Crop yield prediction method and system based on optimized HATCN-MLP

Abstract

The application discloses a crop yield prediction method and system based on an optimized HATCN-MLP, relates to the technical field of crop yield prediction, and comprises the following steps: acquiring historical remote sensing time series data, historical yield data and growth cycle information of crops in a target area during a growth period, and constructing a data set; building an MLP baseline model, configuring multiple structure combinations of a hidden layer of the model, training and testing the model by using the data set, and screening an optimal configuration of the structure combinations according to a test result; introducing a hybrid time sequence attention mechanism into a time sequence convolution network, constructing an HATCN model, and then combining the HATCN model with the optimized MLP baseline model to form a composite prediction model; configuring multiple structure combinations of a hidden layer of the HATCN model, training and testing the HATCN model, and screening an optimal configuration of the structure combinations; acquiring remote sensing time series data of crops in the target area during a current growth period, inputting the remote sensing time series data into the optimized HATCN-MLP composite prediction model, and outputting an accurate crop yield prediction result.

Description

Technical Field

[0001] This invention relates to the field of crop yield prediction technology, and in particular to a crop yield prediction method and system based on optimized HATCN-MLP. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Crop yield forecasting is a core component of precision agriculture and is of great significance for food security planning and the optimal allocation of agricultural resources. Currently, crop yield forecasting methods mainly rely on basic models such as Multilayer Perceptrons (MLPs). These models use fully connected layers to map input features onto remote sensing data of the crop growth period to achieve yield prediction. However, existing crop yield forecasting methods have certain limitations: Firstly, traditional MLP models rely solely on fully connected mapping and lack temporal modeling capabilities. They cannot explicitly mine phenological dynamics and long-term dependencies in remote sensing time-series data of crop growth stages, resulting in insufficient ability to capture temporal features of key crop growth stages and limited nonlinear representation capabilities. Consequently, the models exhibit low prediction accuracy and poor stability. Secondly, the model structure... The design lacks specificity for remote sensing time-series data. Low parameter count (<3K) models have insufficient capacity, are prone to underfitting, and have poor prediction accuracy and stability. While high parameter count (>8K) can improve prediction accuracy to some extent, the model has parameter redundancy, is less efficient, and the marginal benefit of increasing the number of parameters to improve accuracy diminishes, making it impossible to achieve an optimal balance between accuracy and efficiency. This can easily lead to redundant training and performance saturation. In addition, the decision-making mechanism of existing models is vague, and it is impossible to clearly define the contribution weight of remote sensing data at different growth stages to yield prediction. This makes it difficult to optimize remote sensing monitoring strategies in a targeted manner and fails to fully support the needs of precision agriculture monitoring. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a crop yield prediction method and system based on an optimized HATCN-MLP. Based on the characteristics of remote sensing time-series data during the crop growth period, it optimizes the design of a multilayer perceptron (MLP) baseline model and a hybrid attention temporal convolutional network (HATCN) model, combining the two optimized models to construct a composite prediction model. This achieves more efficient and accurate crop yield prediction. Furthermore, by visualizing the attention weights of the HATCN in the model, the interpretability of the model is enhanced, solving the problem that existing crop yield prediction schemes struggle to balance accuracy and efficiency while lacking interpretability.

[0005] In a first aspect, the present invention provides a crop yield prediction method based on optimized HATCN-MLP.

[0006] A crop yield prediction method based on optimized HATCN-MLP includes: Acquire historical remote sensing time-series data, historical yield data, and growth cycle information of crops in the target area during their growth period, and construct a dataset; Build an MLP baseline model, configure various structural combinations of the model's hidden layers, train and test the model using a dataset, and select the optimal configuration of the structural combinations based on the test results; A hybrid temporal attention mechanism is introduced into the temporal convolutional network to construct the HATCN model, which is then combined with the optimized MLP baseline model to form a composite prediction model. Multiple structural combinations of the hidden layers of the HATCN model are configured, and the composite prediction model is trained and tested using a dataset. The optimal configuration of the structural combination is selected based on the test results. The remote sensing time series data of the target area during the current crop growth period are obtained, input into the optimized HATCN-MLP composite prediction model, and the crop yield prediction results are output.

[0007] Further technical solutions include crop growth cycle information, which includes different growth stages of the crop; In the process of crop yield prediction, the attention weight of each time step is calculated by the hybrid temporal attention mechanism of the HATCN model in the composite prediction model. Combined with the different growth stages of the crop, the contribution of different growth stages to yield prediction is visualized and quantified.

[0008] A further technical solution is that the MLP baseline model consists of an input layer, several hidden layers, and an output layer, with all layers connected using a fully connected method; the optimization process of the MLP baseline model is as follows: Configure various structural combinations of several hidden layers in the MLP baseline model and fix the hyperparameter configuration; among them, the structural combinations include equal-width structures and progressively shrinking structures with 2-4 layers of different numbers of hidden neurons; The dataset, which includes historical remote sensing time-series data and historical yield data, is divided into training and testing sets. MLP baseline models with different structural combinations are trained and tested, and the values ​​of multiple evaluation metrics for model testing are recorded. Among them, multiple evaluation metrics include R², rRMSE, rMAE and parameter count. Using different structural combinations as candidate parameter combinations, the Pareto front multi-objective optimization method is adopted to map the candidate parameter combinations to a multi-dimensional objective space corresponding to multiple evaluation indicators, identify non-dominated solutions, and construct a comprehensive scoring function to calculate the comprehensive score of each non-dominated solution. The candidate solution with the highest comprehensive score is selected as the optimal configuration of the MLP baseline model.

[0009] A further technical solution is to construct a basic network based on the causal convolution and dilated convolution mechanisms of temporal convolutional networks, and then introduce a hybrid temporal attention mechanism into the basic network to construct the HATCN model. The hybrid temporal attention mechanism includes a multi-scale stage attention calculation module, a global attention calculation module, and a gated fusion module. The multi-scale stage attention calculation module calculates the attention weights for each stage and aggregates the temporal features within each stage. Based on the aggregated features, it integrates the information from each stage through learnable stage fusion weights to generate multi-scale features. The global attention calculation module calculates long-term features across stages to generate global features. The gated fusion module fuses multi-scale features and global features through a gating mechanism to obtain the final attention weights. Remote sensing time-series data is input into the HATCN model. Temporal features are first extracted through a temporal convolutional network. The temporal features are then used to generate attention weights through a hybrid temporal attention mechanism. The temporal features at the corresponding time steps are weighted and summed according to the attention weights to generate weighted temporal features.

