Machine learning-based plot-level planting management traceability system

Abstract

The invention discloses a plot-level planting management traceability system based on machine learning, and the system comprises a data collection and preprocessing module which collects multi-source environment data and processes the multi-source environment data to generate a standardized agricultural data set; the DiffPool model processing module is used for outputting a neighborhood collaborative agriculture feature map by adopting adaptive pooling depth and space-time perception convolution processing; the agricultural management ontology mapping module is used for mapping and constructing a space time semantic tensor; the inter-plot association feature extraction module is used for constructing a plot-level planting management twin model; the reinforcement learning reasoning module is used for executing dynamic node selection and multi-objective optimization to generate an agricultural management decision; the virtual evolution and consistency verification module is used for performing virtual evolution and executing logic consistency verification to generate space-time fingerprints; the traceability voucher generation module packages the traceability voucher and writes the And the traceability report generation module generates an agricultural production traceability report. According to the invention, efficient tracing of agricultural management behaviors is realized, and the accuracy and real-time performance of agricultural tracing are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural management and traceability technology, and in particular to a plot-level planting management traceability system based on machine learning. Background Technology

[0002] In agricultural production management, with the expansion of crop planting scale and the diversification of planting methods, traditional agricultural management methods have gradually revealed their inability to adapt to the needs of modern agriculture. Traditional agricultural management relies on manual experience and lacks accurate data support and systematic decision-making basis, resulting in low management efficiency, serious waste of resources, and difficulty in tracing various data and decisions in the management and production processes. Under such circumstances, the crop growth process, fertilization, irrigation and other operations often lack scientific planning and are difficult to achieve precise management.

[0003] Currently, with the continuous development of information technology, the application of emerging technologies such as agricultural Internet of Things, remote sensing technology, and big data analysis has brought new opportunities to agricultural management. By collecting and analyzing multi-source environmental data, it is possible to provide certain data support for agricultural management. However, most existing technologies focus on the monitoring and management of single factors, such as environmental factors for crop growth and soil moisture, but rarely consider the synergistic effects between plots and the comprehensive factors of space and time. Existing agricultural decision-making systems have not yet achieved dynamic, real-time management based on crop growth cycles and environmental changes, and most decisions still rely on static data models.

[0004] In terms of traceability in agricultural management, although some blockchain-based traceability systems have been applied to the field of food safety, these systems often struggle to handle the complex spatiotemporal dynamics of agricultural management and lack real-time verification and credibility checks of management decision-making processes. Therefore, how to combine multi-source environmental data, crop management decisions, and traceability needs to build an agricultural management traceability system capable of real-time adjustments and cross-temporal and spatial analysis and management has become an urgent problem to be solved in modern agricultural management. Summary of the Invention

[0005] One objective of this invention is to propose a plot-level planting management traceability system based on machine learning. This invention fully utilizes the improved DiffPool model, reinforcement learning inference structure, and spatiotemporal data processing technology to describe in detail how to dynamically optimize management based on multi-source environmental data and agricultural management decisions, thereby achieving spatiotemporal traceability verification of agricultural decisions. This method has the advantages of accurate decision-making, dynamic adjustment, resource optimization, and traceability, which can improve agricultural management efficiency and reduce resource waste.

[0006] According to an embodiment of the present invention, a plot-level planting management traceability system based on machine learning includes: The data acquisition and preprocessing module collects multi-source environmental data within agricultural plots and performs spatiotemporal layering processing to generate a standardized agricultural dataset. The DiffPool model processing module inputs a standardized agricultural dataset into the improved DiffPool model, employs adaptive pooling depth and spatiotemporally aware convolution processing, and outputs a neighborhood collaborative agricultural feature map. The agricultural management ontology mapping module maps neighborhood collaborative agricultural feature maps to a predefined agricultural management ontology structure, and constructs a space-time semantic tensor. The inter-plot association feature extraction module extracts spatiotemporal association features between plots based on the spatial-temporal semantic tensor, constructs a plot-level planting management twin model, and generates the expected growth baseline flow. The reinforcement learning inference module takes the spatial-temporal semantic tensor and the expected growth baseline flow as input, performs dynamic node selection and multi-objective optimization, and generates agricultural management decisions. The virtual evolution and consistency verification module inputs agricultural management decisions into a plot-level planting management twin model for virtual evolution, performs logical consistency verification, and generates a spatiotemporal fingerprint. The traceability certificate generation module encapsulates agricultural management decisions, authenticity contribution, and spatiotemporal fingerprints into spatiotemporally encrypted traceability certificates, which are then written into the traceability database. The traceability report generation module generates agricultural production traceability reports based on traceability certificates, which users can query and verify.

[0007] Optionally, modules can be integrated using the following methods: Collect multi-source environmental data within agricultural plots, perform spatiotemporal stratification processing, and generate a standardized agricultural dataset. The standardized agricultural dataset is input into the improved DiffPool model, which is processed using adaptive pooling depth and spatiotemporally aware convolution to output a neighborhood collaborative agricultural feature map. Map the feature map of neighborhood collaborative agriculture to a predefined agricultural management ontology structure to construct a space-time semantic tensor. Based on the spatial-temporal semantic tensor, spatiotemporal correlation features between plots are extracted, a plot-level planting management twin model is constructed, and the expected growth baseline flow is generated using the plot-level planting management twin model. Using the spatial-temporal semantic tensor and the expected growth baseline flow as input, a reinforcement learning inference structure is employed to perform dynamic node selection and multi-objective optimization, generating agricultural management decisions that include irrigation amount, fertilizer amount and pest and disease control amount; Agricultural management decisions are input into a plot-level planting management twin model for virtual evolution. The evolutionary prediction features are logically consistent with the measured standardized agricultural data at the next moment, the authenticity contribution is calculated, and a spatiotemporal fingerprint is generated. Agricultural management decisions, authenticity contributions, and spatiotemporal fingerprints are encapsulated as spatiotemporally encrypted traceability credentials, written into the traceability database, and agricultural production traceability reports are output.

[0008] Optionally, the multi-source environmental data includes soil moisture, soil temperature, meteorological parameters, canopy illumination, image information, and operational behavior data; the spatiotemporal stratification processing includes performing spatial grid division based on plot number, performing time period division based on planting cycle, completing spatiotemporal registration of agricultural sensing data, performing normalization and scale unification on the spatiotemporally registered agricultural sensing data, and generating a standardized agricultural data set.

