Silicon fertilizer application parameter optimization method and system based on big data

Abstract

The application discloses a silicon fertilizer application parameter optimization method and system based on big data, and the method comprises agricultural data acquisition, agricultural data preprocessing, production area division, parameter optimization model construction and silicon fertilizer application parameter optimization. The application relates to the technical field of intelligent parameter optimization of agricultural fertilizer application, and particularly discloses a silicon fertilizer application parameter optimization method and system based on big data. The application obtains original data through data acquisition. The agricultural data preprocessing method adopts spectral data processing, spatial data processing, time series missing value processing, label data processing and data set segmentation. The production area division is performed by adopting a neighborhood constraint clustering model, the production area of farmland with similar production potential can be recognized, and the common law mining of the area is strengthened through partition feature extraction. The hierarchical constraint reinforcement learning model is adopted as the parameter optimization model, so that the fertilization decision can respond to the biological and environmental changes, and the multi-objective collaborative optimization and physical constraint guarantee are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent parameter optimization technology for agricultural fertilizer application, specifically to a method and system for optimizing silicon fertilizer application parameters based on big data. Background Technology

[0002] Silicon fertilizer application parameter optimization refers to the process of collecting and analyzing large amounts of agricultural data, including soil composition, climate conditions, crop types, and crop growth stages, and using data mining and machine learning technologies to precisely adjust the application amount, timing, and method of silicon fertilizer. It can provide personalized fertilization plans based on different regions, soil conditions, and crop needs, optimize the use of silicon fertilizer, improve crop yield and quality, and reduce resource waste and environmental pollution.

[0003] However, traditional methods for optimizing silicon fertilizer application parameters suffer from several drawbacks. They neglect the spatial non-uniformity of soil silicon content distribution within farmland, the differences in water and fertilizer migration caused by slope and elevation, and the dynamic impact of microclimate zones on crop silicon absorption efficiency. This leads to excessive or insufficient fertilization in local areas. Furthermore, they fail to establish topological relationships between plots, making it impossible to identify continuous areas with similar production characteristics. This forces the optimization model to learn mixed features, reducing the reliability of decision-making. Traditional methods for optimizing silicon fertilizer application parameters also suffer from the problem that the conventional static models used cannot respond to temporal variables such as climate changes and soil silicon content during the crop growth cycle. This results in a misalignment between the timing of fertilization and the critical period for crop silicon requirements. Moreover, they do not consider the limitations or prohibitions on agricultural operation precision, making the optimization scheme difficult to implement. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for optimizing silicon fertilizer application parameters based on big data. Traditional silicon fertilizer application parameter optimization methods neglect the spatial non-uniformity of soil silicon content distribution within farmland, the differences in water and fertilizer migration caused by slope and elevation, and the dynamic influence of microclimate zones on crop silicon absorption efficiency. This leads to excessive or insufficient fertilizer application in local areas. Furthermore, the methods fail to establish topological relationships between plots, making it impossible to identify continuous areas with similar production characteristics. This forces the optimization model to learn mixed features, reducing decision reliability. This solution creatively employs a neighborhood-constrained clustering model for production area division. Based on cross-modal feature alignment and density-adaptive clustering from multi-source data, it can identify farmland production areas with similar production potential. This allows subsequent parameter optimization to be personalized for the characteristics of different areas. Simultaneously, it strengthens regional feature extraction... This approach leverages the mining of common regional patterns to enhance the adaptability of silicon fertilizer application. Addressing the technical challenges of traditional silicon fertilizer application parameter optimization methods, which rely on static models that fail to respond to temporal variables such as climate shifts and soil silicon content during crop growth cycles, leading to misalignment between fertilization timing and critical silicon requirements, and neglecting limitations or restrictions on agricultural operational precision, thus hindering implementation, this solution creatively employs a hierarchical constraint reinforcement learning model as the parameter optimization model. By integrating dynamic state encoding, constraint generation networks, and action feasibility filtering, it perceives crop growth stages and climate fluctuations in real time. This allows fertilization decisions to dynamically respond to biological and environmental changes while the constraint generation network predicts silicon fertilizer application thresholds, time windows, and application method feasibility. The action filtering layer then enforces compliance with agricultural safety requirements, achieving multi-objective collaborative optimization and robust physical constraints.

[0005] The technical solution adopted by this invention is as follows: The method for optimizing silicon fertilizer application parameters based on big data provided by this invention includes the following steps:

[0006] Step S1: Agricultural data collection;

[0007] Step S2: Agricultural data preprocessing;

[0008] Step S3: Production area division;

[0009] Step S4: Parameter optimization model construction;

[0010] Step S5: Optimize silicon fertilizer application parameters.

[0011] Further, in step S1, the agricultural data acquisition is used to collect the raw data required for optimizing silicon fertilizer application parameters. Specifically, through data acquisition, a parameter optimization raw dataset is obtained. The parameter optimization raw dataset specifically includes a past farmland raw dataset and a real-time farmland raw dataset. Both the past farmland raw dataset and the real-time farmland raw dataset include soil spectral data, spatial attribute data, climate time series data, and crop growth data. The past farmland raw dataset also includes historical silicon fertilizer application record data and crop yield and quality data under the corresponding fertilization records.

[0012] Further, in step S2, the agricultural data preprocessing is used to preprocess the collected raw data, specifically including the following steps:

[0013] Step S21: Spectral data processing, specifically, processing soil spectral data by employing noise filtering, scattering correction, and characteristic band extraction;

[0014] Step S22: Spatial data processing, specifically, processing spatial attribute data by adopting coordinate unification, topological relationship construction and elevation normalization;

[0015] Step S23: Handling missing time series values, specifically by processing the time series data through linear interpolation in the time direction and aligning it to a unified time axis;

[0016] Step S24: Tag data processing, specifically, processing historical silicon fertilizer application record data by using application rate normalization, application time coding, and application method coding;

[0017] Step S25: Dataset splitting, used to obtain training data and test data, specifically splitting the original dataset of past farmland into an optimized training set and an optimized test set.

[0018] The real-time farmland dataset is preprocessed using the spectral data processing, spatial data processing, temporal missing value processing, and label data processing to obtain a dataset to be optimized. The past farmland dataset is preprocessed using the spectral data processing, spatial data processing, temporal missing value processing, label data processing, and dataset segmentation to obtain an optimized training set and an optimized test set.

