Crop yield prediction method and system based on dynamic-static coordination temporal attention, and medium

Abstract

The application discloses a crop yield prediction method and system based on dynamic and static time sequence attention, and a medium, relates to the technical field of intelligent agricultural data mining and remote sensing monitoring, and the scheme is characterized in that: by fusing dynamic variable data and static covariant data, a static geographic context vector is obtained by performing nonlinear mapping on the static covariant data based on a multilayer perception machine encoder, and the static geographic context vector is used as prior knowledge to eliminate noise variables in the dynamic variable data, so that the physical consistency and continuity of the data are ensured, on this basis, the growth evolution law of crops changing with time and the environmental effect caused by the geographic position are extracted respectively, and a deep feature representation is generated based on a multi-head self-attention mechanism to capture the correlation between different variables in the damaged dynamic sequence, and by capturing deep nonlinear features in the space-time dimension, the robust deduction and high-precision prediction of crop yield in different geographic environments are realized.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture data mining and remote sensing monitoring technology, specifically to a method, system, and medium for predicting crop yield based on dynamic and static collaborative temporal attention. Background Technology

[0002] With the rapid development of precision agriculture and remote sensing big data technology, using satellite remote sensing data for large-scale, low-cost crop yield prediction has become an important means of ensuring food security. Existing technical solutions typically utilize multispectral satellites (such as Sentinel-2 and Landsat) to obtain the spectral reflectance of crops during their growth cycle, and then construct vegetation index time series to invert crop growth and yield.

[0003] The existing technologies can be mainly divided into the following two categories, but both have certain limitations in practical applications.

[0004] The first category consists of shallow regression methods based on traditional machine learning. Existing solutions often employ algorithms such as Random Forest (RF), Support Vector Regression (SVR), or Ridge Regression. These methods typically treat spectral data from different time points as independent feature inputs, or extract single image features from key crop growth stages (such as heading and grain-filling stages) for modeling. However, these methods neglect the dynamic characteristics of crop growth, fail to capture the evolution of spectral data over time, and lack the ability to extract contextual information from time series data. When faced with shifts or fluctuations in the growth cycle caused by climate anomalies, the model's generalization ability is poor, and the prediction accuracy is insufficient to meet the needs of refined management.

[0005] The second category comprises temporal deep learning methods based on recurrent neural networks and their variants. To address the temporal dependency problem, some technical solutions have introduced deep learning models such as long short-term memory networks or gated recurrent units, utilizing the recurrent structure to extract the trend features of crop growth to a certain extent. However, these methods are essentially serial computing architectures, making it difficult to utilize hardware parallel acceleration, resulting in low training efficiency on large-scale data. Furthermore, when dealing with long crop growth cycles, gradient vanishing or information forgetting easily occurs, making it difficult to effectively capture long-distance dependencies between the early growth stage and maturity stage. In addition, existing technologies, when fusing static geographic environmental data (such as latitude and longitude) with dynamic temporal spectral data, often only employ simple feature stitching strategies, lacking deep interaction mechanisms. This fails to effectively characterize the nonlinear regulatory effect of geographic location on the crop growth process, thus limiting the model's generalization ability and prediction accuracy in complex terrain areas. Summary of the Invention

[0006] To address the challenges of existing crop yield prediction technologies, such as the difficulty in parallel modeling long-sequence satellite spectral data, insufficient fusion of static geographic constraints and dynamic temporal features, lack of adaptive screening of multi-source and multi-band variables, and inadequate data preprocessing and quality control, this invention provides a crop yield prediction method based on dynamic and static co-attention. This method utilizes the homogeneity of geospatial data to correct the discreteness of temporal data, dividing it into dynamic variable data and static covariate data. By fusing these two data, a multi-layer perceptron encoder is used to perform nonlinear mapping on the static covariate data to obtain a static geographic context vector. This static geographic context vector is then used as prior knowledge to eliminate noise variables in the dynamic variable data, ensuring the physical consistency and continuity of the data. Based on this, the growth and evolution patterns of crops over time and the environmental effects brought about by geographical location are extracted. A multi-head self-attention mechanism is used to capture the correlations between different variables in the damaged dynamic sequence, generating deep feature representations. By capturing deep nonlinear features in the spatiotemporal dimension, robust extrapolation and high-precision prediction of crop yields under different geographic environments are achieved.

[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0008] This solution provides a crop yield prediction method based on dynamic and static collaborative temporal attention, the method comprising:

[0009] A multivariate remote sensing time-series dataset of the target area is collected, and the multivariate remote sensing time-series dataset is preprocessed to obtain dynamic variable data and static covariate data; the multivariate remote sensing time-series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data describing the spatial location of the monitored object;

[0010] A static geographic context vector is obtained by nonlinear mapping of the static covariate data based on a multilayer perceptron encoder, and the static geographic context vector is used as prior knowledge to remove noise variables in the dynamic variable data.

[0011] The dynamic variable data after removing noise variables is masked to obtain a damaged dynamic sequence. The correlation between different variables in the damaged dynamic sequence is captured based on a multi-head self-attention mechanism to generate a deep feature representation. The masked part in the dynamic variable data is reconstructed based on the deep feature representation.

[0012] Global features are concatenated in chronological order, and a time attention mechanism using static geographic context vectors as queries is introduced to extract static sensitive aggregated features that focus on crop growth stages. The crop yield of the target area is then mapped by combining the static sensitive aggregated features and the static geographic context vectors.

[0013] A further optimized solution is that the preprocessing includes:

[0014] Using each monitored object as a sample, the multivariate remote sensing time series dataset is mapped into a continuous tensor space, and the static geographic attribute data of each sample is extracted.

[0015] Based on the static geographic attribute data, local geographic constraints are constructed, the whole sample is mapped into multiple regions with the same geographic properties, the time series tensors of all samples within each region are aggregated, and the statistical distribution parameters of each time series tensor on each spectral channel are calculated, thereby establishing the crop growth standard baseline for each region.

[0016] The temporal tensor of individual samples is evaluated point by point based on the crop growth standard baseline of each region to identify outliers or missing data points. The outliers or missing data points are then corrected based on a spatiotemporal co-interpolation strategy. The spatiotemporal co-interpolation strategy includes: weighting and fusing the local trend of individual samples in the time dimension with the mean of the crop growth standard baseline of the current region in the spatial dimension to generate a correction value.