[0009] Optionally, the generation process of the neighborhood collaborative agriculture feature map includes: Extract the multi-dimensional attribute sequence of each agricultural plot grid cell in the standardized agricultural dataset within a continuous time window, and construct an initial agricultural graph structure. In the initial agricultural graph structure, the nodes are agricultural plot grid cells, the node features are the time-series data vectors of the corresponding grid, and the edges are the spatial adjacency relationships between adjacent grids. Calculate the local feature change rate and spatial gradient of each node, take the plot feature difference degree and spatial coupling degree as input parameters, generate the first layer of binary pooling mask, and automatically divide the initial agricultural map structure into several candidate plot clusters; Within each candidate plot cluster, the information entropy is used to measure the node distribution density, the pooling factor of each plot cluster is calculated, and graph pooling operation is performed to fuse the node features within the same plot cluster into a cluster center feature vector, generating the first layer of nested graph structure. In each nested graph structure, a spatiotemporally aware convolutional kernel is embedded, and a similarity change function is used to calculate the feature correlation between each node and its neighboring nodes in the historical time series. Based on the correlation calculation results, the spatial edge weights are dynamically adjusted to achieve the time dynamic modulation of local spatial relationships. During graph convolution propagation, the location of intervention nodes corresponding to operational behavior data in the standardized agricultural dataset is identified. A Gaussian weighted window is constructed at the intervention time through linear interpolation. Weighted propagation of feature information is performed within the Gaussian weighted window to preserve the traces of the impact of management behavior on the state of the plot. The pooling and convolution operations are repeated until the information density fluctuation rate of the global parcel cluster is less than the preset threshold. The information density fluctuation rate is determined by calculating the ratio of the average mutual information difference between the feature sets of the current layer and the previous layer to the average mutual information of the previous layer. By performing multi-scale connections on the output node representations of each level in the multi-layer nested graph structure, the final neighborhood collaborative agriculture feature map is constructed.

[0010] Optionally, the process of constructing the spatial-temporal semantic tensor includes: constructing an agricultural ontology structure map containing crop type, soil type, application method, growth stage, plot location, and management entity; matching corresponding agricultural ontology concept nodes and relation edges based on the position, feature vector, and timestamp information of agricultural plot grid units in the neighborhood collaborative agriculture feature map, performing semantic alignment operations, and generating a semantic mapping relationship set; based on the semantic mapping relationship set, expanding the hierarchical concept relationships of the matched nodes in the agricultural ontology map into an ontology nested structure, constructing a three-dimensional coordinate mapping function using the spatial adjacency relationship and time series index between nodes, and establishing a spatial-temporal projection between the agricultural ontology nested structure and the neighborhood collaborative agriculture feature map; and based on the projection result, integrating the semantic labels of the ontology nested structure with the plot temporal features in the neighborhood collaborative agriculture feature map to construct a spatial-temporal semantic tensor containing semantic, spatial, and temporal dimensions.

[0011] Optionally, the construction process of the plot-level planting management twin model includes: Extract the slice vector of each agricultural plot grid cell from the spatial-temporal semantic tensor as a semantic state vector; Calculate the cosine distance between semantic state vectors of adjacent agricultural plot grid cells, and construct a grid state transition graph with grid cells as nodes and cosine distance as edge weights; Graph encoding is performed on the grid state transition graph to aggregate the historical semantic state of nodes and the state differences of neighboring nodes. After processing by a fully connected layer, the evolution feature vector of each plot at the current time step is generated. The operation behavior data is acquired and transformed into behavior influence vectors. The behavior influence vectors are concatenated with the evolution feature vectors and input into a Siamese comparison network to generate predicted land parcel state trajectories. In the twin contrast network, the state trajectory of the same plot in a continuous time period is used as a positive sample pair, and the state trajectory of plots with different operation behavior data but spatial distance within a preset range is used as a negative sample pair. A spatiotemporal consistency contrast loss function is constructed, and the parameters of the twin contrast network are optimized by gradient descent algorithm. The latest semantic state vector of the target plot and the operation behavior vector to be verified are input into the optimized Siamese contrast network to infer and generate the expected growth baseline flow of the target plot.

[0012] Optionally, the specific steps for generating agricultural management decisions are as follows: The semantic state vector of the agricultural plot grid cell is concatenated with the corresponding expected growth baseline flow in multiple dimensions and used as a joint state input to construct the reinforcement learning state space. Using each agricultural plot grid unit as the set of decision nodes, the policy probability distribution of each decision node is calculated based on the current reinforcement learning state space. The decision nodes are sorted from high to low according to their policy probability values, and the top P decision nodes are extracted as the subset of target plot nodes currently participating in the decision. For each node in the target plot node subset, a multi-dimensional continuous action space is constructed, and the action vector includes irrigation amount adjustment value, fertilizer amount adjustment value, and pest and disease control amount adjustment value. The action vector is input into the plot-level planting management twin model, and the predicted plot state trajectory under the decision-making action is output. In the time dimension, the predicted plot state trajectory is gradually subtracted from the expected growth baseline flow, the state deviation value at each time step is calculated, and the state transition result is generated. Based on the state transition results, a multi-objective reward vector is constructed. A weighted summation operation is performed on the growth deviation penalty term, resource input cost term, and spatial coordination consistency term to obtain the single-step decision reward. Based on the single-step decision reward, the policy parameters in the reinforcement learning inference structure are updated using the policy gradient algorithm. The node selection and action generation process is repeated until the cumulative reward converges, and the final combination of irrigation amount, fertilizer amount and pest and disease control amount is output as the agricultural management decision.

[0013] Optionally, the specific steps of the logical consistency verification include: The agricultural management decisions are input into the plot-level planting management twin model, virtual evolution calculations are performed, and the predicted plot state feature vector corresponding to the next time step is output. Collect multi-source environmental data of the corresponding agricultural plots at the next time step, perform preprocessing, generate measured standardized agricultural data at the next time step, and transform it into a measured plot state feature vector; Perform a one-to-one mapping between the predicted land parcel state feature vector and the measured land parcel state feature vector in the feature dimension to construct a set of state feature aligned vectors; Extract the corresponding dimension components from the state feature alignment vector set and perform subtraction to obtain the difference sequence of each feature dimension. Perform normalization processing based on the preset dimensional scaling factor to generate a logical deviation vector. Extract the component values ​​of each dimension in the logical deviation vector, determine whether they are synchronously within the pre-set consistency threshold window for each dimension, and generate a Boolean logical consistency result. Using logical consistency results as input for calculating the contribution of authenticity, we can characterize the logical credibility of agricultural management decisions at the current time step.

