Smart agricultural planting decision optimization methods and systems that integrate multimodal data

Abstract

This application provides a smart agriculture planting decision optimization method and system that integrates multimodal data, including: using a convolutional neural network to filter noise and extract features based on a multi-source fusion data matrix, processing high-dimensional information to address temporal and spatial differences, and identifying key decision information fragments; obtaining association patterns from historical planting records from the key decision information fragments, optimizing a multi-objective function through a genetic algorithm to balance yield and cost constraints, and obtaining a preliminary set of planting schemes; updating dynamic environmental variables based on real-time weather forecast data according to the optimized scheme set, obtaining adjusted constraints, and determining the globally optimal planting time and irrigation amount; integrating newly collected data through a feedback loop mechanism using the final decision scheme, obtaining continuous monitoring indicators, and determining the scheme execution deviation; determining whether an alarm mechanism needs to be triggered based on the scheme execution deviation, updating the multimodal integration model parameters through preset rules, and obtaining an adaptively enhanced data processing framework.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a smart agricultural planting decision optimization method and system that integrates multimodal data. Background Technology

[0002] With global population growth and intensifying climate change, agricultural production requires precise management and efficient decision-making in a complex and ever-changing environment. Traditional agricultural management methods are struggling to meet the demands of integrating multi-source, heterogeneous data, and are unable to quickly respond to complex decision-making scenarios in dynamic environments. This makes extracting key information from massive and diverse agricultural data and optimizing planting decisions a crucial area for breakthroughs in smart agriculture.

[0003] Currently, agricultural data processing and decision-making methods have significant limitations. Many existing solutions rely on single or limited data source analysis, making it difficult to comprehensively capture the dynamic characteristics of crop growth and environmental changes. Specifically, information sources such as satellite imagery, soil sensor data, and meteorological data are independent and lack effective integration mechanisms, failing to comprehensively capture the spatiotemporal dynamics of crop growth and thus failing to accurately reflect the true state of crops in complex environments. For example, a single data source may ignore the interaction between soil moisture and weather patterns, causing planting plans to fail in the face of extreme weather such as drought or torrential rain. Simultaneously, existing decision-making methods are often based on static models, lacking the ability to respond to real-time environmental changes, such as sudden temperature fluctuations or outbreaks of pests and diseases, making it difficult to dynamically adjust parameters such as irrigation, fertilization, or planting density. In complex and ever-changing agricultural environments, this static decision-making model can easily lead to secondary problems such as yield fluctuations, increased costs, or excessive pesticide use. Furthermore, traditional methods struggle to balance the contradictions between yield, cost, and environmental impact when optimizing planting plans. For example, pursuing high yields may lead to excessive resource consumption or soil degradation, while environmentally friendly solutions may sacrifice economic benefits. How to construct a decision-making framework that can integrate multi-source heterogeneous data, respond to environmental changes in real time, and dynamically generate optimal planting plans under multi-objective constraints has become a core technical challenge that smart agriculture urgently needs to solve. This problem not only involves the accuracy and efficiency of data integration, but also concerns the model's adaptability to dynamic environments and the balance of multi-objective optimization, directly affecting the stability and sustainability of agricultural production.

[0004] In smart agriculture, the efficient extraction and utilization of data is a primary technical challenge. Agricultural data comes from a wide range of sources, including satellite imagery, sensor data streams, and historical planting records, which vary greatly in time, space, and format. The core challenge lies in accurately extracting valuable information fragments for decision-making from this heterogeneous data. For example, satellite imagery can reflect the spatial distribution of crop growth, but its massive data volume contains significant noise, and direct analysis may obscure crucial information. Furthermore, this challenge directly impacts the effectiveness of subsequent decision optimization. Due to the low quality of extracted information fragments, optimization algorithms are prone to deviating from the optimal solution when dealing with multi-objective constraints due to insufficient or inaccurate information. For instance, in selecting crop varieties or formulating irrigation plans, without accurate environmental information, the algorithm may be unable to effectively balance the relationship between yield and cost.

[0005] The complexity of multi-objective optimization is another key technical challenge. Agricultural planting decisions need to find a balance between multiple objectives, such as maximizing yield, minimizing cost, and minimizing environmental impact. However, current technologies are insufficient in global search capabilities, making it difficult to quickly find the optimal solution in high-dimensional constrained spaces. For example, a planting scheme may excel in increasing yield but lead to excessive water consumption or soil degradation, violating sustainable development goals. This conflict between multiple objectives makes the optimization process exceptionally complex, especially in dynamically changing agricultural environments where algorithms need to rapidly adjust schemes based on real-time data.

[0006] Therefore, how to accurately extract key decision-making information from massive amounts of multimodal data and, based on this information, achieve globally optimal planting plans under multi-objective constraints has become a key issue facing smart agriculture. For example, in a typical scenario, farmers need to determine the optimal planting time and irrigation amount for wheat based on current soil moisture, weather forecasts, and historical planting data. If the data extraction is inaccurate, the optimization algorithm may suggest planting too early, causing crop damage due to low temperatures; or insufficient irrigation, affecting yield. How to integrate multi-source data in a dynamic environment and achieve efficient and balanced decision-making has become a key issue for the development of smart agriculture. Summary of the Invention

[0007] To address the above problems, this invention provides a smart agriculture planting decision optimization method that integrates multimodal data, mainly including:

[0008] By fusing satellite imagery and sensor data streams using a pre-defined multimodal integration model, an initial heterogeneous dataset is obtained to capture dynamic characteristics of crop growth, resulting in a multi-source fusion data matrix. Based on this matrix, a convolutional neural network is used for noise filtering and feature extraction, processing high-dimensional information to address temporal and spatial differences and identifying key decision-making information segments. From these key segments, correlation patterns from historical planting records are extracted, and a genetic algorithm is used to optimize a multi-objective function to balance yield and cost constraints, resulting in a preliminary set of planting schemes. If environmental impact indicators in the preliminary planting scheme set exceed preset thresholds, a particle swarm optimization algorithm is used to adjust scheme parameters to minimize soil degradation risk, determining the optimized scheme set. Based on the optimized scheme set, dynamic environmental variables are updated using real-time weather forecast data to obtain adjusted constraints, determining the globally optimal planting time and irrigation amount. Simulated yield predictions are obtained from the globally optimal planting time and irrigation amount. If the predicted value is lower than the target threshold, the process is iterated by re-extracting information segments from the multi-source fusion data matrix to obtain the final decision scheme. Finally, a feedback loop mechanism is used to integrate newly collected data through the final decision scheme, obtaining continuous monitoring indicators and determining the scheme execution deviation. Based on the deviation from the scheme execution, it is determined whether an alarm mechanism needs to be triggered. The parameters of the multimodal integration model are updated through preset rules to obtain an adaptive and enhanced data processing framework.

[0009] This invention also provides a smart agricultural planting decision optimization system that integrates multimodal data, mainly comprising: a multimodal data integration module, used to obtain an initial heterogeneous dataset by fusing satellite imagery and sensor data streams with a preset multimodal integration model to capture dynamic characteristics of crop growth and obtain a multi-source fused data matrix; a feature extraction and denoising module, used to perform noise filtering and feature extraction based on the multi-source fused data matrix using a convolutional neural network, processing high-dimensional information for temporal and spatial differences, and determining key decision information fragments; an association pattern analysis module, used to obtain association patterns of historical planting records from key decision information fragments, and optimize a multi-objective function through a genetic algorithm to balance yield and cost constraints, obtaining a preliminary set of planting schemes; a multi-objective optimization module, used to adjust scheme parameters using a particle swarm optimization algorithm to minimize soil degradation risk if environmental impact indicators in the preliminary set of planting schemes exceed a preset threshold, and to determine the optimized scheme set; and an environmental constraint adjustment module, used to update dynamic environmental variables based on real-time weather forecast data according to the optimized scheme set, obtain adjusted constraints, and determine the globally optimal planting time and irrigation amount. The global decision-making module is used to obtain simulated yield predictions from the globally optimal planting time and irrigation amount. If the prediction is lower than the target threshold, it backtracks to the multi-source fusion data matrix to re-extract information fragments for iterative optimization and obtain the final decision scheme. The iterative optimization module is used to integrate newly collected data through the final decision scheme using a feedback loop mechanism to obtain continuous monitoring indicators and determine the scheme execution deviation. The dynamic monitoring and feedback module is used to determine whether an alarm mechanism needs to be triggered based on the scheme execution deviation, and to update the multimodal fusion model parameters through preset rules to obtain an adaptive data processing framework.

[0010] This invention discloses a smart agriculture planting decision optimization method that integrates multimodal data. Addressing the problems of isolated data and lack of dynamic adaptability in decision-making in traditional agriculture, this invention constructs a multimodal integrated model by fusing satellite imagery and sensor data streams, generating a multi-source fused data matrix to capture dynamic characteristics of crop growth. A convolutional neural network is used for noise filtering and feature extraction, processing high-dimensional spatiotemporal information and extracting key decision fragments. Combining historical planting records, a genetic algorithm is used to optimize a multi-objective function of yield and cost, generating a preliminary planting plan. If environmental impact exceeds limits, a particle swarm optimization algorithm is used to adjust parameters and minimize the risk of soil degradation. Real-time weather data updates are further incorporated as constraints to determine the optimal planting time and irrigation amount. If the predicted yield is insufficient, iterative optimization is performed backtracking to ultimately form a decision plan. This invention continuously monitors deviations through a feedback loop, dynamically updating model parameters and constructing an adaptive processing framework. Its core technical effect lies in achieving accurate, efficient, and sustainable planting decisions, significantly improving agricultural production efficiency and environmental friendliness. Attached Figure Description

[0011] Figure 1 This is a flowchart of the intelligent agricultural planting decision optimization method that integrates multimodal data according to the present invention.

[0012] Figure 2 This is a schematic diagram illustrating data fusion and CNN feature extraction in an embodiment of the present invention.

[0013] Figure 3 This is a schematic diagram of the two-stage collaborative optimization process of the present invention.

