Agricultural product yield prediction method and system based on belief propagation

Abstract

The application provides an agricultural product yield prediction method and system based on confidence propagation. The method comprises: constructing a plot unit space-time index corresponding to a target agricultural product, and mapping multi-source agricultural condition observation data and agricultural management data to the plot unit; generating a staged agricultural condition representation based on the growth stage of the crop corresponding to the plot unit; determining a quality score and establishing a confidence propagation relationship; combining the spatial correlation and management transmission relationship between plot units, inputting the compensated staged agricultural condition representation and confidence propagation result into a graph time series fusion model, generating a yield prediction value, a prediction interval and an influence factor contribution ranking corresponding to the plot unit under uncertainty constraints, and outputting the yield prediction value, the prediction interval and the influence factor contribution ranking as the yield prediction result of the target plot. The application can improve multi-source fusion accuracy, enhance missing data compensation reliability, and improve yield prediction stability.

Description

Technical Field

[0001] This application relates to the field of intelligent agricultural technology, and in particular to a method and system for predicting agricultural product yield based on confidence propagation. Background Technology

[0002] With the development of digitalization, precision, and intelligence in agricultural production, predicting agricultural yields using remote sensing, meteorological, soil, and field sensing data has become an important technical means in agricultural production management, planting decisions, and production-marketing coordination. Current technologies typically estimate crop growth and final yield in target plots or regions by constructing predictive models based on optical remote sensing imagery, meteorological monitoring data, or historical statistical data. Some solutions also incorporate radar remote sensing, soil properties, and agricultural records to improve prediction accuracy.

[0003] However, most existing technologies focus on simple splicing input of multi-source data or unified time series modeling, lacking a unified spatiotemporal organization mechanism for plot units, making it difficult to handle the problems of asynchronous acquisition time series and inconsistent spatial resolution between different data sources. At the same time, for cloud cover, missing data and abnormal observations, interpolation or elimination methods are usually used, failing to combine cross-modal correlations for reliable compensation. In addition, existing technologies do not make sufficient use of agricultural management variables, information on growth stages and data quality differences, resulting in large fluctuations in prediction results, poor generalization ability, and difficulty in providing simultaneous explanations of prediction intervals and influencing factors. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method and system for predicting agricultural product yield based on confidence propagation, in order to solve the problems of difficulty in aligning multi-source data, difficulty in reliably compensating for missing data, and poor stability of yield prediction in the prior art.

[0005] A first aspect of this application provides a method for predicting agricultural product yield based on confidence propagation, comprising: constructing a spatiotemporal index of plot units corresponding to the target agricultural product, and mapping multi-source agricultural observation data and agricultural management data to plot units; performing temporal reorganization, stage slicing, and cross-source alignment on the mapped multi-source data based on the crop growth stage corresponding to the plot unit to generate staged agricultural condition representations; determining quality scores based on the data integrity, source stability, spatiotemporal consistency, and stage matching degree of each staged agricultural condition representation, and establishing a confidence propagation relationship consisting of original observations, stage representations, compensated representations, and fused representations based on the quality scores, and performing cross-source compensation, conflict resolution, and anomaly verification on low-confidence data; combining the spatial correlation and management transfer relationships between plot units, jointly inputting the compensated staged agricultural condition representations and confidence propagation results into a graph-time series fusion model, generating the predicted yield value, prediction interval, and influence factor contribution ranking corresponding to the plot unit under uncertainty constraints, and outputting the predicted yield value, prediction interval, and influence factor contribution ranking as the yield prediction result of the target plot.