[0017] The corrected multivariate remote sensing time series dataset was divided into dynamic variables and static variables, and the dynamic variables and static covariate data were obtained by standardizing the dynamic variables and static covariate data respectively.

[0018] A further optimization scheme is that the method for removing noise variables from the dynamic variable data includes:

[0019] A static geographic context vector is obtained by nonlinear mapping of the static covariate data based on a multilayer perceptron encoder.

[0020] Obtain all dynamic variables and the flattened vector of each dynamic variable from the dynamic variable data;

[0021] The flattened vectors of all dynamic variables and the static geographic context vectors are input into the gated residual modulation network GRMU. A normalized variable selection weight vector is generated based on the softmax layer of the gated residual modulation network GRMU. At the same time, an independent GRMU subunit is configured for each dynamic variable to perform nonlinear transformation.

[0022] A further optimization is that the masking process includes:

[0023] A binary noise mask matrix is ​​generated from the dynamic variable data after removing noise variables; the time series length and variable dimension of the binary noise mask matrix are consistent with the time series length and variable dimension of the dynamic variable data after removing noise variables.

[0024] The damaged dynamic sequence is obtained by performing element-wise product of the binary noise mask matrix and the dynamic variable data after removing the noise variables.

[0025] A further optimized solution is that the method for reconstructing the masked portion of the dynamic variable data based on the deep feature representation includes:

[0026] Based on a multi-head self-attention mechanism using a Transformer encoder, deep feature representations are generated by parallelly capturing the relationships between different variables in damaged dynamic sequences. ,in Indicates time step The eigenvectors; T represents the total number of time steps;

[0027] Connecting a linear mapping layer to the output of the Transformer encoder will... Project back the dimensions of the dynamic variable data after removing noise variables to reconstruct the masked portion of the dynamic variable data.

[0028] In the process of reconstructing the masked portion of dynamic variable data, the mean square error of the masked elements is used as the loss function.

[0029] A further optimized approach is that the method for predicting crop yields in the target area includes:

[0030] Construct time-series samples with output labels based on dynamic variable data and static covariate data;

[0031] Construct an improved Transformer encoder with parameters consistent with the Transformer encoder described above;

[0032] The improved Transformer encoder is trained in a supervised manner based on time-series samples with output labels:

[0033] A static-sensitive full-sequence attention aggregation mechanism is introduced to map static geographic context vectors to global query vectors, constructing corresponding key and value vectors for each encoding block output of the improved Transformer encoder; using scaled dot products as static-sensitive attention weights, global normalization is performed over all time steps to obtain the attention distribution; and the static-sensitive aggregated features are obtained by weighted summation of the value vectors based on the attention distribution.

[0034] After concatenating the static sensitive aggregated features with the static geographic context vector, the concatenated features are mapped to the crop yield space based on the linear regression head to obtain the crop yield prediction results for the target area.

[0035] A further optimization scheme is that the training method of the improved Transformer encoder includes:

[0036] Using static geographic context vectors as input, queries, keys, and values ​​are constructed. All attention heads share the same Value representation, and each attention head only uses independent linear transformations on the query and key. The weights of all attention heads are averaged to obtain a unified temporal weight. By sharing the Value transformation matrix, the different attention patterns of all attention heads on the time axis are aggregated into the same feature space and batch normalization with residual connections is performed to obtain the attention sub-layer output. A feedforward network is introduced to process the attention sub-layer output independently at each time step to obtain the feedforward network output. The attention sub-layer output and the feedforward network output are batch normalized with residual connections to obtain the encoding block output. Multiple encoding block outputs are stacked to obtain the encoder output.

[0037] A further optimization scheme is to use the mean squared error as the loss function during supervised training.

[0038] This solution also provides a crop yield prediction system based on dynamic-static collaborative temporal attention, used to implement the above-mentioned crop yield prediction method based on dynamic-static collaborative temporal attention. The system includes:

[0039] The acquisition module is used to acquire multivariate remote sensing time-series datasets of the target area;

[0040] The preprocessing module is used to preprocess the multivariate remote sensing time series dataset to obtain dynamic variable data and static covariate data; the multivariate remote sensing time series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data describing the spatial location of the monitored object;

[0041] The static covariate processing module is used to perform nonlinear mapping on the static covariate data based on the multilayer perceptron encoder to obtain a static geographic context vector, and use the static geographic context vector as prior knowledge to remove noise variables in the dynamic variable data.

[0042] The dynamic variable processing module is used to mask the dynamic variable data after removing noise variables to obtain a damaged dynamic sequence. Based on the multi-head self-attention mechanism, it captures the correlation between different variables in the damaged dynamic sequence to generate a deep feature representation. Based on the deep feature representation, it reconstructs the masked part in the dynamic variable data.

[0043] The prediction module is used to concatenate global features in chronological order and introduce a time attention mechanism that uses static geographic context vectors as queries to extract static sensitive aggregated features that focus on the crop growth stage. The static sensitive aggregated features and static geographic context vectors are combined to map the crop yield of the target area.

[0044] This solution also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, can implement the crop yield prediction method based on dynamic-static collaborative temporal attention as described above.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention provides a method, system, and medium for predicting crop yield based on dynamic and static covariate attention. It utilizes the homogeneity of geospatial data to correct the discreteness of time-series data, dividing it into dynamic variable data and static covariate data. By fusing the dynamic and static covariate data, a static geographic context vector is obtained by nonlinearly mapping the static covariate data using a multilayer perceptron encoder. This static geographic context vector is then used as prior knowledge to eliminate noise variables in the dynamic variable data, ensuring the physical consistency and continuity of the data. Based on this, the growth and evolution patterns of crops over time and the environmental effects brought about by geographical location are extracted. Furthermore, a multi-head self-attention mechanism is used to capture the correlations between different variables in the damaged dynamic sequence, generating deep feature representations. By capturing deep nonlinear features in the spatiotemporal dimension, robust extrapolation and high-precision prediction of crop yield under different geographical environments are achieved.