[0010] A further technical solution involves constructing a 1-3 layer HATCN model as the front-end temporal feature extraction module. The output of the temporal feature extraction module is then fed into the optimized MLP baseline model, forming a composite prediction model based on HATCN-MLP. The parameter optimization process for the HATCN model is as follows: The hyperparameter configuration is fixed, and various structural combinations of the hidden layers of the HATCN model are configured, including equal-width and progressively shrinking structures with different numbers of hidden neurons, and different convolutional kernel widths are designed. The dataset, which includes historical remote sensing time-series data and historical yield data, is divided into training and testing sets. Composite prediction models with different structural combinations are trained and tested. The values ​​of multiple evaluation indicators of the model tests are recorded, and the optimal parameter configuration of the HATCN model is selected based on the values ​​of multiple evaluation indicators.

[0011] Secondly, the present invention provides a crop yield prediction system based on optimized HATCN-MLP.

[0012] A crop yield prediction system based on optimized HATCN-MLP includes: The data acquisition module is used to acquire historical remote sensing time-series data, historical yield data, and growth cycle information of crops in the target area during their growth period, and to construct a dataset. The MLP baseline model optimization module is used to build an MLP baseline model, configure various structural combinations of the model's hidden layers, train and test the model using a dataset, and select the optimal configuration of the structural combination based on the test results. The composite prediction model construction and optimization module is used to introduce a hybrid temporal attention mechanism into the temporal convolutional network, construct the HATCN model, and then combine it with the optimized MLP baseline model to form a composite prediction model. It configures various structural combinations of the hidden layers of the HATCN model, trains and tests the composite prediction model using a dataset, and selects the optimal configuration of the structural combination based on the test results. The yield prediction module is used to acquire remote sensing time-series data of the current crop growth period in the target area, input it into the optimized HATCN-MLP composite prediction model, and output the crop yield prediction results.

[0013] Further technical solutions also include: The visualization and quantification module is used to calculate the attention weights at each time step in the crop yield prediction process through the hybrid temporal attention mechanism of the HATCN model in the composite prediction model. Combined with the division of different crop growth stages, it visualizes and quantifies the contribution of different growth stages to yield prediction. The crop growth cycle information includes different growth stages of the crop.

[0014] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing executable instructions; and a processor for implementing the above-described crop yield prediction method based on optimized HATCN-MLP when executing the executable instructions stored in the memory.

[0015] Fourthly, the present invention also provides a computer-readable storage medium storing executable instructions for causing a processor to execute the executable instructions to implement the above-described crop yield prediction method based on optimized HATCN-MLP.

[0016] Fifthly, the present invention also provides a computer program product comprising executable instructions stored in a computer-readable storage medium; wherein, when the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the above-described crop yield prediction method based on optimized HATCN-MLP.

[0017] The above one or more technical solutions have the following beneficial effects: 1. This invention provides a crop yield prediction method and system based on an optimized HATCN-MLP. Based on the characteristics of remote sensing time-series data during the crop growth period, an optimized multilayer perceptron (MLP) baseline model is designed. A hybrid attention temporal convolutional network (HATCN) model is constructed by fusing causal convolution, dilated convolution, and a hybrid attention mechanism (HA) within a temporal convolutional network (TCN). This HA mechanism combines multi-scale stage attention and global attention to adaptively extract and enhance multi-scale stage temporal features during the crop growth period, improving the accuracy and robustness of yield prediction. The HATCN model is then combined with the optimized MLP baseline model to construct a composite prediction model, achieving more efficient and accurate crop yield prediction. Furthermore, the attention weights of the HATCN in the model can be visualized to quantify the contribution of remote sensing data at different growth stages to yield prediction, enhancing model interpretability and providing theoretical support for precision agriculture monitoring. This addresses the problem that existing crop yield prediction schemes struggle to balance accuracy and efficiency while lacking interpretability.

[0018] 2. This invention, through structural optimization and the introduction of the HATCN module, can significantly reduce the standard deviation of model metrics, avoiding the underfitting fluctuations of low-parameter models and the redundant interference of high-parameter models. It can maintain stable performance in different training rounds and data from different years, adapting to practical application scenarios. Moreover, the optimal HATCN-MLP composite model has only about 6.3K parameters, which is in the medium parameter range. Compared with high-parameter MLP models, it reduces the number of parameters and has better error metrics. Compared with low-parameter MLP models, it improves prediction performance and stability, and can balance training efficiency and prediction performance.

[0019] 3. The composite prediction model proposed in this invention can adapt to the different growth stage characteristics of different crops such as wheat and corn through adaptive weight adjustment. In particular, considering the characteristics of corn being sensitive to water and heat conditions and having large weight fluctuations, it can reduce interannual prediction errors through key stage monitoring and optimization. Its applicable scope covers major food crops, making it highly practical. At the same time, through the visualization of time-series weights, the differential growth stage contribution patterns of wheat and corn are clarified, providing guidance for precision agriculture. It can also strengthen remote sensing monitoring of key stages in a targeted manner to further improve prediction accuracy.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, 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.

