Crop yield space-time prediction method, system and equipment based on AI-Agent and multiple attention mechanisms

Abstract

The invention provides a crop yield space-time prediction method, system and device based on AI-Agent and multiple attention mechanisms. Aiming at the heterogeneous problem of multi-source spatio-temporal data (vegetation, climate and soil) fusion, the method fuses agricultural expert knowledge and deep learning, and constructs a semantic decoupling multiple attention network (SD-MANet). The core innovation lies in that Al-Agent is introduced, through a knowledge distillation mechanism driven by reinforcement learning, fusion weights of different data modes are adaptively adjusted (for example, the weight of soil water in an arid year can be dynamically improved by about 30%), and the precision is remarkably improved. The method comprises the following steps: decoupling input data into vegetation, climate and soil groups based on agricultural semantics; an AI-Agent driven multi-scale fusion and time attention module is deployed, and self-adaptive feature integration is achieved; and in combination with SHAP attribution analysis and iterative optimization, parameters are finely adjusted in a closed-loop manner, and the model robustness is enhanced. Compared with a static fusion method, the prediction precision is improved by 15%-20%. The method is suitable for grain safety monitoring, agricultural decision making and disaster assessment under climate change.

Description

Technical Field

[0001] This invention belongs to the fields of remote sensing technology and artificial intelligence, specifically relating to a crop yield prediction method that utilizes AI-Agent coordination of semantic decoupling and multiple attention mechanisms for multi-source spatiotemporal data fusion. This technology is primarily applied in areas such as food security monitoring, agricultural management, and climate change assessment. This invention particularly emphasizes the adaptive coordination application of AI-Agent in agricultural remote sensing, differing from traditional multimodal fusion methods in its dynamic knowledge distillation mechanism. Furthermore, this invention integrates subfields of agricultural ecological parameter prediction, such as remote sensing NDVI sequence processing and meteorological reanalysis data fusion, while incorporating machine learning applications, including multi-scale attention architecture, semantic decoupling strategies, and SHAP interpretive analysis. According to the International Patent Classification (IPC), this invention can be classified under G06N 3 / 08 (neural network training algorithms) and G06Q 50 / 02 (agricultural decision support systems) to ensure its standardization and operability in multimodal remote sensing data-driven prediction. Background Technology

[0002] Crop yield forecasting, as a core tool for ensuring food security, plays a crucial role in the context of global climate change. This field faces multiple challenges, including the complex impacts of climate variability, dynamic changes in soil moisture stress, and the nonlinear characteristics of vegetation growth processes. These factors combine to lead to high uncertainty in the crop growth environment, thus placing higher demands on the accuracy and robustness of forecasting models. Traditional methods mainly rely on process-oriented models, such as Decision Support Systems for Agricultural Technology (DSSAT) or Agricultural Production Systems Simulator (APSIM). These models, based on crop physiological and ecological principles, estimate yield by simulating crop growth processes, possessing a certain degree of mechanistic interpretability and clearly reflecting biophysical processes such as heat accumulation, water use, and nutrient dynamics. However, these methods require extensive parameter calibration, including precise calibration of soil type, crop variety, and meteorological inputs, which is often limited in practical applications due to data acquisition difficulties and computational complexity. Furthermore, process models typically ignore spatiotemporal heterogeneity; for example, under complex terrain or variable climate conditions, prediction errors may fluctuate within the range of 1-2 t / hm² (e.g., according to the validation results of the DSSAT model under variable climate conditions), failing to adequately meet the needs of large-scale regional forecasting.

[0003] With the rise of statistics and data-driven methods, traditional statistical models such as multiple linear regression have been widely used for crop yield prediction. These models establish linear or quasi-linear relationships between environmental variables and yield, enabling estimations based on historical data. For example, using accumulated temperature (GDD) and total precipitation as independent variables can quantify the cumulative impact of climate on yield. Although these methods are computationally efficient and easily integrate ground-based observation data, they struggle to capture the nonlinear heterogeneity of multi-source data. Under conditions of high humidity or extreme drought, the linear assumption often fails, leading to insufficient model generalization ability; for example, prediction bias can reach 20% (refer to the assessment of multiple linear regression on extreme climate datasets), and they cannot effectively handle interactions such as soil moisture stress and vegetation dynamics.