[0019] Furthermore, in step S3, the production area division is used to divide farmland into different production areas so as to facilitate the subsequent optimization of silicon fertilizer application parameters for different production areas. Specifically, the input samples are divided into different production areas by designing a neighborhood-constrained clustering model.

[0020] The division of production areas specifically includes the following steps:

[0021] Step S31: Dual-stream feature extraction, used to extract soil spectral features and spatial attribute features respectively. Specifically, feature enhancement is performed by constructing spectral processing streams and spatial processing streams respectively to obtain spectral enhanced features and spatial enhanced features. The steps include:

[0022] Step S311: Construct a spectral processing stream to extract deep soil spectral features. Specifically, capture the long-range dependencies of soil spectral data through a lightweight transformer model encoder to obtain basic spectral features.

[0023] Step S312: Gated feature enhancement, used to dynamically enhance key soil spectral features, specifically by adaptively fusing basic spectral features and spectral nonlinear transformation features through a gating mechanism to obtain spectral enhancement features;

[0024] Step S313: Construct a spatial processing stream to extract spatial topological features. Specifically, based on the adjacency matrix, a graph convolutional network is used to process the spatial attribute data to obtain spatial enhanced features.

[0025] Step S32: Cross-modal alignment, used to achieve co-alignment of spectral and spatial features. Specifically, it maps spectral enhancement features and spatial enhancement features to a unified space through feature projection and maximizing mutual information loss to obtain aligned and fused features. The steps include:

[0026] Step S321: Feature projection, used to transform heterogeneous features to the same dimensional space. Specifically, it involves projecting spectral enhancement features and spatial enhancement features through a linear transformation with shared weights to obtain spectral projection features and spatial projection features of the same dimension.

[0027] Step S322: Maximize mutual information loss to improve the consistency of cross-modal features. Specifically, optimize the feature alignment process by calculating the mutual information loss of spectral-spatial projection feature pairs.

[0028] Step S323: Obtain alignment and fusion features, which are used to stitch together the aligned features. Specifically, the spectral projection features and spatial projection features after feature alignment are stitched together in the feature dimension to obtain alignment and fusion features.

[0029] Step S33: Design of a neighborhood-constrained clustering model to achieve intelligent division of farmland production areas. Specifically, clustering analysis is performed using a composite distance metric and a density-adaptive neighborhood method. The steps include:

[0030] Step S331: Composite distance metric, used to calculate the comprehensive similarity between farmland plots, specifically by fusing feature distance, geographic distance and Gaussian kernel geographic constraint to obtain the composite distance between samples;

[0031] Step S332: Density adaptive neighborhood, used to identify core farmland areas. Specifically, by calculating local density, samples with local density greater than a preset local density threshold are selected as core points, and the neighborhood radius is calculated based on the local density to obtain the core point set.

[0032] Step S333: Perform density clustering to finally divide farmland production areas. Specifically, non-core points are allocated by expanding clusters from core points to obtain clustering results. Each cluster is a production area, resulting in the production area division results.

[0033] Step S34: Partition feature extraction, used to extract the regional features corresponding to each production area, specifically by using the cluster average alignment fusion features as the regional features of the production area;

[0034] Step S35: Divide farmland areas. Specifically, this involves training a model based on the optimized training set and the optimized test set and verifying its performance. The dataset to be optimized, the optimized training set, and the optimized test set are used as inputs to the neighborhood-constrained clustering model, and the production areas are divided to obtain the divided dataset to be optimized, the divided optimized training set, and the divided optimized test set.

[0035] Further, in step S4, the parameter optimization model construction is used to construct the model required for optimizing silicon fertilizer application parameters in each divided production area, specifically by constructing a hierarchical constraint reinforcement learning model as the parameter optimization model.

[0036] The construction of the parameter optimization model specifically includes the following steps:

[0037] Step S41: State-space modeling, used to construct dynamic state representations, specifically involves processing and fusing regional characteristics of the production area, climate time-series data, and crop growth data through a third-order state encoder to obtain a joint state code. The steps include:

[0038] Step S411: Design of a third-order state encoder to encode three types of input information respectively. Specifically, the time evolution of regional features of the production area is processed by a gated recurrent unit, the climate time series data is processed by a temporal convolutional network, and the crop growth data is processed by a factor decomposition machine to obtain regional state features, climate state features and crop state features.

[0039] Step S412: State feature fusion, used to integrate multi-source state information, specifically by concatenating regional state features, climate state features and crop state features and then performing a linear transformation to obtain joint state coding;

[0040] Step S42: Design a constraint generation network to predict the safety boundary of silicon fertilizer application decisions. Specifically, the constraint generation network maps the constraints of silicon fertilizer application parameters through joint state encoding. The constraint generation network includes continuous constraint branches and discrete constraint branches, both of which are composed of two fully connected layers.

[0041] Step S43: Action filtering strategy network design, used to generate and optimize silicon fertilizer application decision actions. Specifically, after outputting the original actions through the strategy network, action filtering is used to constrain and adjust the original actions to obtain executable actions. The dimensions of the generated actions include silicon fertilizer application amount, silicon fertilizer application time, and silicon fertilizer application method. The steps include:

[0042] Step S431: Generating the original action, which is used to generate the initial silicon fertilizer application decision action. Specifically, the original action is obtained by sampling the probability distribution of silicon fertilizer application amount, silicon fertilizer application time and silicon fertilizer application method by predicting the application rate through joint state coding.

[0043] Step S432: Action filtering design, used to enforce action constraints, specifically, the original action is adjusted based on the silicon fertilizer application amount reference interval, silicon fertilizer application reference time window and available silicon fertilizer application method mask output by the constraint generation network to obtain an executable action;

[0044] Step S44: Design a reward function to evaluate the comprehensive benefits of silicon fertilizer application decisions. Specifically, this involves linearly weighting and integrating yield gain, application accuracy, timing suitability, and cost to obtain a multi-objective reward value.

[0045] Step S45: Constrained policy optimization, used to iteratively optimize the decision policy, specifically by updating the policy network through partitioned experience replay and constrained proximal policy optimization algorithm to obtain the optimized silicon fertilizer application decision. The steps include:

[0046] Step S451: Partitioned experience replay, used to effectively store and utilize the unique experience of each part of the farmland production area. Specifically, it establishes partitioned isolated storage units based on the divided farmland production areas to store the experience tuples of each production area, and adopts a priority sampling mechanism based on reward bias so that experience samples with multi-objective reward values ​​in the production area that are better than the average multi-objective reward value in the production area are more likely to be sampled during training.