[0022] Figure 1 This is an overall flowchart of the crop yield prediction method based on optimized HATCN-MLP in Embodiment 1 of the present invention; Figure 2 This is a graph showing the distribution of indicators and the overall score under different HATCN model configurations in Embodiment 1 of the present invention; Figure 3 This is a visualized line graph showing the contribution weights of different crop growth stages to yield prediction in Embodiment 1 of the present invention; where (a) represents winter wheat and (b) represents summer maize. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and are intended only to describe specific embodiments and to provide further explanation of the invention, and are not intended to limit the scope of exemplary embodiments of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] Example 1 Existing crop yield prediction methods rely on basic models such as Multilayer Perceptrons (MLPs), which only map input features through fully connected layers. This fails to explicitly model phenological dynamics and long-term dependencies in remote sensing time-series data, resulting in insufficient ability to capture temporal features of key crop growth stages, limited prediction accuracy, and poor stability. Especially under high parameter scales, redundant training and performance saturation issues easily arise. This embodiment proposes a crop yield prediction method based on an optimized HATCN-MLP. By optimizing the MLP baseline structure, optimizing the design of the HATCN network for temporal modeling, and combining weight visualization analysis, accurate, stable, and interpretable crop yield estimation and prediction are achieved. The proposed crop yield prediction method specifically includes the following steps: Step S1: Obtain historical remote sensing time-series data, historical yield data, and growth cycle information of crops in the target area during their growth period, and construct a dataset.

[0025] In this embodiment, remote sensing time-series data of crops (wheat: October to June of the following year; maize: May to September) during the crop growing season in the target area over multiple years (e.g., the last 10 years, with a minimum of 5 years to ensure sufficient sample size) are collected. The data source is MODIS (spatial resolution 250m / 500m, temporal resolution 8 days / 16 days). Core variables include: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Land Surface Temperature (LST), and Photosynthetically Active Radiation Absorbance (fPAR). Simultaneously, actual crop yield data (i.e., statistical yearbook) for the corresponding years in the target area are obtained. Preferably, meteorological driving variables for the corresponding time periods, such as precipitation and average temperature, can be collected simultaneously. The data source is ERA5 reanalysis data, with a spatial resolution of 0.25° and a temporal resolution of 1 day. Both the remote sensing data and the meteorological data are time-series data with the same temporal resolution.

[0026] As one implementation method, the collected remote sensing time-series data is preprocessed with noise reduction, missing value imputation, resolution unification, and standardization to obtain standardized remote sensing time-series data; the collected crop actual yield data for multiple years in the target area is preprocessed with outlier removal, normalization, or standardization to lay the foundation for the learning and training of the subsequent prediction model and ensure the accuracy of the final prediction results.

[0027] Furthermore, crop growth cycle information includes different growth stages of the crop. Remote sensing time series data usually contains obvious phenological change patterns, such as temporal fluctuations in key stages like greening, jointing, heading, and grain filling. In this embodiment, the crop (wheat, corn, etc.) growth cycle is divided into four key stages: wheat can be divided into: sowing-overwintering period, greening-jointing period, booting-heading period, and grain filling-maturity period; corn can be divided into: sowing-emergence period, seedling-jointing period, tasseling-flowering period, and grain filling-maturity period.

[0028] Step S2: Build an MLP baseline model, configure various structural combinations of the model's hidden layers, train and test the model using the dataset, and select the optimal configuration of the structural combinations based on the test results.

[0029] Specifically, the Multilayer Perceptron (MLP) is the most fundamental and widely used feedforward neural network structure in deep learning. Its core idea is to learn the complex relationship between input features and the output target through multiple layers of nonlinear mapping. An MLP consists of an input layer, several hidden layers, and an output layer, all connected in a fully connected manner; that is, every neuron in one layer is connected to all neurons in the next layer. For any given layer, the computation process can be represented as: ; in, This is the output of the previous layer. and These are the weight matrix and the bias term, respectively. This is a non-linear activation function. Introducing this activation function allows the network to approximate any continuous function, thus possessing strong feature representation capabilities.

[0030] Model training is achieved by minimizing the loss function. To achieve this, the loss function can be expressed as: ; in, This represents all trainable parameters. These are the model's predicted values. The loss is at the sample level. Parameter updates rely on the backpropagation algorithm, which calculates the gradient of the loss with respect to the parameters of each layer based on the chain rule, and uses an optimizer for iterative updates, allowing the model to gradually improve its fitting ability. Due to its simple structure, computational efficiency, and ease of modeling data patterns of different scales, MLP is widely used in tasks such as remote sensing numerical prediction and image classification.

[0031] The Pareto Front (PF) is a commonly used analytical method in multi-objective optimization, aiming to find the optimal trade-off between multiple competing evaluation metrics. In model design and parameter selection, different settings often simultaneously affect model performance, complexity, and training cost, meaning that the optimal solution for a single metric does not necessarily represent the global optimum. The Pareto Front identifies a set of "non-dominated solutions"—parameter combinations that cannot further improve other metrics without reducing the performance of any one metric—thus providing a systematic characterization of the trade-offs between multiple metrics. Let the parameters to be optimized form the decision variables. The objective function must include optimal prediction performance. Such as the coefficient of determination, error, and the number of model parameters to be minimized. For any two candidate configurations and If satisfied If at least one of them is strictly superior, then it is considered that Dominate ( The sets of parameters that cannot be dominated by any other candidate solution constitute Pareto optimal solutions, and their set is the Pareto front, which can be expressed as: ; in, The set of all feasible solutions (model parameters). and As a candidate solution, To solve Performance vectors across multiple objectives This is the set of objective vectors corresponding to the non-dominated solutions, i.e., the Paruto front.

[0032] In this embodiment, to evaluate the performance of the most basic deep learning structure in the crop yield estimation task, a baseline model MLP-base consisting only of MLPs was first constructed and tested. This MLP baseline model only relies on the fully connected mapping of input features, and therefore can reflect the lower bound performance of the model under the conditions of no deep temporal modeling and no spatial modeling.

[0033] Next, to determine the appropriate network size for this task, various structural combinations of hidden layers in the MLP baseline model were configured based on the number of samples and features used, with fixed hyperparameter configurations, for experimentation. Specifically, the structural combinations included equal-width structures with 2-4 layers of different numbers of hidden neurons and progressively shrinking structures. The equal-width structure (e.g., 64-64) was used to evaluate the expressive power under a fixed capacity, while the progressively shrinking structure (e.g., 64-32) was used to simulate common feature compression processes. To avoid excessive differences in dimensionality between layers leading to training instability, this embodiment did not use a cross-level shrinking structure (e.g., 64-16). Furthermore, in all combinations, only the hidden layer size was adjusted; the output layer was always a single-neuron linear layer (fully connected layer, with a fixed number of neurons of 1). The remaining fixed hyperparameter configurations were: optimizer (Adam), learning rate (1e...). -3 Learning rate scheduler (ReduceLROnPlateau), scheduler adjustment ratio (0.5), scheduler patience value (10), weight decay (1e) -4 ), Maximum number of training rounds (Epoch) (500), Early stop patience value (50).