[0006] A second aspect of this application provides an agricultural product yield prediction system based on confidence propagation, comprising: a construction module for constructing a spatiotemporal index of plot units corresponding to the target agricultural product, and mapping multi-source agricultural observation data and agricultural management data to plot units; a generation module for performing temporal reorganization, stage slicing, and cross-source alignment on the mapped multi-source data based on the crop growth stage corresponding to the plot unit, generating staged agricultural condition representations; a determination module for determining quality scores based on the data integrity, source stability, spatiotemporal consistency, and stage matching degree of each staged agricultural condition representation, and establishing a confidence propagation relationship consisting of original observations, stage representations, compensated representations, and fused representations based on the quality scores, and performing cross-source compensation, conflict resolution, and anomaly verification on low-confidence data; and an output module for combining the spatial association and management transfer relationships between plot units, jointly inputting the compensated staged agricultural condition representations and confidence propagation results into a graph temporal fusion model, generating the yield prediction value, prediction interval, and influence factor contribution ranking corresponding to the plot unit under uncertainty constraints, and outputting the yield prediction value, prediction interval, and influence factor contribution ranking as the yield prediction result of the target plot.

[0007] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By constructing a spatiotemporal index of plot units corresponding to the target agricultural product and mapping multi-source agricultural observation data and agricultural management data to these plot units, and based on the crop growth stage corresponding to the plot unit, temporal reorganization, stage slicing, and cross-source alignment are performed on the mapped multi-source data to generate staged agricultural condition representations. Quality scores are determined based on the data completeness, source stability, spatiotemporal consistency, and stage matching degree of each staged agricultural condition representation. A confidence propagation relationship consisting of original observations, stage representations, compensated representations, and fused representations is established based on the quality scores. Cross-source compensation, conflict resolution, and anomaly verification are performed on low-confidence data. Combining the spatial correlation and management transmission relationships between plot units, the compensated staged agricultural condition representations and confidence propagation results are jointly input into a graph-time series fusion model. Under uncertainty constraints, the model generates the yield prediction value, prediction interval, and impact factor contribution ranking corresponding to the plot unit, and outputs the yield prediction value, prediction interval, and impact factor contribution ranking as the yield prediction result for the target plot. This application can improve the accuracy of multi-source fusion, enhance the reliability of missing data compensation, and improve the stability of yield prediction. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the agricultural product yield prediction method based on confidence propagation provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the agricultural product yield prediction system based on confidence propagation provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] In existing technologies, agricultural yield prediction typically relies on remote sensing data, meteorological data, soil data, and historical statistical data to construct predictive models. Some solutions further incorporate agricultural records or field sensing data to improve prediction accuracy. In practice, most existing solutions stitch together or jointly model multi-source data according to natural time series, and input data from different sources into the prediction model to output yield estimates for the target plot or region. While these technologies can reflect crop growth status to some extent, they generally lack a unified and systematic processing mechanism for differences in collection time, spatial scale, and data quality among multi-source data.

[0012] The main problems with existing technologies are as follows: First, different data sources have significant differences in acquisition frequency, acquisition time, and spatial resolution, making it difficult to accurately align multi-source data on the same plot unit. Second, optical remote sensing is significantly affected by cloud obstruction, and field sensing and agricultural machinery operation data are also prone to missing measurements, anomalies, or discontinuous records. Existing technologies usually use simple interpolation, outlier removal, or direct default processing, which makes it difficult to achieve reliable compensation. In addition, existing technologies do not make sufficient use of crop growth stage information and agricultural management variables, resulting in weak model adaptability to complex production scenarios, which in turn leads to large fluctuations in prediction results and poor stability and generalization ability.

[0013] To address the aforementioned technical issues, this application proposes a method for agricultural product yield prediction based on confidence propagation. This method first constructs a spatiotemporal index of plot units corresponding to the target agricultural product, uniformly mapping multi-source agricultural observation data and agricultural management data to these plot units. Then, using the crop growth stage as a temporal anchor point, it performs temporal reorganization, stage slicing, and cross-source alignment on the mapped multi-source data to generate staged agricultural condition representations. Further, it determines a quality score based on the data integrity, source stability, spatiotemporal consistency, and stage matching degree of the staged agricultural condition representations, and establishes a confidence propagation relationship based on the quality score, consisting of original observations, stage representations, compensated representations, and fused representations. For low-confidence data, it performs cross-source compensation, conflict resolution, and anomaly verification. Finally, it jointly inputs the compensated staged agricultural condition representations and confidence propagation results into a graph-time series fusion model to generate the predicted yield value, prediction interval, and ranking of influencing factors corresponding to the plot units under uncertainty constraints.