[0047] Step S452: Design of a constrained proximal policy optimization algorithm to optimize policy network parameters while satisfying constraints. Specifically, the policy network is updated using the standard proximal policy optimization algorithm, and the policy is corrected by monitoring the constraint satisfaction rate.

[0048] Step S46: Construct and train the model. Specifically, this involves constructing a hierarchical constraint reinforcement learning model by integrating the state space modeling, the constraint generation network design, the action filtering policy network design, the design reward function, and the constraint policy optimization. The model is then trained and its performance is verified based on the partitioned optimized training set and the partitioned optimized test set to obtain the hierarchical constraint reinforcement learning model, which serves as the parameter optimization model.

[0049] Further, in step S5, the optimization of silicon fertilizer application parameters specifically involves using the parameter optimization model to independently optimize the silicon fertilizer application parameters for the production areas centrally divided in the divided dataset to be optimized, obtaining the silicon fertilizer application decision for each production area, and adjusting the actual silicon fertilizer application parameters for each production area based on the silicon fertilizer application decision. The silicon fertilizer application decision specifically includes the silicon fertilizer application amount, silicon fertilizer application time, and silicon fertilizer application method.

[0050] The big data-based silicon fertilizer application parameter optimization system provided by this invention includes an agricultural data acquisition module, an agricultural data preprocessing module, a production area division module, a parameter optimization model construction module, and a silicon fertilizer application parameter optimization module.

[0051] The agricultural data acquisition module is used for agricultural data acquisition. Through agricultural data acquisition, it obtains a parameter-optimized raw dataset and sends the parameter-optimized raw dataset to the agricultural data preprocessing module.

[0052] The agricultural data preprocessing module is used for agricultural data preprocessing. Through agricultural data preprocessing, it obtains a dataset to be optimized, an optimization training set, and an optimization test set, and sends the dataset to be optimized, the optimization training set, and the optimization test set to the production area division module.

[0053] The production area division module is used to divide the production area. Through the production area division, it obtains the divided dataset to be optimized, the divided training set and the divided test set. The divided dataset to be optimized is sent to the silicon fertilizer application parameter optimization module, and the divided training set and the divided test set are sent to the parameter optimization model construction module.

[0054] The parameter optimization model construction module is used to construct a parameter optimization model. It constructs a hierarchical constraint reinforcement learning model as the parameter optimization model and sends the parameter optimization model to the silicon fertilizer application parameter optimization module.

[0055] The silicon fertilizer application parameter optimization module is used to optimize silicon fertilizer application parameters. It obtains silicon fertilizer application decisions by processing data in real time using the parameter optimization model.

[0056] The beneficial effects achieved by the present invention using the above solution are as follows:

[0057] (1) Traditional methods for optimizing silicon fertilizer application parameters neglect the spatial non-uniformity of soil silicon content distribution within farmland, the differences in water and fertilizer migration caused by slope and elevation, and the dynamic influence of microclimate zones on crop silicon absorption efficiency, resulting in excessive or insufficient fertilizer application in local areas. Furthermore, the lack of topological association between plots makes it impossible to identify continuous areas with similar production characteristics, forcing the optimization model to learn mixed features and reducing decision reliability. This solution creatively adopts a neighborhood-constrained clustering model for production area division. Based on cross-modal feature alignment and density adaptive clustering of multi-source data, it can identify farmland production areas with similar production potential, enabling subsequent parameter optimization to be personalized for the characteristics of different areas. At the same time, the extraction of regional features strengthens the mining of common regional patterns and improves the adaptability of silicon fertilizer application.

[0058] (2) In view of the technical problems of traditional silicon fertilizer application parameter optimization methods, the traditional static models used cannot respond to time-series variables such as climate change and soil silicon content during the crop growth cycle, resulting in the misalignment of fertilization timing with the critical period of crop silicon requirement, and the failure to consider the precision limitations or prohibition of agricultural operations, making it difficult to implement the optimization scheme, this scheme creatively adopts a hierarchical constraint reinforcement learning model as the parameter optimization model. By integrating the three-layer mechanism of dynamic state coding, constraint generation network and action feasibility filtering, it can perceive crop growth stage and climate fluctuations in real time, so that the fertilization decision can dynamically respond to biological and environmental changes. At the same time, the constraint generation network predicts the silicon fertilizer application threshold, time window and application method feasibility, and the action filtering layer forces the agricultural safety requirements to be met, thus achieving multi-objective collaborative optimization and physical constraint hard guarantee. Attached Figure Description

[0059] Figure 1 A flowchart illustrating the method for optimizing silicon fertilizer application parameters based on big data provided by this invention;

[0060] Figure 2 A schematic diagram of the modules of the silicon fertilizer application parameter optimization system based on big data provided by the present invention;

[0061] Figure 3 This is a flowchart illustrating the agricultural data preprocessing process in step S2.

[0062] Figure 4 A flowchart illustrating the division of the production area in step S3;

[0063] Figure 5 A schematic diagram of the process for constructing the parameter optimization model in step S4.

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0065] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0066] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0067] Example 1, see Figure 1 The present invention provides a method for optimizing silicon fertilizer application parameters based on big data, the method comprising the following steps:

[0068] Step S1: Agricultural data collection;

[0069] Step S2: Agricultural data preprocessing;

[0070] Step S3: Production area division;

[0071] Step S4: Parameter optimization model construction;

[0072] Step S5: Optimize silicon fertilizer application parameters.

[0073] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the agricultural data acquisition is used to collect the raw data required for the optimization of silicon fertilizer application parameters. Specifically, through data acquisition, the parameter optimization raw dataset is obtained. The parameter optimization raw dataset specifically includes past farmland raw datasets and real-time farmland raw datasets. Both the past farmland raw datasets and the real-time farmland raw datasets include soil spectral data, spatial attribute data, climate time series data, and crop growth data. The past farmland raw datasets also include historical silicon fertilizer application record data and crop yield and quality data under the corresponding fertilization records.

[0074] The soil spectral data specifically includes soil silicon content data, soil pH value data, soil organic matter data, and continuous spectral reflectance data of soil texture type;

[0075] The spatial attribute data specifically includes farmland location coordinate data, farmland elevation data, farmland slope data, and farmland plot boundary vector data;

[0076] The climate time series data specifically includes farmland environmental temperature and humidity data, farmland light intensity data, and farmland precipitation data;

[0077] The crop growth data specifically includes crop growth stage label data, crop growth status label data, crop leaf area index data, crop plant height data, and crop chlorophyll content data.