[0047] 2. This invention provides a method, system, and medium for crop yield prediction based on dynamic and static collaborative temporal attention. The preprocessing and dynamic / static variable grouping of multi-source remote sensing time-series data driven by static geographic constraints simultaneously improves data consistency and spatial heterogeneity modeling capabilities. Before the data enters the prediction model, a preprocessing and quality control process for multi-source remote sensing time-series data under static geographic information constraints is introduced. By grouping data according to static geographic features / dynamic spectral temporal features, combined with spatiotemporal alignment, radiometric normalization, missing data imputation, and anomaly detection, the bias between multi-source, multi-temporal, and multi-resolution data is significantly reduced. Compared with existing schemes that directly normalize and simply stitch together the original sequences, this invention further maps static features such as latitude, longitude, altitude, slope, and soil into geographic context vectors through a static feature mapping subnet, and deeply integrates them into the variable selection and output aggregation stages. This allows the spectral-yield relationship of the same crop under different ecological zones and terrain conditions to be constrained within a unified model framework.

[0048] 3. This invention provides a method, system, and medium for crop yield prediction based on dynamic and static collaborative temporal attention. It employs a static context-driven variable selection network based on a gated residual modulation network (GRMU) and an interpretable multi-head self-attention Transformer encoder with shared values, resulting in stronger long-distance temporal dependency modeling capabilities, multivariate interaction characterization capabilities, and model interpretability. This invention improves the structure of a Transformer backbone that only includes the encoder. On one hand, the GRMU adaptively generates weights at each time step based on current features and static geographical context, achieving explicit variable selection and weighted noise suppression for multi-source, multi-band inputs. On the other hand, the multi-head self-attention module uses an interpretable attention structure with shared values, sharing the same value projection across different attention heads and averaging the attention matrices of each head to obtain a unified temporal dimension attention distribution, which directly reflects the model's attention level at different reproductive stages. Meanwhile, batch normalization is used in each sub-layer of the encoder instead of traditional layer normalization, which effectively alleviates the impact of outliers and missing tests in long remote sensing time series on training stability and shows better robustness in noisy samples and small sample scenarios.

[0049] 4. This invention provides a method, system, and medium for crop yield prediction based on dynamic and static collaborative temporal attention. An end-to-end regression framework combining static context-driven full-sequence attention aggregation with self-supervised pre-training based on input mask denoising offers advantages such as efficient utilization of complete growing season information, improved performance in small-sample scenarios, and robust generalization under multi-year and multi-regional conditions. In the regression prediction stage, a temporal attention aggregation subnet is constructed, using static geographical context vectors as queries and encoded features at each time step as keys and values. This adaptively generates importance scores for different growth stages, which are then normalized to obtain temporal dimension attention weights. This allows the model to simultaneously focus on local fluctuations and the overall growth trajectory of key growth stages under static geographical constraints, avoiding information loss caused by relying solely on a single time point or simple averaging. Drawing inspiration from multivariate time series Transformer representation learning, a self-supervised pre-training process based on input random mask denoising is designed. Without relying on yield labels, a general representation between spectrum, time, and variables is first learned, followed by supervised fine-tuning on a small number of labeled samples. This achieves better prediction results than pure supervised training even with limited labeled samples. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0051] Figure 1 This is a schematic diagram of the crop yield prediction method based on dynamic and static collaborative temporal attention;

[0052] Figure 2 This is a schematic diagram of the structure of a crop yield prediction system based on dynamic and static collaborative temporal attention;

[0053] Figure 3 This is a schematic diagram of the principle of a crop yield prediction system based on dynamic and static collaborative temporal attention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0055] To address the challenges of parallel modeling of long-sequence satellite spectral data, insufficient fusion of static geographic constraints and dynamic temporal features, lack of adaptive selection of multi-source, multi-band variables, and inadequate data preprocessing and quality control in existing technologies, this invention provides the following embodiments to solve the aforementioned technical problems:

[0056] Example 1

[0057] This embodiment provides a crop yield prediction method based on dynamic and static collaborative temporal attention, such as... Figure 1 and Figure 3 As shown, the method includes:

[0058] Step 1: Collect a multivariate remote sensing time-series dataset of the target area, and preprocess the multivariate remote sensing time-series dataset to obtain dynamic variable data (spectral time-series data) and static covariate data. The multivariate remote sensing time-series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data describing the spatial location of the monitored object. The high-dimensional spectral data mainly includes satellite spectra at different time steps. The static geographic attribute data mainly includes geographic attribute data such as latitude and longitude, soil type, and slope.

[0059] In step one, the preprocessing method specifically includes the following sub-steps:

[0060] S11: Using each monitored object as a sample, multi-source satellite spectral values ​​are collected and spatiotemporally registered and quality monitored (aligning and quality-filtering data from different satellite platforms, times, and bands). The length of the same time series data (filling in missing values ​​to ensure consistent length across all time series for easy model input) constitutes a multivariate remote sensing time series dataset. This dataset is then mapped to a continuous tensor space, and the static geographic attribute data of each sample is extracted. Specifically, each monitored object is defined as a sample. , sample The observation data throughout the entire growth cycle are recorded as follows:

[0061] ;

[0062] in, Represents a time step sequence. Characteristic dimensions representing multi-channel satellite bands and derived spectral indices; Represent real numbers;

[0063] Extracting samples The static geographic attribute data is denoted as: ;in Indicates the dimension of spatial coordinates.

[0064] S12, based on static geographic attribute data, construct local geographic constraints, map the whole sample into multiple regions with the same geographic properties, aggregate the time series tensors of all samples within each region, calculate the statistical distribution parameters of each time series tensor on each spectral channel, and thus establish the crop growth standard baseline for each region.

[0065] S13, evaluate the time series tensor of individual samples point by point based on the crop growth standard baseline of each region, identify outliers or missing data points, and correct the outliers or missing data points based on the spatiotemporal co-interpolation strategy; the spatiotemporal co-interpolation strategy includes: weighting and fusing the local trend of individual samples in the time dimension with the mean of the crop growth standard baseline of the current region in the spatial dimension to generate a correction value.