[0034] Then, the dataset, including historical remote sensing time-series data and historical yield data, was divided into training and testing sets. MLP baseline models with different structural combinations were trained and tested. Multiple evaluation metrics for wheat and maize under different MLP model structures were recorded, including R², rRMSE, rMAE, and parameter counts. The metrics are the average values ​​for both crops, and the mean ± standard deviation was used to record the metrics data for all years. The data were averaged three times, and this average was used as the basis for determining the optimal MLP-base structure. The experimental results are shown in Table 1 below.

[0035] Table 1. Prediction Performance of Models with Different MLP-Base Structures

[0036] The experimental results show that both network depth and width affect the model's predictive performance, but the improvement is limited and exhibits a clear diminishing returns trend. As the structure is expanded from 2 layers to 3-4 layers, the model's average R² increases from approximately 40%-42% to 45%-46%, while rRMSE decreases from approximately 10% to around 9.6%, indicating that appropriately increasing network depth can improve the nonlinear representation capability.

[0037] In the high parameter range (>8K), such as 64-64-32 (9.79K) and 64-64-64 (11.97K), although they achieved the highest R² (46.91% and 46.63%), these structures did not exhibit a significant advantage commensurate with their parameter size. Their performance did not increase linearly with the increase in parameter count; instead, they exhibited a phenomenon of "excessive structural complexity and declining marginal returns." This indicates that overly deep or wide MLP structures do not effectively mine more useful features, and some parameters may be redundant. Therefore, high parameter structures do not offer significant cost-effectiveness, and their predictive improvement is insufficient to support the expansion of their model size.

[0038] In contrast, within the medium parameter range (approximately 5-7K), the model performance exhibits a better balance between efficiency and stability. Among these, the 64-32 (5.50K) structure performs exceptionally well, with an R² of 45.75%, outperforming the similarly parameter-rich 64-32-16 (6.05K, 44.49%) and also exceeding the smaller parameter-rich 32-32 (2.82K, 41.28%) and 32-16 (2.24K, 41.12%). Regarding error metrics, the 64-32 structure has an rRMSE of 9.68%, one of the lowest among all structures with fewer than 7K parameters, while its rMAE of 7.64% is also among the best in this range. Furthermore, the structure exhibits low variability in repeated experiments (R²±9.99, rRMSE±1.62, rMAE±1.47), outperforming high-parameter models and most models with smaller parameters. This demonstrates that it effectively captures the nonlinear contribution of temporal features to output while maintaining moderate complexity. The model size and data size are well-matched, with no excessive stacking in network depth and width, and its prediction accuracy and stability are balanced.

[0039] In the low parameter range (<3K), while the models offer a lightweight advantage, they exhibit significant capacity limitations, resulting in generally low average performance. For example, the R² of the 16-16 (1.15K) and 16-16-16 (1.46K) models drops to 38.05% and 34.88%, respectively. More importantly, the stability of low-parameter models is significantly insufficient, with standard deviations of R² generally reaching ±12-15, such as 32-32 (±15.20), 32-16 (±15.02), and 16-16 (±15.08). This high standard deviation reflects the model's susceptibility to initialization and data perturbations across different training epochs, exhibiting typical underfitting characteristics: the model capacity is insufficient to fully learn the complex nonlinear relationship between remote sensing ecology and yield, thus resulting not only in low average predictive power but also significantly increased output volatility.

[0040] The performance across the three parameter ranges indicates that performance improvement is not driven by a monotonic increase in the number of parameters, but rather by the matching relationship between model capacity and data scale. Building upon this, this embodiment introduces a Pareto front multi-objective optimization method. Different structural combinations are used as candidate parameter combinations, which are then mapped to a multi-dimensional objective space corresponding to multiple evaluation indicators. Non-dominated solutions are identified, and a comprehensive scoring function is constructed to calculate the comprehensive score for each non-dominated solution. The candidate solution with the highest comprehensive score is selected as the optimal configuration for the MLP baseline model.

[0041] Specifically, candidate parameter combinations are mapped to a four-dimensional target space (Ri). 2 The system identifies candidate points that are superior in terms of both accuracy and model complexity and are not dominated by other solutions, based on Pareto dominance, and forms a set of non-dominated solutions as alternative superior structures.

[0042] Furthermore, since the Pareto front typically contains multiple non-dominated solutions, this embodiment constructs a comprehensive scoring function to unify the solutions in order to select a final configuration, defined as: ; in, For candidate solutions The overall score, For the number of indicators, For the first The weight of each indicator, This represents the normalized target indicator value. This embodiment selects four indicators (…). ), and adopt equal weight settings ( This ensures that the final result objectively reflects the overall performance of the model in a multi-objective space.

[0043] Based on the selection of PF and multi-objective comprehensive score, the 64-32 structure model with the highest comprehensive score of 0.77 and corresponding to the medium parameter range is selected. This model can effectively avoid the redundancy and training efficiency decline of high parameter structure while ensuring sufficient expressive power. At the same time, it is also better than similar parameter scale models in terms of standard deviation and stability of the index. Therefore, it is the optimal configuration for building the MLP-base model in the future.

[0044] Step S3: Introduce a hybrid temporal attention mechanism into the temporal convolutional network to construct the HATCN model, and then combine it with the optimized MLP baseline model to form a composite prediction model. Configure various structural combinations of the hidden layers of the HATCN model, train and test the composite prediction model using the dataset, and select the optimal configuration of the structural combination based on the test results.