[0004] In the era of machine learning, algorithms such as random forests and support vector machines have gradually become mainstream due to their powerful nonlinear fitting capabilities. These methods can automatically handle high-dimensional feature spaces, such as extracting vegetation indices (NDVI) and meteorological reanalysis variables from satellite remote sensing data, and improve prediction stability through ensemble learning mechanisms. For example, random forest models can reduce the risk of overfitting by constructing multiple decision trees and integrate soil type and crop distribution data in crop yield prediction. However, these methods are highly dependent on expert experience in the feature extraction stage, requiring manual design of derived variables such as peak NDVI or integral NDVI. Furthermore, simple concatenation strategies used for multi-source data fusion cannot adequately address semantic heterogeneity issues, such as the physical meaning differences between vegetation group features (e.g., photosynthetic indicators) and climate group features (e.g., extreme high-temperature days), leading to insufficient adaptability of the model during dynamic growth.

[0005] In recent years, the field has undergone significant changes with the rapid development of deep learning technology. Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) have been introduced to achieve automatic extraction of spatiotemporal features and sequence modeling. For example, LSTM models excel at capturing temporal dependencies within the growing season and can handle long-term correlations between NDVI sequences and meteorological variables, while CNNs extract spatial patterns, such as regional heterogeneity of crop distribution, through convolutional operations. These methods have improved prediction accuracy to some extent, especially when fusing multimodal data. However, existing "feature-level fusion" strategies typically employ static functions, such as channel concatenation or summation, meaning that the weights of different modalities are fixed during training and cannot adapt to the dynamics of crop growth. For example, in drought years, soil water modalities (such as surface volumetric water content swvl1) should receive higher weights to reflect the impact of root water stress, but static fusion may lead to insufficient generalization of the model under varying climatic conditions.

[0006] To address the limitations of static fusion, attention mechanisms have been introduced to achieve dynamic weighting. Some existing techniques utilize temporal attention to identify key phenological periods or channel attention to highlight variables with high information content. However, these attention methods are often applied in a "flat" feature space, treating all inputs as homogeneous entities and ignoring the semantic structure of multimodal data. This may limit the model's ability to simulate agricultural ecological principles, such as its inability to adaptively balance semantically heterogeneous information when integrating vegetation, climate, and soil groups.

[0007] Existing patents and technical literature further demonstrate the progress in this field. For example, CN110443420B (publication date: November 12, 2019) proposes a machine learning-based crop yield prediction method. This method uses climate characteristics and growth cycle data for estimation and implements nonlinear mapping through algorithms such as support vector machines. Although this patent effectively integrates multi-source data, it lacks a multi-attention mechanism and cannot dynamically identify key time windows.

[0008] Among existing international technologies, US20050234691A1 (publication date: October 20, 2005) describes a crop yield prediction system using piecewise linear regression and breakpoint methods. This system combines weather and agricultural parameters for estimation and has a strong statistical foundation, but it does not involve deep learning or attention mechanisms. Recent literature, such as the AgriTransformer model (2025, MDPI), proposes a multimodal crop yield prediction model based on Transformer and attention mechanisms. This model integrates tabular data and vegetation indices, achieving dynamic fusion through cross-modal attention. Although this model improves predictive adaptability, it does not introduce an AI-Agent or semantic decoupling. Another paper, "Integrating multi-modal remotesensing, deep learning, and attention mechanisms for yield prediction" (2024, Frontiers), proposes a multimodal deep learning architecture that uses LSTM and attention mechanisms to fuse high-resolution hyperspectral images, LiDAR, and environmental data. This paper emphasizes the interpretation of attention weights but does not address semantic decoupling or agent-driven knowledge distillation. The paper "Agentic AI for Crop Yield Prediction on Databricks" (2025, XenonStack paper) describes the use of an AI agent for crop yield prediction on the Databricks platform, integrating soil, weather, and yield data. The paper introduces adaptive tuning of the agent AI but does not incorporate semantic decoupling or multiple attention.

[0009] In summary, while existing technologies have made progress in multi-source data fusion and dynamic weighting, they still have limitations overall. Although existing methods fuse multi-source data, they fail to achieve AI-Agent-driven semantic adaptive coordination, resulting in insufficient generalization under dynamic climatic conditions. These limitations are mainly reflected in insufficient depth of processing multimodal heterogeneity and the lack of adaptive mechanisms to cope with the spatiotemporal variability of crop growth. To address the aforementioned objective status quo of existing technologies, this invention proposes a spatiotemporal AI prediction method and system for crop yield based on AI-Agent and semantic decoupling and multiple attention mechanisms, using a structured fusion strategy to address these challenges. This method integrates semantic decoupling, multiple attention, and AI-Agent knowledge distillation, aiming to provide a more adaptive solution and lay the foundation for further development in the field of precision agriculture. Existing methods have prediction errors of 10%-15% in drought years; this invention, through AI-Agent adaptive coordination, reduces the error to less than 1 t / hm². Summary of the Invention