[0014] Figure 4 This is a schematic diagram of the feedback loop and model adaptive update mechanism of the present invention.

[0015] Figure 5 This is a schematic diagram of the intelligent agricultural planting decision optimization system that integrates multimodal data according to the present invention.

[0016] Figure 6 This is a comparison chart showing the application effects of the method in Embodiment 3 of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0018] Example 1

[0019] like Figure 1 The smart agriculture planting decision optimization method that integrates multimodal data in this embodiment may specifically include:

[0020] S101. By fusing satellite imagery and sensor data streams with a pre-defined multimodal integration model, an initial heterogeneous dataset is obtained to capture the dynamic characteristics of crop growth, resulting in a multi-source fusion data matrix.

[0021] Crop growth dynamics are acquired through satellite imagery. Real-time indicators are extracted from the sensor data stream and fused with a pre-defined multimodal integration model. This model takes satellite imagery and sensor data stream as input and outputs an integrated data sequence, resulting in a preliminary fused sequence. A convolutional neural network is then used on this preliminary fused sequence. This network takes heterogeneous datasets from the preliminary fused sequence as input and outputs extracted feature values. The heterogeneous datasets are processed to capture dynamic features of crop growth, resulting in an enhanced feature set. Based on this enhanced feature set, soil moisture auxiliary data is obtained, extracted from the sensor data stream. If the soil moisture is below a preset threshold, pest and disease monitoring indicators, obtained from satellite imagery, are integrated to determine a multi-source fused data matrix. Regional distribution patterns are extracted from the multi-source fused data matrix to determine dynamic changes in crop growth, resulting in a dynamic feature matrix corresponding to the multi-source fused data matrix.

[0022] like Figure 2 As shown, in one possible implementation, input information is first obtained from multiple data sources, including satellite imagery, soil sensor data, and temperature and rainfall data recorded by weather stations. Satellite imagery provides macroscopic information about the distribution of farmland, such as crop cover and vegetation index, while soil sensor data records microscopic indicators such as soil moisture and nutrient content, and weather station data reflects environmental variables such as temperature and rainfall. It should be noted that the data acquisition devices can be fixed sensors or mobile monitoring devices mounted on drones. For example, fixed soil sensors are typically buried in key locations in farmland to monitor soil moisture in real time, while drones acquire field images using high-resolution cameras. The data processing system cleans the data through preprocessing steps, removing outliers or noise, such as removing satellite imagery distortion caused by cloud cover.

[0023] For example, in one implementation, by fusing satellite imagery and sensor data streams into a pre-defined multimodal integration model, it is first necessary to obtain an initial heterogeneous dataset.

[0024] Specifically, the multimodal fusion model is a pre-defined framework for handling data sources from different modalities, such as large-scale spatial information from satellite imagery and real-time local measurements from sensor data streams. This model fuses data through feature alignment and joint representation learning to ensure the capture of dynamic characteristics of crop growth, such as changes in vegetation indices or water stress indicators. The initial heterogeneous dataset includes image data from satellites and numerical streams from ground sensors. These data may be inconsistent in terms of spatiotemporal dimensions, so the model needs to preprocess them to achieve synchronization. Furthermore, satellite imagery, typically from optical satellites or synthetic aperture radar, provides a macroscopic view of crop fields, such as monitoring changes in chlorophyll content in cornfields. Sensor data streams come from IoT devices deployed in farmland, such as soil moisture sensors and temperature probes. This data is acquired in a continuous stream, capturing real-time dynamics such as daily evapotranspiration rates. The fusion process is achieved through the encoder module of the multimodal fusion model, which converts image data into high-dimensional feature vectors and maps sensor data streams to the same feature space, thereby generating a unified representation.

[0025] It should be noted that acquiring heterogeneous datasets involves data cleaning and normalization steps to eliminate noise and scale differences, ensuring the accuracy of the fusion.

[0026] For example, in a rice cultivation scenario, the multimodal integration model can be pre-defined as a network structure based on an attention mechanism. The specific process includes: first, extracting spectral features from satellite imagery, such as the normalized difference vegetation index (NDVI); then, temporally aligning these features with the humidity sequence in the sensor data stream, and calculating correlation weights using a cross-attention layer; finally, fusing the data to obtain a multi-source fusion data matrix, where each row represents a comprehensive feature vector of a timestamp, used to capture the dynamics of rice growth, such as changes in water requirements during the heading stage. This fusion method can integrate multi-source information and improve the accuracy of crop detection.

[0027] Preferably, in another embodiment, for monitoring wheat fields, the model can adjust parameters to emphasize dynamic temperature characteristics. Satellite imagery captures seasonal growth patterns, while sensor data streams record microclimate changes. The fusion process is detailed below: the model's integration layer employs a gating mechanism to selectively weight the contributions of different modalities, for example, prioritizing sensor streams when satellite data is missing due to cloud cover. In this way, the resulting multi-source fused data matrix has a dimension of time series multiplied by the number of features, capturing dynamic features such as the thermal stress response of wheat ear development. This method supports precision agriculture decision-making in practical applications, such as optimizing irrigation timing.

[0028] Understandably, the pre-set parameters for multimodal ensemble models include a training phase that uses historical datasets to optimize parameters.

[0029] Specifically, the model can be based on convolutional neural networks (CNNs) for image processing and recurrent neural networks (RNNs) for data stream processing, and then output a data matrix through a fully connected layer after fusion. This implementation ensures the effective integration of heterogeneous data, providing a reliable dynamic feature foundation for crop growth monitoring. In one embodiment, the training of the CNN is based on historical planting data. For example, the data processing system collects satellite imagery, soil moisture, and meteorological data from the past five years, annotating the corresponding crop growth status (such as health, water shortage, pests and diseases, etc.), and trains the CNN through supervised learning, enabling it to extract features closely related to crop growth from the initial fused sequence. After training, the enhanced feature set output by the network is a high-dimensional vector set containing dynamic features of crop growth, such as the rate of change of soil moisture and the periodic fluctuations of vegetation indices. These features can reflect the health status of crops at different growth stages, providing a basis for subsequent decision-making.

[0030] During the fusion process, pixel values ​​from satellite imagery, time-series data from sensors, and time-point values ​​from meteorological data are linearly combined to generate a preliminary fusion sequence. This preliminary fusion sequence is a one-dimensional vector sequence arranged chronologically, with each time point containing comprehensive information from all modalities. In one possible implementation, the generation of the preliminary fusion sequence also considers the spatiotemporal characteristics of the data. For example, in a soybean planting scenario, the system downsamples the spatial resolution of the satellite imagery to a grid scale matching the sensor data, ensuring that the fusion sequence simultaneously reflects both the spatial distribution and temporal dynamics of the field. During the fusion process, the system may also introduce smoothing processing to reduce data fluctuations caused by sensor malfunctions or sudden weather changes. For example, after a heavy rain, soil moisture sensors may record abnormally high values; the data processing system will smooth these values ​​using a sliding window averaging method to ensure the stability of the preliminary fusion sequence.

[0031] S102. Based on the multi-source fusion data matrix, a convolutional neural network is used for noise filtering and feature extraction. High-dimensional information is processed to address temporal and spatial differences, and key decision information fragments are determined.

[0032] like Figure 2As shown, crop dynamics are captured by fusing satellite imagery with sensor data using a convolutional neural network. This network takes the multi-source fused data matrix as input and removes noise interference by scanning matrix elements layer by layer through convolutional and pooling layers, outputting a filtered matrix. For temporal difference analysis and spatial distribution processing of the filtered matrix, high-dimensional information optimization sequences are obtained by comparing time-series data and spatial coordinate data frame by frame. Feature values ​​are extracted from these sequences to determine an intermediate feature set. Based on this intermediate feature set, pest and disease early warning indicators are integrated. These indicators are generated by comparing soil moisture values ​​with a preset threshold. If the soil moisture is below the threshold, weather-aided data is fused, extracted from the sensor data stream, to obtain an extended fusion set. Crop growth anomaly patterns are identified from the extended fusion set by scanning areas in the fusion set with deviation values ​​exceeding a threshold. These patterns are processed to capture regional distribution changes and determine key decision-making information fragments.

[0033] For example, in one implementation, noise filtering and feature extraction are performed using a convolutional neural network based on a multi-source fusion data matrix. First, it's necessary to understand the structure of the multi-source fusion data matrix. This matrix integrates satellite imagery and sensor data, presented as a high-dimensional array, where rows represent timestamps and columns correspond to fused features such as vegetation indices and soil moisture. A convolutional neural network is applied to the matrix input, achieving noise filtering through multiple convolutional operations.

[0034] Specifically, the initial layers of the network use convolutional kernels to scan the matrix, capturing local patterns and suppressing random noise, such as filtering cloud interference signals in satellite imagery during crop monitoring. This filtering process depends on the kernel size setting, typically 3x3 or 5x5, to smooth the data without losing crucial dynamic information. Further, feature extraction is accomplished through subsequent convolutional and pooling layers, which progressively abstract higher-order features, such as edge variations in crop growth stages or water distribution patterns, thus providing a basis for decision-making. To handle high-dimensional information with spatiotemporal differences, convolutional neural networks need to incorporate spatiotemporal modules.

[0035] In one embodiment, the model structure of a convolutional neural network (CNN) can specifically be:

[0036] The noise filtering and feature extraction based on the multi-source fused data matrix are performed using a convolutional neural network. Specifically, a three-dimensional convolutional neural network (3D-CNN) can be used because it can effectively handle [T, H, W] data with a time dimension. An exemplary network structure is as follows:

[0037] 1. Input layer: Receives the [N, T, H, W] data tensor constructed above.

[0038] 2. Convolutional Block 1: Contains a 3D convolutional layer (e.g., 32 filters of size 3×3×3 with a stride of 1), followed by a ReLU activation function and a 3D max pooling layer (e.g., a pooling kernel of size 2×2×2). This layer is used to extract the spatiotemporal local features of the lower layer and perform preliminary noise reduction.