[0078] The historical silicon fertilizer application record data specifically includes historical silicon fertilizer application amount data, historical silicon fertilizer application time data, and historical silicon fertilizer application method data.

[0079] The crop yield and quality data specifically include crop yield per unit area data, crop silicon content data, and crop quality label data from the corresponding fertilization records.

[0080] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the agricultural data preprocessing is used to preprocess the collected raw data, specifically including the following steps:

[0081] Step S21: Spectral data processing, specifically, processing soil spectral data by employing noise filtering, scattering correction, and characteristic band extraction;

[0082] The noise filtering is used to eliminate measurement errors of the spectrometer. Specifically, it involves smoothing the spectral curve by sliding window convolution using an SG filter, with the window size divided into five bands.

[0083] The scattering correction is used to eliminate soil particle scattering interference. Specifically, it involves independently standardizing each sample through a standard normal transformation to eliminate baseline drift.

[0084] The feature band extraction is used to focus on silicon-sensitive spectral bands, specifically by retaining bands whose correlation with silicon content is greater than a preset correlation threshold through principal component analysis.

[0085] Step S22: Spatial data processing, specifically, processing spatial attribute data by adopting coordinate unification, topological relationship construction and elevation normalization;

[0086] The coordinate unification is used to resolve the differences in coordinate systems of farmland plots. Specifically, it involves converting the latitude and longitude of farmland location coordinate data into planar coordinates through UTM projection transformation.

[0087] The aforementioned topological relationship construction is used to establish spatial associations of farmland plots, specifically by generating an adjacency matrix through Delaunay triangulation combined with physical distance threshold constraints;

[0088] The elevation normalization is used to eliminate differences in terrain dimensions, specifically by mapping the elevation to a uniform range through a range scaling method.

[0089] Step S23: Handling missing time series values, specifically by processing the time series data through linear interpolation in the time direction and aligning it to a unified time axis;

[0090] Step S24: Tag data processing, specifically, processing historical silicon fertilizer application record data by using application rate normalization, application time coding, and application method coding;

[0091] The application rate normalization is used to adjust the data to fit the model input range. Specifically, it maps the historical silicon fertilizer application rate data to the [0,1] interval using the minimum-maximum normalization method.

[0092] The application time coding is used to represent the optimal time period for silicon fertilizer application. Specifically, it involves coding historical silicon fertilizer application time data using a Gaussian radial basis function to obtain the distribution of silicon fertilizer application time suitability.

[0093] The application method encoding is used to discretize silicon fertilizer application methods, specifically by encoding historical silicon fertilizer application method data through a unique thermal encoding method.

[0094] Step S25: Dataset splitting, used to obtain training data and test data, specifically splitting the original dataset of past farmland into an optimized training set and an optimized test set.

[0095] The real-time farmland dataset is preprocessed using the spectral data processing, spatial data processing, temporal missing value processing, and label data processing to obtain a dataset to be optimized. The past farmland dataset is preprocessed using the spectral data processing, spatial data processing, temporal missing value processing, label data processing, and dataset segmentation to obtain an optimized training set and an optimized test set.

[0096] Example 4, see Figure 1 , Figure 2 and Figure 4This embodiment is based on the above embodiment. In step S3, the production area division is used to divide farmland into different production areas so as to facilitate the subsequent optimization of silicon fertilizer application parameters for different production areas. Specifically, the input samples are divided into different production areas by designing a neighborhood-constrained clustering model.

[0097] The division of production areas specifically includes the following steps:

[0098] Step S31: Dual-stream feature extraction, used to extract soil spectral features and spatial attribute features respectively. Specifically, feature enhancement is performed by constructing spectral processing streams and spatial processing streams respectively to obtain spectral enhanced features and spatial enhanced features. The steps include:

[0099] Step S311: Construct a spectral processing stream to extract deep soil spectral features. Specifically, capture the long-range dependencies of soil spectral data through a lightweight transformer model encoder to obtain basic spectral features.

[0100] Step S312: Gated feature enhancement, used to dynamically enhance key soil spectral features. Specifically, it adaptively fuses basic spectral features and spectral nonlinear transformation features through a gating mechanism to obtain spectral enhancement features. The formula used is as follows:

[0101] ;

[0102] In the formula, Sw represents the gating weight of the spectral processing stream. This represents the sigmoid activation function. Indicates the weights of the spectral gated transform. This represents the bias term of the spectral gated transform. Nf represents the fundamental spectral characteristics, and Nf represents the spectral nonlinear transformation characteristics. This represents the ELU activation function. Represents the weights of the spectral nonlinear transformation. This represents the bias term of the spectral nonlinear transformation. Indicates spectral enhancement features, This represents element-wise multiplication.

[0103] Step S313: Construct a spatial processing stream to extract spatial topological features. Specifically, based on the adjacency matrix, a graph convolutional network is used to process the spatial attribute data to obtain spatial enhanced features.

[0104] Step S32: Cross-modal alignment, used to achieve co-alignment of spectral and spatial features. Specifically, it maps spectral enhancement features and spatial enhancement features to a unified space through feature projection and maximizing mutual information loss to obtain aligned and fused features. The steps include:

[0105] Step S321: Feature projection, used to transform heterogeneous features to the same dimensional space. Specifically, it involves projecting spectral enhancement features and spatial enhancement features through a linear transformation with shared weights to obtain spectral projection features and spatial projection features of the same dimension. The formula used is as follows:

[0106] ;

[0107] In the formula, Indicates spectral projection characteristics, LN represents the spatial projection characteristics, and LN represents the layer normalization function. Indicates spatial enhancement features, Shared projection weights;

[0108] Step S322: Maximize the mutual information loss to improve the consistency of cross-modal features. Specifically, this is achieved by calculating the mutual information loss of spectral-spatial projection feature pairs to optimize the feature alignment process. The formula used is as follows:

[0109] ;

[0110] In the formula, and Represents the distinct elements of the feature similarity matrix. This represents the cosine similarity calculation function, where tem represents the temperature parameter. This represents the spectral projection characteristics of the a-th sample. This represents the spatial projection feature of the b-th sample. The mutual information loss value is represented by A, where A represents the total number of samples.

[0111] Step S323: Obtain alignment and fusion features, which are used to stitch together the aligned features. Specifically, the spectral projection features and spatial projection features after feature alignment are stitched together in the feature dimension to obtain alignment and fusion features.