[0010] To address the aforementioned technical problems, this invention discloses a method, system, and device for spatiotemporal prediction of crop yield based on AI-Agent and a multi-attention mechanism, aiming to solve the problem that the accuracy of crop yield prediction is limited by the heterogeneity of multi-source spatiotemporal data in existing technologies. This invention integrates agricultural expert knowledge with deep learning mechanisms to construct a semantically decoupled multi-attention network (SD-MANet), and introduces an AI-Agent to coordinate multimodal feature extraction, dynamic fusion, and temporal attention, achieving high-precision spatiotemporal prediction of crop yield. Specifically, this invention innovatively employs a semantic decoupling strategy to process heterogeneous data, classifying input features into three major groups based on agricultural ecological principles: vegetation, climate, and soil geography; the multi-attention mechanism enables adaptive weight allocation and key time window identification; and SHAP attribution analysis enhances the interpretability of the model. This method, system, and device are applicable to crops such as US corn, providing reliable data support for food security monitoring, agricultural decision-making, and disaster assessment under climate change.

[0011] To achieve the above objectives, the technical solution provided by this invention is as follows: First, semantic decoupling is performed based on agricultural knowledge to establish modality-specific representations; second, multi-scale extraction is used to verify the impact of input on prediction; third, dynamic fusion and temporal attention are achieved through SD-MANet; finally, SHAP and iterative optimization are employed to improve performance. The core of this solution lies in the coordination of the AI-Agent, where the agent is responsible for knowledge distillation and parameter adjustment to ensure efficient operation. The technical solution of this invention is described in detail below, including its theoretical basis, model architecture, optimization strategies, and system implementation, to ensure the completeness and operability of the solution.

[0012] I. Technical Problems and Solutions

[0013] In existing technologies, crop yield prediction mainly relies on static fusion strategies, which face challenges from the semantic heterogeneity of multi-source data and the dynamic nature of growth processes, leading to a significant increase in prediction errors. For example, under extreme climatic conditions, such as the 2012 drought in the US Midwest, traditional models often ignore the interaction between soil water modalities and vegetation dynamics, resulting in yield underestimations of 10%-15%. This invention proposes the SD-MANet framework based on AI-Agent, where the AI-Agent acts as an intelligent agent, dynamically coordinating semantic decoupling and multiple attention modules through reinforcement learning to achieve deep integration of physical models and data-driven methods. This framework not only simulates the decision-making logic of agricultural experts (such as prioritizing soil moisture stress during drought periods) but also ensures that the prediction results conform to agricultural ecological principles through iterative optimization. The core of this approach is to utilize the adaptive capabilities of the AI-Agent to achieve high-precision spatiotemporal prediction, for example, controlling the root mean square error (RMSE) within 0.269 t / hm² on the test dataset, while simultaneously improving the model's robustness to interannual variability.

[0014] II. Theoretical Derivation of Parameter Prediction

[0015] The theoretical basis of this invention stems from agricultural ecological models, which express crop yield as a composite form f of multimodal functions:

[0016]

[0017] in, This indicates the yield density (unit: t / ha). It represents vegetation group characteristics (such as the photosynthetic intensity reflected by the NDVI sequence). This indicates climate group characteristics (such as growth accumulated temperature (GDD) and quantitative heat accumulation and stress on extreme high-temperature days). The soil geography group represents soil characteristics (e.g., volumetric water content swvl1 / swvl2 characterizes root water availability). This formula reveals the semantic heterogeneity of multimodal data: the vegetation group emphasizes biomass accumulation, the climate group focuses on external stresses, and the soil geography group focuses on intrinsic location dependence. The sensitivity of each mode to yield is assessed using the Jacobian matrix.

[0018]

[0019] in, Represents matrix elements, This represents unknown parameters (such as temperature threshold or emissivity). The matrix quantifies the contribution of each mode and is used to optimize input combinations, such as prioritizing drought-sensitive bands. Validation on simulated datasets shows that Jacobi matrix optimization reduces the mean absolute error (MAE) by 15%. Specifically, in a simulated scenario covering US maize from 2008 to 2022, the MAE decreased from 0.85 t / ha to 0.72 t / ha, confirming the matrix's role in improving mode separation accuracy. This derivation not only provides theoretical support but also guides the design of subsequent model architectures, ensuring that the prediction process conforms to biophysical principles.