[0039] 3. Convolutional Block 2: Contains a 3D convolutional layer (e.g., 64 filters of size 3×3×3), followed by a ReLU activation function and a 3D max pooling layer. This layer is used to combine low-level features to form more complex feature representations.

[0040] 4. Flatten Layer: Unfolds the 3D feature map output by convolution block 2 into a 1D vector.

[0041] 5. Dense Layer: A fully connected layer with 128 neurons, using the ReLU activation function to perform non-linear combination of features.

[0042] 6. Output Layer: Outputs a feature vector, which is the "key decision information fragment". This vector contains a quantitative assessment of the current crop growth status (such as growth level, water stress index, nutrient abundance / deficiency).

[0043] Specifically, when processing multi-source fused data matrices, temporal differences refer to the dynamic changes of data at different timestamps, such as the daily growth rate of crops, while spatial differences involve locational variations within a field, such as the humidity gradient in edge areas. The network can employ a three-dimensional convolutional layer to handle these differences, which convolves simultaneously in both temporal and spatial dimensions to generate spatiotemporal feature maps.

[0044] For example, in a cornfield monitoring scenario, 3D convolution captures the spatial expansion of the seasonal growth cycle, ensuring that high-dimensional information such as multiple feature vectors is effectively compressed. Activation functions such as ReLU further highlight relevant patterns and reduce the influence of irrelevant dimensions. This approach enables the network to extract robust features from complex matrices, supporting subsequent analysis. Preferably...

[0045] In one embodiment, for rice cultivation, noise filtering in a convolutional neural network focuses on burst noise in the sensor data stream, such as outliers caused by equipment malfunctions. The encoder portion of the network first applies batch normalization layers to standardize the input, then filters noise through multiple convolutional blocks, each including convolution, activation, and pooling operations. Specifically, a fusion matrix is ​​input into the network, convolutional layers compute locally weighted sums to smooth noise, and pooling layers downsample to reduce dimensionality. To address temporal and spatial differences, a temporal attention mechanism is introduced, which computes weights between timestamps to emphasize key periods such as the growing season. Simultaneously, a spatial attention layer highlights hotspots in the field, such as areas with insufficient water. Through these steps, high-dimensional information is transformed into a low-dimensional representation, ultimately identifying key decision-making information fragments, such as identifying time windows requiring immediate irrigation. This method improves the accuracy of features in rice cultivation scenarios. Furthermore…

[0046] It should be noted that identifying key decision information fragments depends on the network's output layer. After feature extraction, the network uses fully connected layers or a softmax classifier to select decision-relevant fragments from the processed high-dimensional information.

[0047] Specifically, these segments are the highest-scoring subsets of the feature sequence, and are considered critical based on thresholds such as confidence levels exceeding 0.8.

[0048] For example, in wheat fields, a fragment might correspond to a peak period of thermal stress. After the network processes and fuses the matrix, it outputs a labeled sequence, marking decision points such as adjusting the timing of fertilizer application. This determination process ensures the relevance of the information.

[0049] In another implementation of orchard monitoring, exemplarily, the convolutional neural network is tailored to emphasize spatial variations. The fused data matrix includes canopy images and root zone sensor data. The network's convolutional layers first filter image noise, such as light variations, and then extract features, such as real density distributions. For temporal variations, a recurrent convolutional structure is used to process sequential data, ensuring the continuity of high-dimensional information across seasonal changes. Ultimately, key decision-making information fragments are identified as maturity signals for predicting harvest time. This implementation demonstrates the versatility of the technology, adapting to different crops within the same agricultural sector.

[0050] Understandably, the overall process of the aforementioned convolutional neural network begins with the input fusion matrix, proceeds through noise filtering and feature extraction, then to spatiotemporal difference processing and high-dimensional information management, culminating in the output decision segment. This modular design facilitates parameter optimization, such as adjusting the learning rate to adapt to specific crop dynamics. Through these steps, the technical solution can be effectively implemented in crop growth monitoring.

[0051] S103. Obtain the association patterns of historical planting records from key decision information fragments, optimize the multi-objective function through genetic algorithm to balance yield and cost constraints, and obtain a preliminary set of planting schemes.

[0052] like Figure 3 As shown, historical planting records are obtained from the decision information fragment. The correspondence between yield and cost is determined through association pattern extraction, resulting in an association pattern. This association pattern extraction is based on the matching calculation of yield and cost data in historical records. A multi-objective function is constructed for the association pattern, and a genetic algorithm is used to optimize the multi-objective function to balance yield constraints and cost control factors. If the optimization iteration process converges, a preliminary set of schemes is determined. The multi-objective function uses the correspondence in the association pattern as input, and the genetic algorithm solves the function value through population initialization and crossover mutation iterations. Key variables are analyzed based on the preliminary scheme set, and soil nutrient distribution data is obtained and integrated into the optimization iteration process to obtain adjusted scheme variables. Soil nutrient distribution data is expanded from the decision information fragment and adjusted in combination with key variables. The yield constraint is verified using the adjusted scheme variables, and a preliminary set of planting schemes is determined. Verification is based on a comparison calculation between the scheme variables and the yield constraint.

[0053] For example, in one implementation, obtaining correlation patterns of historical planting records from key decision information fragments first involves collecting and processing agricultural historical data.

[0054] Specifically, these key decision-making information fragments can include crop yield data from the past few years, soil type, fertilizer application rates, irrigation records, and weather conditions. Data mining techniques, such as the Apriori algorithm, are used to mine frequent itemsets and extract association patterns from these fragments.

[0055] For example, in the historical record of rice cultivation, if high yields are often associated with specific amounts of fertilizer and rainfall, patterns such as "yields increase when fertilizer application exceeds a certain threshold and rainfall is sufficient" can be identified. Furthermore, these association patterns can be used to construct the input for a multi-objective function.

[0056] It should be noted that multi-objective functions aim to balance maximizing output and minimizing costs. For example, the function can be defined as f1 = maximizing (expected output) and f2 = minimizing (total cost of seeds, fertilizer, and labor).

[0057] In one possible implementation, association patterns help initialize function parameters, such as setting yield prediction models based on historical patterns. A genetic algorithm is then used to optimize the multi-objective function to obtain a preliminary set of planting schemes. The genetic algorithm is a biomimetic optimization method that simulates the natural selection process, including population initialization, fitness evaluation, selection, crossover, and mutation steps.

[0058] Specifically, a planting scheme population is first initialized, with each scheme representing a crop configuration, such as seed density, fertilizer type, and planting time for maize. Each scheme's chromosome can be encoded as a binary or real-number vector, representing decision variables. During optimization, a fitness function evaluates the multi-objective values ​​of each scheme.

[0059] For example, in wheat cultivation, the projected yield of the calculated scheme is based on historical correlation patterns, such as "yield increases by 20% when soil pH is between 6 and 7 and nitrogen fertilizer is applied in appropriate amounts," while costs such as fertilizer expenditure are estimated. The selection operation uses a roulette wheel method to select a scheme with high fitness; the crossover operation exchanges some genes of the schemes, such as exchanging the fertilization strategies of the two schemes; mutation introduces random changes, such as slightly adjusting the planting density, to avoid local optima.

[0060] For example, in one embodiment, for vegetable cultivation such as tomatoes, a genetic algorithm iterates for 50 generations to generate a set of solutions in the Pareto front that balance high yield (e.g., 5000 kg per acre) and low cost (e.g., total input less than 2000 yuan). This method enables efficient resource allocation.

[0061] Preferably, in fruit tree planting scenarios, such as apple orchards, the association pattern extracts "sufficient sunlight and adequate watering are associated with high fruit quality" from historical records. After optimization by a genetic algorithm, multiple solutions are provided, such as pruning and fertilization combinations for different tree ages. Furthermore, this process can be extended to mixed crop planting, such as corn and soybean rotation. The algorithm considers cost constraints such as land rotation costs and generates a set of solutions to support decision-makers' choices.

[0062] In one embodiment, the multi-objective function of the genetic algorithm (GA) can be: the objective function of optimizing the multi-objective function through the genetic algorithm to balance output and cost constraints can be defined as:

[0063] Optimize F(X) = {max(f_yield(X)), min(f_cost(X))}, where: X is a decision variable vector (chromosome) encoding the planting scheme to be optimized, for example, X = [seed density, nitrogen fertilizer application rate, irrigation strategy parameters...]. f_yield(X) is a yield prediction model (e.g., a regression forest or neural network model trained on historical data) used to predict the expected yield under scheme X. f_cost(X) is a cost calculation function, f_cost(X) = Cost_seed(X) + Cost_fertilizer(X) + Cost_water(X) + ..., used to calculate the total cost of scheme X.

[0064] By solving the problem using a genetic algorithm (such as NSGA-II), a set of optimal solutions in the Pareto front is obtained, which is the initial planting scheme set.

[0065] S104. If the environmental impact indicators in the initial planting scheme set exceed the preset threshold, the particle swarm optimization algorithm is used to adjust the scheme parameters to minimize the risk of soil degradation, and the optimized scheme set is determined.

[0066] like Figure 3 As shown, environmental impact indicators are obtained from a preliminary set of schemes. These indicators are compared with preset thresholds to determine if they exceed the thresholds, resulting in a subset of schemes exceeding the threshold. For this subset, a particle swarm optimization (PSO) algorithm is used to adjust the scheme parameters. Risk factors are extracted from preset soil degradation risk data and integrated into the optimization process. The PSO algorithm takes the scheme parameters as input and outputs adjusted values, resulting in an adjusted parameter set. Soil degradation risk values ​​are calculated based on the adjusted parameter set. These risks are obtained through a weighted summation of risk factors. If the risk value is lower than a preset threshold, an optimized scheme is determined. Cycle factors are obtained from preset crop growth cycle data and integrated into the schemes, resulting in a fused scheme set. The environmental impact indicators are verified using this fused scheme set to determine the optimized scheme set.