[0112] Step S33: Design of a neighborhood-constrained clustering model to achieve intelligent division of farmland production areas. Specifically, clustering analysis is performed using a composite distance metric and a density-adaptive neighborhood method. The steps include:

[0113] Step S331: Composite distance metric, used to calculate the comprehensive similarity between farmland plots. Specifically, it obtains the composite distance between samples by fusing feature distance, geographic distance, and Gaussian kernel geographic constraint. The formula used is as follows:

[0114] ;

[0115] In the formula, This represents the distance weight of the a-th sample. This represents the variance calculation function. This represents the alignment fusion feature of the a-th sample. This represents the composite distance between the a-th sample and the b-th sample. This represents the distance weight of the b-th sample. Represents the feature distance weights. Indicates geographical distance weight. Indicates the Gaussian kernel geographic constraint weights. This represents the alignment and fusion feature of the b-th sample. This represents the position coordinate data of the a-th sample. This represents the position coordinate data of the b-th sample. Indicates the width of the Gaussian kernel. This indicates the calculation of Euclidean distance;

[0116] Step S332: Density adaptive neighborhood, used to identify core farmland areas. Specifically, through local density calculation, samples with local densities greater than a preset local density threshold are selected as core points, and the neighborhood radius is calculated based on the local density to obtain the core point set. The formula used is as follows:

[0117] ;

[0118] In the formula, This represents the local density of the a-th sample. This represents the neighborhood radius of the a-th sample. Represents the reference neighborhood radius. Indicates the maximum local density;

[0119] Step S333: Perform density clustering to finally divide farmland production areas. Specifically, non-core points are allocated by expanding clusters from core points to obtain clustering results. Each cluster is a production area, resulting in the production area division results.

[0120] Step S34: Partition feature extraction, used to extract the regional features corresponding to each production area, specifically by using the cluster average alignment fusion features as the regional features of the production area;

[0121] Step S35: Divide farmland areas. Specifically, this involves training a model based on the optimized training set and the optimized test set and verifying its performance. The dataset to be optimized, the optimized training set, and the optimized test set are used as inputs to the neighborhood-constrained clustering model, and the production areas are divided to obtain the divided dataset to be optimized, the divided optimized training set, and the divided optimized test set.

[0122] By performing the above operations, this solution addresses the technical problems of traditional silicon fertilizer application parameter optimization methods, which neglect the spatial non-uniformity of soil silicon content distribution within farmland, the differences in water and fertilizer migration caused by slope and elevation, and the dynamic influence of microclimate zones on crop silicon absorption efficiency. These problems lead to excessive or insufficient fertilizer application in local areas, and the lack of topological association between plots makes it impossible to identify continuous areas with similar production characteristics, forcing the optimization model to learn mixed features and reducing decision reliability. This solution creatively adopts a neighborhood-constrained clustering model for production area division. Based on cross-modal feature alignment and density adaptive clustering of multi-source data, it can identify farmland production areas with similar production potential, allowing subsequent parameter optimization to be personalized for the characteristics of different areas. At the same time, the extraction of regional features strengthens the mining of common regional patterns and improves the adaptability of silicon fertilizer application.

[0123] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the parameter optimization model is constructed to build the model required for optimizing silicon fertilizer application parameters in each divided production area. Specifically, it is to construct a hierarchical constraint reinforcement learning model as the parameter optimization model.

[0124] The construction of the parameter optimization model specifically includes the following steps:

[0125] Step S41: State-space modeling, used to construct dynamic state representations, specifically involves processing and fusing regional characteristics of the production area, climate time-series data, and crop growth data through a third-order state encoder to obtain a joint state code. The steps include:

[0126] Step S411: Design of a third-order state encoder to encode three types of input information respectively. Specifically, the time evolution of regional features of the production area is processed by a gated recurrent unit, the climate time series data is processed by a temporal convolutional network, and the crop growth data is processed by a factor decomposition machine to obtain regional state features, climate state features and crop state features.

[0127] Step S412: State feature fusion, used to integrate multi-source state information, specifically involves concatenating regional state features, climate state features, and crop state features and then performing a linear transformation to obtain joint state coding. The formula used is as follows:

[0128] ;

[0129] In the formula, Se represents the joint state code. Let Rf represent the ReLU activation function, Cf represent the regional state characteristics, and Cr represent the crop state characteristics. Indicates the weights of the state coding transition;

[0130] Step S42: Design a constraint generation network to predict the safety boundary of silicon fertilizer application decisions. Specifically, the constraint generation network maps the constraints of silicon fertilizer application parameters through joint state encoding. The constraint generation network includes continuous constraint branches and discrete constraint branches, both of which are composed of two fully connected layers.

[0131] The continuous constraint branch processes the joint state encoding through two fully connected layers and outputs a four-dimensional vector, including the maximum reference value of silicon fertilizer application, the minimum reference value of silicon fertilizer application, the maximum reference value of silicon fertilizer application time window, and the minimum reference value of silicon fertilizer application time window, to obtain the silicon fertilizer application reference interval and silicon fertilizer application reference time window.

[0132] The discrete constraint branch processes the joint state encoding through two fully connected layers and combines it with the sigmoid activation function to output a silicon fertilizer application method availability vector. The silicon fertilizer application method availability vector has the number of silicon fertilizer application methods as its dimension and the elements are the normalized values ​​of silicon fertilizer application method availability. By comparing it with the corresponding preset silicon fertilizer application method availability threshold, the silicon fertilizer application method availability vector is binarized to obtain the available silicon fertilizer application method mask.

[0133] Step S43: Action filtering strategy network design, used to generate and optimize silicon fertilizer application decision actions. Specifically, after outputting the original actions through the strategy network, action filtering is used to constrain and adjust the original actions to obtain executable actions. The dimensions of the generated actions include silicon fertilizer application amount, silicon fertilizer application time, and silicon fertilizer application method. The steps include:

[0134] Step S431: Generating the initial action, used to generate preliminary silicon fertilizer application decision actions. Specifically, it involves predicting the probability distribution of silicon fertilizer application amount, application time, and application method through joint state coding, and sampling to obtain the initial action. The formula used is as follows:

[0135] ;

[0136] In the formula, This represents the predicted mean. Indicates the standard deviation of the forecast. Indicates the weights of the mean linear transformation. The standard deviation is represented by the linear transformation weight. This represents the bias term of the linear transformation of the mean. This represents the bias term of the linear transformation of the standard deviation. Indicates the original action. The mean is The standard deviation is The normal distribution;

[0137] Step S432: Action filtering design, used to enforce action constraints, specifically involves adjusting the original action based on the silicon fertilizer application rate reference interval, silicon fertilizer application reference time window, and available silicon fertilizer application method mask output by the constraint generation network to obtain an executable action. The steps include:

[0138] Step S4321: Adjust the constraint of the silicon fertilizer application amount dimension of the action to limit the reasonable range of silicon fertilizer application amount. Specifically, through the trimming operation, the silicon fertilizer application amount dimension value of the original action is constrained within the silicon fertilizer application amount reference range to obtain the silicon fertilizer application amount dimension value of the executable action.