[0066] Considering that crop growth is spatially autocorrelated due to the influence of the geographical environment, this embodiment uses static geographic attribute data to construct local geographic constraints in order to identify environmental noise in the data. Specifically, high-dimensional spatial features can be divided based on unsupervised clustering algorithms, and the whole sample can be mapped to several geographically homogeneous partitions. By establishing specific crop growth standard baselines for each partition, the normal spectral response of the corresponding geographical neighborhood under ideal conditions can be reflected.

[0067] This embodiment utilizes the crop growth standard baseline to perform point-by-point quality assessment and cleaning of the time-series tensors of individual samples; it calculates the deviation between the spectral response curve of each sample and the crop growth standard baseline of its respective region. When the observed value at a certain time step significantly deviates from the distribution range of the crop growth standard baseline and lacks trend support from neighboring time steps, the data point is considered an outlier affected by cloud cover, shadows, or sensor errors. For identified outliers or missing data points, corrections are made based on a spatiotemporal co-interpolation strategy. The correction value is generated by a weighted fusion of the local trend of the sample in the time dimension and the baseline mean of the region in the spatial dimension, as shown in the specific formula:

[0068] ;

[0069] in, Indicates sample At time step The corrected value; This represents the preset hyperparameters used to adjust the weighting of the time trend and the spatial baseline; , Representing samples respectively At time step The effective observations of the previous and next time steps; Indicates that sample j is at time step Valid observations; Represents clusters The total number of samples included.

[0070] S14. The corrected multivariate remote sensing time series dataset is divided into dynamic variables and static variables, and the dynamic variables and static variables are standardized respectively to obtain dynamic variable data and static covariate data.

[0071] Specifically, by calculating the mean and standard deviation of dynamic and static variables, the values ​​of all dimensions are mapped to a unified distribution interval, eliminating dimensional differences between different spectral bands and geographical coordinates. This provides standardized input data for the subsequent construction of a deep learning-based yield prediction model. The specific standardization formula is as follows:

[0072] ;

[0073] in, Indicates sample At time step The Data after 3D feature standardization; Indicates sample At time step The Original data for dimensional features; and These represent the first (i) of all the original data. The mean and standard deviation of the dimensional features; This represents a very small constant used to ensure training stability.

[0074] The preprocessing process aims to address the temporal discreteness and environmental noise interference issues in multi-source remote sensing monitoring data. In this embodiment, discrete observation records are mapped into a continuous tensor space, and static spatial properties are introduced as constraints to achieve standardized reconstruction of dynamic growth characteristics and correction of outliers.

[0075] Step 2: Based on the multilayer perceptron encoder, the static covariate data is nonlinearly mapped to obtain a static geographic context vector, and the static geographic context vector is used as prior knowledge to remove noise variables in the dynamic variable data.

[0076] In step two, the methods for removing noise variables from dynamic variable data include:

[0077] S21, Based on the multilayer perceptron encoder, the static covariate data is nonlinearly mapped to obtain a static geographic context vector;

[0078] For static covariate data that do not change over time The background environment for crop growth is defined, which has a global constraint on yield. To extract its high-dimensional features and generate context vectors for subsequent model adjustment, this embodiment uses a multilayer perceptron encoder (MLP Encoder) to perform a non-linear mapping on the static covariate data, mapping low-dimensional static attributes to dimensions corresponding to the model's hidden layers. Consistent static geographic context vectors :

[0079] ;

[0080] in, This indicates that the encoding and vector extraction operations are performed by a multilayer perceptron encoder.

[0081] In the nonlinear mapping process, to improve the ability to handle nonlinear relationships, this embodiment designs a gated residual modulation network (GRMU) to skip nonlinear processing when necessary (degenerating to linear or identity mapping), thereby adapting to the complexity of different features. The GRMU operation process is represented as GRMU( , Its input includes the main input. and optional context input The gated residual modulation network (GRMU) contains two fully connected layers. and and in the fully connected layer Using the ELU activation function in fully connected layers The Dropout regularization strategy is used to avoid overfitting. The gated linear unit (GLU) in the gated residual modulation network (GRMU) controls the propagation ratio of values ​​through sigmoid function activation. The specific formulas for the operation of the GRMU are as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] in, , , , This represents the weight matrix that can be updated through backpropagation for different network layers; This represents the input of the gated linear unit (GLU); GLU() represents the GLU operation. This represents the sigmoid activation function; This indicates element-wise multiplication; b1 and b2 represent the bias vectors corresponding to the first and second layer linear transformations, respectively; b3 and b4 represent the bias vectors corresponding to the gated branches and feature branches in the gated linear unit (GLU), respectively; Layernorm() represents the layer normalization function.

[0086] S22, obtain all dynamic variables and the flattened vector of each dynamic variable from the dynamic variable data;

[0087] S23, input the flattened vectors of all dynamic variables and the static geographic context vectors into the gated residual modulation network GRMU, generate a normalized variable selection weight vector based on the Softmax layer of the gated residual modulation network GRMU, and configure an independent gated residual modulation subunit for nonlinear transformation for each dynamic variable.

[0088] Specifically, set time steps The set of dynamic variables is , which includes Each dynamic variable will be a dynamic variable Independently mapped through a linear layer to Dimensions; subsequently, to determine the importance weight of each dynamic variable, a flattened vector of all dynamic variables is generated. With static geographic context vectors The input is fed into a gated residual modulation network (GRMU), and a normalized variable selection weight vector is generated by the softmax layer. for:

[0089] ;

[0090] Where Softmax() represents the Softmax activation function;

[0091] Variable selection weight vector The j-th element Represents the dynamic variables in the geographical environment at time t. The importance of this; at the same time, each dynamic variable in the set of dynamic variables will also undergo nonlinear transformations through its independent GRMU unit, such as The j-th dynamic variable at time j Nonlinear characteristics obtained by performing nonlinear transformation Finally, combining the nonlinear characteristics of all dynamic variables by weighted aggregation using a variable selection weight vector, we have:

[0092] ;

[0093] This embodiment utilizes static geographical prior knowledge to eliminate noise variables in the dynamic sequence, retaining only the feature combinations most critical for yield prediction. These features are then fed into the Transformer encoder in step three for time series modeling, providing global static environmental constraints for the model and achieving effective coordination between dynamic and static features.