[0020] III. SD-MANet Model Architecture

[0021] The core of this invention is the AI-Agent-based SD-MANet architecture, which employs semantic decoupling and a multi-attention mechanism to process 45 channels (27 dynamic channels + 18 static channels) of input data. First, the input is grouped into vegetation (12 channels), climate (18 channels), and soil geography (15 channels), and processed by a dedicated 3×3 convolutional encoder to learn modality-specific representations. The AI-Agent coordinates a multi-scale fusion attention module to achieve dynamic weight allocation.

[0022]

[0023] in, Indicates fused output. Enter information. Let represent the weights of the k-th group, calculated using a 1×1 convolution and the sigmoid function. The fused features are then fed into a ConvLSTM unit to capture temporal dependencies. The state update formula is:

[0024]

[0025] in, , , These are the hidden state, the output gate, and the cell state (the input gate and the forget gate are updated through a similar mechanism).

[0026] The time attention module further weights the key window:

[0027]

[0028] Among them Spatiotemporal attention score. The AI-Agent transmits parameters through knowledge distillation to ensure deployment efficiency. The pseudocode for the specific coordination algorithm is as follows:

[0029] text

[0030] copy

[0031] # AI-Agent Coordination Algorithm

[0032] Initialize the Q-value table Q(s, a) = 0 # State-action value function

[0033] For each episode:

[0034] Choose an initial state s (current modal feature).

[0035] However, it did not end:

[0036] Choose action a (adjust weights or distillation parameters) based on the ε-greedy strategy.

[0037] Perform the action, observe the new state s' and the reward R (based on accuracy and robustness).

[0038] Update Q(s, a) = Q(s, a) + α [R + γ max Q(s', a') - Q(s, a)] # α is the learning rate (typically 0.1), γ is the discount factor (typically 0.9)

[0039] Finish

[0040] The algorithm initializes the Q-value table and iteratively updates it based on the reward R, ensuring optimized fusion process in resource-constrained environments. The architecture is experimentally validated, achieving R²=0.950 on the US corn test set, significantly outperforming the static fusion model.

[0041] IV. SHAP Attribution and Parameter Fine-Tuning Strategies

[0042] We introduce SHAP (SHapley Additive exPlanations) to analyze feature contributions and quantify the marginal impact of each mode:

[0043]

[0044] in, Let denot SHAP value, S represent the feature subset, and v represent the model prediction function. This analysis reveals that the number of extreme high-temperature days contributes up to 25% to drought years, enhancing the model's interpretability. Iterative optimization uses the AdamW optimizer, combined with the mean squared error loss function, for parameter updates. The label refinement strategy employs:

[0045]

[0046] Where y′ is the adjusted label and λ is the relaxation factor (range 0.4-0.6). AI-Agent-driven closed-loop iteration ensures physical consistency; for example, in simulating extreme events, the model bias is reduced by 12% after iteration. This strategy, through multiple validations, improves the model's generalization ability under varying climate conditions.

[0047] V. System Implementation

[0048] This invention provides a system comprising a data acquisition module (for acquiring NDVI, meteorological, and soil data), a semantic decoupling module (for classification and characterization extraction), an AI-Agent coordination module (for driving dynamic fusion), and an optimization and interpretation module (for SHAP analysis and parameter fine-tuning). Furthermore, electronic devices, such as a processor and memory, are provided to store computer program instructions to implement the aforementioned methods. The system implementation process is as follows: Acquire a crop spatiotemporal dataset (covering US maize from 2008 to 2022, with approximately 119,235 samples); Deploy an AI-Agent system and dynamically optimize the network architecture; Utilize the AI-Agent to coordinate knowledge transfer and form an adaptive model; Fuse the model to predict crop yield in real time.

[0049] The optimal deployment of an AI-Agent system for dynamic search includes: Defining the agent search space:

[0050]

[0051] Where L is the number of levels, U is the unit size, D is the decision function, and M is the mutation probability. Reinforcement learning is used to guide the search, and state similarity is calculated as follows:

[0052]

[0053] The reward function R evaluates efficiency, and the optimal architecture is:

[0054]

[0055] θ is a surrogate parameter, and this process was optimized experimentally, reducing training time by 20% in a GPU environment.