[0067] For example, in one implementation, the initial set of planting schemes is derived from a preliminary optimization process, and these schemes include parameters such as crop type, planting density, and fertilization strategy. First, the environmental impact indicators of each scheme need to be evaluated, such as soil erosion rate, nutrient loss, and carbon footprint. These indicators are calculated using historical data and simulation models, with preset thresholds set based on agricultural standards, such as a soil erosion rate not exceeding 2 tons per hectare per year. If any scheme's indicator exceeds the threshold, an adjustment mechanism is triggered.

[0068] Specifically, when environmental impact indicators exceed thresholds, a particle swarm optimization (PSO) algorithm is used to adjust parameters. This algorithm simulates bird flock foraging behavior, using particles to represent potential solutions. Each particle has a position and velocity, with the position corresponding to solution parameters such as fertilizer application rate and irrigation frequency. After initializing the particle swarm, the algorithm iteratively updates the position of each particle, calculating the velocity based on the globally optimal and individually optimal positions, thereby moving towards the direction that minimizes soil degradation risk. Soil degradation risk is defined as a comprehensive indicator, including the reduction of soil organic matter and the degree of acidification. The risk value is quantified by a function equal to the erosion rate multiplied by the nutrient loss coefficient. Further, during the PSO optimization process, a fitness function evaluates the performance of each particle. The specific steps include: first, randomly initializing the particle swarm, with each particle encoded as a vector representing the solution parameters; then, calculating the fitness, i.e., the soil degradation risk value, with the goal of minimizing it while maintaining yield constraints; next, adjusting the particle position using a velocity update formula, for example, velocity equals the inertia weight multiplied by the previous velocity plus cognitive and social components; after multiple iterations, convergence to the optimized parameter set.

[0069] In one possible implementation, the objective function of particle swarm optimization (PSO) can be defined as follows: The objective function for adjusting the scheme parameters using the particle swarm optimization algorithm to minimize the risk of soil degradation can be defined as:

[0070] Minimize G(Y) subject to f_yield(Y) = Yield_threshold, where Y is the scheme parameter (particle) to be fine-tuned, for example, Y = [nitrogen fertilizer application rate, irrigation amount]. G(Y) is the soil degradation risk index function, for example, G(Y) = w1 * Salinization_Risk(Y) + w2 * Compaction_Risk(Y), where w1 and w2 are weighting coefficients. The risk function can be established based on parameters such as changes in soil EC value and irrigation intensity. The constraint f_yield(Y) = Yield_threshold ensures that when optimizing environmental indicators, the expected yield cannot be lower than an acceptable threshold (e.g., 90% of the target yield).

[0071] For example, in corn cultivation, if the initial fertilizer application causes nutrient loss exceeding a threshold, the algorithm can adjust to reduce the amount of nitrogen fertilizer while optimizing planting density to maintain yield.

[0072] Preferably, the algorithm is applied in vegetable planting scenarios, such as for the tomato program. When environmental indicators show a high risk, the particle swarm adjusts irrigation parameters through 50 iterations to reduce the risk of soil salinization.

[0073] It should be noted that the optimization process considers multiple dimensions, such as combining weather data to predict risk changes, to ensure the sustainability of the solution.

[0074] In one possible implementation, the optimized decision set involves verifying whether the risk value of each solution is below a threshold and calculating a comprehensive score.

[0075] For example, in rice cultivation, the criteria include minimizing risk and controlling costs, and a subset of schemes that meet the conditions are selected.

[0076] Understandably, this adjustment mechanism is extended to fruit tree planting, such as the apple orchard scheme. If excessive watering in the initial set causes soil quality degradation, the algorithm optimizes and generates a low-risk scheme, allowing farmers to choose.

[0077] For example, in mixed crop rotation, such as combining corn and soybeans, the particle swarm optimization algorithm adjusts the rotation cycle parameters to minimize long-term soil degradation and ensure the diversity of the scheme set. Furthermore, through the above steps, the initial scheme can be finely adjusted, improving the environmental friendliness of agricultural practices.

[0078] S105. Based on the optimized scheme set, update the dynamic environmental variables through real-time weather forecast data, obtain the adjusted constraints, and determine the globally optimal planting time and irrigation amount.

[0079] Temperature and rainfall indices are obtained from real-time weather forecast data, and dynamic environmental variables are updated to obtain an adjusted variable set. Constraint boundary values ​​are calculated for this adjusted variable set, whereby they are obtained by weighted averaging of temperature and rainfall indices. The applicability of the boundary values ​​is determined by comparing them against preset thresholds, thus defining the adjusted constraints. Crop type data, extracted from a preset crop database, is integrated based on these adjusted constraints. A genetic algorithm is then used to solve a multi-objective function, where the genetic algorithm takes the constraints and crop type data as input and outputs an optimized solution to obtain the globally optimal planting time and irrigation amount.

[0080] For example, in one implementation, an optimized set of scenarios serves as a basis for updating dynamic environmental variables by incorporating real-time weather forecast data. These variables include factors such as soil moisture, temperature changes, and rainfall probability, designed to reflect current environmental conditions.

[0081] Specifically, the system obtains hourly updated weather data, such as temperature forecasts and wind speed information, from a meteorological service interface, and then maps this data to environmental variables. If the forecast indicates impending rainfall, the soil moisture variable will increase accordingly, thus influencing subsequent decision-making. This update mechanism ensures that the scheme set can adapt to short-term environmental fluctuations, supporting sustainable agricultural practices. Furthermore, the updating process of dynamic environmental variables involves a data fusion step. First, historical weather data is collected as a baseline, and then compared with real-time forecasts to calculate the deviation value.

[0082] For example, in a corn growing scenario, if the forecast temperature is 5 degrees Celsius higher than the historical average, the system adjusts the temperature variable to reflect the potential increase in evaporation.

[0083] It should be noted that this fusion employs a weighted average method, with real-time data receiving higher weight to prioritize current information. This approach keeps environmental variables dynamic, avoiding the risks associated with static solutions. Based on the updated environmental variables, adjusted constraints are obtained. These constraints include window limits for planting time and upper limits for irrigation volume, aiming to prevent overwatering or delayed planting.

[0084] Specifically, constraint adjustments are implemented through a rule engine. For example, if soil moisture exceeds a threshold of 80%, the irrigation constraint is reduced to 50 cubic meters per hectare per day. Simultaneously, the crop growth cycle is considered to ensure the constraints do not affect yield targets. This adjustment mechanism is widely used in vegetable cultivation; for example, for tomato crops, the constraints dynamically set the watering interval based on humidity variables.

[0085] Preferably, a search algorithm is used to traverse possible combinations when determining the globally optimal planting time and irrigation amount. Global optimum refers to minimizing resource consumption while maximizing expected yield under constraints.

[0086] For example, the algorithm evaluates multiple time points, such as sowing one day earlier or later, and calculates the corresponding irrigation needs. Through iterative comparison, it selects the solution that satisfies all constraints. In the rice planting example, if the weather forecast indicates an extended drought period, the system adjusts the planting time to before the start of the rainy season and optimizes the irrigation amount to 200 cubic meters per hectare per week to ensure water use efficiency.

[0087] For example, this method is also applicable to fruit tree planting scenarios such as apple orchards. After real-time weather data updates environmental variables, constraints may limit irrigation to avoid soil waterlogging. The algorithm then determines the optimal planting time as mid-spring and calculates the irrigation amount based on tree age and soil type. This application demonstrates the versatility of the technology.

[0088] Understandably, the logic of this process forms a closed loop from data updates to optimization decisions. In the example of mixed crop rotation, such as soybean and corn, if the weather forecast predicts high temperatures, the updated environmental variables will adjust the constraints, leading to a one-week delay in planting and a 15% reduction in irrigation, thereby lowering the risk of soil degradation. Furthermore...

[0089] In one possible implementation, the system integrates sensor data to assist in weather forecast verification.

[0090] For example, field humidity sensors provide real-time calibration, ensuring the accuracy of variables. This enhances the reliability of constraint adjustments. In another embodiment, for greenhouse vegetable cultivation, dynamic updates focus on light variables. If the forecast of increased cloudy days increases, the irrigation upper limit is relaxed to compensate for reduced photosynthesis. Algorithms determine the optimal time and values, supporting fine-grained management.

[0091] It's important to note that the seamless integration of these steps ensures continuous optimization of the solution set. In practice, farmers can adjust their practices based on the outputs to improve overall environmental friendliness.

[0092] S106. Obtain the simulated yield prediction value from the globally optimal planting time and irrigation amount. If the prediction value is lower than the target threshold, backtrack to the multi-source fusion data matrix to re-extract information fragments to iteratively optimize the process and obtain the final decision scheme.

[0093] like Figure 4 As shown, initial simulation parameters are obtained by setting the optimal planting time and irrigation amount. A crop growth model is used to calculate the predicted yield value. The crop growth model is pre-established based on a crop growth cycle equation. The inputs are the initial simulation parameters, including soil type and meteorological conditions, and the output is the simulated crop yield value. The predicted yield value is compared with a threshold. If it is lower than the threshold, the data is backtracked from the multi-source data matrix to extract information fragments to determine the iterative input data. These information fragments include soil moisture fragments and crop adaptation attributes. The soil moisture fragments and crop adaptation attributes are fused based on the iterative input data to obtain optimized irrigation amount settings and planting time adjustments. This fusion uses a weighted average method to integrate the fragments and generate adjusted values. The optimized irrigation amount settings and planting time adjustments are repeatedly calculated to obtain a final decision scheme. The repeated calculations use the crop growth model to re-input the adjusted parameters to obtain new predicted yield values ​​until the threshold is met.

[0094] For example, in one implementation, starting with the globally optimal planting time and irrigation amount, it is first necessary to understand the process of obtaining the globally optimal planting time and irrigation amount. These parameters are derived from historical data and real-time monitoring through optimization algorithms.

[0095] For example, in a rice cultivation scenario, the globally optimal planting time might be determined based on a comprehensive analysis of soil moisture, temperature, and photoperiod, while irrigation amount takes into account the water requirements of the crop at different growth stages. Using these parameters, a simulated environment can be constructed to predict yield.