[0045] Temporal Convolutional Networks (TCNs) are a class of deep learning architectures specifically designed for processing time-series data. Their core idea is to replace recursive structures with one-dimensional convolutions, achieving more stable and efficient temporal feature extraction and modeling through parallel computation. Two key mechanisms of TCNs are causal convolution and dilated convolution. Causal convolution ensures that the network's prediction of the current output depends only on current and past inputs, thus maintaining strict temporal dependencies. Dilated convolution further expands the receptive field, capturing dependencies over a longer time span by inserting holes between convolutional sampling positions, allowing the convolutional kernel to cover a longer sequence span without increasing the number of parameters. The feature extraction and computation of temporal convolutional networks can be represented as: ; in, To output the value of the feature at time step t, where t is the current time index. The kernel size is [size]. k This refers to the position index within the convolution kernel. For convolution kernel parameters, For the input sequence at time td The value of k, the rate of expansion As the number of network layers increases exponentially, TCNs can model long-term dependencies with relatively shallow depth. TCNs often incorporate residual structures to enhance information flow, mitigate gradient vanishing, and preserve cross-layer features, thereby improving overall training stability. Thanks to the large receptive field provided by dilated convolutions and the multi-scale feature representations formed by multiple convolutional layers, TCNs can capture short-term fluctuations and long-term dependencies with fewer parameters, thus outperforming traditional temporal deep feature extraction methods such as 1D-CNNs in feature extraction capabilities. Based on these advantages, TCNs are widely used in many tasks requiring long-sequence pattern recognition and are often used to extract stable and interpretable key features from complex temporal data.

[0046] To further enhance the feature extraction capability and interpretability of the TCN model, this embodiment constructs a base network based on the causal convolution and dilated convolution mechanisms of the temporal convolutional network (TCN). A hybrid temporal attention mechanism (Hybrid Attention, HA) is introduced into this base network to construct the HATCN model, which can combine multi-scale stage attention and global attention to adaptively learn the importance of different growth stages for yield prediction.

[0047] Specifically, the preprocessed remote sensing time-series data is input into the HATCN model, and temporal features are first extracted through the temporal convolutional network TCN. The temporal features are then used to generate attention weights through a hybrid temporal attention mechanism (HA). The temporal features at the corresponding time steps are then weighted and summed based on the attention weights to generate weighted temporal features.

[0048] The hybrid temporal attention mechanism includes a multi-scale stage attention calculation module, a global attention calculation module, and a gating fusion module. In the multi-scale stage attention calculation module, attention weights are calculated for each stage, and the temporal features within each stage are aggregated to obtain the aggregated features. , can be represented as: ; In the above formula, For TCN at time step The output characteristics, For the first The set of time steps contained in each growth stage. In the first The first growth stage Attention score at each time step and These represent the current time step of interest and all reference time steps used for normalization within the same phase, respectively.

[0049] Among them, attention score Represented as: ; In the above formula, The output features of TCN at time step t; For activation functions; and It consists of learnable weight matrices and biases. This represents a learnable weight vector, whose parameters are learned and optimized through model training.

[0050] Subsequently, based on the aggregated features, information from each stage is integrated through learnable stage fusion weights to generate multi-scale features. , can be represented as: ; In the above formula, For the first The weights of each growth stage in the final multi-scale features can be used to train scalar parameters, which can be adaptively adjusted by the optimizer. This represents the total number of growth stages. and These represent the current growth stage of interest and all stages used as a reference for normalization, respectively.

[0051] While computing multi-scale features, the global attention computation module introduces global attention to capture long-term features across stages and generate global features. , can be represented as: ; In the above formula, T This represents the total number of time steps, i.e., the total number of time points across all growth stages. Let t be the global attention score at time step t.

[0052] Similarly, this global attention score It can be represented as: ; In the above formula, The output features of TCN at time step t; For activation functions; and It consists of learnable weight matrices and biases. This represents a learnable weight vector, whose parameters are learned and optimized through model training.

[0053] Finally, in the gating fusion module, multi-scale features and global features are fused through a gating mechanism to obtain the final attention weights. , can be represented as: ; ; In the above formula, This is a learnable weight matrix used to calculate the gating coefficients. For learnable bias vectors, This indicates element-wise multiplication. This represents vector concatenation. For each feature dimension, there is an activation function and a corresponding gating weight.

[0054] Thanks to the aforementioned staged modeling, multi-scale feature fusion, and global attention mechanism, HA can enhance the temporal feature extraction capability of TCN, making it superior to single attention mechanisms or traditional TCN in capturing key time steps and long-term dependencies. In this embodiment, this mechanism is combined with TCN to construct the HATCN model, which is used to efficiently extract multi-scale temporal features during crop growth.

[0055] Building upon this, considering that the MLP-based model relies solely on the original input features for mapping, its structure cannot explicitly capture the temporal dependencies within the crop's growth cycle. Remote sensing time series typically contain clear phenological change patterns, such as temporal fluctuations in key stages like greening, jointing, heading, and grain filling. These dynamic features are often highly correlated with the final crop yield, and relying solely on MLP mapping makes it difficult to fully utilize the internal structure of the time series. To address this issue, this embodiment introduces a Hybrid Attention Temporal Convolutional Network (HATCN) as a temporal feature extraction module before the MLP-based model. HATCN can effectively model long-term dependencies through causal convolution and dilated convolution, enabling it to capture multi-scale phenological dynamics while maintaining high efficiency. It also possesses the advantages of hybrid attention, adaptively highlighting features of key growth stages and important time steps, enhancing the perception of local fluctuations and global trends, thereby improving the expressiveness and discriminability of time series features. By combining HATCN with MLP, the model can simultaneously possess the ability to model time structure and cross-variable fusion, thereby making fuller use of the deep dynamic information in remote sensing time series. This solves the problem that existing models cannot explicitly model phenological dynamics and long-term dependencies in remote sensing time series data, improves the model's ability to capture the temporal characteristics of key stages of crop growth, and thus improves the accuracy and stability of predictions.

[0056] Furthermore, similar to the optimization method of MLP-base, in this embodiment, a model is constructed using 1-3 layers of HATCN as the front-end temporal feature (i.e., deep feature) extraction model. The deep temporal features output by HATCN are then fed into a determined MLP-base (64-32), forming a composite prediction model based on HATCN-MLP. The parameter optimization process of the introduced HATCN model is as follows: First, the hyperparameter configuration is fixed, that is, the training strategies such as Adam optimizer, learning rate setting, scheduler, early stopping strategy and index recording method are the same as in the previous step and remain unchanged. At the same time, various structural combinations of the hidden layers of the HATCN model are configured, including equal-width and progressively shrinking structures with different numbers of hidden neurons (16, 32, 64), and different convolutional kernel widths k=2-4 are designed.

