A smart agricultural data processing method based on big data analysis

Abstract

The application discloses a kind of wisdom agricultural data processing methods based on big data analysis, it is related to big data analysis technical field, including: acquisition agricultural production integrated data, pre-processing obtains agricultural multi-source time series dataset;Cross-source consistency correction and space-time continuity verification are carried out, and are converted into fuzzy state data;Fuzzy association rule mining is utilized, and agricultural fuzzy association rule base is obtained;Characteristic enhancement is carried out, and enhanced agricultural time series dataset is obtained;Improved SegRNN model is constructed, and agricultural state prediction result is obtained;Joint verification is carried out, and comprehensive agricultural condition state is formed;Wisdom agricultural data processing result is generated, and farm scheduling decision is output.The application realizes the unified management of multi-source heterogeneous agricultural data and the accurate prediction of agricultural state by the collaborative processing of fuzzy association rule mining and improved SegRNN model.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a smart agriculture data processing method based on big data analytics. Background Technology

[0002] As agricultural production moves towards digitalization, intelligence, and precision, smart agriculture is increasingly deploying a large number of sensors, image acquisition devices, agricultural machinery operation terminals, and agricultural platforms to collect data on the agricultural environment, soil moisture, crop growth, pests and diseases, agricultural operations, and historical yields. Existing smart agriculture data processing methods can typically store, statistically analyze, and display data from a single source or a small number of data types. However, agricultural production data is characterized by dispersed sources, inconsistent sampling frequencies, significant differences in data formats, complex spatial relationships, and obvious temporal variations. This makes it difficult to collaboratively process different data under a unified timeline, unified plot dimension, and unified crop growth stage. Agricultural data collection is susceptible to sensor drift, communication interruptions, environmental disturbances, and errors in manual recording, leading to missing, anomaly, duplication, and low reliability issues, thus reducing the accuracy and stability of the analysis results.

[0003] Existing agricultural data analysis methods largely rely on fixed rules, empirical thresholds, or conventional prediction models, failing to adequately explore the dynamic correlations between agricultural indicators. This makes it difficult to accurately identify the continuous evolutionary relationships between soil, weather, crop growth, pests and diseases, and agricultural operations. Due to the distinct stages, lags, and regional differences in agricultural production processes, changes in agricultural indicators often do not immediately manifest as changes in crop status. Instead, they are gradually transmitted through multiple time segments and related factors. Existing methods struggle to extract stable and interpretable agricultural correlation rules from complex data. In crop growth status prediction and agricultural risk early warning, traditional time-series models typically process data in a fixed chronological order, lacking effective handling of redundant time steps within segments, hidden state error propagation, and long-term prediction error accumulation. This makes it difficult to simultaneously address short-term anomalies and long-term growth trends, resulting in untimely, inaccurate, and uninterpretable identification of changes in crop growth, soil moisture, pest and disease risk, and yield trends.

[0004] Therefore, how to provide a smart agriculture data processing method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a smart agriculture data processing method based on big data analytics. This invention comprehensively utilizes multi-source agricultural data governance, fuzzy association rule mining, and an improved SegRNN model for data analysis and intelligent prediction. It details the entire process from collecting comprehensive agricultural production data from agricultural production areas, through preprocessing, crop condition fingerprint generation, fuzzy association rule mining, feature enhancement, agricultural condition prediction, rule joint verification, to agricultural decision output. In terms of data processing and model structure, it innovatively employs rotating grid coupling filtering, dynamic potential domain partitioning, time-crossing minimum rearrangement, hidden state diffusion, and exponentially hierarchical parallel extrapolation. This achieves unified processing of multi-source heterogeneous agricultural data, dynamic crop condition correlation mining, accurate crop condition identification, and agricultural production decision generation. Compared with existing technologies, this invention has advantages such as strong data fusion capabilities, accurate crop condition correlation identification, high stability of time-series prediction, strong interpretability of decision results, and outstanding adaptability to complex agricultural production scenarios.

[0006] A smart agriculture data processing method based on big data analysis according to an embodiment of the present invention includes: Collect comprehensive agricultural production data within the agricultural production area, preprocess the comprehensive agricultural production data, and obtain a multi-source time-series agricultural dataset; Cross-source consistency correction and spatiotemporal continuity verification are performed on agricultural multi-source time series datasets to generate an agricultural status matrix. Based on the dynamic characteristics of agricultural conditions in the agricultural status matrix, an agricultural status fingerprint is formed, and the agricultural status fingerprint is converted into fuzzy status data. By utilizing fuzzy association rule mining, agricultural relationship aggregation is performed on fuzzy state data. Candidates are extracted using rotating grid coupling screening, and fragment relationships are analyzed through dynamic segmentation of the potential domain to obtain an agricultural fuzzy association rule library. Based on the agricultural fuzzy association rule base, feature enhancement is performed on the agricultural multi-source time series dataset. The association rule results are transformed into agricultural condition status labels, risk category labels, and segment labels to obtain the enhanced agricultural time series dataset. An improved SegRNN model is constructed to integrate segment features of the enhanced agricultural time series dataset. The time step order is adjusted based on the minimum time crossover. The hidden state diffusion solution is introduced to eliminate state errors. The prediction is generated by exponentially hierarchical parallel extrapolation. The agricultural state prediction results are obtained. The agricultural status prediction results are jointly verified with the agricultural fuzzy association rule base to determine the main related factors corresponding to the changes in agricultural status and form a comprehensive agricultural status. Based on the comprehensive agricultural situation, smart agriculture data processing results are generated, and irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions corresponding to agricultural production are output.

[0007] Optionally, the comprehensive agricultural production data includes agricultural environmental data, soil moisture data, crop growth data, pest and disease data, agricultural operation data, and historical yield data.

[0008] Optionally, the preprocessing of comprehensive agricultural production data includes data cleaning, missing data completion, anomaly removal, time alignment, unit unification, and index construction.

[0009] Optionally, the process of obtaining the agricultural multi-source time-series dataset includes: Collect comprehensive agricultural production data according to the plot number, sensor node number, crop type and growth stage within the agricultural production area, and record the data source, collection time, spatial location and data attributes of each data point to form the original agricultural data table; Data cleaning was performed on the original agricultural data table, deleting duplicate records and records with missing fields. Missing fields in continuous time series with data collection interruptions were filled in. Anomalies were removed based on the calibration range of the sensor equipment, the variation range of adjacent sampling, and the distribution of data from the same period of similar plots. The cleaned data was time-aligned according to a uniform sampling time interval, and the data of different dimensions were processed to unify the dimensions. A data index was constructed based on the plot number, crop type, growth stage, collection time and data source to obtain an agricultural multi-source time series dataset.

[0010] Optionally, the conversion of agricultural condition fingerprints into fuzzy condition data includes: Agricultural baseline units are constructed based on the same plot of land, the same crop type, the same growth stage, and the same collection period. Data from multiple agricultural time series datasets are assigned to the corresponding agricultural baseline units. Data from different sources of the same agricultural indicator within the same agricultural baseline unit are recorded as source data groups. Cross-source consistency correction is performed on the source data group. The source data of the same agricultural indicator within the same agricultural baseline unit are sorted according to their numerical values, and the median value of the sort is taken as the correction value of the agricultural indicator. When the number of source data is even, the arithmetic mean of the two middle values ​​is taken as the correction value of the agricultural indicator. After cross-source consistency correction, the spatiotemporal continuity of agricultural indicators is verified. The absolute value of the time difference between adjacent collection periods for the same agricultural indicator is calculated, and the absolute value of the spatial difference between adjacent plots for the same agricultural indicator is calculated. An agricultural status matrix is ​​generated according to the correspondence between agricultural benchmark units and agricultural indicators. Based on the agricultural condition status matrix, dynamic features of agricultural conditions are extracted. The direction of change, number of consecutive changes and cross-source response order of agricultural indicators within the same agricultural condition benchmark unit are combined to form an agricultural condition status fingerprint. Based on the value range of agricultural indicators corresponding to the agricultural condition status fingerprint, the agricultural condition status fingerprint is converted into fuzzy state data.