[0096] Specifically, the globally optimal planting time refers to the starting date that maximizes crop growth under given climatic conditions, while irrigation amount refers to the total water supply per unit area. These are calculated using genetic algorithms or particle swarm optimization methods to ensure the accuracy of the simulation. Furthermore, the calculation process for the simulated yield prediction involves the application of crop growth models.

[0097] For example, in corn cultivation, the model considers factors such as photosynthesis, transpiration, and nutrient absorption to calculate the expected yield. This process includes data input, model execution, and output generation. Internally, the model simulates daily growth increments and accumulates them to obtain the total yield prediction.

[0098] It should be noted that without these parameters, the predicted values ​​may be significantly off, so this step ensures the scientific nature of the decision-making process.

[0099] For example, if the simulated yield forecast is lower than a target threshold, a backtracking mechanism is triggered. The target threshold can be a user-defined minimum yield standard, such as 5000 kg per hectare. In the corn planting example, if the forecast is 4800 kg, which is lower than the threshold, the system backtracks to a multi-source fusion data matrix. A multi-source fusion data matrix is ​​a structured matrix that integrates multiple data sources, such as soil sensor data, satellite remote sensing images, weather forecasts, and historical yield records, forming a multi-dimensional array where each dimension represents a different data type, such as time series or spatial distribution. This matrix is ​​constructed through data cleaning, normalization, and fusion algorithms, such as using weighted averages or principal component analysis, to transform heterogeneous data into a unified format for easier subsequent extraction.

[0100] In one embodiment, the process of re-extracting information fragments after tracing back to the multi-source fused data matrix is ​​key to iterative optimization.

[0101] Specifically, information fragments refer to subsets of data in the matrix that are relevant to the current parameters. For example, extracting temperature sequences related to planting time or rainfall data fragments related to irrigation amount. This extraction can be achieved through clustering algorithms or attention mechanisms. First, the causes of low yields are identified, such as insufficient moisture, and then the corresponding fragments are selected from the matrix for analysis. For instance, in a wheat planting scenario, if the simulated yield is low, soil moisture fragments are extracted backtracking, their mismatch with irrigation amount is analyzed, and parameters are adjusted. This process iterates multiple times until the predicted value reaches a threshold. This mechanism effectively improves the robustness of decision-making and, in practical agricultural applications, can reduce resource waste and improve yield stability.

[0102] Preferably, the iterative optimization process includes multiple loop steps.

[0103] In one possible implementation, each iteration begins after re-extracting information fragments, and then uses optimization algorithms such as gradient descent to update the planting time and irrigation parameters.

[0104] For example, in vegetable cultivation, if the initial simulated yield falls below a threshold, weather data fragments are extracted to analyze potential drought risks. Then, irrigation is fine-tuned by increasing it by 10%, and the simulation is repeated. If the yield remains below the threshold, iteration continues, extracting soil nutrient fragments for adjustment. The principle behind this process is to gradually approach the optimal solution through a feedback loop, avoiding the local optima problem caused by one-time optimization. Specifically, the number of iterations can be capped, such as 10. In each round, the difference between the predicted value and the threshold is calculated as a loss function to guide parameter updates, thus forming a closed-loop optimization. In agricultural practice, this method is applicable to different crops, such as rice or corn, ensuring the adaptability of the decision-making scheme. Furthermore, the detailed construction process of the multi-source fusion data matrix deserves further explanation. The multi-source fusion data matrix is ​​a high-dimensional data structure designed to integrate heterogeneous information from sensor networks, satellite imagery, and databases. The specific process first collects raw data, such as soil pH values ​​obtained from ground sensors and rainfall data from weather stations, and then fuses them using time alignment and spatial interpolation methods.

[0105] For example, during the fusion phase, Krige interpolation is used to handle spatial data inhomogeneity, mapping all data onto a unified grid matrix, where each element represents the fused value at a specific spatiotemporal point. This matrix may be m rows by n columns, where m is the time step and n is the number of features, such as temperature and humidity. When re-extracting information fragments, the system queries sub-blocks of the matrix, for example, selecting the humidity column from the past 30 days as a fragment for analyzing irrigation demand. This design makes the backtracking process efficient, quickly locating the source of problems during iterations, thereby optimizing planting parameters. In rice paddy implementation, this matrix helps identify flood-risk segments, adjust planting times to avoid peak rainy seasons, and ultimately increase simulated yields above a threshold.

[0106] Understandably, obtaining the final decision-making solution is the output stage of the entire process. After iterative optimization, if the simulated yield prediction reaches or exceeds the target threshold, a decision-making solution is generated, including recommended planting time, irrigation amount, and potential risk warnings.

[0107] For example, in one corn planting embodiment, after three iterations, the predicted yield increased from an initial 4500 kg to 5200 kg, and the solution outputs "the optimal planting time is April 15th, and the weekly irrigation amount is 200 cubic meters per hectare." This solution, based on optimization results, ensures the feasibility of practical application. In another implementation, the application is extended to different crop scenarios, such as soybean planting. Similarly, the yield is simulated from globally optimal parameters; if it falls below a threshold, nitrogen-related segments in the matrix are extracted backtracked for iterative adjustments. This variant demonstrates the versatility of the technology, requiring no changes to the core framework.

[0108] Specifically, the technical effect of iterative optimization is to improve prediction accuracy. Through the backtracking mechanism, the system can adaptively adjust under data-driven conditions, reducing losses caused by uncertainty in the agricultural field.

[0109] For example, in greenhouse vegetable cultivation, this method can control yield prediction errors to within 5%, supporting precision agriculture decision-making.

[0110] S107. By integrating newly collected data through a feedback loop mechanism in the final decision-making scheme, continuous monitoring indicators are obtained, and deviations in scheme execution are determined.

[0111] like Figure 4 As shown, through the feedback loop mechanism, continuous monitoring indicators are obtained from the newly collected data. These continuous monitoring indicators are then compared numerically with the final decision-making scheme to determine a preliminary deviation value. The deviation magnitude is obtained through difference calculation in this numerical comparison. Based on the preliminary deviation value, the newly collected data results and the deviation magnitude are fused using the data integration process to obtain adjusted monitoring indicators. This fusion is achieved through a weighted summation method. Using the adjusted monitoring indicators, the implementation deviation of the scheme is determined, and deviation correction parameters are obtained. The deviation is determined using a threshold comparison method.

[0112] For example, in one implementation, a feedback loop mechanism is used to apply the final decision-making scheme to the agricultural planting process, such as implementing recommended planting times and irrigation amounts in rice paddies. This mechanism is designed to continuously integrate newly collected data to dynamically adjust the scheme.

[0113] Specifically, the feedback loop mechanism refers to a closed-loop process in which the data acquisition module is activated immediately after the plan is implemented to collect field information such as changes in soil moisture and crop growth status. This data is transmitted to the central processing system in real time through a wireless sensor network.

[0114] It's important to note that this mechanism works by cyclically comparing expected and actual data to identify potential problems, thereby maintaining the stability of the planting process. In the corn planting scenario, after the plan is implemented, temperature and rainfall data are collected daily to form a time-series dataset for subsequent integration. Furthermore, the process of integrating newly collected data involves a data fusion step. First, raw data, such as soil moisture content and leaf area index, is acquired from sensors and satellite remote sensing. This data undergoes cleaning and time synchronization processing before being integrated into the existing decision-making model.

[0115] For example, in a wheat planting embodiment, newly collected data includes crop images taken by drones. These images are used to extract feature values, such as plant height, through image recognition algorithms. These features are then combined with the expected irrigation amount parameters in the plan to form an updated data matrix. This integration employs a weighted fusion method, combining new data with historical data to ensure data comprehensiveness.

[0116] Preferably, the fusion process considers data weighting, for example, assigning higher weights to real-time sensor data to reflect current environmental changes, thereby providing an accurate basis for monitoring. In a vegetable greenhouse scenario, this step helps integrate light intensity data and adjust irrigation plans within the scheme.

[0117] For example, continuous monitoring indicators are obtained by processing and integrating data. The specific process includes extracting key indicators, such as crop growth rate and water use efficiency, from the integrated data. These indicators are calculated through statistical analysis, for example, by calculating daily growth increments as monitoring values.

[0118] In one possible implementation, for soybean cultivation, continuous monitoring indicators include yield forecasting indicators, with the expected harvest amount derived by simulating current data using a model. This acquisition process is based on time series analysis, where new data is input into the monitoring model to generate continuous curves showing how the indicators change over time.

[0119] Understandably, this step ensures the continuity of monitoring, avoids the loss of intermittent data, and supports timely intervention in practical applications.

[0120] In one embodiment, determining the deviation from the plan's execution involves comparing the monitoring indicators with the target values ​​in the decision-making plan.

[0121] Specifically, deviation calculation is achieved through difference analysis. For example, the deviation is obtained by subtracting the actual soil moisture index from the recommended irrigation threshold. If the deviation exceeds a preset range, such as a moisture deviation greater than 10%, a cyclical adjustment is triggered. Furthermore, in rice cultivation, this determination process analyzes multiple indicators, such as combining yield and moisture indicators, to calculate a comprehensive deviation score, and uses a threshold to judge the implementation effect. The advantage of this method is that it quantifies the gap between the plan and reality, supporting subsequent optimization. In cornfield implementation, after the deviation is determined, the system generates a report listing the specific sources of deviation, such as growth delays caused by insufficient irrigation, thereby guiding farmers to make adjustments.

[0122] Preferably, the feedback loop mechanism, after determining the deviation, loops back to the integration step to form an iterative process.

[0123] For example, in wheat cultivation, if deviations indicate a lower-than-expected yield, a feedback loop mechanism re-collects and integrates data, updating the program parameters. This loop repeats until the deviation is minimized, ensuring the accuracy of program execution. In another implementation, for vegetable cultivation, the loop mechanism integrates weather forecast data, identifies deviations caused by weather changes, and adjusts irrigation amounts to match actual needs.