[0057] The dataset, including historical remote sensing time-series data and historical yield data, was divided into training and testing sets. Composite prediction models with different structural combinations were trained and tested. Multiple evaluation metrics for the model tests were recorded. The final experimental results are as follows: Figure 2 As shown, the optimal parameter configuration of the HATCN model is then selected based on multiple evaluation index values.

[0058] The experimental results show that HATCN-MLP exhibits a more consistent and systematic performance advantage compared to MLP-base. Specifically, in terms of accuracy, the introduction of HATCN improves the overall R² level from the medium range of 34.88%-46.91% to the higher range of 49.71%-59.33%, with the average increasing from 42.47% to 57.14%, an improvement of approximately 15%. Simultaneously, rRMSE and rMAE both decrease by an average of about 1%, indicating that the model has a stronger ability to capture effective information from crop growth timelines. This result demonstrates that the convolutional receptive field provided by HATCN can more fully utilize temporal signals, giving the model a stronger overall expressive power. Similarly, the introduction of HATCN also improves stability. Under different parameter combinations, the model's standard deviation is generally lower than that of MLP-base; for example, R² decreases by about 4%, and both error metrics decrease by about 0.3%. This phenomenon indicates that the model is no longer highly dependent on specific structural settings. The introduction of HATCN not only improves the average level of predictions but also reduces performance fluctuations, making the model more controllable in parameter selection and practical applications.

[0059] In terms of model structure configuration, neural networks with high R² values ​​typically exhibit low depth and high width characteristics, meaning fewer layers but a relatively large number of neurons per layer. Specifically, the best-performing structures are often single-layer high-width or multi-layer structures with equal width, with the configuration of 64 neurons in the first layer and 32 in the last layer showing particularly outstanding performance. Regarding kernel width, no clear hierarchical pattern is observed; instead, a certain degree of random distribution is observed. This indicates that under specific neuron combinations, there is a stable optimal matching characteristic for kernel width, meaning that certain kernel widths exhibit better fit in specific network structures. In the comparison of power factor (PF) and overall score, the configuration with 16 neurons in a single layer and a kernel length of 2 (approximately 6.3K parameters) ranked first in the PF set with an overall score of 53.74, demonstrating that it maintains optimal performance while keeping the number of parameters relatively low. This configuration performs particularly well in models with fewer than 10K parameters, ranking second in performance, only slightly behind a three-layer 16-16-16 multilayer structure with a kernel length of 2 and equal width, with an R² difference of only 0.15%. Both error metrics are even slightly better than the three-layer structure. Considering the overall performance and efficiency tradeoff, this configuration outperforms other models with similar parameter sizes.

[0060] Through specific analysis, this embodiment ultimately selects a configuration of 16 neurons in a single layer and a convolutional kernel length of 2 as the standard structure of the HATCN model. Preferably, the Pareto front multi-objective optimization method and comprehensive scoring function can also be used to determine the optimal configuration of the model during the optimization process of the above-mentioned MLP baseline model. The process is similar to the steps described above and will not be repeated here.

[0061] Step S4: Obtain remote sensing time series data of the target area during the current crop growth period, input it into the optimized HATCN-MLP composite prediction model, and output the crop yield prediction results.

[0062] As another implementation method, to address the problem that existing model decision-making mechanisms are fuzzy and cannot clearly define the contribution weight of remote sensing data at different reproductive stages to yield prediction, the method proposed in this embodiment further includes: Step S5: In the crop yield prediction process, the attention weight of each time step is calculated through the hybrid temporal attention mechanism of the HATCN model in the composite prediction model. Combined with the different growth stages of the crop, the contribution of different growth stages to yield prediction is visualized and quantified.

[0063] Specifically, to gain a deeper understanding of the model's decision-making mechanism and reveal the contribution of different growth stages to yield prediction, this embodiment uses the hybrid temporal attention mechanism in the HATCN model to perform temporal weight visualization analysis. The crop growth cycle is divided into four key stages, and the importance of different growth stages is quantified by calculating the attention weight at each time step. The contribution weight of different stages to yield prediction is shown below. Figure 3 As shown.

[0064] Figure 3 Table (a) illustrates the evolution of the weights of temporal features during the winter wheat growth period. Overall, the weights show an approximately monotonically increasing trend over time: gradually increasing from a low value (approximately 0.012) during the sowing-overwintering period, and reaching a maximum value (approximately 0.020) during the grain-filling-maturity period. This trend reflects the differences in the observability of remote sensing features at different growth stages: as plant height, leaf area index, and canopy structure develop, spectral and structural information in the middle and later stages more stably reflects the plant's growth status and final yield potential; therefore, the model tends to assign higher weights in the later stages of growth. Notably, the weight curves show obvious upward inflection points at the transitions between the overwintering and greening stages and between the heading and grain-filling stages. The first transition corresponds to the plant entering a rapid growth phase from a low-metabolic dormant state, during which canopy coverage, leaf area index, and related spectral characteristics all increase rapidly, enhancing remote sensing's ability to identify differences in growth status. The second transition marks the critical transition from ear formation to grain filling, at which point canopy photosynthetic capacity reaches its peak, and the rate of biomass accumulation accelerates, further enhancing the sensitivity of related remote sensing features to yield. Overall, the two stages of weight changes are highly consistent with two important changes in canopy structure and photosynthetic performance during winter wheat growth, indicating that the model is more sensitive to these physiological transitions in the temporal changes of input features and utilizes the spectral and structural information of these stages to improve the accuracy of yield prediction.