[0094] Step 3: Mask the dynamic variable data after removing noise variables to obtain the damaged dynamic sequence. Capture the correlation between different variables in the damaged dynamic sequence based on the multi-head self-attention mechanism to generate a deep feature representation. Reconstruct the masked part in the dynamic variable data based on the deep feature representation.

[0095] In step three, the masking process includes:

[0096] S31, Generate a binary noise mask matrix for the dynamic variable data after removing noise variables; The time series length and variable dimension of the binary noise mask matrix are consistent with the time series length and variable dimension of the dynamic variable data after removing noise variables;

[0097] S32, the damaged dynamic sequence is obtained by performing element-wise multiplication of the binary noise mask matrix with the dynamic variable data after removing the noise variables.

[0098] Specifically, the dynamic variable data after removing noise variables is used as the original input data. (in The length of the time series. (For the variable dimension), generate a binary noise mask matrix. The lengths of the masked consecutive segments follow a geometric distribution, ensuring that data cannot be recovered through simple local interpolation alone, but must rely on the global context and the interdependencies between variables, which is beneficial for learning deep features; the damaged dynamic sequence is obtained by element-wise multiplying the binary noise mask matrix with the dynamic variable data after removing noise variables. :

[0099] ;

[0100] in, This represents the Hadamard product, which is the element-wise multiplication of the corresponding matrices. In masked positions, the input value is set to 0; in non-masked positions, the original feature values ​​are preserved.

[0101] In step three, the masked portion of the dynamic variable data is reconstructed based on the deep feature representation, specifically including the following methods:

[0102] T31, based on the multi-head self-attention mechanism of the Transformer encoder, captures the correlation between different variables in damaged dynamic sequences in parallel to generate deep feature representations: ,in, This represents the feature vector at time step 1; Indicates time step eigenvectors; The eigenvectors represent time steps T; T represents the total number of time steps.

[0103] T32, connect a linear mapping layer to the output of the Transformer encoder to... Projecting back the dimensions of the dynamic variable data after removing noise variables, and reconstructing the masked portion of the dynamic variable data using a linear reconstruction head; thus obtaining the reconstructed time step. Dynamic variable data :

[0104] ;

[0105] in, The weight matrix represents the output linear mapping layer (reconstruction head), which is used to linearly project the encoded representation of the time step back into the dynamic variable space; This represents the bias vector of the linear mapping layer, which performs a dimension-wise translation correction on the projected result.

[0106] T33 uses the mean square error of the masked elements as the loss function during the reconstruction of the masked portion of dynamic variable data.

[0107] Unlike traditional loss functions, this embodiment calculates the loss only for the elements covered by the mask, rather than the entire sequence. This strategy improves the ability to infer missing information and avoids simply copying known information. The set of indices of the masked elements is defined as follows: Where t represents the time step (phase) index, i.e., the t-th time point in the sequence; i represents the variable (feature dimension) index; m t,i Represents a binary mask matrix At position (t, i), m t,i =0 indicates that the variable is masked at this time step, and the loss is calculated here; m t,i =1 indicates that it is not covered by the mask;

[0108] Then the loss function for:

[0109] ;

[0110] in, Represents the original time step Dynamic variable data;

[0111] By minimizing this loss function, the Transformer encoder can learn the inherent temporal evolution patterns and relationships between variables in the spectral data during the pre-training stage, providing high-quality pre-training initial weights for subsequent downstream regression tasks.

[0112] This embodiment constructs a multi-layer Transformer structure containing only the encoder as the backbone for feature extraction. Each encoder block, based on the existing multi-head self-attention, residual connections, and batch normalization, introduces static geographic context vectors and multivariate features as input to a gated residual modulation network (GRMU). The GRMU's gated nonlinear transformation capability generates scoring vectors for each spectral band and covariate. Softmax normalization is then used to obtain variable selection weights, which are then weighted by sample and by time step to highlight variables more critical for yield prediction and suppress noise. In the multi-head self-attention, all attention heads share the same value projection, and the attention weight matrices of different heads are averaged to obtain a unified temporal attention distribution, achieving good interpretability while capturing long-distance temporal dependencies and multivariate interactions. Furthermore, this embodiment uses batch normalization instead of traditional layer normalization in each encoder sub-layer, effectively mitigating the impact of outliers and missing data in long-sequence remote sensing data on training stability.

[0113] Step four: Concatenate global features in chronological order, and introduce a time attention mechanism that uses static geographic context vectors as queries to extract static sensitive aggregated features that focus on the crop growth stage. Combine the static sensitive aggregated features and static geographic context vectors to map the crop yield of the target area.

[0114] In step four, the method for predicting crop yield in the target area includes:

[0115] S41, Construct time-series samples with output labels based on dynamic variable data and static covariate data;

[0116] S42, Construct an improved Transformer encoder with parameters consistent with the Transformer encoder;

[0117] S43, Supervised training of the improved Transformer encoder is performed based on time-series samples with output labels; this step specifically includes the following methods:

[0118] Using static geographic context vectors as input, queries, keys, and values ​​are constructed. All attention heads share the same Value representation, and each attention head only uses independent linear transformations on the query and key. The weights of all attention heads are averaged to obtain a unified temporal weight. By sharing the Value transformation matrix, the different attention patterns of all attention heads on the time axis are aggregated into the same feature space and batch normalization with residual connections is performed to obtain the attention sub-layer output. A feedforward network is introduced to process the attention sub-layer output independently at each time step to obtain the feedforward network output. The attention sub-layer output and the feedforward network output are batch normalized with residual connections to obtain the encoding block output. Multiple encoding block outputs are stacked to obtain the encoder output.

[0119] Specifically, the improved Transformer encoder receives a static geographic context vector as input. , The dimension of the static geographic context vector is represented; then the first... The input of the layer encoder is First, a query is constructed using a linear mapping. ,key Sum :

[0120] ;

[0121] in, They represent the first Layer encoder query ,key Sum The learnable weight matrix;

[0122] In this embodiment, the multi-head self-attention sublayer adopts an interpretable multi-head self-attention shared value structure, that is, for the first... Each attention head uses an independent linear transformation only for the query and the key, while all attention heads share the same value representation.