[0014] Optionally, the specific steps for calculating the authenticity contribution and generating the spatiotemporal fingerprint include: Extract the logical consistency result vector, count the number of elements with a Boolean value of "true" in the vector, and calculate the ratio of the number of elements with a value of "true" to the total dimension of the vector to obtain the logical consistency ratio at the current time step. Obtain the weight coefficients corresponding to the crop type, growth stage, and environmental sensitivity level of the plot, and use a weighted average algorithm to fuse the logical consistency ratio with the weight coefficients to calculate the authenticity contribution of the current agricultural management decision to the target plot. The land parcel identifier, timestamp, management behavior code, and authenticity contribution are concatenated into a string in a preset order. A hash function is then used to perform hash processing on the concatenated string to generate a unique spatiotemporal fingerprint of the land parcel-level management behavior.

[0015] Optionally, the traceability report includes plot identifiers, agricultural management decision content, corresponding authenticity contribution, generated spatiotemporal fingerprints, logical consistency verification results, and key environmental indicators. The traceability report displays the plot status evolution trend and decision traceability path in a visual structure, and supports user query, verification, and export.

[0016] The beneficial effects of this invention are: (1) To achieve precision agricultural management, this invention combines multi-source environmental data with an improved DiffPool model to achieve refined management of agricultural plots. It can dynamically adjust agricultural management decisions based on factors such as real-time environmental changes and crop growth status, such as irrigation, fertilization and pest control, thereby improving resource utilization, reducing overuse, and improving the precision and efficiency of agricultural management. (2) Dynamic spatiotemporal traceability and credibility verification: This invention realizes real-time traceability and credibility verification of agricultural management decisions through virtual evolution and logical consistency verification mechanism, ensuring that every management decision has traceable evidence and rationality guarantee. By generating spatiotemporal fingerprints and authenticity contribution, it improves the transparency and verifiability of agricultural management decisions and solves the problems of data opacity and difficulty in traceability in traditional agricultural management. (3) Multi-objective optimization decision-making and collaborative management: Through reinforcement learning reasoning structure and multi-objective optimization, this invention can comprehensively consider growth deviation, resource consumption and spatial coordination consistency to generate optimal agricultural management decisions. While considering multiple factors, it can ensure the optimality and stability of the decision, improve the collaborative management capability of plots in the region, avoid resource waste, and optimize the overall benefits of agricultural production. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a module connection structure diagram of a plot-level planting management traceability system based on machine learning proposed in this invention; Figure 2 This is an overall flowchart of a plot-level planting management traceability system based on machine learning proposed in this invention; Figure 3 This is a structural diagram of the improved DiffPool model processing module of a plot-level planting management traceability system based on machine learning proposed in this invention. Figure 4 This is a diagram showing the structure of a plot-level planting management twin model for a plot-level planting management traceability system based on machine learning proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1-4 A plot-level planting management traceability system based on machine learning includes: The data acquisition and preprocessing module collects multi-source environmental data within agricultural plots and performs spatiotemporal layering processing to generate a standardized agricultural dataset. The DiffPool model processing module inputs a standardized agricultural dataset into the improved DiffPool model, employs adaptive pooling depth and spatiotemporally aware convolution processing, and outputs a neighborhood collaborative agricultural feature map. The agricultural management ontology mapping module maps neighborhood collaborative agricultural feature maps to a predefined agricultural management ontology structure, and constructs a space-time semantic tensor. The inter-plot association feature extraction module extracts spatiotemporal association features between plots based on the spatial-temporal semantic tensor, constructs a plot-level planting management twin model, and generates the expected growth baseline flow. The reinforcement learning inference module takes the spatial-temporal semantic tensor and the expected growth baseline flow as input, performs dynamic node selection and multi-objective optimization, and generates agricultural management decisions. The virtual evolution and consistency verification module inputs agricultural management decisions into a plot-level planting management twin model for virtual evolution, performs logical consistency verification, and generates a spatiotemporal fingerprint. The traceability certificate generation module encapsulates agricultural management decisions, authenticity contribution, and spatiotemporal fingerprints into spatiotemporally encrypted traceability certificates, which are then written into the traceability database. The traceability report generation module generates agricultural production traceability reports based on traceability certificates, which users can query and verify.

[0020] In this embodiment, the modules are interconnected using the following method: Collect multi-source environmental data within agricultural plots, perform spatiotemporal stratification processing, and generate a standardized agricultural dataset. The standardized agricultural dataset is input into the improved DiffPool model, which is processed using adaptive pooling depth and spatiotemporally aware convolution to output a neighborhood collaborative agricultural feature map. Map the feature map of neighborhood collaborative agriculture to a predefined agricultural management ontology structure to construct a space-time semantic tensor. Based on the spatial-temporal semantic tensor, spatiotemporal correlation features between plots are extracted, a plot-level planting management twin model is constructed, and the expected growth baseline flow is generated using the plot-level planting management twin model. Using the spatial-temporal semantic tensor and the expected growth baseline flow as input, a reinforcement learning inference structure is employed to perform dynamic node selection and multi-objective optimization, generating agricultural management decisions that include irrigation amount, fertilizer amount and pest and disease control amount; Agricultural management decisions are input into a plot-level planting management twin model for virtual evolution. The evolutionary prediction features are logically consistent with the measured standardized agricultural data at the next moment, the authenticity contribution is calculated, and a spatiotemporal fingerprint is generated. Agricultural management decisions, authenticity contributions, and spatiotemporal fingerprints are encapsulated as spatiotemporally encrypted traceability credentials, written into the traceability database, and agricultural production traceability reports are output.

[0021] In this embodiment, the multi-source environmental data includes soil moisture, soil temperature, meteorological parameters, canopy illumination, image information, and operational behavior data; the spatiotemporal stratification processing includes performing spatial grid division based on plot number, performing time period division based on planting cycle, completing spatiotemporal registration of agricultural sensing data, performing normalization and scale unification on the spatiotemporally registered agricultural sensing data, and generating a standardized agricultural data set.