[0124] S108. Determine whether an alarm mechanism needs to be triggered based on the deviation of the scheme execution, and update the parameters of the multimodal integrated model through preset rules to obtain an adaptive data processing framework.

[0125] like Figure 4 As shown, deviation correction parameters are obtained from the execution monitoring indicators through deviation value comparison. These parameters are calculated by subtracting a benchmark value from the monitoring indicators using a difference calculation method. A threshold comparison is performed between the deviation correction parameters and a preset threshold. If the deviation correction parameter exceeds the preset threshold, an alarm mechanism is triggered, resulting in an alarm trigger judgment result. Based on the alarm trigger judgment result, the parameters in the model integration optimization are updated using preset rules in the rule parameter update. These parameters include weight coefficients adjusted using a weighted average method. Data from the multimodal fusion mechanism is integrated to determine the updated model parameters. Using the updated model parameters, an adaptive adjustment process and data framework enhancement are integrated. Integration is achieved by combining the data from the adjustment process in a layer-by-layer stacking manner to obtain enhanced framework components. Parameter verification is performed on these framework components by comparing the consistency between components, resulting in a verified framework structure. Based on the verified framework structure, deviation correction and rule parameter updates are applied, processed using an iterative correction method. The execution monitoring indicators are then integrated to determine the final adaptively enhanced data processing framework.

[0126] For example, in one implementation, the process of determining whether to trigger an alarm mechanism based on deviations in the execution of the plan first assesses the degree of deviation by comparing the actual monitoring indicators with preset target values.

[0127] Specifically, this judgment is based on deviation calculations. For example, in rice cultivation, if the actual soil moisture differs from the recommended value by more than a preset threshold, such as 8%, the system automatically assesses whether an alarm condition has been met. The principle behind this mechanism is to promptly identify potential risks and ensure continuous adjustments throughout the planting process.

[0128] The paper describes an adaptive data processing framework that updates the parameters of a multimodal ensemble model using preset rules. These preset rules are "condition-action" pairs obtained through expert knowledge or historical data mining. Three specific rule examples are provided below:

[0129] Rule Example 1: Dynamically Adjusting Input Weights Based on the Source of Bias

[0130] IF: Through the feedback loop mechanism adopted by the final decision scheme, it is detected that the actual crop growth of a certain plot is consistently lower than expected, and the deviation analysis module attributes the deviation to "persistent water stress" (for example, the soil moisture sensor reading of the plot is lower than 80% of the model's recommended value for 7 consecutive days).

[0131] THEN: The system determines that the current multimodal ensemble model is insufficiently sensitive to soil moisture. An update action is performed: at the spatial location corresponding to this plot, the input weight w_moisture of the soil moisture sensor data features is increased. The update formula can be: w_moisture_new = w_moisture_old * (1 + α), where α is a learning rate, for example, 0.05. This makes the model pay more attention to changes in soil moisture at this plot in future decisions.

[0132] Rule Example 2: Model Structure Modification Based on External Unexpected Events

[0133] IF: Monitoring data (e.g., drone hyperspectral image analysis) show that new pests and diseases have broken out in farmland that were not covered by historical data, leading to significant deviations in yield forecasts.

[0134] THEN: The system determines that the existing features cannot capture this anomaly. Execute the update action: After feature extraction by the CNN, dynamically load a pre-trained dedicated identification sub-network for this new pest and disease, and integrate its output "disease probability" as a new feature dimension into the "key decision information fragment".

[0135] Rule Example 3: Adjusting the Optimization Objective Function Based on Decision Outcomes

[0136] IF: At the end of a planting cycle, the final actual yield is significantly higher than expected, but the consumption of soil nutrients (such as available nitrogen) also far exceeds the model prediction, leading to a decline in soil fertility.

[0137] THEN: The system determines that the penalty weight for cost (especially resource consumption reflected by fertilizer cost) in the multi-objective function of the genetic algorithm is too low. An update action is performed: In the optimization of the next planting cycle, the weight coefficient β of the cost function f_cost(X) in the multi-objective optimization evaluation is automatically increased, for example, β_new = β_old * 1.1, to guide the algorithm to generate a solution that pays more attention to resource sustainability.

[0138] It's important to note that the alarm mechanism is designed with a multi-level response, including a primary alarm that notifies farmers via SMS and a secondary alarm that activates automatic irrigation equipment. In the corn planting scenario, deviation assessment is combined with temperature indicators; if the growth rate deviation exceeds the expected 5%, an alarm is triggered to prevent yield loss. This assessment process emphasizes real-time performance, using embedded algorithms to process the data stream and form a reliable decision-making basis. Further, updating the parameters of the multimodal fusion model through preset rules is a crucial step. A multimodal fusion model refers to a system that processes multiple data types, such as fusing sensor data like soil moisture, satellite imagery like crop cover, and meteorological data like rainfall forecasts. These data sources are integrated into a unified representation by the model for agricultural decision-making. The update process involves preset rules, such as adjusting model weights based on the magnitude of the deviation. For example, if the deviation originates from changes in humidity, the rule increases the parameter weight of the sensor data to 0.6.

[0139] Specifically, in the wheat planting example, the preset rules include a gradient descent method to progressively optimize model parameters to minimize prediction errors. This update ensures that the model is more sensitive to environmental changes, such as automatically prioritizing the processing of water-related parameters during drought periods.

[0140] Understandably, the application of this rule forms an iterative cycle, supporting the continuous learning of the model.

[0141] For example, the adaptively enhanced data processing framework is the overall structure formed by the above updates. This framework includes a data input layer, an integration layer, and an output layer, where the integration layer processes multimodal data using updated model parameters. In the vegetable greenhouse scenario, the framework's adaptability is reflected in dynamically adjusting parameters to cope with changes in light intensity; for example, the framework automatically incorporates newly acquired temperature data to generate optimized irrigation recommendations.

[0142] In one possible implementation, this framework is extended to soybean cultivation. After an alarm mechanism detects a deviation, rules update the model parameters, such as adjusting image feature parameters from initial values ​​to reflect the current crop state. This enhances the framework's adaptability to seasonal changes; for example, it increases the weighting of rainfall data during the rainy season, resulting in more accurate yield forecasts. The business objective of this process is to reduce resource waste and provide adjustable decision support through the framework. In another embodiment, for deviation detection in rice paddies, if an alarm is triggered, the system updates the model using rules such as threshold adaptive algorithms, for example, changing parameters from fixed values ​​to dynamic values ​​based on historical deviations. The resulting framework integrates these updates and outputs enhanced monitoring reports, supporting real-time farmer intervention. This framework's structure emphasizes modularity, facilitating expansion to similar crop scenarios. Furthermore, the parameter update rules for the multimodal integrated model can include conditional branches; for example, in corn cultivation, if the deviation type is nutrient deficiency, the rules prioritize updating the parameters of relevant chemical indicators. This update process is explained in detail as follows: first, deviation data is collected; then, rules are applied to calculate new parameter values, such as multiplying weights by a deviation factor; finally, the update effect is verified. In this way, the framework gains adaptability to handle diverse agricultural environments.

[0143] For example, in wheat field implementation, the alarm mechanism is triggered based on a comprehensive deviation score; it is activated if the score exceeds 20. The adaptability of the updated framework is reflected in the generation of personalized planting plans, such as adjusting fertilizer parameters to match actual soil conditions. This framework's business process, from deviation input to output framework, ensures the coherence of the entire chain.

[0144] Understandably, this technical solution is versatile in vegetable cultivation. By processing multimodal data such as humidity and light intensity through a framework, it updates parameters and improves decision-making accuracy. In practice, the framework supports multiple iterations, gradually reducing the impact of biases.

[0145] In one possible implementation, the generation of the decision-making framework combines a balance of long-term and short-term data. For example, the system comprehensively analyzes historical data from the past three years and recent real-time data to ensure that the decision-making framework is both stable and responsive to changes. In a corn planting case, the updated decision-making framework recommends a dynamically adjusted irrigation scheme, reducing irrigation during the rainy season and increasing irrigation during the dry season, thereby improving crop adaptability and yield stability.

[0146] Example 2

[0147] like Figure 5As shown, this embodiment of the invention provides a smart agricultural planting decision optimization system that integrates multimodal data, mainly including: a multimodal data integration module, used to obtain an initial heterogeneous dataset by fusing satellite imagery and sensor data streams with a preset multimodal integration model to capture the dynamic characteristics of crop growth and obtain a multi-source fused data matrix; a feature extraction and noise reduction module, used to perform noise filtering and feature extraction using a convolutional neural network based on the multi-source fused data matrix, processing high-dimensional information for temporal and spatial differences, and determining key decision information fragments; an association pattern analysis module, used to obtain association patterns of historical planting records from key decision information fragments, optimize a multi-objective function using a genetic algorithm to balance yield and cost constraints, and obtain a preliminary planting scheme set; a multi-objective optimization module, used to adjust scheme parameters using a particle swarm optimization algorithm to minimize soil degradation risk if environmental impact indicators in the preliminary planting scheme set exceed a preset threshold, and determine the optimized scheme set; and an environmental constraint adjustment module, used to update dynamic environmental variables based on real-time weather forecast data according to the optimized scheme set, obtain adjusted constraints, and determine the globally optimal planting time and irrigation amount. The global decision-making module is used to obtain simulated yield predictions from the globally optimal planting time and irrigation amount. If the prediction is lower than the target threshold, it backtracks to the multi-source fusion data matrix to re-extract information fragments for iterative optimization and obtain the final decision scheme. The iterative optimization module is used to integrate newly collected data through the final decision scheme using a feedback loop mechanism to obtain continuous monitoring indicators and determine the scheme execution deviation. The dynamic monitoring and feedback module is used to determine whether an alarm mechanism needs to be triggered based on the scheme execution deviation, and to update the multimodal fusion model parameters through preset rules to obtain an adaptive data processing framework.