[0056] VI. Beneficial Effects

[0057] This invention significantly improves prediction accuracy: achieving a coefficient of determination (R²) of 0.950 and a root mean square error (RMSE) of 0.269 t / ha on the test set. Compared to traditional methods, it enhances generalization ability and supports extreme climate applications, such as predicting a bias of less than 1 t / ha in the drought year of 2012. AI-Agent fusion reduces computational costs by 20% and is suitable for real-time prediction. The following comparison table summarizes the performance improvements:

[0058]

[0059] This table, validated using a US corn dataset, demonstrates the non-obvious advantages of this invention in nonlinear heterogeneous processing, providing an efficient tool for agricultural decision-making. Attached Figure Description

[0060] Figure 1 This is a flowchart of the spatiotemporal AI prediction method for crop yield based on AI-Agent according to an embodiment of the present invention, showing the sequence of steps of the overall technical solution, including data acquisition, semantic decoupling, AI-Agent coordination and optimization processes.

[0061] Figure 2 This is a flowchart of multimodal data processing, showing the complete pipeline from spatiotemporal alignment of NDVI sequences, meteorological variables, and soil moisture data to Z-score normalization.

[0062] Figure 3 The illustrations for multi-scale feature extraction illustrate the dynamic and static feature construction process at the local (3×3), context (9×9), and region (21×21) levels.

[0063] Figure 4 This is a schematic diagram of the gating fusion mechanism, showing the process of AI-Agent adjusting modal weights, including 1×1 convolution calculation and dynamic weighting of Sigmoid activation; in the diagram, circles represent 1×1 convolutions and arrows represent the weight adjustment process.

[0064] Figure 5 This is a schematic diagram of the temporal attention module, illustrating the mechanism by which composite temporal attention calculates spatiotemporal attention scores, used to identify key windows during the growing season.

[0065] Figure 6 This is a schematic diagram of the overall model structure, showing the end-to-end process of the SD-MANet architecture, including semantic decoupling, ConvLSTM timing capture, and AI-Agent coordination modules.

[0066] Figure 7 This is a scatter plot used to compare the distribution of model predictions with the true labels, and is used to evaluate the coefficient of determination (R²) and correlation.

[0067] Figure 8 The residual distribution plot shows the histogram of prediction errors and statistical indicators such as root mean square error (RMSE).

[0068] Figure 9 To predict the spatial distribution of errors, we map the spatial pattern of errors in the US corn region, highlighting the biases in drought years.

[0069] Figure 10This is an interannual variation tracking map that tracks changes in environmental factors, such as NDVI and soil moisture, in typical grid cells in different years (e.g., the drought year of 2012).

[0070] Figure 11 This is a spatial distribution map of production output, comparing the spatial patterns of model-predicted output with actual data, used for regional validation (comparison of prediction and reality).

[0071] Figure 12 The feature importance map is generated by quantifying the contribution rate of each mode (such as the number of extreme high-temperature days) using SHAP values.

[0072] Figure 13 This is a scale-cumulative importance map, with bar charts showing the cumulative impact of multi-scale features (such as local vs. regional) on the prediction.

[0073] Figure 14 To illustrate the ablation experiment performance, we analyze the performance gap after removing specific modules (such as multiple attention). Detailed Implementation

[0074] This invention provides a spatiotemporal AI prediction method, system, and device for crop yield based on AI-Agent and semantic decoupling and a multi-attention mechanism, applicable to the US maize dataset (2008-2022). This method achieves high-precision yield prediction by integrating multi-source spatiotemporal data, making it suitable for agricultural applications under climate change. The hardware environment includes an Intel Xeon Platinum CPU (multi-core configuration), an RTX 4090 GPU (24GB VRAM), and 80GB RAM; the software environment is based on the Ubuntu 22.04 operating system, Python 3.10, and the deep learning framework PyTorch 2.0. This implementation is not limited to specific hardware. The following details the method steps, model architecture, system modules, and verification results to ensure that those skilled in the art can implement this invention based on this description. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0075] Example 1

[0076] This embodiment provides a spatiotemporal AI prediction method for crop yield based on AI-Agent and semantic decoupling and multiple attention mechanisms, such as... Figure 1 As shown, it includes the following steps:

[0077] S1: Acquire the dataset, including NDVI sequences (from the AVHRR GIMMS-3G+ dataset, spatial resolution 0.0833°, covering the growing season from March to October 2008-2022), meteorological variables (from the ECMWF ERA5 reanalysis product, including accumulated temperature (GDD), extreme high-temperature days, total precipitation, etc., spatial resolution 0.1°), soil moisture data (volume content of topsoil swvl10⁻⁷ cm and subsoil swvl27⁻²28 cm), and crop distribution labels (from the USDA CDL dataset, original resolution 30 m). Data preprocessing includes spatiotemporal alignment (unified to EPSG:4326 coordinate system), resampling to a 0.0833° grid (using bilinear interpolation), and Z-score normalization, such as... Figure 2 As shown. The dataset covers US corn from 2008 to 2022, with approximately 119,235 samples (7,949 grid cells × 15 years). A corn percentage threshold of >0.02 was used to exclude non-dominant planting areas.

[0078] S2: Semantic decoupling, extracting multi-scale features (27 dynamic channels, including 9 variables × 3 scales such as NDVI and GDD; 18 static channels, including 6 variables × 3 scales such as peak NDVI and integral NDVI), such as... Figure 3 As shown. Based on agricultural knowledge, the input is classified into three groups: vegetation (12 channels), climate (18 channels), and soil geography (15 channels). A dedicated 3×3 convolutional encoder is used to learn modality-specific representations, such as... Figure 4 As shown. The encoding process uses the ReLU activation function and a batch size of 32 to ensure non-linear feature capture.

[0079] S3: Introduce an AI-Agent, use reinforcement learning to adjust weights, and optimize the mean squared error (MSE) loss, such as... Figure 5 As shown. Training used the AdamW optimizer with an initial learning rate of 0.001, weight decay of 0.01, and an early stopping threshold of 10 epochs (based on no improvement in MSE on the validation set). The AI-Agent dynamically optimized the attention parameters using the Q-learning algorithm, and the iteration efficiency was evaluated based on a custom reward function R (R = 0.6·accuracy + 0.4·robustness), as shown. Figure 6 As shown. The training set accounts for 70%, the validation set 15%, and the test set 15%. Five-fold cross-validation is used to ensure generalization.

[0080] S4: Joint optimization to achieve prediction; verification on the independent test set shows a coefficient of determination (R²) of 0.950 and a root mean square error (RMSE) of 0.269 t / hm². Figure 7 As shown.

[0081] Theoretical verification: The impact of multi-scale combinations was evaluated using a simulated dataset (10,000 samples generated based on a radiative transfer model). Results showed that introducing a regional scale (21×21) improved prediction accuracy by 12%, with the regional scale dominating the contribution at 45%, confirming the necessity of multi-scale extraction. Figure 8 As shown.

[0082] Regional validation: The drought year of 2012 (representing an extreme event) and the bumper year of 2016 were selected as case studies, and MODIS products were used for cross-comparison. The spatial distribution predicted by the model is consistent with the actual yield trend, with an error of <1 t / hm² in the drought year (MAE=0.85 t / hm²). Nighttime predictions are better than daytime predictions (due to reduced radiation interference), as shown in the example. Figure 9 As shown.

[0083] Ground validation: Comparison was conducted using ISMN station data (approximately 500 stations in southern North America). Station data underwent outlier removal via box plot (IQR threshold 1.5 times), and spatiotemporal interpolation was synchronized to satellite observations. Results showed a daytime PCC of 0.994 and a nighttime PCC of 0.994, with an overall correlation coefficient improvement of 0.1 compared to the baseline product. Figure 10 As shown.

[0084] The beneficial effects of this embodiment are: through AI-Agent adaptive coordination, the accuracy is superior to existing technologies, providing a novel paradigm that integrates semantic decoupling and multiple attention, supporting agricultural decision-making under complex climate conditions, such as... Figure 11 As shown.

[0085] Example 2

[0086] This embodiment provides a spatiotemporal AI prediction system for crop yield based on AI-Agent, semantic decoupling, and multiple attention mechanisms, used to implement the method described in Embodiment 1. The system includes:

[0087] The data acquisition module is used to acquire multi-source spatiotemporal data and reference labels. Specifically, this module supports automatic downloading and preprocessing of data from AVHRR, ECMWF, and USDA databases, including resampling and filtering, to ensure the spatiotemporal consistency of the input data, such as... Figure 12 As shown.

[0088] The semantic decoupling module is used to classify input data and extract modal representations. This module implements grouping based on hard-coded indexes (12 channels for vegetation, 18 channels for climate, and 15 channels for soil geography) and processes feature maps using a dedicated convolutional encoder. It supports parallel computation to improve efficiency. Figure 13 As shown.