[0011] Optionally, the process of obtaining the agricultural fuzzy association rule base includes: The fuzzy state data is generated into a fuzzy state grid according to agricultural indicators, plot numbers and collection time. The grid cell value of the fuzzy state grid is the membership degree of the agricultural state fingerprint in the corresponding fuzzy state. Candidate selection is performed using a rotating mesh coupling screening method. Taking the fuzzy state mesh of agricultural indicators as the processing object, each fuzzy state mesh is rotated and aligned according to a preset angle sequence, and the candidate score is calculated. The agricultural situation potential domain is constructed based on the fuzzy state grid. The potential domain value is calculated for each agricultural situation reference unit. The segment relationship is analyzed by dynamic segmentation of the potential domain. The agricultural situation segments are divided according to the continuous change relationship of the potential domain value in the time and space directions. The intra-segment consistency, inter-segment mutation and inter-segment transmission direction of each agricultural situation segment are calculated to obtain the key agricultural situation segments and continuous agricultural situation evolution segments. Using candidate association terms as the combined source of rule antecedents and consequents, and key agricultural condition segments and continuous agricultural condition evolution segments as the rule generation range, fuzzy support and fuzzy confidence are calculated to obtain an agricultural fuzzy association rule base.

[0012] Optionally, the obtained enhanced agricultural time-series dataset includes: Read the rule antecedent, rule consequent, fuzzy support, fuzzy confidence and corresponding agricultural condition segments from the agricultural fuzzy association rule base, match the fuzzy terms in the rule antecedent with the agricultural condition benchmark units in the agricultural multi-source time series dataset, and determine the matching rules; Extract the association rule results according to the matching rules, take the agricultural indicator combination corresponding to the antecedent of the matching rule as the association indicator combination, take the agricultural status corresponding to the consequent of the matching rule as the rule output result, and write the rule output result into the corresponding agricultural condition benchmark unit. Based on the rule output results, generate crop condition status labels and risk category labels. Convert the rule consequents representing crop growth, soil moisture, pest and disease risk, and yield trend into crop condition status labels, and convert the rule consequents representing normal, mild risk, moderate risk, and severe risk into risk category labels. Based on the key agricultural condition segments and continuous agricultural condition evolution segments corresponding to the matching rules, segment labels are generated. The associated indicator combinations, agricultural condition status labels, risk category labels and segment labels are added to the corresponding time positions in the agricultural multi-source time series dataset to obtain the enhanced agricultural time series dataset.

[0013] Optionally, obtaining the agricultural state prediction results includes: An improved SegRNN model is constructed, which includes time-interleaved minimum rearrangement, hidden state diffusion solution, and exponentially ordered parallel extrapolation. Time-interleaved minimum rearrangement is set between the segmented input layer and the recurrent unit, hidden state diffusion solution is set between the output of the recurrent unit and the inter-segment state transfer end, and exponentially ordered parallel extrapolation is set at the prediction output end. The time step order is adjusted based on the minimum time crossover rearrangement. The enhanced agricultural time series dataset is divided into agricultural time series segments according to the segment label. The crossover cost is calculated for the segment input vectors of any two time steps within the same agricultural time series segment. The crossover cost matrix is ​​constructed based on the crossover cost between all time steps. The time step rearrangement sequence within the segment is determined according to minimizing the sum of the crossover costs of adjacent processing time steps. The hidden state diffusion solution is introduced to resolve the state error. The time step rearrangement sequence within the segment is input into the cyclic unit to obtain the rearranged hidden state sequence. Each hidden state in the rearranged hidden state sequence is used as a state node. The state difference between any two state nodes is calculated. A state diffusion relation matrix is ​​generated based on the state difference. The state diffusion relation matrix is ​​used to iteratively diffuse and update each state node to obtain the diffused hidden state sequence. The prediction generation is performed by exponential hierarchical parallel extrapolation. The segment state transition matrix is ​​calculated based on the terminal diffusion hidden state of continuous agricultural time series segments. The segment state transition matrix is ​​multiplied continuously in hierarchical order to obtain the hierarchical transition matrix. The terminal diffusion hidden state of the latest agricultural time series segment is extrapolated with the hierarchical transition matrix in parallel to obtain the agricultural state prediction result. To train the improved SegRNN model, historical samples from the enhanced agricultural time-series dataset are used as training inputs, and the actual agricultural state results corresponding to future collection periods are used as training labels. The training loss is calculated based on the agricultural state prediction results and the actual agricultural state results. The model parameters of the improved SegRNN model are updated in reverse according to the training loss until the training loss meets the convergence condition, thus obtaining the trained improved SegRNN model.

[0014] Optionally, the formation of a comprehensive agricultural situation includes: The agricultural status prediction results are archived according to plot number, crop type, growth stage, prediction period and segment label to obtain prediction verification units; Based on the prediction verification unit, the association rules corresponding to the prediction results of agricultural state are retrieved from the agricultural fuzzy association rule base to obtain the matching association rule group; Consistency calculation is performed on the matching association rule group. The fuzzy support and fuzzy confidence of each matching association rule are multiplied to obtain the rule matching quantity. The rule matching quantities of all matching association rules in the same prediction verification unit are summed to obtain the prediction rule consistency quantity. The rule matching quantities of the same association factor in different matching association rules are accumulated to obtain the contribution quantity of the association factor. The main related factors are determined in descending order of their contribution. The agricultural status prediction results, the consistency of the prediction rules, and the main related factors are combined to generate a comprehensive agricultural status.

[0015] Optionally, the generation of smart agriculture data processing results includes: Read the comprehensive agricultural situation status and generate agricultural decision input units according to plot number, crop type, growth stage, forecast period, agricultural status category, status change direction, risk level and main related factors; Based on the soil moisture status, crop growth status, pest and disease risk status and yield trend status in the agricultural decision input unit, irrigation decision, fertilization decision, plant protection decision and agricultural scheduling decision are generated respectively. The risk level value, prediction time urgency value and related factor contribution of each decision within the same plot are added together to obtain the decision priority. The decision execution order is determined by prioritizing decisions from highest to lowest. The decision execution order, corresponding plot number, crop type, growth stage, forecast period, main related factors, and comprehensive agricultural status are written into the smart agriculture data processing results, and irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions corresponding to agricultural production are output.

[0016] The beneficial effects of this invention are: This invention improves the standardized processing capabilities and analytical quality of multi-source heterogeneous agricultural data by uniformly collecting, preprocessing, performing cross-source consistency correction, and verifying spatiotemporal continuity of comprehensive agricultural production data. Compared with existing methods relying on single data sources, manual experience analysis, or simple statistical processing, this invention can uniformly transform agricultural environmental data, soil moisture data, crop growth data, pest and disease data, agricultural operation data, and historical yield data into a multi-source time-series agricultural dataset. Furthermore, it generates an agricultural condition matrix and agricultural condition fingerprint, enabling collaborative processing of data from different sources, with different frequencies and scales, under unified plots, crop stages, and collection periods. This enhances the comparability, continuity, and usability of agricultural data. By employing rotating grid coupling filtering and dynamic potential domain segmentation in fuzzy association rule mining, this invention can extract candidate association terms and key agricultural condition fragments across time and indicators from fuzzy state data. This enhances the ability to identify dynamic relationships between soil, weather, crop growth, pests and diseases, and agricultural operations, making the agricultural fuzzy association rule base better reflect the actual agricultural condition evolution process and providing stable and interpretable rule basis for predictive analysis and agricultural decision-making.