[0139] Step S4322: Constraint adjustment of the silicon fertilizer application time dimension of the action, used to adjust the effectiveness of silicon fertilizer application time. Specifically, by using the sigmoid activation function and combining the relative position of the silicon fertilizer application time dimension value of the original action within the silicon fertilizer application reference time window, the silicon fertilizer application time dimension value of the original action is scaled to obtain the silicon fertilizer application time dimension value of the executable action. The formula used is as follows:

[0140] ;

[0141] In the formula, The value representing the time dimension of silicon fertilizer application indicates the action that can be performed. The value representing the time dimension of silicon fertilizer application in the original action. This indicates the minimum reference value for the time window for silicon fertilizer application. This represents the maximum reference value for the silicon fertilizer application time window, and fa represents the time dimension scaling factor.

[0142] Step S4323: Adjust the constraint of the action silicon fertilizer application method dimension to ensure the selection of feasible silicon fertilizer application methods. Specifically, when the available silicon fertilizer application method mask is valid, select the silicon fertilizer application method with the largest silicon fertilizer application method availability normalization value. Otherwise, resample and generate actions until the silicon fertilizer application methods allowed by the available silicon fertilizer application method mask are obtained, and obtain the silicon fertilizer application method dimension value of the executable action.

[0143] Step S44: Design a reward function to evaluate the overall benefits of silicon fertilizer application decisions. Specifically, this involves linearly weighting and integrating yield gain, application accuracy, timing suitability, and cost to obtain a multi-objective reward value. The formula used is as follows:

[0144] ;

[0145] In the formula, Re represents the multi-objective reward value. Indicates output weight. Indicates the weight of application amount. Indicates the application time weight. Indicates the cost weight of the application method. Indicates the rate of increase in production. This indicates the baseline yield without the application of silicon fertilizer. Values ​​representing the silicon fertilizer application rate dimension indicating executable actions. This represents the theoretically optimal amount of silicon fertilizer to apply. This represents the theoretically optimal time for silicon fertilizer application, and AC represents the cost coefficient of silicon fertilizer application method.

[0146] Step S45: Constrained policy optimization, used to iteratively optimize the decision policy, specifically by updating the policy network through partitioned experience replay and constrained proximal policy optimization algorithm to obtain the optimized silicon fertilizer application decision. The steps include:

[0147] Step S451: Partitioned experience replay, used to effectively store and utilize the unique experience of each part of the farmland production area. Specifically, it establishes partitioned isolated storage units based on the divided farmland production areas to store the experience tuples of each production area, and adopts a priority sampling mechanism based on reward bias so that experience samples with multi-objective reward values ​​in the production area that are better than the average multi-objective reward value in the production area are more likely to be sampled during training.

[0148] The empirical tuple is represented as follows:

[0149] ;

[0150] In the formula, This represents the experience tuple for the c-th production region. This represents the joint state code of the c-th production region at time step t. This represents the set of executable actions for the c-th production region. This represents the set of multi-objective reward values ​​for the c-th production region at time step t. The joint state code of the c-th production region at time step t+1, the joint state code of time step t+1, any executable action in the set of executable actions, and the multi-objective reward value corresponding to this executable action in the set of multi-objective reward values ​​at time step t, constitute an experience sample.

[0151] The priority sampling mechanism based on reward bias uses the following formula:

[0152] ;

[0153] In the formula, This represents the sampling probability of the i-th empirical sample. Let represent the multi-objective reward value of the i-th experience sample. This represents the average multi-objective reward value for the production area. This represents the multi-objective reward value of the j-th experience sample. This represents the minimum value where the probability of prevention is 0.

[0154] Step S452: Design of a constraint-based near-end policy optimization algorithm to optimize policy network parameters while satisfying constraints. Specifically, the policy network is updated using a standard near-end policy optimization algorithm, and policy correction is performed using constraint satisfaction rate monitoring. The constraint satisfaction rate monitoring uses the following formula:

[0155] ;

[0156] In the formula, This represents the constraint satisfaction rate, and T represents the total number of time steps. This indicates the minimum reference value for silicon fertilizer application. This indicates the maximum reference value for silicon fertilizer application. This represents the value of the silicon fertilizer application amount dimension for the action generated by the policy network. This represents an indicator function. Its value is 1 when the silicon fertilizer application rate dimension value of the action generated by the policy network is within the reference range of silicon fertilizer application rate, and its value is 0 otherwise.

[0157] Step S46: Construct and train the model. Specifically, this involves constructing a hierarchical constraint reinforcement learning model by integrating the state space modeling, the constraint generation network design, the action filtering policy network design, the design reward function, and the constraint policy optimization. The model is then trained and its performance is verified based on the partitioned optimized training set and the partitioned optimized test set to obtain the hierarchical constraint reinforcement learning model, which serves as the parameter optimization model.

[0158] By performing the above operations, this solution addresses the technical problems of traditional silicon fertilizer application parameter optimization methods. These problems include the inability of traditional static models to respond to temporal variables such as climate changes and soil silicon content during the crop growth cycle, leading to a misalignment between fertilization timing and the critical period for crop silicon requirements. Furthermore, the method fails to consider the limitations or prohibitions on agricultural operation precision, making it difficult to implement the optimization scheme. This solution creatively adopts a hierarchical constraint reinforcement learning model as the parameter optimization model. By integrating a three-layer mechanism of dynamic state encoding, constraint generation network, and action feasibility filtering, it can perceive factors such as crop growth stages and climate fluctuations in real time. This allows fertilization decisions to dynamically respond to biological and environmental changes. At the same time, the constraint generation network predicts the silicon fertilizer application threshold, time window, and application method feasibility, and the action filtering layer forces compliance with agricultural safety requirements, thus achieving multi-objective collaborative optimization and hard physical constraint guarantees.