[0065] Figure 3Figure (b) shows the evolution of the weights of temporal features during the summer maize growth period. Unlike the increasing pattern of winter wheat, maize weights exhibit a typical "bimodal" characteristic. The first peak occurs during the seedling-jointing stage, while the second, higher peak occurs during the grain-filling-maturity stage. This bimodal structure reflects the differences in remote sensing features between the rapid vegetative growth stage and the grain-filling stage in maize yield formation. The early peak corresponds to the stage of rapid canopy expansion and photosynthetic capacity establishment, with significant relative changes in leaf area index and canopy coverage. Remote sensing features show obvious differences during this stage and therefore contribute significantly to the model. The later peak corresponds to the grain-filling stage, where rapid accumulation of dry matter into the grains directly affects the final yield, thus the relevant remote sensing features have the highest weights in the model. It is worth noting that the first important period is not the tasseling-flowering period in agronomic cognition. The model's focus is subtly different from the importance of agronomy. This may be because although this period is relatively sensitive, it may not be reflected in remote sensing features. The model tends to focus on periods with significant changes in remote sensing features and those directly related to yield, rather than the critical growth period defined in biology.

[0066] Comparing the evolution of temporal feature weights for winter wheat and summer maize during their growth periods reveals significant differences, while also exhibiting some consistency. The weight curve for winter wheat shows an increasing pattern. As the canopy gradually closes from the greening stage to the grain-filling stage, the accumulated remote sensing features form a smooth upward trend, thus the weights of features do not show a significant decreasing trend after the greening stage. In contrast, summer maize exhibits a bimodal pattern. The rapid vegetative growth period from the seedling stage to the jointing stage causes rapid leaf expansion. At this time, remote sensing features such as leaf area index may tend to saturate and their fluctuation range decreases in the short term. Subsequently, plant growth enters a relatively stable stage, leading to a decrease in weights after the first peak. Furthermore, regarding interannual differences, the weight curve for winter wheat shows high stability, while the first peak for maize fluctuates significantly across different years, especially in years like 2018 when ecological quality declined, with a lag in the timing of the first peak. This reflects the sensitivity of maize to environmental conditions such as water and temperature changes. However, both are assigned high weights from the grain-filling to maturity stage, indicating the importance of spectral information for yield prediction during this stage and demonstrating that the model can robustly utilize crop growth information in the later stages to reflect yield differences.

[0067] Based on the above visualization results, key time windows can be identified, providing guidance for precision agriculture: strengthening remote sensing monitoring during the greening and grain-filling stages of wheat and the jointing and grain-filling stages of corn can more effectively capture key information affecting yield. For example, key stages play a decisive role in yield prediction, and subsequent yield prediction can be directly made using monitoring data from key stages, avoiding data redundancy. At the same time, the cross-year weight stability provides support for the reliability of the model, enhancing its stability in actual production.

[0068] Example 2 This embodiment provides a crop yield prediction system based on an optimized HATCN-MLP, specifically including: The data acquisition module is used to acquire historical remote sensing time-series data, historical yield data, and growth cycle information of crops in the target area during their growth period, and to construct a dataset. The MLP baseline model optimization module is used to build an MLP baseline model, configure various structural combinations of the model's hidden layers, train and test the model using a dataset, and select the optimal configuration of the structural combination based on the test results. The composite prediction model construction and optimization module is used to introduce a hybrid temporal attention mechanism into the temporal convolutional network, construct the HATCN model, and then combine it with the optimized MLP baseline model to form a composite prediction model. It configures various structural combinations of the hidden layers of the HATCN model, trains and tests the composite prediction model using a dataset, and selects the optimal configuration of the structural combination based on the test results. The yield prediction module is used to acquire remote sensing time-series data of the current crop growth period in the target area, input it into the optimized HATCN-MLP composite prediction model, and output the crop yield prediction results.

[0069] As a further implementation method, it also includes: The visualization and quantification module is used to calculate the attention weights at each time step in the crop yield prediction process through the hybrid temporal attention mechanism of the HATCN model in the composite prediction model. Combined with the division of different crop growth stages, it visualizes and quantifies the contribution of different growth stages to yield prediction. The crop growth cycle information includes different growth stages of the crop.

[0070] Example 3 This embodiment provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in this embodiment.

[0071] Example 4 This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, will cause the processor to execute the method described above in this embodiment.

[0072] Example 5 This embodiment provides a computer program product including executable instructions, which are computer instructions; the executable instructions are stored in a computer-readable storage medium. When the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the electronic device performs the method described in this embodiment.

[0073] The steps and methods involved in Embodiments 2 to 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0074] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0075] The above description is only a preferred embodiment of the present invention. Although the specific implementation of the present invention has been described in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the present invention.

Claims

1. A crop yield prediction method based on optimized HATCN-MLP, characterized in that, include: Acquire historical remote sensing time-series data, historical yield data, and growth cycle information of crops in the target area during their growth period, and construct a dataset; Build an MLP baseline model, configure various structural combinations of the model's hidden layers, train and test the model using a dataset, and select the optimal configuration of the structural combinations based on the test results; A hybrid temporal attention mechanism is introduced into the temporal convolutional network to construct the HATCN model, which is then combined with the optimized MLP baseline model to form a composite prediction model. Multiple structural combinations of the hidden layers of the HATCN model are configured, and the composite prediction model is trained and tested using a dataset. The optimal configuration of the structural combination is selected based on the test results. Acquire remote sensing time-series data of the target area during the current crop growth period, input it into the optimized HATCN-MLP composite prediction model, and output the crop yield prediction results. The MLP baseline model consists of an input layer, several hidden layers, and an output layer, all connected by a fully connected mechanism. The optimization process of the MLP baseline model is as follows: Configure various structural combinations of several hidden layers in the MLP baseline model and fix the hyperparameter configuration; among them, the structural combinations include equal-width structures and progressively shrinking structures with 2-4 layers of different numbers of hidden neurons; The dataset, which includes historical remote sensing time-series data and historical yield data, is divided into training and testing sets. MLP baseline models with different structural combinations are trained and tested, and the values ​​of multiple evaluation metrics for model testing are recorded. Among them, multiple evaluation metrics include R², rRMSE, rMAE and parameter count. Using different structural combinations as candidate parameter combinations, the Pareto front multi-objective optimization method is adopted to map the candidate parameter combinations to a multi-dimensional objective space corresponding to multiple evaluation indicators, identify non-dominated solutions, and construct a comprehensive scoring function to calculate the comprehensive score of each non-dominated solution. The candidate solution with the highest comprehensive score is selected as the optimal configuration of the MLP baseline model. Based on the causal convolution and dilated convolution mechanisms of temporal convolutional networks, a basic network is constructed, and a hybrid temporal attention mechanism is introduced into the basic network to construct the HATCN model. The hybrid temporal attention mechanism includes a multi-scale stage attention calculation module, a global attention calculation module, and a gated fusion module. The multi-scale stage attention calculation module calculates the attention weights for each stage and aggregates the temporal features within each stage. Based on the aggregated features, it integrates the information from each stage through learnable stage fusion weights to generate multi-scale features. The global attention calculation module calculates long-term features across stages to generate global features. The gated fusion module fuses multi-scale features and global features through a gating mechanism to obtain the final attention weights. Remote sensing time-series data is input into the HATCN model. Temporal features are first extracted through a temporal convolutional network. The temporal features are then used to generate attention weights through a hybrid temporal attention mechanism. The temporal features at the corresponding time steps are weighted and summed according to the attention weights to generate weighted temporal features.