[0017] This invention improves the accuracy, stability, and real-time performance of crop growth status identification, soil moisture prediction, pest and disease risk assessment, and yield trend analysis by constructing an improved SegRNN model to integrate segment features, adjust time step order, resolve state errors, and generate parallel predictions for enhanced agricultural time-series datasets. The time-crossing minimum rearrangement reduces the interference of redundant time steps and outlier sampling points within segments on cyclic computation, making the processing order of agricultural time-series segments more consistent with the changing patterns of agricultural conditions. Hidden state diffusion solution reduces the impact of hidden state error propagation on prediction results while preserving information on crop growth changes, slow soil moisture decline, and weak changes in early pest and disease risks. Exponential hierarchical parallel extrapolation reduces the accumulation of gradual errors in long-term predictions, improving the consistency of prediction results across multiple future data collection periods. By jointly validating the agricultural state prediction results with an agricultural fuzzy association rule base, this invention can determine the main correlation factors corresponding to changes in agricultural state and form a comprehensive agricultural state, thereby outputting irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions. This invention overcomes the problems of unstable agricultural data quality, insufficient correlation mining, weak interpretability of prediction results, and insufficient real-time decision-making in the prior art, and improves the precision, automation and intelligence of smart agricultural production management. Attached Figure Description

[0018] 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 flowchart of a smart agriculture data processing method based on big data analysis proposed in this invention; Figure 2 This is a structural block diagram of fuzzy association rule mining in a smart agriculture data processing method based on big data analysis proposed in this invention. Figure 3 This is a functional diagram of the improved SegRNN model for a smart agriculture data processing method based on big data analysis proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A smart agriculture data processing method based on big data analytics includes: Collect comprehensive agricultural production data within the agricultural production area, preprocess the comprehensive agricultural production data, and obtain a multi-source time-series agricultural dataset; Cross-source consistency correction and spatiotemporal continuity verification are performed on agricultural multi-source time series datasets to generate an agricultural status matrix. Based on the dynamic characteristics of agricultural conditions in the agricultural status matrix, an agricultural status fingerprint is formed, and the agricultural status fingerprint is converted into fuzzy status data. By utilizing fuzzy association rule mining, agricultural relationship aggregation is performed on fuzzy state data. Candidates are extracted using rotating grid coupling screening, and fragment relationships are analyzed through dynamic segmentation of the potential domain to obtain an agricultural fuzzy association rule library. Based on the agricultural fuzzy association rule base, feature enhancement is performed on the agricultural multi-source time series dataset. The association rule results are transformed into agricultural condition status labels, risk category labels, and segment labels to obtain the enhanced agricultural time series dataset. An improved SegRNN model is constructed to integrate segment features of the enhanced agricultural time series dataset. The time step order is adjusted based on the minimum time crossover. The hidden state diffusion solution is introduced to eliminate state errors. The prediction is generated by exponentially hierarchical parallel extrapolation. The agricultural state prediction results are obtained. The agricultural status prediction results are jointly verified with the agricultural fuzzy association rule base to determine the main related factors corresponding to the changes in agricultural status and form a comprehensive agricultural status. Based on the comprehensive agricultural situation, smart agriculture data processing results are generated, and irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions corresponding to agricultural production are output.

[0021] In this embodiment, the comprehensive agricultural production data includes agricultural environmental data, soil moisture data, crop growth data, pest and disease data, agricultural operation data, and historical yield data.

[0022] In this embodiment, the preprocessing of comprehensive agricultural production data includes data cleaning, missing data completion, anomaly removal, time alignment, unit unification, and index construction.

[0023] In this embodiment, obtaining the agricultural multi-source time-series dataset includes: Collect comprehensive agricultural production data according to the plot number, sensor node number, crop type and growth stage within the agricultural production area, and record the data source, collection time, spatial location and data attributes of each data point to form the original agricultural data table; Data cleaning was performed on the original agricultural data table, deleting duplicate records and records with missing fields. Missing fields in continuous time series with data collection interruptions were filled in. Anomalies were removed based on the calibration range of the sensor equipment, the variation range of adjacent sampling, and the distribution of data from the same period of similar plots. The cleaned data was time-aligned according to a uniform sampling time interval, and data of different dimensions were standardized. A data index was constructed based on plot number, crop type, growth stage, collection time, and data source to obtain an agricultural multi-source time-series dataset. Specifically, the data index was constructed based on plot number, crop type, growth stage, collection time, and data source as follows: Each plot, crop type, growth stage, and data source is assigned a unique code, and a uniform sampling interval is set to 15 minutes. Starting from 0:00 every day, the collection time of each data is assigned to the corresponding 15-minute time slice, and a time slice number is generated. A composite index key is generated in a fixed order of plot code, crop code, growth stage code, time slice number, and data source code. When there are multiple numerical data under the same composite index key, the arithmetic mean of all values ​​is calculated as the data value corresponding to the composite index key. When multiple non-numerical data points exist under the same composite index key, the data point whose collection time is closest to the center of the 15-minute time slice is selected as the record value. The cleaned data is then sorted, archived, and stored according to the composite index key, so that each piece of agricultural data can be associated with a specific plot, crop type, growth stage, collection time slice, and data source, thus obtaining a multi-source time-series agricultural dataset.

[0024] In this embodiment, converting the agricultural condition fingerprint into fuzzy state data includes: Agricultural baseline units are constructed based on the same plot of land, crop type, growth stage, and data collection period. Data from multiple agricultural time-series datasets are grouped into corresponding agricultural baseline units. Data from different sources for the same agricultural indicator within the same agricultural baseline unit are categorized as source data groups. Agricultural benchmark units are constructed based on the same plot of land, crop type, growth stage, and collection time period. Data from the agricultural multi-source time series dataset are assigned to the corresponding agricultural benchmark units. Specifically, the plot number, crop type, growth stage, and collection time slice of each record in the agricultural multi-source time series dataset are read. The collection time slice is unified to 15 minutes. Data with the same plot number, crop type, growth stage, and collection time slice are grouped together. A unique agricultural benchmark unit number is generated for the current group. When assigning data, if the collection time is within the same 15-minute time slice, it is archived according to the agricultural benchmark unit corresponding to the time slice. If the same plot of land has multiple crop types or multiple growth stages within the same time slice, different agricultural benchmark units are established for each, so that each agricultural benchmark unit corresponds to only one specific plot of land, crop type, growth stage, and collection time period. Data from different sources for the same agricultural indicator within the same agricultural baseline unit are categorized as source data groups. Specifically, within each agricultural baseline unit, data are reclassified according to the name of the agricultural indicator. Data representing the same agricultural meaning are grouped under the same agricultural indicator name. The data source identifier, collection time, data value, and data quality label corresponding to each data point are retained. For multiple numerical data points of the same agricultural indicator from the same data source, the arithmetic mean of all data values ​​is calculated as the indicator value under that source. For multiple non-numerical data points of the same agricultural indicator from the same data source, the data with the collection time closest to the center of the 15-minute time slice is selected as the indicator value under that source. The source indicator values ​​of the same agricultural indicator within the same agricultural baseline unit are summarized to form the source data group of the agricultural indicator. Cross-source consistency correction is performed on the source data group. The source data of the same agricultural indicator within the same agricultural baseline unit are sorted according to their numerical values, and the median value of the sort is taken as the correction value of the agricultural indicator. When the number of source data is even, the arithmetic mean of the two middle values ​​is taken as the correction value of the agricultural indicator. After cross-source consistency correction, the agricultural indicators are subjected to spatiotemporal continuity verification. The absolute value of the time difference between adjacent data collection periods for the same agricultural indicator is calculated, as is the absolute value of the spatial difference between adjacent plots for the same agricultural indicator. An agricultural condition status matrix is ​​generated according to the correspondence between agricultural condition benchmark units and agricultural indicators. The specific steps for generating the agricultural condition status matrix are as follows: After cross-source consistency correction, the data is sorted according to agricultural baseline units, so that each row corresponds to one agricultural baseline unit and each column corresponds to one agricultural indicator. The correction value of the corresponding agricultural indicator in the agricultural baseline unit is written into the corresponding position of the matrix. For the absolute value of the time difference, the correction value of the same agricultural indicator in the adjacent previous collection period under the same plot, the same crop type, and the same growth stage is selected. The current correction value is subtracted from the correction value of the previous collection period and the absolute value is taken. If there is no data in the current collection period, the correction value of the adjacent next collection period is used for calculation. For the absolute value of spatial difference, first determine the adjacent plots of the current plot according to the plot adjacency table, calculate the arithmetic mean of the correction values ​​of the same agricultural indicator of the adjacent plots under the same crop type, the same growth stage, and the same collection period, and take the absolute value after subtracting the correction value of the current plot from the arithmetic mean. When there is no valid data of adjacent plots of the current plot, the absolute value of spatial difference is recorded as 0. Write the correction value, absolute value of time difference and absolute value of spatial difference of each agricultural indicator into the matrix elements to form an agricultural condition status matrix that can simultaneously represent the current status of the indicator, the degree of time change and the degree of spatial difference. Based on the agricultural condition status matrix, dynamic features of agricultural conditions are extracted. The direction of change, number of consecutive changes, and cross-source response order of agricultural indicators within the same agricultural condition baseline unit are combined to form an agricultural condition status fingerprint. The agricultural condition status fingerprint is then converted into fuzzy state data based on the value range of the corresponding agricultural indicators. Specifically, the extraction of dynamic features of agricultural conditions from the agricultural condition status matrix includes: Read the agricultural indicator correction value, absolute value of time difference, and absolute value of spatial difference in the corresponding row of each agricultural baseline unit. Arrange the same agricultural indicator for the same plot, the same crop type, and the same growth stage in the order of the collection time. Compare the agricultural indicator correction value of the current collection time with the agricultural indicator correction value of the previous collection time. If the current value is greater than the previous value, it is recorded as an upward direction. If the current value is less than the previous value, it is recorded as a downward direction. If the current value is equal to the previous value, it is recorded as a stable direction. By tracing back from the current collection period to each subsequent collection period, the number of consecutive collection periods in which agricultural indicators maintain the same direction of change is counted, thus obtaining the number of consecutive changes. At the same time, the agricultural indicators corresponding to different data sources within the same agricultural baseline unit are sorted according to the order in which they change, with the data source showing the earliest change in the absolute value of the time difference ranked first. If multiple data sources change in the same collection period, they are sorted from largest to smallest by the absolute value of the time difference to obtain the cross-source response order. The direction of change, the number of consecutive changes, and the cross-source response order are used together as the dynamic characteristics of agricultural indicators.