[0159] Example 6, see Figure 1 and Figure 2This embodiment is based on the above embodiment. In step S5, the optimization of silicon fertilizer application parameters specifically involves using the parameter optimization model to independently optimize the silicon fertilizer application parameters for the production areas in the divided dataset to be optimized, obtaining the silicon fertilizer application decision for each production area, and adjusting the actual silicon fertilizer application parameters for each production area based on the silicon fertilizer application decision. The silicon fertilizer application decision specifically includes the silicon fertilizer application amount, silicon fertilizer application time, and silicon fertilizer application method.

[0160] Example 7, see Figure 1 and Figure 2 This embodiment is based on the above embodiments. The silicon fertilizer application parameter optimization system based on big data provided by the present invention includes an agricultural data acquisition module, an agricultural data preprocessing module, a production area division module, a parameter optimization model construction module, and a silicon fertilizer application parameter optimization module.

[0161] The agricultural data acquisition module is used for agricultural data acquisition. Through agricultural data acquisition, it obtains a parameter-optimized raw dataset and sends the parameter-optimized raw dataset to the agricultural data preprocessing module.

[0162] The agricultural data preprocessing module is used for agricultural data preprocessing. Through agricultural data preprocessing, it obtains a dataset to be optimized, an optimization training set, and an optimization test set, and sends the dataset to be optimized, the optimization training set, and the optimization test set to the production area division module.

[0163] The production area division module is used to divide the production area. Through the production area division, it obtains the divided dataset to be optimized, the divided training set and the divided test set. The divided dataset to be optimized is sent to the silicon fertilizer application parameter optimization module, and the divided training set and the divided test set are sent to the parameter optimization model construction module.

[0164] The parameter optimization model construction module is used to construct a parameter optimization model. It constructs a hierarchical constraint reinforcement learning model as the parameter optimization model and sends the parameter optimization model to the silicon fertilizer application parameter optimization module.

[0165] The silicon fertilizer application parameter optimization module is used to optimize silicon fertilizer application parameters. It obtains silicon fertilizer application decisions by processing data in real time using the parameter optimization model.

[0166] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0168] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for optimizing silicon fertilizer application parameters based on big data, characterized in that: The method includes the following steps: S1: Agricultural data collection, which involves collecting data to obtain a parameter-optimized raw dataset, specifically including past farmland raw datasets and real-time farmland raw datasets; S2: Agricultural data preprocessing, which involves preprocessing the collected raw data to obtain the dataset to be optimized, the training set for optimization, and the test set for optimization; S3: Production area division, used to divide farmland into different production areas, so as to facilitate the subsequent personalized optimization of silicon fertilizer application parameters for different production areas. Specifically, by designing a neighborhood-constrained clustering model, the input samples are divided into different production areas, resulting in a divided dataset to be optimized, a divided training set for optimization, and a divided test set for optimization. S4: Parameter optimization model construction, used to build the model required for optimizing silicon fertilizer application parameters in each divided production area, specifically to build a layer-constrained reinforcement learning model as the parameter optimization model; S5: Optimization of silicon fertilizer application parameters, specifically, using the parameter optimization model to independently optimize the silicon fertilizer application parameters for the divided production areas, obtaining the silicon fertilizer application decision for each production area, and adjusting the actual silicon fertilizer application parameters for each production area based on the silicon fertilizer application decision. The silicon fertilizer application decision specifically includes silicon fertilizer application amount, silicon fertilizer application time, and silicon fertilizer application method. The construction of the parameter optimization model specifically includes the following steps: Step S41: State space modeling, used to construct dynamic state representation, specifically by processing and fusing regional features, climate time series data and crop growth data of the production area through a third-order state encoder to obtain joint state coding; Step S42: Design a constraint generation network to predict the safety boundary of silicon fertilizer application decisions. Specifically, the constraint generation network maps the constraints of silicon fertilizer application parameters through joint state encoding. The constraint generation network includes continuous constraint branches and discrete constraint branches, both of which are composed of two fully connected layers. Step S43: Action filtering strategy network design, used to generate and optimize silicon fertilizer application decision actions. Specifically, after the original actions are output through the strategy network, action filtering is used to constrain and adjust the original actions to obtain executable actions. The dimensions of the generated actions include silicon fertilizer application amount, silicon fertilizer application time, and silicon fertilizer application method. Step S44: Design a reward function to evaluate the comprehensive benefits of silicon fertilizer application decisions. Specifically, this involves linearly weighting and integrating yield gain, application accuracy, timing suitability, and cost to obtain a multi-objective reward value. Step S45: Constrained policy optimization, used to iteratively optimize the decision policy, specifically by updating the policy network through partitioned experience replay and constrained proximal policy optimization algorithm to obtain the optimized silicon fertilizer application decision; Step S46: Construct and train the model. Specifically, this involves constructing a hierarchical constraint reinforcement learning model by integrating the state space modeling, the constraint generation network design, the action filtering policy network design, the design reward function, and the constraint policy optimization. The model is then trained and its performance is verified based on the partitioned optimized training set and the partitioned optimized test set to obtain the hierarchical constraint reinforcement learning model, which serves as the parameter optimization model.