[0089] The AI-Agent Coordination Module is used for adaptive fusion and temporal attention based on reinforcement learning-driven SD-MANet. This module defines the agent search space:

[0090]

[0091] Where L is the number of levels (typically 4-8), U is the unit size (128-512 hidden nodes), D is the decision function (ε-greedy policy), and M is the mutation probability (0.1-0.3). Reinforcement learning is used to guide the search, and the state similarity function is calculated based on cosine similarity.

[0092]

[0093] To ensure a 20% improvement in exploration efficiency, the Q-value table is iteratively updated using Q-learning to achieve dynamic adjustment of modality weights. Optimal parameters:

[0094]

[0095] Where R is the reward function. This module supports real-time adaptation, such as automatically increasing the soil group weight in drought scenarios, etc. Figure 14 As shown.

[0096] The optimization and interpretation module is used for SHAP attribution analysis and parameter fine-tuning. This module calculates feature contributions and refines labels through iterative closed-loop processing, supporting visual outputs such as importance graphs.

[0097] The system in this embodiment is deployed in a hardware environment, supports distributed computing, is suitable for processing large-scale datasets, and ensures real-time prediction.

[0098] Example 3

[0099] This embodiment provides an electronic device comprising: a processor (e.g., an Intel Xeon series) and a memory (e.g., an SSD storage medium) storing computer program instructions. When the processor executes the computer program instructions, it implements the spatiotemporal AI prediction method for crop yield based on AI-Agent and semantic decoupling and multiple attention mechanisms as described in any one of claims 1-5. Specifically, this device is suitable for real-time agricultural decision-making, supports GPU acceleration (utilizing the CUDA 11.8 framework), and integrates a network interface to obtain online data updates. This implementation ensures the portability of the method, with training time not exceeding 24 hours under standard computing resources.

[0100] Example 4

[0101] This embodiment provides comparative experiments to verify the advantages of the proposed solution over the benchmark model. The benchmark model adopts a standard CNN-LSTM architecture (without AI-Agent and semantic decoupling) and uses the same US corn dataset (2008-2022, 119,235 samples). The training settings are consistent: AdamW optimizer, learning rate 0.001, early stopping threshold 10 epochs, batch size 32.

[0102] On the overall test set, our solution achieved R²=0.950 and RMSE=0.269 t / ha; the baseline LSTM model achieved R²=0.900 and RMSE=0.400 t / ha, demonstrating a significant performance improvement (R² increased by 0.05 and RMSE decreased by 33%).

[0103] For the subset of the drought year 2012 (extreme high temperature days > 20 days, approximately 8000 samples), our scheme has a MAE of 0.85 t / ha and an RMSE of 1.20 t / ha; the baseline model has a MAE of 1.00 t / ha and an RMSE of 1.35 t / ha. Our scheme reduces the RMSE by 0.15 t / ha (a decrease of 11%), thanks to the AI-Agent dynamically adjusting the soil water mode weights (an increase of 25%).

[0104] In a high-yield year (such as 2016), the proposed scheme achieved a PCC of 0.975, while the baseline model achieved a PCC of 0.900, representing an improvement of 0.075. This demonstrates the superiority of semantic decoupling in capturing nonlinear interactions.

[0105] Ablation experiments further confirmed that removing the multiple attention module reduced the R² of this scheme to 0.920; removing the AI-Agent coordination reduced the R² to 0.910, confirming the necessity of each component. This comparative experiment was based on 5-fold cross-validation with a standard deviation of <0.02, ensuring the reliability of the results.

[0106] The above are preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A spatiotemporal prediction method for crop yield based on AI-Agent and multiple attention mechanisms, characterized in that, Includes the following steps: Multi-source spatiotemporal data are acquired as input datasets; semantic decoupling is performed based on agricultural knowledge, and the input data is classified into vegetation, climate, and soil geography groups, and deep representations of each group are extracted; an AI-Agent system is deployed, and reinforcement learning is used to coordinate the semantically decoupled multi-attention network to achieve adaptive adjustment of modality weights and crop yield prediction; SHAP attribution analysis and parameter fine-tuning strategies are introduced to optimize the model.