[0025] In this embodiment, obtaining the agricultural fuzzy association rule base includes: The fuzzy state data is used to generate a fuzzy state grid based on agricultural indicators, plot numbers, and collection time. The grid cell value of the fuzzy state grid is the membership degree of the agricultural state fingerprint in the corresponding fuzzy state. The generation of the fuzzy state grid is specifically as follows: The fuzzy state data is classified according to agricultural indicators. A two-dimensional grid is established for each agricultural indicator, with the plot number as the row and the collection time as the column. Grid layers are established according to the fuzzy state type corresponding to the agricultural indicator, so that the same agricultural indicator has an independent data carrying position in different fuzzy states. The agricultural state fingerprint of each plot at each collection time is read, and the correction value, change direction, number of consecutive changes and cross-source response order of the corresponding agricultural indicator are extracted. The membership degree is calculated according to the pre-defined fuzzy state value range of the agricultural indicator. When the correction value of the agricultural indicator is located in the central interval of the corresponding fuzzy state, the membership degree is recorded as 1. When the correction value of the agricultural indicator is located in the transition interval between adjacent fuzzy states, the membership degree is calculated linearly decreasing according to the distance of the correction value from the boundary of the central interval. When the correction value of the agricultural indicator exceeds the corresponding fuzzy state value range, the membership degree is recorded as 0. If the same plot of land, the same collection time, and the same agricultural indicator have multiple sources of membership, calculate the arithmetic mean of the membership of all sources as the grid cell value, and write each grid cell value into the corresponding agricultural indicator and the corresponding fuzzy state grid layer according to the plot number and collection time to obtain the fuzzy state grid. Candidate extraction is performed using a rotated mesh coupled screening method. Taking the fuzzy state mesh of agricultural indicators as the processing object, each fuzzy state mesh is rotated and aligned according to a preset angle sequence, and candidate scores are calculated. Specifically, the rotated mesh coupled screening method for candidate extraction is as follows: From the fuzzy state grid, select the grid layers corresponding to different agricultural indicators to form the screening combination. Take one random grid layer as the reference grid layer, and rotate the remaining grid layers according to the preset angle sequence. The preset angle sequence is set to range from -45 degrees to 45 degrees with an interval of 5 degrees. When rotating, the center point of the grid is used as the rotation center. The rotated grid cells fall back into the plot row and collection time column corresponding to the reference grid layer. For each rotation angle, compare the membership degree of the reference grid layer and the rotated grid layer at the same grid position. The smaller membership degree at the same position is taken as the overlap of the grid position. The total overlap of all grid positions is obtained by summing the overlap of all grid positions. The maximum continuous length of non-zero overlap in the continuous collection time column is counted to obtain the continuous overlap time length. The number of times the same combination to be screened has a non-zero overlap in all rotation angles is counted to obtain the number of repeated overlaps. The total number of angle overlaps, the continuous overlap time length, and the number of repeated overlaps are multiplied to obtain the candidate score. The combinations to be screened are sorted from largest to smallest according to the candidate scores, and the agricultural indicator combinations with the highest ranking are determined as candidate related items. A potential domain for agricultural conditions is constructed based on a fuzzy state grid. The potential domain value is calculated for each agricultural condition reference unit. Segment relationships are analyzed through dynamic segmentation of the potential domain. Agricultural conditions are divided into segments according to the continuous changes in the potential domain value in the temporal and spatial directions. The intra-segment consistency, inter-segment mutation, and inter-segment propagation direction of each agricultural condition segment are calculated to obtain key agricultural condition segments and continuous agricultural condition evolution segments. Specifically, the key agricultural condition segments and continuous agricultural condition evolution segments are obtained as follows: The potential domain value is calculated based on the change in membership degree of each agricultural baseline unit in the fuzzy state grid. The potential domain value is obtained by summing the change in fuzzy state membership degree of the same agricultural baseline unit relative to the previous collection period, the difference relative to adjacent plots in the same collection period, and the change in agricultural indicator state relative to adjacent plots. The potential domain value is continuously traversed according to the collection time sequence and the adjacency relationship of plots. Agricultural baseline units with the same direction of change and adjacent potential domain value differences not exceeding 50% of the standard deviation of all potential domain values ​​are grouped into the same agricultural segment. For each agricultural condition segment, the intra-segment consistency, inter-segment mutation, and inter-segment transmission direction are calculated. The intra-segment consistency is calculated by dividing the number of agricultural condition reference units with the same direction of change within the segment by the total number of agricultural condition reference units within the segment. The inter-segment mutation is calculated by the absolute value of the difference between the average potential values ​​of two adjacent agricultural condition segments. The inter-segment transmission direction is determined according to the chronological order and the order of adjacent plots of potential value increases. Agricultural condition segments with an intra-segment consistency of not less than 0.7 and an average potential value of not less than the average potential value of all segments are identified as key agricultural condition segments. Segment sequences in which adjacent key agricultural condition segments have the same inter-segment transmission direction, an inter-segment mutation that is not less than the average of all inter-segment mutations, and a duration of not less than two data collection periods are identified as continuous agricultural condition evolution segments. Using candidate association terms as the combined source of rule antecedents and consequents, and key agricultural condition segments and continuous agricultural condition evolution segments as the rule generation scope, fuzzy support and fuzzy confidence are calculated to obtain an agricultural fuzzy association rule base. Specifically, the agricultural fuzzy association rule base is as follows: The agricultural indicators and fuzzy states in each candidate association item are split into rules. Fuzzy items that occur earlier or are located upstream in the inter-segment transmission direction are taken as rule antecedents, and fuzzy items that occur later or are located downstream in the inter-segment transmission direction are taken as rule consequents. The grid cells in key agricultural condition segments and continuous agricultural condition evolution segments are taken as rule statistical ranges. The membership degree of the grid cells corresponding to the rule antecedent and rule consequent is read in each statistical range. The smaller value of the membership degree of the antecedent and the membership degree of the consequent in the same grid position is taken as the rule matching value. The arithmetic mean of all rule matching values ​​is taken as the fuzzy support. The minimum membership degree of each fuzzy term in the antecedent of the rule at the same grid position is taken as the antecedent matching value. The arithmetic mean of all antecedent matching values ​​is taken as the antecedent support. The fuzzy confidence is obtained by dividing the fuzzy support by the antecedent support. Rules with fuzzy support not less than 0.2 and fuzzy confidence not less than 0.6 are retained. The antecedent, consequent, corresponding plot, corresponding collection time period, corresponding agricultural information segment, fuzzy support, and fuzzy confidence are recorded. After sorting the fuzzy confidence from largest to smallest, an agricultural fuzzy association rule base is formed.