[0022] In this embodiment, the generation process of the neighborhood collaborative agriculture feature map includes: Extract the multi-dimensional attribute sequence of each agricultural plot grid cell in the standardized agricultural dataset within a continuous time window. The multi-dimensional attribute sequence includes the time-series data of soil moisture, soil temperature, meteorological parameters, canopy illumination, image information, and operational behavior collected by the plot grid cell within the continuous time window. Construct an initial agricultural graph structure, in which nodes are agricultural plot grid cells, node features are the time-series data vectors of the corresponding grid, and edges are the spatial adjacency relationships between adjacent grids. Calculate the local feature change rate and spatial gradient of each node, take the plot feature difference degree and spatial coupling degree as input parameters, generate the first layer of binary pooling mask, and automatically divide the initial agricultural map structure into several candidate plot clusters; Specifically, the local feature change rate of a node is obtained by calculating the difference magnitude of the feature vectors of the node at adjacent time points within a continuous time window; the spatial gradient is obtained by calculating the difference magnitude between the feature vector of the node and the feature vectors of its spatial neighbors; the plot feature difference degree is obtained by calculating the weighted Euclidean distance between the local feature change rates and spatial gradients of candidate node pairs; the spatial coupling degree is measured based on the connection strength and edge weight sparsity of the spatial adjacent edges between nodes; the plot feature difference degree and the spatial coupling degree are normalized and then weighted and fused to construct a plot aggregation discrimination matrix; the aggregation discrimination matrix is ​​binarized according to a preset threshold to generate a first-layer binarized pooling mask; the mask is used to indicate the set of nodes that can be aggregated in the initial agricultural map structure, and the initial agricultural map structure is automatically divided into several candidate plot clusters accordingly; Within each candidate plot cluster, the information entropy is used to measure the node distribution density, the pooling factor of each plot cluster is calculated, and graph pooling operation is performed to fuse the node features within the same plot cluster into a cluster center feature vector, generating the first layer of nested graph structure. Specifically, the calculation process of the information entropy metric node distribution density is as follows: In each candidate land parcel cluster, the frequency of the time-series data vectors of all nodes in the cluster is statistically analyzed in each feature dimension. Based on the normalized frequency distribution, a feature probability distribution function is constructed, and the Shannon information entropy is calculated as the node distribution density index of the candidate land parcel cluster. The pooling factor is the normalized ratio of the information entropy value to the mean of the information entropy of all candidate land parcel clusters. The ratio is used to control the feature compression ratio of the land parcel cluster. Based on the pooling factor, the attention-weighted graph pooling method is adopted. In each candidate land parcel cluster, attention weights are assigned to the temporal data vectors of the nodes. The attention weights are determined based on the feature similarity between the node and other nodes in the cluster. The weighted node features are then weighted and averaged to obtain the cluster center feature vector, which replaces all nodes in the original cluster, completing the graph pooling operation and generating the first-level nested graph structure containing the cluster center node. In each nested graph structure, a spatiotemporally aware convolutional kernel is embedded, and a similarity change function is used to calculate the feature correlation between each node and its neighboring nodes in the historical time series. Based on the correlation calculation results, the spatial edge weights are dynamically adjusted to achieve the time dynamic modulation of local spatial relationships. The similarity change function is a weighted cosine similarity function based on a time window. Specifically, within a preset historical time window, the time-series data vectors of a certain node and its neighboring nodes are extracted respectively. The cosine similarity sequence between the vectors corresponding to each time step is calculated. The weights are assigned by combining the time decay factor, and a weighted average is performed on the similarity sequence to obtain the feature similarity between nodes. The feature correlation is the rate of change of the output of the above similarity change function in different time windows. The similarity slope is obtained by using the sliding window difference method and normalization is performed to characterize the dynamic fluctuation of the association strength between node pairs. The method of dynamically adjusting spatial edge weights is as follows: using feature correlation as the modulation coefficient, combined with the spatial adjacency weights of node pairs in the original graph structure, a modulation function is constructed to reconstruct the edge weights. Specifically, the modulation function is the edge weight multiplied by the feature correlation exponential decay factor, which enhances the connection strength between highly stable correlated nodes and suppresses the spatial propagation influence between low-stability correlated nodes, thereby realizing the dynamic modulation of spatial relationships in the time dimension. During graph convolutional propagation, the location of intervention nodes corresponding to operational behavior data in the standardized agricultural dataset is identified. A Gaussian weighted window is constructed at the intervention time using linear interpolation. Weighted propagation of feature information is performed within the Gaussian weighted window to preserve the traces of the impact of management behavior on the plot status. The intervention node refers to the plot grid unit in the standardized agricultural dataset that has records of operational behaviors such as irrigation, fertilization, and pest and disease control at a specific time step. For each intervention node, a Gaussian weighted window centered on the intervention time step is constructed. Weighted fusion operation is performed on its historical time series features. The weighted result is used for graph convolutional propagation to enhance the expression of feature evolution by management behavior. The pooling and convolution operations are repeated until the information density fluctuation rate of the global parcel cluster is less than the preset threshold. The information density fluctuation rate is determined by calculating the ratio of the average mutual information difference between the feature sets of the current layer and the previous layer to the average mutual information of the previous layer. By performing multi-scale connections on the output node representations of each level in the multi-layer nested graph structure, the final neighborhood collaborative agriculture feature map is constructed.

[0023] In this embodiment, the process of constructing the spatial-temporal semantic tensor includes: constructing an agricultural ontology structure map containing crop type, soil type, application method, growth stage, plot location, and management entity; matching corresponding agricultural ontology concept nodes and relation edges based on the position, feature vector, and timestamp information of agricultural plot grid units in the neighborhood collaborative agriculture feature map, performing semantic alignment operations, and generating a semantic mapping relationship set; based on the semantic mapping relationship set, expanding the hierarchical concept relationships of the matched nodes in the agricultural ontology map into an ontology nested structure, constructing a three-dimensional coordinate mapping function using the spatial adjacency relationship and time series index between nodes, and establishing a spatial-temporal projection between the agricultural ontology nested structure and the neighborhood collaborative agriculture feature map; based on the projection result, fusing the semantic labels of the ontology nested structure with the plot temporal features in the neighborhood collaborative agriculture feature map, constructing a spatial-temporal semantic tensor containing semantic, spatial, and temporal dimensions; the three-dimensional coordinate mapping function refers to the spatial position index of agricultural plot grid units in the neighborhood collaborative agriculture feature map. Its corresponding time series step size Construct mapping function ,in These represent the concept node number and hierarchical path code in the agricultural management ontology structure, respectively. The spatial and temporal semantic positioning and attribution of specific grid nodes in the agricultural feature map in the agricultural ontology structure are realized through a three-dimensional coordinate mapping function.