2. The crop yield prediction method based on optimized HATCN-MLP as described in claim 1, characterized in that, Crop growth cycle information includes different growth stages of the crop; In the process of crop yield prediction, the attention weight of each time step is calculated by the hybrid temporal attention mechanism of the HATCN model in the composite prediction model. Combined with the different growth stages of the crop, the contribution of different growth stages to yield prediction is visualized and quantified.

3. The crop yield prediction method based on optimized HATCN-MLP as described in claim 1, characterized in that, A 1-3 layer HATCN model is constructed as the front-end temporal feature extraction module. The output of the temporal feature extraction module is then fed into the optimized MLP baseline model to form a composite prediction model based on HATCN-MLP. The parameter optimization process of the HATCN model is as follows: The hyperparameter configuration is fixed, and various structural combinations of the hidden layers of the HATCN model are configured, including equal-width and progressively shrinking structures with different numbers of hidden neurons, and different convolutional kernel widths are designed. The dataset, which includes historical remote sensing time-series data and historical yield data, is divided into training and testing sets. Composite prediction models with different structural combinations are trained and tested. The values ​​of multiple evaluation indicators of the model tests are recorded, and the optimal parameter configuration of the HATCN model is selected based on the values ​​of multiple evaluation indicators.

4. A crop yield prediction system based on optimized HATCN-MLP, characterized in that, include: The data acquisition module is used to acquire historical remote sensing time-series data, historical yield data, and growth cycle information of crops in the target area during their growth period, and to construct a dataset. The MLP baseline model optimization module is used to build an MLP baseline model, configure various structural combinations of the model's hidden layers, train and test the model using a dataset, and select the optimal configuration of the structural combination based on the test results. The composite prediction model construction and optimization module is used to introduce a hybrid temporal attention mechanism into the temporal convolutional network, construct the HATCN model, and then combine it with the optimized MLP baseline model to form a composite prediction model. It configures various structural combinations of the hidden layers of the HATCN model, trains and tests the composite prediction model using a dataset, and selects the optimal configuration of the structural combination based on the test results. The yield prediction module is used to acquire remote sensing time-series data of the current crop growth period in the target area, input it into the optimized HATCN-MLP composite prediction model, and output the crop yield prediction results. The MLP baseline model consists of an input layer, several hidden layers, and an output layer, all connected by a fully connected mechanism. The optimization process of the MLP baseline model is as follows: Configure various structural combinations of several hidden layers in the MLP baseline model and fix the hyperparameter configuration; among them, the structural combinations include equal-width structures and progressively shrinking structures with 2-4 layers of different numbers of hidden neurons; The dataset, which includes historical remote sensing time-series data and historical yield data, is divided into training and testing sets. MLP baseline models with different structural combinations are trained and tested, and the values ​​of multiple evaluation metrics for model testing are recorded. Among them, multiple evaluation metrics include R², rRMSE, rMAE and parameter count. Using different structural combinations as candidate parameter combinations, the Pareto front multi-objective optimization method is adopted to map the candidate parameter combinations to a multi-dimensional objective space corresponding to multiple evaluation indicators, identify non-dominated solutions, and construct a comprehensive scoring function to calculate the comprehensive score of each non-dominated solution. The candidate solution with the highest comprehensive score is selected as the optimal configuration of the MLP baseline model. Based on the causal convolution and dilated convolution mechanisms of temporal convolutional networks, a basic network is constructed, and a hybrid temporal attention mechanism is introduced into the basic network to construct the HATCN model. The hybrid temporal attention mechanism includes a multi-scale stage attention calculation module, a global attention calculation module, and a gated fusion module. The multi-scale stage attention calculation module calculates the attention weights for each stage and aggregates the temporal features within each stage. Based on the aggregated features, it integrates the information from each stage through learnable stage fusion weights to generate multi-scale features. The global attention calculation module calculates long-term features across stages to generate global features. The gated fusion module fuses multi-scale features and global features through a gating mechanism to obtain the final attention weights. Remote sensing time-series data is input into the HATCN model. Temporal features are first extracted through a temporal convolutional network. The temporal features are then used to generate attention weights through a hybrid temporal attention mechanism. The temporal features at the corresponding time steps are weighted and summed according to the attention weights to generate weighted temporal features.

5. The crop yield prediction system based on optimized HATCN-MLP as described in claim 4, characterized in that, Also includes: The visualization and quantification module is used to calculate the attention weights at each time step in the crop yield prediction process through the hybrid temporal attention mechanism of the HATCN model in the composite prediction model. Combined with the division of different crop growth stages, it visualizes and quantifies the contribution of different growth stages to yield prediction. The crop growth cycle information includes different growth stages of the crop.

6. An electronic device, characterized in that, include: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the crop yield prediction method based on optimized HATCN-MLP as described in any one of claims 1-3.

7. A computer-readable storage medium, characterized in that, The system stores executable instructions that, when executed by a processor, implement the crop yield prediction method based on optimized HATCN-MLP as described in any one of claims 1-3.

8. A computer program product, characterized in that, The computer program product includes executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the crop yield prediction method based on optimized HATCN-MLP as described in any one of claims 1-3.

Images