2. The spatiotemporal AI prediction method for crop yield based on AI-Agent and multiple attention mechanisms according to claim 1, characterized in that, The steps for semantic decoupling based on agricultural knowledge include: grouping NDVI, peak NDVI, and integral NDVI into the vegetation group to reflect crop photosynthesis and growth rhythm; grouping variables such as accumulated temperature (GDD), extreme high-temperature days, and total precipitation into the climate group to quantify heat accumulation and water stress; grouping soil volumetric water content (swvl1, swvl2), crop proportion, and normalized latitude / longitude into the soil geography group to characterize root water and geographical location; and processing the feature maps of each group using a dedicated convolutional encoder to learn modality-specific representations, where the formula is: ,in, This represents the encoded feature map of the i-th group. This indicates a dedicated convolution operation. This represents the i-th set of input data; this encoding is used to handle the nonlinear characteristics of vegetation groups under drought conditions, ensuring improved prediction robustness.

3. The spatiotemporal AI prediction method for crop yield based on AI-Agent and multiple attention mechanisms according to claim 1, characterized in that, The steps for deploying the AI-Agent system to coordinate semantically decoupled multi-attention networks include: The input data is extracted using multiple scales and then standardized using Z-score, where the formula is: ,in, Represents the standardized value. Represents the original value. Let σ represent the mean and σ represent the standard deviation. The multi-scale fusion attention module calculates group weights, where the formula is: ,in, Let σ represent the weight of the k-th group, and let σ represent the Sigmoid function. Represents a 1×1 convolution. Represents the k-th feature map; The fused features are fed into a ConvLSTM to capture temporal dependencies, where the state update formula is: ,in, , , These are the hidden state, the output gate, and the cell state (the input gate and the forget gate are updated through a similar mechanism). The composite temporal attention module calculates the temporal weights, where the formula is: ,in, It is a spatiotemporal attention score. K is the query vector, and K is the key matrix (the asymptotic weights and learnable coefficients are optimized through training). AI-Agent dynamically adjusts attention parameters using the Q-learning algorithm and optimizes the fusion process based on the state similarity function to achieve adaptive fusion.

4. The spatiotemporal AI prediction method for crop yield based on AI-Agent and multiple attention mechanisms according to claim 1, characterized in that, The steps for introducing SHAP attribution analysis and parameter fine-tuning strategy include: Update parameters using the AdamW optimizer, with the loss function being the mean squared error (MSE): ,in, It is a predicted value. These are real labels; The SHAP value calculates the marginal contribution of each feature, where the formula is: .in, denoted by SHAP value, S represents the feature subset, N represents the complete feature set, and v represents the model prediction function; The tag refinement strategy adopted is as follows: Where y′ is the adjusted label and λ is the relaxation factor (range 0.4-0.6). This adjustment ensures consistency with extreme heat events and improves the accuracy of interpretation; AI-Agent drives closed-loop iteration to ensure physical consistency.

5. The spatiotemporal AI prediction method for crop yield based on AI-Agent and multiple attention mechanisms according to claim 1, characterized in that, It also includes a reward function R, used to evaluate the exploration efficiency of the agent architecture; the agent parameters of the optimal network architecture are: Where R is the reward metric function and θ is the surrogate parameter; the reward function R = α·accuracy + β·robustness, where accuracy is the coefficient of determination R², robustness is the stability of the root mean square error RMSE under extreme climatic conditions, and α and β are adjustable coefficients.

6. The spatiotemporal AI prediction method for crop yield based on AI-Agent and multiple attention mechanisms according to claim 1, characterized in that, The method is applicable to a variety of crop datasets, including but not limited to corn, soybean, and wheat datasets, to improve the generality of predictions.

7. A spatiotemporal AI prediction system for crop yield based on AI-Agent and multiple attention mechanisms, characterized in that, The system for implementing the method according to any one of claims 1-6 comprises: The data acquisition module is used to acquire multi-source spatiotemporal data and reference labels; The semantic decoupling module is used to classify input data and extract modal representations, and is connected to the data acquisition module to receive input data. The AI-Agent coordination module is used for adaptive fusion and temporal attention based on reinforcement learning-driven semantic decoupling multi-attention network. It is connected to the semantic decoupling module and is used to adjust the decoupling parameters based on reinforcement learning output. The optimization and interpretation module is used for SHAP attribution analysis and parameter fine-tuning, and connects with the AI-Agent coordination module to achieve iterative optimization.

8. An electronic device, characterized in that, The device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the spatiotemporal AI prediction method for crop yield based on AI-Agent and semantic decoupling and multiple attention mechanisms as described in any one of claims 1-6; the device is suitable for real-time agricultural decision-making and supports GPU acceleration.

Images