2. The method for optimizing silicon fertilizer application parameters based on big data according to claim 1, characterized in that: The division of production areas specifically includes the following steps: Step S31: Dual-stream feature extraction, used to extract soil spectral features and spatial attribute features respectively. Specifically, feature enhancement is performed by constructing spectral processing streams and spatial processing streams respectively to obtain spectral enhanced features and spatial enhanced features. The steps include: Step S311: Construct a spectral processing stream to extract deep soil spectral features. Specifically, capture the long-range dependencies of soil spectral data through a lightweight transformer model encoder to obtain basic spectral features. Step S312: Gated feature enhancement, used to dynamically enhance key soil spectral features, specifically by adaptively fusing basic spectral features and spectral nonlinear transformation features through a gating mechanism to obtain spectral enhancement features; Step S313: Construct a spatial processing stream to extract spatial topological features. Specifically, based on the adjacency matrix, a graph convolutional network is used to process the spatial attribute data to obtain spatial enhanced features. Step S32: Cross-modal alignment, used to achieve co-alignment of spectral and spatial features. Specifically, it maps spectral enhancement features and spatial enhancement features to a unified space through feature projection and maximizing mutual information loss to obtain aligned and fused features. The steps include: Step S321: Feature projection, used to transform heterogeneous features to the same dimensional space. Specifically, it involves projecting spectral enhancement features and spatial enhancement features through a linear transformation with shared weights to obtain spectral projection features and spatial projection features of the same dimension. Step S322: Maximize mutual information loss to improve the consistency of cross-modal features. Specifically, optimize the feature alignment process by calculating the mutual information loss of spectral-spatial projection feature pairs. Step S323: Obtain alignment and fusion features, which are used to stitch together the aligned features. Specifically, the spectral projection features and spatial projection features after feature alignment are stitched together in the feature dimension to obtain alignment and fusion features. Step S33: Design of a neighborhood-constrained clustering model to achieve intelligent division of farmland production areas. Specifically, clustering analysis is performed using a composite distance metric and a density-adaptive neighborhood method. The steps include: Step S331: Composite distance metric, used to calculate the comprehensive similarity between farmland plots, specifically by fusing feature distance, geographic distance and Gaussian kernel geographic constraint to obtain the composite distance between samples; Step S332: Density adaptive neighborhood, used to identify core farmland areas. Specifically, by calculating local density, samples with local density greater than a preset local density threshold are selected as core points, and the neighborhood radius is calculated based on the local density to obtain the core point set. Step S333: Perform density clustering to finally divide farmland production areas. Specifically, non-core points are allocated by expanding clusters from core points to obtain clustering results. Each cluster is a production area, resulting in the production area division results. Step S34: Partition feature extraction, used to extract the regional features corresponding to each production area, specifically by using the cluster average alignment fusion features as the regional features of the production area; Step S35: Divide farmland areas. Specifically, this involves training a model based on the optimized training set and the optimized test set and verifying its performance. The dataset to be optimized, the optimized training set, and the optimized test set are used as inputs to the neighborhood-constrained clustering model, and the production areas are divided.

3. The method for optimizing silicon fertilizer application parameters based on big data according to claim 1, characterized in that: The state space modeling process includes the following steps: Step S411: Design of a third-order state encoder to encode three types of input information respectively. Specifically, the time evolution of regional features of the production area is processed by a gated recurrent unit, the climate time series data is processed by a temporal convolutional network, and the crop growth data is processed by a factor decomposition machine to obtain regional state features, climate state features and crop state features. Step S412: State feature fusion, used to integrate multi-source state information, specifically by concatenating regional state features, climate state features and crop state features and then performing a linear transformation to obtain joint state coding; The action filtering strategy network design includes the following steps: Step S431: Generating the original action, which is used to generate the initial silicon fertilizer application decision action. Specifically, the original action is obtained by sampling the probability distribution of silicon fertilizer application amount, silicon fertilizer application time and silicon fertilizer application method by predicting the application rate through joint state coding. Step S432: Action filtering design, used to enforce action constraints, specifically, the original action is adjusted based on the silicon fertilizer application amount reference interval, silicon fertilizer application reference time window and available silicon fertilizer application method mask output by the constraint generation network to obtain an executable action; The constraint strategy optimization includes the following steps: Step S451: Partitioned experience replay, used to effectively store and utilize the unique experience of each part of the farmland production area. Specifically, it establishes partitioned isolated storage units based on the divided farmland production areas to store the experience tuples of each production area, and adopts a priority sampling mechanism based on reward bias so that experience samples with multi-objective reward values ​​in the production area that are better than the average multi-objective reward value in the production area are more likely to be sampled during training. Step S452: Design of a constrained proximal policy optimization algorithm to optimize policy network parameters while satisfying constraints. Specifically, the policy network is updated using the standard proximal policy optimization algorithm, and the policy is corrected by monitoring the constraint satisfaction rate.

4. The method for optimizing silicon fertilizer application parameters based on big data according to claim 1, characterized in that: Both the historical farmland raw dataset and the real-time farmland raw dataset include soil spectral data, spatial attribute data, climate time series data, and crop growth data. The historical farmland raw dataset also includes historical silicon fertilizer application records and crop yield and quality data corresponding to the fertilization records.

5. The method for optimizing silicon fertilizer application parameters based on big data according to claim 1, characterized in that: The agricultural data preprocessing specifically includes the following steps: Step S21: Spectral data processing, specifically, processing the soil spectral data by employing noise filtering, scattering correction, and characteristic band extraction; Step S22: Spatial data processing, specifically, processing spatial attribute data by adopting coordinate unification, topological relationship construction, and elevation normalization; Step S23: Handling missing time series values, specifically by processing the time series data through linear interpolation in the time direction and aligning it to a unified time axis; Step S24: Tag data processing, specifically, processing historical silicon fertilizer application record data by using application rate normalization, application time coding, and application method coding; Step S25: Dataset splitting, used to obtain training data and test data, specifically splitting the original dataset of past farmland into an optimized training set and an optimized test set; The real-time farmland raw dataset is preprocessed through the spectral data processing, spatial data processing, temporal missing value processing, and label data processing to obtain a dataset to be optimized; the past farmland raw dataset is preprocessed through the spectral data processing, spatial data processing, temporal missing value processing, label data processing, and dataset segmentation to obtain an optimized training set and an optimized test set.

6. A big data-based silicon fertilizer application parameter optimization system, used to implement the big data-based silicon fertilizer application parameter optimization method as described in any one of claims 1-5, characterized in that: It includes an agricultural data acquisition module, an agricultural data preprocessing module, a production area division module, a parameter optimization model construction module, and a silicon fertilizer application parameter optimization module; The agricultural data acquisition module is used for agricultural data acquisition. Through agricultural data acquisition, it obtains a parameter-optimized raw dataset and sends the parameter-optimized raw dataset to the agricultural data preprocessing module. The agricultural data preprocessing module is used for agricultural data preprocessing. Through agricultural data preprocessing, it obtains a dataset to be optimized, an optimization training set, and an optimization test set, and sends the dataset to be optimized, the optimization training set, and the optimization test set to the production area division module. The production area division module is used to divide the production area. Through the production area division, it obtains the divided dataset to be optimized, the divided training set and the divided test set. The divided dataset to be optimized is sent to the silicon fertilizer application parameter optimization module, and the divided training set and the divided test set are sent to the parameter optimization model construction module. The parameter optimization model construction module is used to construct a parameter optimization model. It constructs a hierarchical constraint reinforcement learning model as the parameter optimization model and sends the parameter optimization model to the silicon fertilizer application parameter optimization module. The silicon fertilizer application parameter optimization module is used to optimize silicon fertilizer application parameters. It obtains silicon fertilizer application decisions by processing data in real time using the parameter optimization model.

Images