[0024] In this embodiment, the construction process of the plot-level planting management twin model includes: Extract the slice vector of each agricultural plot grid cell from the spatial-temporal semantic tensor as a semantic state vector; Calculate the cosine distance between semantic state vectors of adjacent agricultural plot grid cells, and construct a grid state transition graph with grid cells as nodes and cosine distance as edge weights; Graph encoding is performed on the grid state transition graph to aggregate the historical semantic state of nodes and the state differences of neighboring nodes. After processing by a fully connected layer, the evolution feature vector of each plot at the current time step is generated. The operation behavior data is acquired and transformed into behavior influence vectors. The behavior influence vectors are concatenated with the evolution feature vectors and input into a Siamese comparison network to generate predicted land parcel state trajectories. In the twin contrast network, the state trajectory of the same plot in a continuous time period is used as a positive sample pair, and the state trajectory of plots with different operation behavior data but spatial distance within a preset range is used as a negative sample pair. A spatiotemporal consistency contrast loss function is constructed, and the parameters of the twin contrast network are optimized by gradient descent algorithm. The spatiotemporal consistency contrastive loss function is used to guide the Siamese contrastive network to simultaneously learn the evolution of land parcel status over time (temporal consistency) and the ability to respond to differences in management behavior under spatial proximity (spatial discriminability) during training. Let the first... Each training sample represents the state trajectory of the target plot, denoted as: ;in, Indicates time step The semantic state vector of the land parcel at that time.

[0025] Define a positive sample pair as the state trajectory of the same plot of land over two consecutive time periods. Negative sample pairs are defined as pairs of land parcels that are spatially adjacent but have significantly different management behaviors. ,in It is spatial distance. The spatially weighted contrastive loss function is constructed as follows: ; in, The encoding vector generated by the Siamese network for positive sample pairs. Similarity measurement functions (such as cosine similarity). The negative sample set consists of spatially adjacent but behaviorally different land parcels. Spatial behavior weighting factor Temperature coefficient; The latest semantic state vector of the target plot and the operation behavior vector to be verified are input into the optimized Siamese contrast network to infer and generate the expected growth baseline flow of the target plot.

[0026] In this embodiment, the specific steps for generating agricultural management decisions are as follows: The semantic state vector of the agricultural plot grid cell is concatenated with the corresponding expected growth baseline flow in multiple dimensions and used as a joint state input to construct the reinforcement learning state space. Using each agricultural plot grid unit as the set of decision nodes, the policy probability distribution of each decision node is calculated based on the current reinforcement learning state space. The decision nodes are sorted from high to low according to their policy probability values, and the top P decision nodes are extracted as the subset of target plot nodes currently participating in the decision. In the agricultural management decision-making process, it is necessary to dynamically select a subset of nodes that have a key impact on the overall management efficiency from multiple candidate agricultural plots based on the differences and evolution trends of the performance of plots in the current spatial and temporal semantic state. The system first concatenates the semantic state vector of each agricultural plot grid unit with its corresponding expected growth baseline flow to construct a joint state representation in the reinforcement learning state space. The joint state representation serves as the input to the policy network, which can be a feedforward neural network, graph attention network, or parameterized Gaussian policy model, and is used to output the optimal action probability distribution of each plot node in the current state. Specifically, the policy network reception dimension is joint state vector For each decision node Execution strategy mapping function Generate the probability distribution of nodes in the action space, where For policy network parameters, Represents an action vector that satisfies To reduce computational complexity, this system normalizes the policy probability vector and then takes the value corresponding to the highest probability. The priority value for the current plot node; the set of priority values ​​for all decision nodes. Sort in descending order, extract the first... The highest probability plot nodes constitute the target plot nodes for the current agricultural management action generation and decision simulation, thereby improving the computational efficiency and convergence speed of reinforcement learning on large-scale agricultural graph structures.

[0027] For each node in the target plot node subset, a multi-dimensional continuous action space is constructed, and the action vector includes irrigation amount adjustment value, fertilizer amount adjustment value, and pest and disease control amount adjustment value. The action vector is input into the plot-level planting management twin model, and the predicted plot state trajectory under the decision-making action is output. In the time dimension, the predicted plot state trajectory is gradually subtracted from the expected growth baseline flow, the state deviation value at each time step is calculated, and the state transition result is generated. Based on the state transition results, a multi-objective reward vector is constructed. A weighted summation operation is performed on the growth deviation penalty term, resource input cost term, and spatial coordination consistency term to obtain the single-step decision reward. In the reinforcement learning inference structure, to achieve a comprehensive quantitative evaluation of the effectiveness of agricultural management strategies, the system constructs a multi-objective reward vector based on the plot state transition results. Weighted fusion of each objective sub-item is performed to form a single-step decision reward for strategy optimization. The multi-objective reward vector comprehensively considers three factors: growth deviation, resource input, and spatial consistency. The single-step reward function... The design is as follows: ; in, This represents the growth deviation penalty term for each target plot at the current moment, used to measure the difference between the predicted state trajectory and the expected growth baseline flow; This represents the resource input cost item, used to assess the resource consumption level of current management actions in the three dimensions of irrigation, fertilization, and pest and disease control; This represents a spatial coordination consistency term, used to penalize situations where there are excessive differences in management strategies between spatially adjacent plots. , , These are the weighting coefficients for each sub-item; The calculation method for the growth deviation penalty is as follows: ; in, Indicates the first A plot of agricultural land is taking shape in the present moment and into the future. The predicted state sequence within each time step This represents the expected growth baseline flow for the corresponding plot of land. This indicates the number of target land parcels currently involved in the decision-making process. It is a Euclidean distance metric; The calculation method for resource input cost items is as follows: ; in, , and They represent the first The adjustment values ​​for irrigation, fertilizer, and pest and disease control for each plot of land at the current time step. , , This refers to the unit cost coefficient for the corresponding resource type; The spatial consistency term is calculated as follows: ; in, Indicates land parcel Spatial adjacent block set, Indicates land parcel With the plot of land Spatial adjacency weights between them Indicates land parcel At time step Agricultural management action vector; By constructing a multi-objective reward function that simultaneously constrains crop growth consistency, resource utilization efficiency, and spatial coordination within the same reinforcement learning framework, the reinforcement learning strategy can gradually converge to the optimal decision path for plot-level planting management under multi-objective conditions, ultimately yielding a single-step decision reward. As an input of real-time feedback signals into the strategy update algorithm, it is used to drive the continuous optimization of agricultural management decision-making strategies.

[0028] Based on the single-step decision reward, the policy parameters in the reinforcement learning inference structure are updated using the policy gradient algorithm. The node selection and action generation process is repeated until the cumulative reward converges, and the final combination of irrigation amount, fertilizer amount and pest and disease control amount is output as the agricultural management decision.