[0148] Example 3

[0149] This embodiment, taking into account the specific scenario of water scarcity and high wheat planting density in North China, details the implementation process of this application.

[0150] S301 Multimodal Data Fusion and Matrix Construction

[0151] (a) Data Acquisition and Preprocessing

[0152] In this embodiment, the multi-source data includes at least:

[0153] 1. Satellite remote sensing data: For example, multispectral images acquired through the "Gaofen-2" satellite, with a temporal resolution of 5 days and a spatial resolution of 10 meters. Atmospheric correction, radiometric calibration, and cloud detection and removal are then performed on the raw images.

[0154] 2. Sensor Data Stream: Deploy IoT soil monitoring stations in a grid pattern (e.g., one per hectare) within the farmland to collect parameters including soil volumetric water content, soil temperature, and electrical conductivity (EC value) at a depth of 0-40 cm, with a time resolution of 1 hour. Perform outlier removal (e.g., using the 3σ criterion) and smoothing filtering (e.g., using a moving average method) on the sensor data to eliminate noise.

[0155] 3. Meteorological data: Daily meteorological data from farm weather stations or public meteorological service interfaces, including maximum / minimum temperature, rainfall, sunshine hours, and wind speed.

[0156] 4. Historical planting data: A historical database of the past 3-5 years, recording crop varieties, sowing dates, irrigation / fertilization records, pest and disease occurrences, and final yield data by plot.

[0157] (ii) Spatiotemporal alignment and normalization: To fuse heterogeneous data, it is necessary to unify them onto the same spatiotemporal reference.

[0158] 1. Spatial Alignment: A 10m x 10m grid is used as the basic spatial unit, consistent with the resolution of satellite imagery. Sensor data and plot-level historical data are mapped onto each grid using spatial interpolation methods (such as inverse distance weighted interpolation or kriging interpolation).

[0159] 2. Time Alignment: Using "day" as the basic time unit. Hourly sensor data is aggregated into daily averages / daily cumulative values; satellite data updated every 5 days is filled into daily data using linear or spline interpolation methods.

[0160] 3. Data normalization: To eliminate the influence of different feature dimensions, the min-max scaling method is used for all feature data to scale the values ​​to the interval.

[0161] (III) Feature Engineering and Matrix Construction

[0162] Derived features are calculated from preprocessed data, such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Leaf Area Index (LAI) based on satellite imagery.

[0163] Finally, a four-dimensional multi-source fusion data tensor is constructed as the data matrix. The tensor has dimensions [N, T, H, W], where:

[0164] N is the number of feature channels. For example, N=10, and the channels correspond to 10 features such as NDVI, soil moisture, daily average temperature, and rainfall.

[0165] T is the length of the time series. For example, T=30 represents data from the past 30 days.

[0166] H and W are the height and width of the farmland area after it has been rasterized.

[0167] This structured data matrix completely preserves the spatiotemporal dynamic information of the crop growth environment.

[0168] Specific scenario: A 100-hectare wheat field on a farm in Hebei Province aims to save irrigation water to the greatest extent possible while ensuring a yield of no less than 600 kg / mu.

[0169] Data sources include:

[0170] Satellite data: Weekly Gaofen-2 satellite imagery, from which Normalized Difference Vegetation Index (NDVI) and Leaf Area Index (LAI) are extracted.

[0171] Sensor data: 50 IoT soil moisture sensors deployed in the field upload the volumetric water content of the 0-20cm and 20-40cm soil layers every hour.

[0172] Meteorological data: Daily temperature, rainfall, wind speed, and sunshine hours provided by the farm's self-built weather station.

[0173] Historical data: planting density, irrigation time and amount, fertilization plan and final yield per acre for the plot over the past five years.

[0174] The system aligns the above data in time and space, using "day" as the time unit and a 100x100 meter grid as the spatial unit, to construct a multi-dimensional data matrix.

[0175] S302 Feature Extraction Based on CNN

[0176] The data matrix above was processed using a CNN model containing 3D convolutional layers. The 3D convolutional kernels can simultaneously capture features in both the temporal (crop growth cycle) and spatial (plot differences) dimensions. The model output, "key decision information fragments," consists of feature maps labeled with the current wheat growth rating (excellent / good / medium / poor) and potential drought stress level (high / medium / low) for each grid cell.

[0177] S303 GA Optimized Preliminary Planting Plan

[0178] Define a multi-objective optimization function: F(X) = {max(predicted yield per acre), min(total irrigation amount)}. The decision variables X include: sowing time, irrigation amount during the greening stage, and irrigation amount during the jointing stage. The genetic algorithm iterates through 200 generations to generate a Pareto optimal solution set, for example:

[0179] Option A: The predicted yield is 650 kg per mu, and the total irrigation volume is 200 cubic meters per mu.

[0180] Option B: The predicted yield is 620 kg per mu, and the total irrigation volume is 160 cubic meters per mu.

[0181] Option C: The predicted yield is 600 kg per mu, and the total irrigation volume is 140 cubic meters per mu.

[0182] S304 PSO Optimization Environmental Impact

[0183] In this scenario, the environmental impact indicator is set as "deep soil water deficit risk." The system determines that scheme A will lead to excessive consumption of deep soil water, exceeding the preset -5% threshold. At this point, the system locks the yield target at 630 kg / mu and activates the PSO algorithm to fine-tune the irrigation timing of scheme A with the goal of "minimizing deep soil water deficit," ultimately obtaining the optimized scheme A': predicted yield of 635 kg / mu, total irrigation volume of 185 cubic meters / mu, and deep water deficit risk within the threshold. The system selects scheme A', schemes B, and C as the initial planting scheme set.

[0184] S305 Dynamic Decision Making Based on Real-Time Weather Forecasts

[0185] Before the wheat jointing stage, the system obtains the weather forecast for the next 7 days, indicating that there will be one effective rainfall. Based on this, the system updates the dynamic environmental variables, recalculates, and determines the globally optimal solution as: adopting the irrigation strategy of scheme C, and postponing the jointing irrigation time by 3 days, thereby utilizing natural precipitation and further conserving water.

[0186] S306 Production Forecasting and Iterative Backtracking

[0187] Based on the finalized plan, the crop growth model predicts a final yield of 605 kg per mu (approximately 0.067 hectares), which is higher than the target threshold of 600 kg per mu, and the plan is approved. If the predicted value is lower than 600 kg per mu, the system will backtrack to S302, adjust the attention weights of the CNN, pay more attention to features such as historical change rate, re-extract information segments, and perform optimization iterations.

[0188] S307-S308 Execution Monitoring and Adaptive Updates

[0189] During implementation, the system used a drone equipped with a thermal infrared camera to patrol the fields weekly. An anomaly in temperature was detected in the southwest of the plot. Combined with soil sensor data, it was determined that irrigation sprinklers in that area were clogged, causing the deviation. The system immediately sent an alert to the farmer's mobile app, along with the problem location. After the issue was resolved, the data for this deviation event (deviation type, cause, and impact) was recorded. Through preset rules, the system slightly increased the weight of the drone's thermal infrared data in that area in the next decision, achieving adaptive enhancement of the data processing framework.

[0190] like Figure 6As shown, compared with the old method that relies on traditional experience and fixed irrigation plans, by adopting the method of Embodiment 3 of this application, the farm can save an average of 45 cubic meters of irrigation water per mu while maintaining the same yield (an average yield of 610 kg per mu), thus improving water-saving efficiency by 25%. At the same time, it avoids the risk of yield reduction caused by local irrigation problems, demonstrating significant technical advantages.

[0191] In this invention, it should be understood that if the implemented module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0192] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort. The above are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the protection scope of the present invention.

Claims

1. A smart agriculture planting decision optimization method integrating multimodal data, characterized in that, The method includes: By fusing satellite imagery and sensor data streams using a pre-defined multimodal integration model, an initial heterogeneous dataset is obtained to capture the dynamic characteristics of crop growth, resulting in a multi-source fusion data matrix. Based on this matrix, a convolutional neural network is used for noise filtering and feature extraction, processing high-dimensional information to address temporal and spatial differences and identifying key decision-making information fragments. From these fragments, correlation patterns from historical planting records are obtained, and a genetic algorithm is used to optimize a multi-objective function to balance yield and cost constraints, resulting in a preliminary set of planting schemes. Environmental impact indicators are then obtained from this preliminary set, and comparisons with pre-defined thresholds determine whether they exceed the thresholds, yielding a subset of schemes exceeding the thresholds. For the subset of schemes exceeding the threshold, a particle swarm optimization algorithm is used to adjust the scheme parameters. Risk factors are extracted from preset soil degradation risk data and integrated into the optimization process. The particle swarm optimization algorithm takes the scheme parameters as input and outputs adjusted values ​​to obtain an adjusted parameter set. Soil degradation risk values ​​are calculated based on the adjusted parameter set, where the soil degradation risk value is obtained by weighted summation of risk factors. If the risk value is lower than a preset threshold, an optimized scheme is determined. Cycle factors are obtained from preset crop growth cycle data and integrated into the scheme to obtain a fused scheme set. Based on the optimized scheme set, dynamic environmental variables are updated using real-time weather forecast data to obtain adjusted constraints. The process involves determining the optimal planting time and irrigation amount globally; obtaining initial simulation parameters based on these optimal planting time and irrigation amount settings; calculating the predicted yield using a crop growth model (pre-established based on a crop growth cycle equation), where the inputs are the initial simulation parameters including soil type and meteorological conditions, and the output is the simulated crop yield value; comparing the predicted yield value with a threshold, if it falls below the threshold, backtracking from the multi-source data matrix to extract information fragments to determine iterative input data, including soil moisture fragments and crop adaptation attributes; fusing the soil moisture fragments and crop adaptation attributes based on the iterative input data to obtain optimized irrigation amount settings and planting time adjustments, using a weighted average method to integrate the fragments and generate adjusted values; repeatedly calculating the optimized irrigation amount settings and planting time adjustments to determine the final decision scheme, using the crop growth model to re-input the adjusted parameters to obtain new predicted yield values ​​until the threshold is met; integrating newly collected data using a feedback loop mechanism based on the final decision scheme to obtain continuous monitoring indicators and determine the scheme execution deviation; determining whether an alarm mechanism needs to be triggered based on the scheme execution deviation, updating the multimodal integration model parameters according to preset rules to obtain an adaptively enhanced data processing framework. Specifically, the noise filtering and feature extraction based on the multi-source fusion data matrix using a convolutional neural network are employed. A three-dimensional convolutional neural network (3D-CNN) is used to construct a four-dimensional multi-source fusion data tensor as the data matrix. The dimensions of this tensor are [N, T, H, W], where N is the number of feature channels, T is the time series length, and H and W are the height and width of the farmland area after rasterization.