[0123] Remember the number of attention points. Each key has a dimension of , No. The weight matrix for each attention head is as follows:

[0124] ;

[0125] in, Indicates the first Individual attention head query s and keys The learnable weight matrix; ( ) T This indicates transpose.

[0126] A uniform time weight is obtained by averaging the weights of all attention heads:

[0127] ;

[0128] Subsequently, a single Value transformation matrix is ​​used. By aggregating the different attention patterns of all attention heads along the time axis into the same feature space, the final multi-head self-attention output is obtained. :

[0129] ;

[0130] Furthermore, to adapt to multi-source temporal data and enhance training stability, the encoder uses batch normalization after the residual connections instead of layer normalization in the classic Transformer. The batch normalization operation including the residual connections yields the attention sublayer output. :

[0131] ;

[0132] in, This represents the batch normalization operator that calculates the mean and variance along both the batch and time dimensions.

[0133] Batch normalization effectively suppresses numerical scale differences between different samples and the impact of outliers on the training process, making it more suitable for the amplitude distribution characteristics of satellite remote sensing time series. After the attention sublayer, a feedforward network is introduced to enhance the nonlinear modeling capability of the representation. The feedforward network acts independently at each time step, and its form consists of a two-layer fully connected network and a nonlinear activation function, denoted as... If the output of the feedforward network is:

[0134] ;

[0135] in, , They represent the first The learnable parameter matrix of the layer encoder feedforward network; Indicates the first The bias vector of the layer encoder feedforward network; Indicates the first The bias vector of the activation function of the layer encoder feedforward network;

[0136] Subsequently, the output of the attention sublayer is residually concatenated with the output of the feedforward network and batch normalized to obtain the output of the encoded block. :

[0137] ;

[0138] By stacking The aforementioned encoding blocks, supported by shared value multi-head self-attention, can simultaneously capture high-order interaction features between long-range temporal dependencies and multiple variables, and maintain numerical stability under batch normalization adjustment. The final encoder output... :

[0139] .

[0140] S44 introduces a static-sensitive full-sequence attention aggregation mechanism to map static geographic context vectors to global query vectors, constructing corresponding key and value vectors for each encoding block output of the improved Transformer encoder; using scaled dot products as static-sensitive attention weights, global normalization is performed over all time steps to obtain the attention distribution; and the static-sensitive aggregated features are obtained by weighted summation of the value vectors based on the attention distribution.

[0141] To highlight the modulating effect of static geographic environment on the entire reproductive timeline, a static-sensitive full-sequence attention aggregation mechanism is introduced. This mechanism aggregates static geographic context vectors... The mapping is used as a global query vector to perform a cross-time attention retrieval on the encoded feature sequence along the time axis.

[0142] Static geographic vectors are encoded into queries through linear transformation. :

[0143] ;

[0144] in, These represent the learnable weight matrix that linearly projects the static geographic context vector onto the query space, and the learnable bias vector corresponding to the nature mapping layer, respectively. Both can be updated through backpropagation during training. W represents the learnable weight matrix q The dimension;

[0145] Encoded features for each time step Construct the corresponding key vector and value vector:

[0146] ;

[0147] in, , , and These represent the weight matrix that linearly maps the encoded features to the key vector, the bias vector when generating the key vector, the weight matrix that linearly maps the encoded features to the value vector, and the bias vector when generating the value vector, respectively. They are all linear layer parameters used in the attention aggregation module to construct the key and value, and are updated through backpropagation during training.

[0148] Using scaled dot products as statically sensitive attention weights, a global normalization is performed over all time steps to obtain the attention distribution. :

[0149] ;

[0150] in, Represents an exponential form with base e; express The dimension of the vector; T represents the total number of time steps in the time series; This represents the key at the j-th time step;

[0151] The static sensitive aggregated features are obtained by weighted summation of the value vectors. :

[0152] ;

[0153] In this aggregation process, the query vector is entirely determined by static geographic conditions, while the keys and values ​​are derived from dynamic temporal encoding, thus enabling the filtering of key reproductive stages on the timeline based on static information. Under static climate and soil constraints, the attention weight automatically increases the contribution of the time segment that best matches the current geographic environment.

[0154] S45, after concatenating the static sensitive aggregated features with the static geographic context vector, the concatenated features are mapped to the crop yield space based on the linear regression head to obtain the crop yield prediction results for the target area.

[0155] Static sensitive aggregation features Compared with the original static geographic context vector Together they constitute the final multimodal fusion representation, which is sensitive to static aggregated features. With the original static geographic context Vector concatenation is performed to obtain fused features. :

[0156] ;

[0157] Wherein, Concat() represents the vector concatenation operation;

[0158] Subsequently, the fused features are mapped to the crop yield space based on the linear regression head to obtain the yield prediction results. :

[0159] ;

[0160] in Indicates the weights of the output layer; This indicates the bias of the output layer.

[0161] For a batch with size The sample set, the actual output label is denoted as The corresponding prediction result is The time-series regression branch uses mean squared error as the loss function. :

[0162] ;

[0163] This loss function updates all parameters in the Transformer encoder, the statically sensitive full-sequence aggregation module, and the regression head through backpropagation, and forms a unified optimization process with the encoder parameters shared in the unsupervised pre-training stage. This allows the model to maintain its ability to model long-distance dependencies while adapting to the output labels in a targeted manner, thereby obtaining high-precision regression prediction capabilities applicable to multi-source remote sensing time series.

[0164] Example 2

[0165] This embodiment provides a crop yield prediction system based on dynamic and static collaborative temporal attention, such as... Figure 2 As shown, the system for implementing the crop yield prediction method based on dynamic-static collaborative temporal attention as described in Example 1 includes:

[0166] The acquisition module is used to acquire multivariate remote sensing time-series datasets of the target area;

[0167] The preprocessing module is used to preprocess the multivariate remote sensing time series dataset to obtain dynamic variable data and static covariate data; the multivariate remote sensing time series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data describing the spatial location of the monitored object;

[0168] The static covariate processing module is used to perform nonlinear mapping on the static covariate data based on the multilayer perceptron encoder to obtain a static geographic context vector, and use the static geographic context vector as prior knowledge to remove noise variables in the dynamic variable data.