[0029] In this embodiment, the specific steps of the logical consistency verification include: The agricultural management decisions are input into the plot-level planting management twin model, virtual evolution calculations are performed, and the predicted plot state feature vector corresponding to the next time step is output. Collect multi-source environmental data of the corresponding agricultural plots at the next time step, perform preprocessing, generate measured standardized agricultural data at the next time step, and transform it into a measured plot state feature vector; Perform a one-to-one mapping between the predicted land parcel state feature vector and the measured land parcel state feature vector in the feature dimension to construct a set of state feature aligned vectors; Extract the corresponding dimension components from the state feature alignment vector set and perform subtraction to obtain the difference sequence of each feature dimension. Perform normalization processing based on the preset dimensional scaling factor to generate a logical deviation vector. In this embodiment, a preset dimensional scaling factor is used to standardize the differences in state features across different dimensions, eliminating the bias caused by differences in units, scales, or dimensions, and ensuring the scientific validity and comparability of logical consistency judgments. Each feature dimension corresponds to a dimensional scaling factor, which is typically set in the following three ways: For feature dimensions with sufficient historical observation data, scaling factors can be automatically generated based on their maximum value, average fluctuation range or statistical standard deviation in actual planting scenarios, so that the normalized difference falls within a standardized range, avoiding interference with consistency judgment due to excessively large values ​​of individual dimensions. For some indicators with attributes unique to the agricultural field (such as growth activity index, pest and disease severity, etc.), if there is a lack of stable historical statistical distribution, agricultural experts set upper and lower limits based on experience, and then determine the allowable fluctuation range of the feature based on this, and set a scaling factor to reflect the tolerance threshold of feature fluctuation. For some indicators that are significantly affected by the external environment but need to be included in the consistency judgment (such as soil moisture content, light index, etc.), the scaling factor can be dynamically adjusted in combination with contextual information such as season, plot geographical conditions and crop variety to ensure universality and adaptability in different management scenarios. In summary, the preset dimensional scaling factor is not only used to achieve scale uniformity of multidimensional agricultural characteristics, but also serves as an important basic parameter in the logical consistency verification mechanism to ensure the scientificity and stability of difference assessment. The setting method is adaptable, professional and interpretable, and has wide applicability and engineering feasibility in actual planting management.

[0030] Extract the component values ​​of each dimension in the logical deviation vector, determine whether they are synchronously within the pre-set consistency threshold window for each dimension, and generate a Boolean logical consistency result. Using logical consistency results as input for calculating the contribution of authenticity, we can characterize the logical credibility of agricultural management decisions at the current time step.

[0031] In this embodiment, the specific steps for calculating the authenticity contribution and generating the spatiotemporal fingerprint include: Extract the logical consistency result vector, count the number of elements with a Boolean value of "true" in the vector, and calculate the ratio of the number of elements with a value of "true" to the total dimension of the vector to obtain the logical consistency ratio at the current time step. Obtain the weight coefficients corresponding to the crop type, growth stage, and environmental sensitivity level of the plot, and use a weighted average algorithm to fuse the logical consistency ratio with the weight coefficients to calculate the authenticity contribution of the current agricultural management decision to the target plot. The land parcel identifier, timestamp, management behavior code, and authenticity contribution are concatenated into a string in a preset order. The concatenated string is then hashed using a hash function to generate a unique spatiotemporal fingerprint of the land parcel-level management behavior. The hashing process uses a salted hash algorithm, where the concatenated original string is connected to a randomly generated salt value and then input into the hash function to ensure that the output spatiotemporal fingerprint is collision-resistant and unique.

[0032] In this embodiment, the traceability report includes plot identifiers, agricultural management decision content, corresponding authenticity contribution, generated spatiotemporal fingerprints, logical consistency verification results, and key environmental indicators. The traceability report displays the plot status evolution trend and decision traceability path in a visual structure, and supports user query, verification, and export.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical agricultural planting area. To address issues such as missing records of management activities across multiple plots, difficulty in verifying data traceability, and delayed agricultural decision-making responses, the area was divided into 350 plot grid units, each covering approximately 0.5 hectares of land. Crop types included rice, corn, and tomatoes. Sensors were deployed to collect data on soil moisture, temperature, meteorological factors, canopy light intensity, image information, and operational activities for three consecutive months, with a sampling frequency of four times per day.

[0034] The collected multi-source agricultural environmental data is input into the system's standardized processing module and uniformly converted into a data format with consistent units, time alignment, and missing data completion, forming a standardized agricultural data set. Through the improved DiffPool module, the agricultural graph structure is continuously aggregated in time and space and graph convolution propagation is performed to construct a multi-scale neighborhood collaborative agricultural feature graph, fully exploring the inherent coupling relationship between plots in the time and space dimensions.

[0035] The feature map is mapped to a predefined agricultural management ontology structure to generate a spatial-temporal semantic tensor. This tensor is then input into a plot-level planting management twin model to perform virtual evolution simulation and output the expected crop growth state. The expected results output by the model are compared with the standardized agricultural data measured in the next time step for consistency. Based on the verification results, the authenticity contribution of each agricultural management behavior is calculated. Combining the plot identifier and timestamp, a unique spatiotemporal fingerprint is generated through hashing.

[0036] The system automatically encapsulates agricultural management decision-making results, their contribution to authenticity, and the generated spatiotemporal fingerprints, and writes them into the traceability database, ultimately generating an agricultural production traceability report. This report not only includes all parameters and criteria from the decision-making process but also comes with complete traceability documentation for use in agricultural supervision and industry traceability. To verify the system's effectiveness, a comparison was made with traditional manual records, focusing on analyzing indicators such as traceability accuracy, anomaly detection capability, and data coverage.

[0037] During 90 consecutive days of application, compared with traditional recording methods, the accuracy of real operation event identification of the system of this invention increased from 78.3% to 93.5%; the abnormal data identification rate increased from 65.1% to 88.7%; the spatiotemporal integrity rate of plot-level data increased from 58.4% to 96.2%; the rate of traceable key management behaviors from the generated traceability reports reached over 94%; and the consistency of traceability results for similar management behaviors across multiple plots and time periods reached 92.8%, verifying the system's strong robustness and traceability capabilities in heterogeneous data environments.

[0038] Table 1: Performance Comparison of the Invention System and Traditional Methods The comparison results show that the system of this invention outperforms traditional manual recording methods in several key performance indicators. Regarding the accuracy of management behavior recognition, the system improves the accuracy of recognition from 78.3% of traditional methods to 93.5% through unified modeling and feature extraction of multi-source heterogeneous data, reducing the occurrence of missed and false alarms. In terms of abnormal data recognition rate, this invention employs multi-scale graph structure analysis and feature stability measurement mechanisms, enabling the system to more sensitively capture abrupt changes and inconsistencies in the data, thereby increasing the recognition rate from 65.1% to 88.7% and enhancing the ability to monitor data anomalies.