2. The intelligent agricultural planting decision optimization method integrating multimodal data according to claim 1, characterized in that, The process involves fusing satellite imagery and sensor data streams using a pre-defined multimodal integration model to obtain an initial heterogeneous dataset for capturing dynamic characteristics of crop growth, resulting in a multi-source fused data matrix, including: Crop growth dynamics are collected by satellite imagery, real-time indicators are extracted from sensor data streams, and a pre-set multimodal integration model is fused. The model takes satellite imagery and sensor data streams as inputs and outputs an integrated data sequence to obtain a preliminary fusion sequence. For the preliminary fusion sequence, a convolutional neural network is used. The input of this network is the heterogeneous dataset in the preliminary fusion sequence, and the output is the extracted feature values. The heterogeneous dataset is processed to capture the dynamic features of crop growth and obtain an enhanced feature set. Based on the enhanced feature set, soil moisture auxiliary data is obtained, which is extracted from the sensor data stream. If the soil moisture is lower than a preset threshold, pest and disease monitoring indicators are integrated, which are obtained from satellite imagery, to determine a multi-source fusion data matrix. Regional distribution patterns are extracted from the multi-source fusion data matrix to determine the dynamic changes in crop growth, thereby obtaining the dynamic feature matrix corresponding to the multi-source fusion data matrix.

3. The intelligent agricultural planting decision optimization method integrating multimodal data according to claim 1, characterized in that, The process involves using a convolutional neural network to filter noise and extract features from a multi-source fused data matrix, processing high-dimensional information to address temporal and spatial differences, and identifying key decision information fragments, including: Crop dynamics are captured by fusing satellite imagery with sensor data. A convolutional neural network is used, which takes the multi-source fused data matrix as input and removes noise interference by scanning matrix elements layer by layer through convolutional and pooling layers. The output is a filtered matrix. For the temporal difference analysis and spatial distribution processing of the filtered matrix, the high-dimensional information optimization sequence is obtained by comparing the time series data and spatial coordinate data in the matrix frame by frame, and feature values ​​are extracted from the sequence to determine the intermediate feature set. The pest and disease early warning index is integrated based on the intermediate feature set. This index is generated by comparing the soil moisture value with a preset threshold. If the soil moisture is lower than the threshold, weather auxiliary data is fused. This data is extracted from the sensor data stream to obtain an extended fusion set. The extended fusion set is used to determine abnormal crop growth patterns. These patterns are identified by scanning regions in the fusion set where the deviation value is higher than a threshold. The patterns are then processed to capture changes in regional distribution and determine key decision information fragments.

4. The intelligent agricultural planting decision optimization method integrating multimodal data according to claim 1, characterized in that, The process involves obtaining association patterns from historical planting records based on key decision information fragments, optimizing a multi-objective function using a genetic algorithm to balance yield and cost constraints, and obtaining a preliminary set of planting schemes, including: Historical planting records are obtained from the decision information fragments. The correspondence between yield and cost is determined by the correlation pattern extraction to obtain the correlation pattern. The correlation pattern extraction is based on the matching calculation of the correspondence between yield data and cost data in the historical records. A multi-objective function is constructed for the aforementioned association pattern. A genetic algorithm is used to optimize the multi-objective function to balance output constraints and cost control factors. If the optimization iteration process converges, a preliminary set of solutions is determined. The multi-objective function takes the corresponding relationship in the association pattern as input, and the genetic algorithm solves the function value through population initialization and crossover mutation iteration. Based on the preliminary set of schemes, key variables are analyzed, soil nutrient distribution data is obtained and integrated into the optimization iteration process to obtain adjusted scheme variables. Among them, soil nutrient distribution data is obtained by expanding the decision information fragments and combined with key variables for adjustment. The yield constraints are verified by the adjusted scheme variables to determine the initial set of planting schemes, wherein the verification is based on the comparison calculation between the scheme variables and the yield constraints.

5. The intelligent agricultural planting decision optimization method integrating multimodal data according to claim 1, characterized in that, The step of updating dynamic environmental variables based on the optimized scheme set using real-time weather forecast data, obtaining adjusted constraints, and determining the globally optimal planting time and irrigation amount includes: Temperature and rainfall indices are obtained from real-time weather forecast data, dynamic environmental variables are updated, and an adjusted variable set is obtained. For the adjusted variable set, the constraint boundary values ​​are calculated, wherein the constraint boundary values ​​are obtained by weighted average of temperature index and rainfall index, and the applicability of the boundary is judged by comparison with a preset threshold to determine the adjusted constraint conditions; The crop type data is integrated according to the adjusted constraints, wherein the crop type data is extracted from a preset crop database, and a genetic algorithm is used to solve a multi-objective function. The genetic algorithm takes the constraints and crop type data as input and outputs an optimized solution to obtain the globally optimal planting time and irrigation amount.

6. The intelligent agricultural planting decision optimization method integrating multimodal data according to claim 1, characterized in that, The process of integrating newly collected data through a feedback loop mechanism in the final decision-making scheme to obtain continuous monitoring indicators and determine the deviation of the scheme execution includes: Through the feedback loop mechanism, continuous monitoring indicators are obtained from the newly collected data. The continuous monitoring indicators are compared with the final decision scheme to determine the preliminary deviation value. The deviation magnitude is obtained by difference calculation. Based on the initial deviation value, the new data collection results are fused with the deviation amplitude using the data integration process to obtain the adjusted monitoring indicators, wherein the fusion is performed by integrating the results through a weighted summation method. The deviation in the implementation of the scheme is determined by the adjusted monitoring indicators, and the deviation correction parameters are obtained, wherein the determination is made by threshold comparison.

7. The intelligent agricultural planting decision optimization method integrating multimodal data according to claim 1, characterized in that, The step of determining whether an alarm mechanism needs to be triggered based on deviations in the scheme execution, and updating the parameters of the multimodal integration model through preset rules to obtain an adaptively enhanced data processing framework includes: By comparing the deviation values, deviation correction parameters are obtained from the monitoring indicators. The deviation correction parameters are obtained by subtracting the benchmark value from the monitoring indicators through difference calculation. The deviation correction parameters are compared with a preset threshold. If the deviation correction parameters exceed the preset threshold, it is determined that an alarm mechanism needs to be triggered, and an alarm triggering judgment result is obtained. Based on the alarm trigger judgment result, the parameters in the model integration optimization are updated using the preset rules in the rule parameter update. The parameters include weight coefficients that are adjusted by weighted averaging. Data from the multimodal fusion mechanism is integrated to determine the updated model parameters. The updated model parameters are used to integrate the adaptive adjustment process with the data framework enhancement. The integration is achieved by combining the data from the adjustment process in a layer-by-layer manner to obtain the enhanced framework components. The parameters of the framework components are then verified by comparing the consistency between the components to obtain the verified framework structure. Based on the verified framework structure, deviation correction is applied to obtain and rule parameters are updated. The application is processed through iterative correction, and the execution monitoring indicators are integrated to determine the final adaptive data processing framework.

8. A smart agricultural planting decision optimization system integrating multimodal data, used to implement the method described in any one of claims 1-7, characterized in that, The system includes: The multimodal data integration module is used to fuse satellite imagery and sensor data streams into a pre-defined multimodal integration model to obtain an initial heterogeneous dataset to capture the dynamic characteristics of crop growth and obtain a multi-source fused data matrix; the feature extraction and noise reduction module is used to perform noise filtering and feature extraction based on the multi-source fused data matrix using a convolutional neural network, to process high-dimensional information for temporal and spatial differences, and to determine key decision information fragments. The association pattern analysis module is used to obtain association patterns of historical planting records from key decision information fragments, and optimizes a multi-objective function through a genetic algorithm to balance yield and cost constraints, thereby obtaining a preliminary set of planting schemes. The multi-objective optimization module is used to adjust the scheme parameters to minimize the risk of soil degradation by using the particle swarm optimization algorithm if the environmental impact indicators in the initial planting scheme set exceed the preset threshold, and to judge the optimized scheme set. The environmental constraint adjustment module is used to update dynamic environmental variables based on the optimized scheme set through real-time weather forecast data, obtain the adjusted constraint conditions, and determine the globally optimal planting time and irrigation amount. The global decision module is used to obtain simulated yield predictions from the global optimal planting time and irrigation amount. If the prediction is lower than the target threshold, it will backtrack to the multi-source fusion data matrix to re-extract information fragments to iteratively optimize the process and obtain the final decision scheme. The iterative optimization module is used to integrate newly collected data through a feedback loop mechanism based on the final decision scheme, obtain continuous monitoring indicators, and determine the deviation of the scheme execution. The dynamic monitoring and feedback module is used to determine whether an alarm mechanism needs to be triggered based on deviations in the implementation of the plan, and to update the parameters of the multimodal integrated model through preset rules to obtain an adaptive data processing framework. Specifically, the noise filtering and feature extraction based on the multi-source fusion data matrix using a convolutional neural network are employed. A three-dimensional convolutional neural network (3D-CNN) is used to construct a four-dimensional multi-source fusion data tensor as the data matrix. The dimensions of this tensor are [N, T, H, W], where N is the number of feature channels, T is the time series length, and H and W are the height and width of the farmland area after rasterization.

Images