[0026] In this embodiment, obtaining the enhanced agricultural time-series dataset includes: Read the rule antecedents, rule consequents, fuzzy support, fuzzy confidence, and corresponding agricultural condition segments from the agricultural fuzzy relational rule base. Match the fuzzy terms in the rule antecedents with the agricultural condition benchmark units in the agricultural multi-source time series dataset to determine the matching rules. Specifically, the matching rules are determined as follows: Based on the plot number, crop type, growth stage, and collection period of the current agricultural baseline unit, rules that match the agricultural condition segment to which the current agricultural baseline unit belongs are selected from the agricultural fuzzy association rule base. The names of agricultural indicators and fuzzy states in the antecedent of each rule are read one by one. The same agricultural indicators are searched in the corresponding agricultural baseline unit in the agricultural multi-source time series dataset. The membership degree of the agricultural indicators in the corresponding fuzzy states is read. When the membership degree of the antecedent fuzzy item is not less than 0.5, it is determined that the antecedent fuzzy item matches the current agricultural baseline unit. When all fuzzy items in the antecedent of the same rule are matched, the minimum value among the membership degrees of all antecedent fuzzy items is taken as the antecedent matching degree of the rule. Rules with an antecedent matching degree of not less than 0.6, and fuzzy support and fuzzy confidence of not less than 0.2 and not less than 0.6 are determined as matching rules. If multiple rules are matched for the same agricultural baseline unit, they are sorted from largest to smallest according to the result of multiplying the antecedent matching degree and fuzzy confidence, and the rule with the highest ranking is retained as the matching rule for that agricultural baseline unit. Extract the association rule results according to the matching rules, take the agricultural indicator combination corresponding to the antecedent of the matching rule as the association indicator combination, take the agricultural status corresponding to the consequent of the matching rule as the rule output result, and write the rule output result into the corresponding agricultural condition benchmark unit. Based on the rule output, crop condition status labels and risk category labels are generated. Rule consequents representing crop growth, soil moisture, pest and disease risk, and yield trends are converted into crop condition status labels, and rule consequents representing normal, mild, moderate, and severe risk are converted into risk category labels. Specifically, the generation of crop condition status labels and risk category labels based on the rule output is as follows: Read the rule consequents of the matching rules, identify the agricultural status type corresponding to the rule consequents, and classify the consequents pointing to crop growth into the crop growth label, the consequents pointing to soil moisture into the soil moisture label, the consequents pointing to pest and disease risk into the pest and disease risk label, and the consequents pointing to yield trend into the yield trend label. Calculate the label strength for each label. The label strength is the product of the membership degree, fuzzy support degree, and fuzzy confidence degree of the rule consequent. If there are multiple labels of the same type within the same agricultural condition baseline unit, select the label with the highest label strength as the agricultural condition status label. Calculate the risk score according to the risk-related rule consequent. The risk score is the arithmetic mean of the label strengths of all risk categories within the same agricultural condition baseline unit. A normal label is generated when the risk score is less than 0.25, a slightly risky label is generated when the risk score is not less than 0.25 and less than 0.5, a moderately risky label is generated when the risk score is not less than 0.5 and less than 0.75, and a severely risky label is generated when the risk score is not less than 0.75. Write the agricultural condition status label and the risk category label into the corresponding agricultural condition baseline unit. Based on the key agricultural condition segments and continuous agricultural condition evolution segments corresponding to the matching rules, segment labels are generated. The associated indicator combinations, agricultural condition status labels, risk category labels, and segment labels are then added to the corresponding time positions in the multi-source agricultural time-series dataset to obtain an enhanced agricultural time-series dataset. Specifically, the enhanced agricultural time-series dataset is as follows: Read the key agricultural condition segments and continuous agricultural condition evolution segments corresponding to each matching rule. Determine the corresponding time position of the segment in the agricultural multi-source time series dataset according to the plot number, collection start and end time and agricultural indicator range of the segment. Generate key segment labels for key agricultural condition segments and evolution segment labels for continuous agricultural condition evolution segments. Write the combination of agricultural indicators in the antecedent of the matching rule into the corresponding time position as the associated indicator combination. If there are multiple segment labels at the same time position, prioritize retaining the segment labels with higher risk categories. When the risk categories are the same, the fragment labels corresponding to the matching rules with higher fuzzy confidence are retained. The agricultural status labels, risk category labels, and fragment labels are merged with the original agricultural indicator values ​​using the same composite index key. The data are then reordered according to plot number, crop type, growth stage, and collection time to obtain an enhanced agricultural time-series dataset.

[0027] In this embodiment, obtaining the agricultural state prediction result includes: An improved SegRNN model is constructed, which includes time-interleaved minimum rearrangement, hidden state diffusion solution, and exponentially ordered parallel extrapolation. Time-interleaved minimum rearrangement is set between the segmented input layer and the recurrent unit, hidden state diffusion solution is set between the output of the recurrent unit and the inter-segment state transfer end, and exponentially ordered parallel extrapolation is set at the prediction output end. The time-step order adjustment is based on the minimum time-cross rearrangement. The enhanced agricultural time-series dataset is divided into agricultural time-series segments according to segment labels. Cross-cost is calculated for the segment input vectors of any two time steps within the same agricultural time-series segment. A cross-cost matrix is ​​constructed based on the cross-costs between all time steps. The time-step rearrangement sequence within a segment is determined by minimizing the sum of the cross-costs of adjacent processing time steps. Specifically, the time-step order adjustment based on the minimum time-cross rearrangement is as follows: The enhanced agricultural time-series dataset is divided into a set of agricultural time-series segments according to the segment labels. Each segment label corresponds to one agricultural time-series segment. The agricultural indicator values, associated indicator combinations, agricultural condition labels, and risk category labels of each collection time slice within each agricultural time-series segment are organized into a segment input vector. Any two time steps within the same agricultural time-series segment are compared pairwise, and the cross cost between the two time steps is calculated. The cross cost consists of a numerical difference term, a label difference term, and an association difference term. The numerical difference term is the average of the absolute values ​​of the normalized numerical differences of the same agricultural indicators in the two time steps. The label difference term is the number of inconsistent agricultural condition labels and risk category labels. The association difference term is the proportion of the number of different indicators in the associated indicator combination to the total number of associated indicators. The three terms are added together to obtain the cross cost between the two time steps. The cross-cost between all time steps is written into the cross-cost matrix according to the time step correspondence. Each element in the matrix represents the degree of information cross-processing when two adjacent time steps are processed. When determining the time step rearrangement sequence within the segment, the first and last time steps of the agricultural time series segment are kept unchanged. The intermediate time steps are permuted and combined for calculation. The cross-cost between adjacent processing time steps is accumulated one by one. The permutation with the smallest cumulative cross-cost is determined as the time step rearrangement sequence within the segment. When the number of intermediate time steps exceeds 12, the time step with the smallest cross-cost among the unsorted time steps is selected sequentially from the first time step to form the initial sequence. The adjacent positions are swapped and corrected until the cumulative cross-cost no longer decreases after two consecutive rounds of swapping. The time step rearrangement sequence within the segment is obtained. A hidden state diffusion solution is introduced to resolve state errors. The time-step rearrangement sequence within a segment is input into a recurrent unit to obtain a rearranged hidden state sequence. Each hidden state in the rearranged hidden state sequence is used as a state node. The state difference between any two state nodes is calculated. A state diffusion relation matrix is ​​generated based on the state difference. The state diffusion relation matrix is ​​used to iteratively diffuse and update each state node to obtain the diffused hidden state sequence. Specifically, the hidden state diffusion solution is introduced to resolve state errors as follows: The time-step rearrangement sequence within the segment is input into the recurrent unit of the improved SegRNN model. The hidden states corresponding to each time step are calculated sequentially according to the rearrangement sequence to obtain the rearranged hidden state sequence. Each hidden state in the rearranged hidden state sequence is set as a state node. The state difference is calculated for any two state nodes. The state difference is the sum of the absolute values ​​of the differences between the corresponding components in the two hidden state vectors and divided by the number of hidden state components to obtain the normalized node difference value. A state diffusion relationship matrix is ​​generated based on the state difference. The diffusion relationship value between two state nodes is obtained by dividing 1 by the sum of the state difference and 1. All diffusion relationship values ​​corresponding to the same state node are normalized so that the sum of the diffusion relationship values ​​in the same row is 1. During iterative diffusion update, each state node first reads its own hidden state from the previous round, then reads the diffusion state obtained by other state nodes in the same segment according to the state diffusion relationship matrix, and adds 70% of its own hidden state to 30% of the diffusion state as the hidden state after this round of update. This process is repeated for 5 rounds, or the iteration stops when the average change of the hidden state of all state nodes in two adjacent rounds is less than 0.001, thus obtaining the diffusion hidden state sequence. Exponentially hierarchical parallel extrapolation is used for prediction generation. The segment state transition matrix is ​​calculated based on the terminal diffusion hidden state of consecutive agricultural time series segments. The segment state transition matrices are then multiplied sequentially in hierarchical order to obtain the hierarchical transition matrix. The terminal diffusion hidden state of the latest agricultural time series segment is then extrapolated in parallel with the hierarchical transition matrix to obtain the agricultural state prediction result. Specifically, the exponentially hierarchical parallel extrapolation method is used for prediction generation. Read the terminal diffusion hidden state of continuous agricultural time series segments, arrange them in the order of collection time, calculate the segment state transition matrix for two adjacent agricultural time series segments, calculate the difference between each component of the terminal diffusion hidden state of the next agricultural time series segment and the corresponding component of the terminal diffusion hidden state of the previous agricultural time series segment, divide the difference by the absolute value of the corresponding component of the terminal diffusion hidden state of the previous agricultural time series segment and add 1 to obtain the state transition value of the component, and write all the state transition values ​​into the segment state transition matrix in the order of hidden state components. The state transition matrices of multiple segments are calculated sequentially according to time order. The state transition matrix of the first segment is multiplied by the state transition matrix of the second segment to obtain the first-order sequential transition matrix. The first-order sequential transition matrix is ​​then multiplied by the state transition matrices of subsequent segments to obtain the final sequential transition matrix. The terminal diffusion hidden state of the latest agricultural time series segment is multiplied by the sequential transition matrix corresponding to different prediction step sizes to obtain the predicted hidden state for each future collection period. The predicted hidden state is input into the prediction output terminal, which consists of a fully connected computation layer and a normalized output layer. The fully connected computation layer converts the predicted hidden state into crop growth value, soil moisture value, pest and disease risk value, and yield trend value. The normalized output layer organizes each value into the value range of the corresponding agricultural state to generate the agricultural state prediction result. The improved SegRNN model is trained by using historical samples from the enhanced agricultural time-series dataset as training input and the actual agricultural state results corresponding to future collection periods as training labels. The training loss is calculated based on the predicted and actual agricultural state results. The model parameters of the improved SegRNN model are then updated in reverse order based on the training loss until the training loss meets the convergence condition, resulting in the trained improved SegRNN model. Specifically, the training loss is calculated based on the predicted and actual agricultural state results as follows: The crop growth value, soil moisture value, and yield trend value are uniformly normalized to the range of 0 to 1. The pest and disease risk status is converted into four probability outputs: normal, mild risk, moderate risk, and severe risk. The numerical prediction loss and the category prediction loss are calculated separately. The numerical prediction loss is the square of the difference between the predicted crop growth value, predicted soil moisture value, and predicted yield trend value for each prediction collection period and the corresponding actual value. The arithmetic mean of all squared results is calculated. The category prediction loss is the negative of the natural logarithm of the predicted probability corresponding to the actual pest and disease risk category. The arithmetic mean of all prediction collection periods is then calculated. The agricultural condition label loss is calculated as follows: when the predicted agricultural condition label is inconsistent with the true label, it is recorded as 1, and when they are consistent, it is recorded as 0. The arithmetic mean of all label results is calculated. The numerical prediction loss, the category prediction loss, and the agricultural condition label loss are added together in proportions of 0.5, 0.3, and 0.2, respectively, to obtain the training loss. The convergence condition is set as follows: the decrease in training loss is less than 0.001 in 10 consecutive training rounds, or the number of training rounds reaches 200 rounds.