[0169] The dynamic variable processing module is used to mask the dynamic variable data after removing noise variables to obtain a damaged dynamic sequence. Based on the multi-head self-attention mechanism, it captures the correlation between different variables in the damaged dynamic sequence to generate a deep feature representation. Based on the deep feature representation, it reconstructs the masked part in the dynamic variable data.

[0170] The prediction module is used to concatenate global features in chronological order and introduce a time attention mechanism that uses static geographic context vectors as queries to extract static sensitive aggregated features that focus on the crop growth stage. The static sensitive aggregated features and static geographic context vectors are combined to map the crop yield of the target area.

[0171] Example 3

[0172] This embodiment provides a crop yield prediction system based on dynamic and static collaborative temporal attention, and its working process is as follows: Figure 3 As shown, it includes multi-source satellite data preprocessing and sample construction process, spectral encoder pre-training process, static covariate encoding process, time series regression prediction process, etc.

[0173] Multi-source satellite data preprocessing and sample construction workflow: Geometric correction, spatial resampling, and radiometric normalization are performed on multivariate remote sensing time-series datasets from different satellite platforms and imaging times. Cloud noise and outlier detection are conducted within static geographic partitions based on typical spectral time-series baseline curves, and invalid observations are marked. Subsequently, spectral sequences of each plot are resampled at equal intervals according to a unified crop growing season time axis, and interpolation padding is used to map unequal-length sequences to a preset length. Based on this, the data is further divided into two groups: dynamic variables (multi-temporal, multi-band reflectance, etc.) and static variables (latitude, longitude, altitude, slope, soil type, etc.), providing input for subsequent static covariate encoding and dynamic-static feature co-modeling.

[0174] The spectral encoder pre-training process involves pre-training multi-source satellite spectral time series data in an unsupervised manner. Masks are randomly generated for the input dynamic spectral time series based on variables and time. The masked position features are set to zero and then fed into the variable selection subnet and Transformer encoder. A linear reconstruction head is connected to the output, and the reconstruction error is calculated only at the mask position and the parameters are updated in reverse. Through the mask reconstruction task, the model can learn the time dependence and multivariate correlation of the spectral sequence without using output labels.

[0175] Static covariate encoding branch: This module is responsible for uniformly representing time-independent features such as latitude and longitude, altitude, slope, soil type, and variety coding. The above static features are first normalized and then input into the multilayer perceptron (MLP) static encoding subnet to obtain a fixed-length static geographic context vector. This vector is used in the variable selection subnet and static sensitive full sequence aggregation to provide the model with global prior constraints that reflect regional environmental differences.

[0176] The static context-driven variable selection subnetwork: This module adaptively filters multi-source, multi-band features at each time step. The dynamic feature vector of the current time step and the static geographic context vector are input into a gating network centered on a gated residual modulation network (GRMU) to generate importance scores for each spectral band and covariate, which are then normalized to obtain variable weights. The input features are weighted and denoised before being fed into the Transformer encoder, enabling the model to highlight key variables with stronger yield relationships during both pre-training and supervised training phases, thus improving the effectiveness and robustness of temporal feature extraction.

[0177] The Transformer encoding and statically sensitive full-sequence aggregation module: This module first inputs the time series data after variable selection into a multi-layer Transformer encoder. Through a multi-head self-attention and feedforward network sharing values, long-distance temporal dependencies and multivariate interaction features are extracted to obtain high-dimensional representations of each time step. Then, a statically sensitive full-sequence aggregation strategy is employed, using the static geographic context vector as the query to map the features of each time step to keys and values. The temporal dimension attention weights are calculated, and the value vectors are weighted and summed to obtain global temporal features focusing on key reproductive stages. Simultaneously, all time-step features are sequentially concatenated to form a complete growing season representation, which, along with the statically sensitive global features and the static geographic context vector, is fed into the regression output layer to obtain the predicted plot yield.

[0178] Example 4

[0179] This embodiment provides a computer-readable medium storing a computer program, which, when executed by a processor, can implement the crop yield prediction method based on dynamic-static collaborative temporal attention as described in Embodiment 1; specifically, it performs the following steps:

[0180] Step 1: Collect a multivariate remote sensing time-series dataset of the target area, and preprocess the multivariate remote sensing time-series dataset to obtain dynamic variable data and static covariate data; the multivariate remote sensing time-series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data that describes the spatial location of the monitored object;

[0181] Step 2: Based on the multilayer perceptron encoder, the static covariate data is nonlinearly mapped to obtain a static geographic context vector, and the static geographic context vector is used as prior knowledge to remove noise variables in the dynamic variable data.

[0182] Step 3: Mask the dynamic variable data after removing noise variables to obtain the damaged dynamic sequence. Capture the correlation between different variables in the damaged dynamic sequence based on the multi-head self-attention mechanism to generate a deep feature representation. Reconstruct the masked part in the dynamic variable data based on the deep feature representation.

[0183] Step four: Concatenate global features in chronological order, and introduce a time attention mechanism that uses static geographic context vectors as queries to extract static sensitive aggregated features that focus on the crop growth stage. Combine the static sensitive aggregated features and static geographic context vectors to map the crop yield of the target area.

