use a drone equipped with a multispectral camera to acquire multi-temporal remote sensing images of maize throughout its entire growth period in the target area, perform preprocessing, and obtain multi-temporal remote sensing images of each plot. S2. Based on the measured yield data on the ground, a yield raster map of the target area is generated using the inverse distance weighting method; The yield raster map is segmented using the plot boundaries as a mask, and the average value of all pixels within each plot is taken as the yield value. The yield value of each plot is paired one-to-one with the multi-temporal remote sensing images of each plot to construct an image-yield pairing sample set. S3. Using the image-output paired sample set, train the three-dimensional convolutional neural network model to obtain the trained three-dimensional convolutional neural network model. S4. The trained three-dimensional convolutional neural network model is used to identify the multi-temporal and multispectral remote sensing images to be estimated, obtain standardized yield values, and then restore them to the true yield through destandardization, so as to realize plot-level corn prediction.

2. The method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks according to claim 1, characterized in that, The bands of the multi-temporal multispectral remote sensing image include: blue, green, red, infrared, and near-infrared.

3. The method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks according to claim 1, characterized in that, In S1, the process of preprocessing multi-temporal remote sensing images of maize throughout its entire growth period in the target area includes: using Pix4Dmapper and ENVI software to perform aerial triangulation, stitching, orthorectification, and cropping to obtain multi-temporal images of each plot.

4. The method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks according to claim 1, characterized in that, In S3, the three-dimensional convolutional neural network model includes: The 3D convolution module includes a three-layer Conv3D structure with convolution kernel sizes of 1×3×3, 3×3×3, and 3×3×3, respectively. It combines BatchNorm3D and LeakyReLU activation functions to extract spatiotemporal features. The channel attention module is used to obtain the statistical description of each spectral channel through global average pooling, and then learn the weight coefficients of each channel through a two-layer fully connected structure, thereby weighting and adjusting different spectral channels. The temporal attention module receives the feature sequence after channel weighting, takes the feature vectors at different time steps as the input sequence, and calculates the temporal dependencies in parallel in multiple subspaces through a multi-head attention mechanism. The fully connected regression module is used to flatten and reduce the dimensionality of the high-dimensional spatiotemporal features output by the temporal attention module. It achieves the mapping from the deep feature space to the output value through a series of fully connected layers and nonlinear activation functions.

5. The method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks according to claim 4, characterized in that, The temporal attention module MHA uses four attention heads, and the calculation formula is as follows: MHA(Q,K,V)=Concat(head1,head2,...,head h )×W (1) （2） Where i = 1, 2, ..., h; , , These represent the query, key, and value of the input features, respectively. , , For the mapping parameters of the i-th attention head; The feature dimensions for each head.

6. The method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks according to claim 1, characterized in that, In step S4, the training process of the three-dimensional convolutional neural network model includes: The Adam optimizer is used, with a batch size of 16. The model's fit is measured by calculating the squared difference between the predicted and actual values, using the mean squared error as the loss function. A Dropout random deactivation mechanism with a ratio of 0.3–0.5 is introduced into the convolutional and fully connected layers, and a batch normalization operation is added to alleviate the gradient vanishing and distribution drift problems. An early termination mechanism is enabled, and training is terminated when the verification loss does not decrease for 20 consecutive rounds.

7. The method for estimating maize yield based on multispectral remote sensing and three-dimensional convolutional neural networks according to claim 1, characterized in that, In step S4, the true output is restored through destandardization; the specific formula for destandardization is as follows: （3） （4） Where T is the time series length, C is the number of spectral channels, and H and W are the spatial dimensions, respectively. Indicates by parameters The determined mapping function of the deep learning model, The average output of the training set. The standard deviation is denoted as .

8. An electronic device, characterized in that, include: Memory, which stores computer programs; The processor, when executing the computer program, implements the corn yield estimation method based on multispectral remote sensing and three-dimensional convolutional neural networks as described in any one of claims 1-7.

Images