[0028] In this embodiment, the formation of a comprehensive agricultural situation includes: The agricultural status prediction results are archived according to plot number, crop type, growth stage, prediction period and segment label to obtain prediction verification units; Based on the prediction verification unit, the association rules corresponding to the prediction results of agricultural state are retrieved from the agricultural fuzzy association rule base to obtain the matching association rule group; Consistency calculation is performed on the matching association rule group. The fuzzy support and fuzzy confidence of each matching association rule are multiplied to obtain the rule matching quantity. The rule matching quantities of all matching association rules in the same prediction verification unit are summed to obtain the prediction rule consistency quantity. The rule matching quantities of the same association factor in different matching association rules are accumulated to obtain the contribution quantity of the association factor. The main related factors are determined in descending order of their contribution. The agricultural status prediction results, the consistency of the prediction rules, and the main related factors are combined to generate a comprehensive agricultural status.

[0029] In this embodiment, generating smart agriculture data processing results includes: Read the comprehensive agricultural situation status and generate agricultural decision input units according to plot number, crop type, growth stage, forecast period, agricultural status category, status change direction, risk level and main related factors; Based on the soil moisture status, crop growth status, pest and disease risk status, and yield trend status in the agricultural decision input unit, irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions are generated respectively. The decision priority is obtained by summing the risk level, prediction time urgency, and contribution of related factors for each decision within the same plot. Specifically, the irrigation, fertilization, plant protection, and agricultural scheduling decisions are generated as follows: The system reads soil moisture status, crop growth status, pest and disease risk status, yield trend status, risk level, and main related factors from the agricultural decision input unit. The risk level is converted into a numerical value: normal is recorded as 0, mild risk as 1, moderate risk as 2, and severe risk as 3. When the soil moisture status is low or continues to decline, and the main related factors include decreased soil moisture, insufficient rainfall, or increased sunlight, an irrigation decision is generated. The irrigation amount is calculated by multiplying the difference between the target moisture value and the current moisture value by the plot area. The target moisture value is set as the historical average suitable moisture value for the current growth stage of the crop. When the crop growth status is weak or insufficient recovery, and the main related factors include insufficient nutrients, excessively long fertilization intervals, or continuous decline in crop growth, a fertilization decision is generated. The amount of fertilizer is calculated by multiplying the difference between the target growth value and the current growth value by the crop stage fertilizer requirement coefficient. The target growth value is set as the historical average of normal growth at the current growth stage of the crop. When the pest and disease risk status is mild, moderate, or severe, and the main related factors include increased air humidity, continuous rainfall, weak light, or similar segments of historical diseases, a plant protection decision is generated. The plant protection level is determined according to the risk level value corresponding to the pest and disease risk status. When the yield trend is declining or multiple plots are in a state of moderate or higher risk at the same time, an agricultural scheduling decision is generated. The scheduling object is calculated by adding the risk level value of each plot, the urgency value of the forecast period, and the contribution of related factors. When the forecast period is no more than 1 collection period away from the current time, the urgency value is recorded as 3; when it is more than 1 but no more than 4 collection periods away, it is recorded as 2; when it is more than 4 collection periods away, it is recorded as 1. The execution order of irrigation, fertilization, plant protection and agricultural scheduling is determined according to the decision priority value from large to small. The decision execution order is determined by prioritizing decisions from highest to lowest. The decision execution order, corresponding plot number, crop type, growth stage, forecast period, main related factors, and comprehensive agricultural status are written into the smart agriculture data processing results, and irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions corresponding to agricultural production are output.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart agriculture continuous monitoring scenario. The agricultural production area was divided into 12 plots, each equipped with data collection terminals for soil moisture, agricultural environment, crop growth, pest and disease identification, and agricultural operation records. Within one monitoring cycle, a total of 1680 pieces of comprehensive agricultural production data, 360 pieces of soil moisture data, 480 pieces of agricultural environment data, 300 pieces of crop growth data, 180 pieces of pest and disease data, 160 pieces of agricultural operation data, and 200 pieces of historical yield-related data were received. The raw data exhibited issues such as inconsistent sampling intervals, duplicate reporting, partial missing data, and inconsistent recording formats. The initial missing data rate was 4.6%, the duplicate rate was 2.1%, and the anomaly rate was 1.5%.

[0031] After the data enters the processing flow, the comprehensive agricultural production data undergoes preprocessing. 36 duplicate records were deleted, and 77 missing records due to data collection interruptions were completed using the arithmetic mean of adjacent data collection periods for the same plot. When both adjacent data collection periods were missing, the average of similar indicators from adjacent plots was used. 25 abnormal records were removed based on the sensor range and the magnitude of changes in adjacent sampling. After preprocessing, 1542 valid records were obtained, reducing the missing rate to 0.4% and the abnormal rate to 0.3%.