[0184] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting crop yield based on dynamic and static collaborative temporal attention, characterized in that, The method includes: A multivariate remote sensing time-series dataset of the target area is collected, and the multivariate remote sensing time-series dataset is preprocessed to obtain dynamic variable data and static covariate data; the multivariate remote sensing time-series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data describing the spatial location of the monitored object; A static geographic context vector is obtained by nonlinear mapping of the static covariate data based on a multilayer perceptron encoder, and the static geographic context vector is used as prior knowledge to remove noise variables in the dynamic variable data. The process involves masking the dynamic variable data after removing noise variables to obtain a damaged dynamic sequence. A deep feature representation is generated by capturing the correlation between different variables in the damaged dynamic sequence using a multi-head self-attention mechanism. The masked portion of the dynamic variable data is then reconstructed based on this deep feature representation. The reconstructing of the masked portion of the dynamic variable data based on the deep feature representation includes the following method: A multi-head self-attention mechanism based on a Transformer encoder is used to capture the correlation between different variables in a damaged dynamic sequence in parallel, generating a deep feature representation Z: ;in, Indicates time step The eigenvectors; t=1,2,…,t,…,T; T represents the total number of time steps; Connecting a linear mapping layer to the output of the Transformer encoder will... Project back the dimensions of the dynamic variable data after removing noise variables to reconstruct the masked portion of the dynamic variable data. In the process of reconstructing the masked part in dynamic variable data, the mean square error of the masked elements is used as the loss function; Global features are concatenated in chronological order, and a time attention mechanism based on static geographic context vectors is introduced to extract static sensitive aggregated features that focus on crop growth stages. The crop yield of the target area is then mapped by combining the static sensitive aggregated features and the static geographic context vectors. The methods for predicting crop yields in the target area include: Construct time-series samples with output labels based on dynamic variable data and static covariate data; Construct an improved Transformer encoder with parameters consistent with the Transformer encoder described above; The improved Transformer encoder is trained in a supervised manner based on time-series samples with output labels: A static-sensitive full-sequence attention aggregation mechanism is introduced to map static geographic context vectors to global query vectors, constructing corresponding key and value vectors for each encoding block output of the improved Transformer encoder; using scaled dot products as static-sensitive attention weights, global normalization is performed over all time steps to obtain the attention distribution; and the static-sensitive aggregated features are obtained by weighted summation of the value vectors based on the attention distribution. After concatenating the static sensitive aggregated features with the static geographic context vector, the concatenated features are mapped to the crop yield space based on the linear regression head to obtain the crop yield prediction results for the target area.

2. The crop yield prediction method based on dynamic-static collaborative temporal attention according to claim 1, characterized in that, The preprocessing includes: Using each monitored object as a sample, the multivariate remote sensing time series dataset is mapped into a continuous tensor space, and the static geographic attribute data of each sample is extracted. Based on the static geographic attribute data, local geographic constraints are constructed, the whole sample is mapped into multiple regions with the same geographic properties, the time series tensors of all samples within each region are aggregated, and the statistical distribution parameters of each time series tensor on each spectral channel are calculated, thereby establishing the crop growth standard baseline for each region. The temporal tensor of individual samples is evaluated point by point based on the crop growth standard baseline of each region to identify outliers or missing data points. The outliers or missing data points are then corrected based on a spatiotemporal co-interpolation strategy. The spatiotemporal co-interpolation strategy includes: weighting and fusing the local trend of individual samples in the time dimension with the mean of the crop growth standard baseline of the current region in the spatial dimension to generate a correction value. The corrected multivariate remote sensing time series dataset was divided into dynamic variables and static variables, and the dynamic variables and static covariate data were obtained by standardizing the dynamic variables and static covariate data respectively.

3. The crop yield prediction method based on dynamic-static collaborative temporal attention according to claim 1, characterized in that, The methods for removing noise variables from the dynamic variable data include: A static geographic context vector is obtained by nonlinear mapping of the static covariate data based on a multilayer perceptron encoder. Obtain all dynamic variables and the flattened vector of each dynamic variable from the dynamic variable data; The flattened vectors of all dynamic variables and the static geographic context vectors are input into the gated residual modulation network GRMU. A normalized variable selection weight vector is generated based on the softmax layer of the gated residual modulation network GRMU. At the same time, an independent GRMU subunit is configured for each dynamic variable to perform nonlinear transformation.

4. The crop yield prediction method based on dynamic-static collaborative temporal attention according to claim 3, characterized in that, The masking method includes: A binary noise mask matrix is ​​generated from the dynamic variable data after removing noise variables; the time series length and variable dimension of the binary noise mask matrix are consistent with the time series length and variable dimension of the dynamic variable data after removing noise variables. The damaged dynamic sequence is obtained by performing element-wise product of the binary noise mask matrix and the dynamic variable data after removing the noise variables.

5. The crop yield prediction method based on dynamic-static collaborative temporal attention according to claim 1, characterized in that, The training method for the improved Transformer encoder includes: Using static geographic context vectors as input, queries, keys, and values ​​are constructed. All attention heads share the same Value representation, and each attention head only uses independent linear transformations on the query and key. The weights of all attention heads are averaged to obtain a unified temporal weight. By sharing the Value transformation matrix, the different attention patterns of all attention heads on the time axis are aggregated into the same feature space and batch normalization with residual connections is performed to obtain the attention sub-layer output. A feedforward network is introduced to process the attention sub-layer output independently at each time step to obtain the feedforward network output. The attention sub-layer output and the feedforward network output are batch normalized with residual connections to obtain the encoding block output. Multiple encoding block outputs are stacked to obtain the encoder output.

6. The crop yield prediction method based on dynamic-static collaborative temporal attention according to claim 1, characterized in that, Mean squared error is used as the loss function during supervised training.

7. A crop yield prediction system based on dynamic and static collaborative temporal attention, characterized in that, The system is used to implement the crop yield prediction method based on dynamic-static collaborative temporal attention as described in any one of claims 1-6, the system comprising: The acquisition module is used to acquire multivariate remote sensing time-series datasets of the target area; The preprocessing module is used to preprocess the multivariate remote sensing time series dataset to obtain dynamic variable data and static covariate data; the multivariate remote sensing time series dataset includes high-dimensional spectral data that changes dynamically over time and static geographic attribute data describing the spatial location of the monitored object; The static covariate processing module is used to perform nonlinear mapping on the static covariate data based on the multilayer perceptron encoder to obtain a static geographic context vector, and use the static geographic context vector as prior knowledge to remove noise variables in the dynamic variable data. The dynamic variable processing module is used to mask the dynamic variable data after removing noise variables to obtain a damaged dynamic sequence. Based on the multi-head self-attention mechanism, it captures the correlation between different variables in the damaged dynamic sequence to generate a deep feature representation. Based on the deep feature representation, it reconstructs the masked part in the dynamic variable data. The prediction module is used to concatenate global features in chronological order and introduce a time attention mechanism that uses static geographic context vectors as queries to extract static sensitive aggregated features that focus on the crop growth stage. The static sensitive aggregated features and static geographic context vectors are combined to map the crop yield of the target area.

8. A computer-readable medium having a computer program stored thereon, characterized in that, The computer program, when executed by a processor, can implement the crop yield prediction method based on dynamic-static collaborative temporal attention as described in any one of claims 1-6.

Images