[0039] Regarding the spatiotemporal completeness of land parcel-level data, this system utilizes an improved DiffPool structure to effectively aggregate node information and preserve features in the graph structure, alleviating data gaps caused by missing or asynchronous acquisition. This increases the completeness rate from 58.4% to 96.2%, providing a more continuous and realistic spatiotemporal data foundation for modeling and analysis. In terms of the proportion of traceable management behaviors, the system improves the coverage of traceable events based on the evaluation mechanism for the authenticity contribution of management behaviors and the hash fingerprint generation strategy, expanding the traceability of behaviors from 71.6% to 94.1%.

[0040] To ensure consistency in identifying management behaviors across plots and time periods, this system constructs a unified agricultural semantic tensor structure to achieve standardized matching at the semantic level, improving the consistency of management results across plots from 64.2% in the traditional method to 92.8%, and enhancing the system's generalization ability in different application scenarios. In terms of processing efficiency, this system can automatically process an average of 4,400 data entries per day, more than twice that of traditional manual recording, significantly shortening the work cycle. In terms of the traceability report generation cycle, the system has shortened it from the traditional 12 hours to 30 minutes, improving the real-time response capability of agricultural management.

[0041] In summary, the system of the present invention demonstrates significant advantages in terms of recognition accuracy, data integrity, anomaly detection capability, result consistency, and processing efficiency, fully verifying its application value and promotion prospects in actual agricultural production scenarios.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A plot-level planting management traceability system based on machine learning, characterized in that, include: The data acquisition and preprocessing module collects multi-source environmental data within agricultural plots and performs spatiotemporal layering processing to generate a standardized agricultural dataset. The DiffPool model processing module inputs a standardized agricultural dataset into the improved DiffPool model, employs adaptive pooling depth and spatiotemporally aware convolution processing, and outputs a neighborhood collaborative agricultural feature map. The agricultural management ontology mapping module maps neighborhood collaborative agricultural feature maps to a predefined agricultural management ontology structure, and constructs a space-time semantic tensor. The inter-plot association feature extraction module extracts spatiotemporal association features between plots based on the spatial-temporal semantic tensor, constructs a plot-level planting management twin model, and generates the expected growth baseline flow. The reinforcement learning inference module takes the spatial-temporal semantic tensor and the expected growth baseline flow as input, performs dynamic node selection and multi-objective optimization, and generates agricultural management decisions. The virtual evolution and consistency verification module inputs agricultural management decisions into a plot-level planting management twin model for virtual evolution, performs logical consistency verification, and generates a spatiotemporal fingerprint. The traceability certificate generation module encapsulates agricultural management decisions, authenticity contribution, and spatiotemporal fingerprints into spatiotemporally encrypted traceability certificates, which are then written into the traceability database. The traceability report generation module generates agricultural production traceability reports based on traceability certificates, which users can query and verify.

2. The plot-level planting management traceability system based on machine learning according to claim 1, characterized in that, The modules are connected in the following way: Collect multi-source environmental data within agricultural plots, perform spatiotemporal stratification processing, and generate a standardized agricultural dataset. The standardized agricultural dataset is input into the improved DiffPool model, which is processed using adaptive pooling depth and spatiotemporally aware convolution to output a neighborhood collaborative agricultural feature map. Map the feature map of neighborhood collaborative agriculture to a predefined agricultural management ontology structure to construct a space-time semantic tensor. Based on the spatial-temporal semantic tensor, spatiotemporal correlation features between plots are extracted, a plot-level planting management twin model is constructed, and the expected growth baseline flow is generated using the plot-level planting management twin model. Using the spatial-temporal semantic tensor and the expected growth baseline flow as input, a reinforcement learning inference structure is employed to perform dynamic node selection and multi-objective optimization, generating agricultural management decisions that include irrigation amount, fertilizer amount and pest and disease control amount; Agricultural management decisions are input into a plot-level planting management twin model for virtual evolution. The evolutionary prediction features are logically consistent with the measured standardized agricultural data at the next moment, the authenticity contribution is calculated, and a spatiotemporal fingerprint is generated. Agricultural management decisions, authenticity contributions, and spatiotemporal fingerprints are encapsulated as spatiotemporally encrypted traceability credentials, written into the traceability database, and agricultural production traceability reports are output.

3. The plot-level planting management traceability system based on machine learning according to claim 2, characterized in that, The multi-source environmental data includes soil moisture, soil temperature, meteorological parameters, canopy illumination, image information, and operational behavior data; the spatiotemporal stratification processing includes performing spatial grid division based on plot number, performing time period division based on planting cycle, completing spatiotemporal registration of agricultural sensing data, performing normalization and scale unification on the spatiotemporally registered agricultural sensing data, and generating a standardized agricultural data set.

4. The plot-level planting management traceability system based on machine learning according to claim 3, characterized in that, The generation process of the neighborhood collaborative agriculture feature map includes: Extract the multi-dimensional attribute sequence of each agricultural plot grid cell in the standardized agricultural dataset within a continuous time window, and construct an initial agricultural graph structure. In the initial agricultural graph structure, the nodes are agricultural plot grid cells, the node features are the time-series data vectors of the corresponding grid, and the edges are the spatial adjacency relationships between adjacent grids. Calculate the local feature change rate and spatial gradient of each node, take the plot feature difference degree and spatial coupling degree as input parameters, generate the first layer of binary pooling mask, and automatically divide the initial agricultural map structure into several candidate plot clusters; Within each candidate plot cluster, the information entropy is used to measure the node distribution density, the pooling factor of each plot cluster is calculated, and graph pooling operation is performed to fuse the node features within the same plot cluster into a cluster center feature vector, generating the first layer of nested graph structure. In each nested graph structure, a spatiotemporally aware convolutional kernel is embedded, and a similarity change function is used to calculate the feature correlation between each node and its neighboring nodes in the historical time series. Based on the correlation calculation results, the spatial edge weights are dynamically adjusted to achieve the time dynamic modulation of local spatial relationships. During graph convolution propagation, the location of intervention nodes corresponding to operational behavior data in the standardized agricultural dataset is identified. A Gaussian weighted window is constructed at the intervention time through linear interpolation. Weighted propagation of feature information is performed within the Gaussian weighted window to preserve the traces of the impact of management behavior on the state of the plot. The pooling and convolution operations are repeated until the information density fluctuation rate of the global parcel cluster is less than the preset threshold. The information density fluctuation rate is determined by calculating the ratio of the average mutual information difference between the feature sets of the current layer and the previous layer to the average mutual information of the previous layer. By performing multi-scale connections on the output node representations of each level in the multi-layer nested graph structure, the final neighborhood collaborative agriculture feature map is constructed.