[0032] A composite index was constructed based on plot number, crop type, growth stage, collection time, and data source. Each record was assigned to a corresponding 15-minute time slice. When multiple numerical data points exist under the same composite index key, the arithmetic mean was used as the data value; when multiple non-numerical data points exist under the same composite index key, the data point closest to the center of the time slice was used as the record value. The processed dataset forms an agricultural multi-source time-series dataset containing 1542 valid index records. Each plot forms a continuous time-series sequence, with an average of 128 records per plot.

[0033] In the agricultural baseline unit construction phase, agricultural multi-source time-series datasets were assigned to corresponding agricultural baseline units based on the same plot, crop type, growth stage, and data collection period, resulting in 1440 agricultural baseline units. For different source data of the same agricultural indicator within the same agricultural baseline unit, source data groups were established. After cross-source consistency correction, spatiotemporal continuity verification was performed, and an agricultural status matrix was generated. The average soil moisture of adjacent plots during the same period was 27.4%, with an absolute spatial difference of 2.5 percentage points. The corrected values, absolute temporal differences, and absolute spatial differences were written into the matrix elements to generate a 1440-row, 18-column agricultural status matrix.

[0034] Dynamic features of agricultural conditions were extracted based on the agricultural condition status matrix. For soil moisture sequences of the same plot, the correction values ​​of the current collection period and the previous collection period were compared to identify the direction of increase, decrease, or stability, and the number of consecutive collection periods with the same direction of change was traced back. In the simulated data, five plots showed a continuous decrease in soil moisture, with the longest consecutive decrease being six times and an average decrease of 1.0 percentage point. After combining the cross-source response sequence, an agricultural condition status fingerprint was formed and converted into fuzzy status data of suitable soil moisture, low soil moisture, weak growth, and mild risk of pests and diseases.

[0035] In the fuzzy association rule mining stage, a fuzzy state grid was generated based on agricultural indicators, plot numbers, and data collection time. Taking the low soil moisture state as an example, a 12-row, 120-column two-dimensional grid was established, with each grid cell value representing the membership degree of the corresponding plot and data collection time. At the same time, the membership degree for increased air humidity was 0.74, and the membership degree for weak crop growth was 0.68. After candidate item extraction using a rotating grid coupling screening method, the candidate score for the combination of low soil moisture and weak crop growth was 1685.4, and it was identified as a candidate association item.

[0036] In the dynamic segmentation and solution stage of the potential domain, the agricultural potential domain value is calculated based on the fuzzy state grid. Agricultural segments are then divided according to the continuous changes in the potential domain value in both time and space, resulting in 38 key agricultural segments and 16 continuous agricultural evolution segments. Using candidate association terms as the source of rule antecedents and consequents, 118 agricultural fuzzy association rules are ultimately obtained, with 87 effective rules, an average fuzzy support of 0.33, and an average fuzzy confidence of 0.73. Traditional association rule mining yields 61 effective rules with an average fuzzy confidence of 0.60.

[0037] Feature enhancement was performed on a multi-source time-series agricultural dataset based on an agricultural fuzzy association rule base. The rule consequent membership degree was 0.69, fuzzy support was 0.31, fuzzy confidence was 0.76, and label strength was 0.162. Time locations were labeled as weak crop growth and mild risk. The enhanced dataset retains 1542 sample records and adds four new enhanced fields: association indicator combinations, crop condition labels, risk category labels, and segment labels. The effective features per sample increased from 18 dimensions to 31 dimensions.

[0038] In the training phase of the improved SegRNN model, 1200 samples from the enhanced agricultural time-series dataset were used as training input, 200 samples as validation input, and 142 samples as test input. Each agricultural time-series segment contains 16 collection time steps, and each time step contains 31-dimensional input features. The prediction targets are crop growth values, soil moisture values, pest and disease risk categories, and yield trend values ​​for the next 8 collection time steps. During training, the time step order was adjusted based on the minimum time step crossover. The total crossover cost of the original adjacent time steps of the agricultural time-series segment was 8.74, which decreased to 6.21 after crossover, reducing the interference of redundant time steps within the segment.

[0039] In the hidden state diffusion solution process, each hidden state in the rearranged hidden state sequence is treated as a state node, and the state difference is calculated to generate a state diffusion relation matrix. The average state difference of the 16 state nodes within a segment decreased from 0.184 to 0.121. After five consecutive iterations, the average change in hidden states decreased to 0.0008. After state error reduction, the mean hidden state response of the segment with slowly decreasing soil moisture increased from 0.42 to 0.57, and the mean hidden state response of the segment with mild pest and disease risk increased from 0.36 to 0.51. Exponential parallel extrapolation was used to generate agricultural state prediction results for the next eight data collection periods.

[0040] In the comparative experiment, the traditional method used conventional data cleaning, ordinary association rule mining, and the original SegRNN model, while the method of this invention adopted a complete processing flow. Both methods used the same 1200 training samples, 200 validation samples, and 142 test samples. The test results showed that the average absolute error of soil moisture prediction decreased from 4.9 percentage points to 2.4 percentage points, the average absolute error of crop growth prediction decreased from 0.118 to 0.066, the accuracy of pest and disease risk identification increased from 81.6% to 92.4%, the yield trend prediction error decreased from 9.3% to 5.1%, the ability to identify moderate to high-risk segments in advance increased from an average of 2 collection periods to an average of 6 collection periods, and the consistency rate between prediction results and rule verification increased from 68.5% to 86.9%.

[0041] In the agricultural decision-making output stage, the agricultural state prediction results are jointly verified with the agricultural fuzzy association rule base to form a comprehensive agricultural state and generate irrigation, fertilization, plant protection, and agricultural scheduling decisions. In the test sample, there were 39 instances of plots requiring irrigation, 24 instances requiring fertilization, 18 instances requiring plant protection, and 13 instances requiring scheduling. Traditional methods made 70 correct decisions, with 16 false alarms and 8 missed alarms; this invention made 89 correct decisions, with 5 false alarms and 2 missed alarms. The accuracy rate of irrigation decisions increased from 78.0% to 92.3%, fertilization decisions from 72.0% to 87.5%, plant protection decisions from 72.2% to 88.9%, and agricultural scheduling decisions from 69.2% to 84.6%. This embodiment demonstrates that this invention outperforms traditional methods in terms of data quality, association rule reliability, prediction accuracy, and decision accuracy, validating its feasibility and effectiveness in smart agriculture data processing 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 smart agriculture data processing method based on big data analytics, characterized in that, include: Collect comprehensive agricultural production data within the agricultural production area, preprocess the comprehensive agricultural production data, and obtain a multi-source time-series agricultural dataset; Cross-source consistency correction and spatiotemporal continuity verification are performed on agricultural multi-source time series datasets to generate an agricultural status matrix. Based on the dynamic characteristics of agricultural conditions in the agricultural status matrix, an agricultural status fingerprint is formed, and the agricultural status fingerprint is converted into fuzzy status data. By utilizing fuzzy association rule mining, agricultural relationship aggregation is performed on fuzzy state data. Candidates are extracted using rotating grid coupling screening, and fragment relationships are analyzed through dynamic segmentation of the potential domain to obtain an agricultural fuzzy association rule library. Based on the agricultural fuzzy association rule base, feature enhancement is performed on the agricultural multi-source time series dataset. The association rule results are transformed into agricultural condition status labels, risk category labels, and segment labels to obtain the enhanced agricultural time series dataset. An improved SegRNN model is constructed to integrate segment features of the enhanced agricultural time series dataset. The time step order is adjusted based on the minimum time crossover. The hidden state diffusion solution is introduced to eliminate state errors. The prediction is generated by exponentially hierarchical parallel extrapolation. The agricultural state prediction results are obtained. The agricultural status prediction results are jointly verified with the agricultural fuzzy association rule base to determine the main related factors corresponding to the changes in agricultural status and form a comprehensive agricultural status. Based on the comprehensive agricultural situation, smart agriculture data processing results are generated, and irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions corresponding to agricultural production are output.

2. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The comprehensive agricultural production data includes agricultural environmental data, soil moisture data, crop growth data, pest and disease data, agricultural operation data, and historical yield data.

3. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The preprocessing of comprehensive agricultural production data includes data cleaning, missing data completion, anomaly removal, time alignment, unit unification, and index construction.

4. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The obtained agricultural multi-source time-series dataset includes: Collect comprehensive agricultural production data according to the plot number, sensor node number, crop type and growth stage within the agricultural production area, and record the data source, collection time, spatial location and data attributes of each data point to form the original agricultural data table; Data cleaning was performed on the original agricultural data table, deleting duplicate records and records with missing fields. Missing fields in continuous time series with data collection interruptions were filled in. Anomalies were removed based on the calibration range of the sensor equipment, the variation range of adjacent sampling, and the distribution of data from the same period of similar plots. The cleaned data was time-aligned according to a uniform sampling time interval, and the data of different dimensions were processed to unify the dimensions. A data index was constructed based on the plot number, crop type, growth stage, collection time and data source to obtain an agricultural multi-source time series dataset.

5. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The process of converting agricultural condition fingerprints into fuzzy state data includes: Agricultural baseline units are constructed based on the same plot of land, the same crop type, the same growth stage, and the same collection period. Data from multiple agricultural time series datasets are assigned to the corresponding agricultural baseline units. Data from different sources of the same agricultural indicator within the same agricultural baseline unit are recorded as source data groups. Cross-source consistency correction is performed on the source data group. The source data of the same agricultural indicator within the same agricultural baseline unit are sorted according to their numerical values, and the median value of the sort is taken as the correction value of the agricultural indicator. When the number of source data is even, the arithmetic mean of the two middle values ​​is taken as the correction value of the agricultural indicator. After cross-source consistency correction, the spatiotemporal continuity of agricultural indicators is verified. The absolute value of the time difference between adjacent collection periods for the same agricultural indicator is calculated, and the absolute value of the spatial difference between adjacent plots for the same agricultural indicator is calculated. An agricultural status matrix is ​​generated according to the correspondence between agricultural benchmark units and agricultural indicators. Based on the agricultural condition status matrix, dynamic features of agricultural conditions are extracted. The direction of change, number of consecutive changes and cross-source response order of agricultural indicators within the same agricultural condition benchmark unit are combined to form an agricultural condition status fingerprint. Based on the value range of agricultural indicators corresponding to the agricultural condition status fingerprint, the agricultural condition status fingerprint is converted into fuzzy state data.

6. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The obtained agricultural fuzzy association rule base includes: The fuzzy state data is generated into a fuzzy state grid according to agricultural indicators, plot numbers and collection time. The grid cell value of the fuzzy state grid is the membership degree of the agricultural state fingerprint in the corresponding fuzzy state. Candidate selection is performed using a rotating mesh coupling screening method. Taking the fuzzy state mesh of agricultural indicators as the processing object, each fuzzy state mesh is rotated and aligned according to a preset angle sequence, and the candidate score is calculated. The agricultural situation potential domain is constructed based on the fuzzy state grid. The potential domain value is calculated for each agricultural situation reference unit. The segment relationship is analyzed by dynamic segmentation of the potential domain. The agricultural situation segments are divided according to the continuous change relationship of the potential domain value in the time and space directions. The intra-segment consistency, inter-segment mutation and inter-segment transmission direction of each agricultural situation segment are calculated to obtain the key agricultural situation segments and continuous agricultural situation evolution segments. Using candidate association terms as the combined source of rule antecedents and consequents, and key agricultural condition segments and continuous agricultural condition evolution segments as the rule generation range, fuzzy support and fuzzy confidence are calculated to obtain an agricultural fuzzy association rule base.

7. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The obtained enhanced agricultural time-series dataset includes: Read the rule antecedent, rule consequent, fuzzy support, fuzzy confidence and corresponding agricultural condition segments from the agricultural fuzzy association rule base, match the fuzzy terms in the rule antecedent with the agricultural condition benchmark units in the agricultural multi-source time series dataset, and determine the matching rules; Extract the association rule results according to the matching rules, take the agricultural indicator combination corresponding to the antecedent of the matching rule as the association indicator combination, take the agricultural status corresponding to the consequent of the matching rule as the rule output result, and write the rule output result into the corresponding agricultural condition benchmark unit. Based on the rule output results, generate crop condition status labels and risk category labels. Convert the rule consequents representing crop growth, soil moisture, pest and disease risk, and yield trend into crop condition status labels, and convert the rule consequents representing normal, mild risk, moderate risk, and severe risk into risk category labels. Based on the key agricultural condition segments and continuous agricultural condition evolution segments corresponding to the matching rules, segment labels are generated. The associated indicator combinations, agricultural condition status labels, risk category labels and segment labels are added to the corresponding time positions in the agricultural multi-source time series dataset to obtain the enhanced agricultural time series dataset.

8. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The obtained agricultural state prediction results include: An improved SegRNN model is constructed, which includes time-interleaved minimum rearrangement, hidden state diffusion solution, and exponentially ordered parallel extrapolation. Time-interleaved minimum rearrangement is set between the segmented input layer and the recurrent unit, hidden state diffusion solution is set between the output of the recurrent unit and the inter-segment state transfer end, and exponentially ordered parallel extrapolation is set at the prediction output end. The time step order is adjusted based on the minimum time crossover rearrangement. The enhanced agricultural time series dataset is divided into agricultural time series segments according to the segment label. The crossover cost is calculated for the segment input vectors of any two time steps within the same agricultural time series segment. The crossover cost matrix is ​​constructed based on the crossover cost between all time steps. The time step rearrangement sequence within the segment is determined according to minimizing the sum of the crossover costs of adjacent processing time steps. The hidden state diffusion solution is introduced to resolve the state error. The time step rearrangement sequence within the segment is input into the cyclic unit to obtain the rearranged hidden state sequence. Each hidden state in the rearranged hidden state sequence is used as a state node. The state difference between any two state nodes is calculated. A state diffusion relation matrix is ​​generated based on the state difference. The state diffusion relation matrix is ​​used to iteratively diffuse and update each state node to obtain the diffused hidden state sequence. The prediction generation is performed by exponential hierarchical parallel extrapolation. The segment state transition matrix is ​​calculated based on the terminal diffusion hidden state of continuous agricultural time series segments. The segment state transition matrix is ​​multiplied continuously in hierarchical order to obtain the hierarchical transition matrix. The terminal diffusion hidden state of the latest agricultural time series segment is extrapolated with the hierarchical transition matrix in parallel to obtain the agricultural state prediction result. To train the improved SegRNN model, historical samples from the enhanced agricultural time-series dataset are used as training inputs, and the actual agricultural state results corresponding to future collection periods are used as training labels. The training loss is calculated based on the agricultural state prediction results and the actual agricultural state results. The model parameters of the improved SegRNN model are updated in reverse according to the training loss until the training loss meets the convergence condition, thus obtaining the trained improved SegRNN model.

9. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The formation of a comprehensive agricultural situation includes: The agricultural status prediction results are archived according to plot number, crop type, growth stage, prediction period and segment label to obtain prediction verification units; Based on the prediction verification unit, the association rules corresponding to the prediction results of agricultural state are retrieved from the agricultural fuzzy association rule base to obtain the matching association rule group; Consistency calculation is performed on the matching association rule group. The fuzzy support and fuzzy confidence of each matching association rule are multiplied to obtain the rule matching quantity. The rule matching quantities of all matching association rules in the same prediction verification unit are summed to obtain the prediction rule consistency quantity. The rule matching quantities of the same association factor in different matching association rules are accumulated to obtain the contribution quantity of the association factor. The main related factors are determined in descending order of their contribution. The agricultural status prediction results, the consistency of the prediction rules, and the main related factors are combined to generate a comprehensive agricultural status.

10. The smart agriculture data processing method based on big data analysis according to claim 1, characterized in that, The generated smart agriculture data processing results include: Read the comprehensive agricultural situation status and generate agricultural decision input units according to plot number, crop type, growth stage, forecast period, agricultural status category, status change direction, risk level and main related factors; Based on the soil moisture status, crop growth status, pest and disease risk status and yield trend status in the agricultural decision input unit, irrigation decision, fertilization decision, plant protection decision and agricultural scheduling decision are generated respectively. The risk level value, prediction time urgency value and related factor contribution of each decision within the same plot are added together to obtain the decision priority. The decision execution order is determined by prioritizing decisions from highest to lowest. The decision execution order, corresponding plot number, crop type, growth stage, forecast period, main related factors, and comprehensive agricultural status are written into the smart agriculture data processing results, and irrigation decisions, fertilization decisions, plant protection decisions, and agricultural scheduling decisions corresponding to agricultural production are output.

Images