5. A plot-level planting management traceability system based on machine learning according to claim 4, characterized in that, The process of constructing the spatial-temporal semantic tensor includes: constructing an agricultural ontology structure map containing crop type, soil type, application method, growth stage, plot location, and management entity; matching corresponding agricultural ontology concept nodes and relation edges based on the position, feature vector, and timestamp information of agricultural plot grid units in the neighborhood collaborative agriculture feature map, performing semantic alignment operations, and generating a semantic mapping relationship set; based on the semantic mapping relationship set, expanding the hierarchical concept relationships of the matched nodes in the agricultural ontology map into an ontology nested structure, constructing a three-dimensional coordinate mapping function using the spatial adjacency relationship and time series index between nodes, and establishing a spatial-temporal projection between the agricultural ontology nested structure and the neighborhood collaborative agriculture feature map; and based on the projection result, integrating the semantic labels of the ontology nested structure with the plot temporal features in the neighborhood collaborative agriculture feature map to construct a spatial-temporal semantic tensor containing semantic, spatial, and temporal dimensions.

6. A plot-level planting management traceability system based on machine learning according to claim 5, characterized in that, The construction process of the plot-level planting management twin model includes: Extract the slice vector of each agricultural plot grid cell from the spatial-temporal semantic tensor as a semantic state vector; Calculate the cosine distance between semantic state vectors of adjacent agricultural plot grid cells, and construct a grid state transition graph with grid cells as nodes and cosine distance as edge weights; Graph encoding is performed on the grid state transition graph to aggregate the historical semantic state of nodes and the state differences of neighboring nodes. After processing by a fully connected layer, the evolution feature vector of each plot at the current time step is generated. The operation behavior data is acquired and transformed into behavior influence vectors. The behavior influence vectors are concatenated with the evolution feature vectors and input into a Siamese comparison network to generate predicted land parcel state trajectories. In the twin contrast network, the state trajectory of the same plot in a continuous time period is used as a positive sample pair, and the state trajectory of plots with different operation behavior data but spatial distance within a preset range is used as a negative sample pair. A spatiotemporal consistency contrast loss function is constructed, and the parameters of the twin contrast network are optimized by gradient descent algorithm. The latest semantic state vector of the target plot and the operation behavior vector to be verified are input into the optimized Siamese contrast network to infer and generate the expected growth baseline flow of the target plot.

7. A plot-level planting management traceability system based on machine learning according to claim 6, characterized in that, The specific steps for generating agricultural management decisions are as follows: The semantic state vector of the agricultural plot grid cell is concatenated with the corresponding expected growth baseline flow in multiple dimensions and used as a joint state input to construct the reinforcement learning state space. Using each agricultural plot grid unit as the set of decision nodes, the policy probability distribution of each decision node is calculated based on the current reinforcement learning state space. The decision nodes are sorted from high to low according to their policy probability values, and the top P decision nodes are extracted as the subset of target plot nodes currently participating in the decision. For each node in the target plot node subset, a multi-dimensional continuous action space is constructed, and the action vector includes irrigation amount adjustment value, fertilizer amount adjustment value, and pest and disease control amount adjustment value. The action vector is input into the plot-level planting management twin model, and the predicted plot state trajectory under the decision-making action is output. In the time dimension, the predicted plot state trajectory is gradually subtracted from the expected growth baseline flow, the state deviation value at each time step is calculated, and the state transition result is generated. Based on the state transition results, a multi-objective reward vector is constructed. A weighted summation operation is performed on the growth deviation penalty term, resource input cost term, and spatial coordination consistency term to obtain the single-step decision reward. Based on the single-step decision reward, the policy parameters in the reinforcement learning inference structure are updated using the policy gradient algorithm. The node selection and action generation process is repeated until the cumulative reward converges, and the final combination of irrigation amount, fertilizer amount and pest and disease control amount is output as the agricultural management decision.

8. A plot-level planting management traceability system based on machine learning according to claim 7, characterized in that, The specific steps of the logical consistency verification include: The agricultural management decisions are input into the plot-level planting management twin model, virtual evolution calculations are performed, and the predicted plot state feature vector corresponding to the next time step is output. Collect multi-source environmental data of the corresponding agricultural plots at the next time step, perform preprocessing, generate measured standardized agricultural data at the next time step, and transform it into a measured plot state feature vector; Perform a one-to-one mapping between the predicted land parcel state feature vector and the measured land parcel state feature vector in the feature dimension to construct a set of state feature aligned vectors; Extract the corresponding dimension components from the state feature alignment vector set and perform subtraction to obtain the difference sequence of each feature dimension. Perform normalization processing based on the preset dimensional scaling factor to generate a logical deviation vector. Extract the component values ​​of each dimension in the logical deviation vector, determine whether they are synchronously within the pre-set consistency threshold window for each dimension, and generate a Boolean logical consistency result. Using logical consistency results as input for calculating the contribution of authenticity, we can characterize the logical credibility of agricultural management decisions at the current time step.

9. A plot-level planting management traceability system based on machine learning according to claim 8, characterized in that, The specific steps for calculating the authenticity contribution and generating the spatiotemporal fingerprint include: Extract the logical consistency result vector, count the number of elements with a Boolean value of "true" in the vector, and calculate the ratio of the number of elements with a value of "true" to the total dimension of the vector to obtain the logical consistency ratio at the current time step. Obtain the weight coefficients corresponding to the crop type, growth stage, and environmental sensitivity level of the plot, and use a weighted average algorithm to fuse the logical consistency ratio with the weight coefficients to calculate the authenticity contribution of the current agricultural management decision to the target plot. The land parcel identifier, timestamp, management behavior code, and authenticity contribution are concatenated into a string in a preset order. A hash function is then used to perform hash processing on the concatenated string to generate a unique spatiotemporal fingerprint of the land parcel-level management behavior.

10. A plot-level planting management traceability system based on machine learning according to claim 9, characterized in that, The traceability report includes plot identification, agricultural management decision content, corresponding authenticity contribution, generated spatiotemporal fingerprint, logical consistency verification results, and key environmental indicators. The traceability report displays the plot status evolution trend and decision traceability path in a visual structure, and supports user query, verification, and export.

Images