[0014] Through the above technical solution, this application no longer adopts the processing method of simply splicing and directly modeling multi-source data in the prior art. Instead, it constructs a prediction link that connects the plot units, growth period anchor points, quality scores, and confidence propagation, thereby improving the accuracy of multi-source fusion, enhancing the reliability of missing data compensation, and improving the stability of yield prediction. At the same time, it can also provide the prediction interval and the ranking of the contribution of influencing factors on the basis of outputting the yield prediction value, thereby improving the credibility and interpretability of the prediction results.

[0015] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0016] Figure 1 This is a flowchart illustrating the agricultural product yield prediction method based on confidence propagation provided in an embodiment of this application. Figure 1 As shown, the method may specifically include: S101, construct a spatiotemporal index of plot units corresponding to the target agricultural products, and map multi-source agricultural observation data and agricultural management data to plot units; S102, based on the crop growth stage corresponding to the plot unit, performs temporal reorganization, stage slicing and cross-source alignment on the mapped multi-source data to generate staged crop condition representation; S103. Based on the data integrity, source stability, spatiotemporal consistency and stage matching degree of agricultural condition representation at each stage, a quality score is determined. Based on the quality score, a confidence propagation relationship consisting of original observations, stage representations, compensation representations and fusion representations is established. Cross-source compensation, conflict resolution and anomaly verification are performed on low-confidence data. S104 combines the spatial relationships and management transmission relationships between plot units, inputs the compensated staged agricultural condition representation and confidence propagation results into the time series fusion model, generates the yield forecast, forecast interval and influence factor contribution ranking corresponding to the plot unit under uncertainty constraints, and outputs the yield forecast, forecast interval and influence factor contribution ranking as the yield forecast result of the target plot.

[0017] In some embodiments, a spatiotemporal index of land parcel units corresponding to the target agricultural product is constructed, and multi-source agricultural observation data and agricultural management data are mapped to land parcel units, including: Extract the plot boundaries, planting season sequence, and data source identifiers corresponding to the target planting area, construct a unified index key for the associated plot identifier, time phase identifier, and source identifier, and form a spatiotemporal index of plot units; Based on the spatiotemporal index of land parcel units, spatial attribution matching, time reference conversion, and scale normalization are performed on multi-source agricultural observation data and agricultural management data to generate index mapping results corresponding to each land parcel unit.

[0018] Specifically, this embodiment first uses agricultural plots within the target planting area as the basic organizational objects, uniformly extracting and standardizing plot boundaries, planting season start and end times, and data source identities to form a plot unit spatiotemporal index that can simultaneously represent spatial affiliation, temporal location, and data source relationships. Subsequently, based on this plot unit spatiotemporal index, spatial affiliation matching, time reference conversion, and scale normalization are performed on multi-source agricultural observation data and agricultural management data, so that data from different collection methods, collection frequencies, and spatial scales are uniformly mapped to the corresponding plot units, generating an index mapping result for plot-level yield prediction, providing a consistent data entry point for subsequent growth period alignment, quality scoring, and confidence propagation.

[0019] Among them, the plot boundary can be a cadastral boundary, a planting management boundary, or a planting unit boundary obtained based on remote sensing segmentation. The planting season time series is used to characterize the sowing time, key agricultural events, and harvest window of the same plot within the current planting season. The data source identifier is used to distinguish the data identity corresponding to different observation and management sources. The unified index key is not a simple numbering of various types of data, but rather an associative encapsulation of plot identifiers, time phase identifiers, and source identifiers, so that subsequent access to remote sensing observations, meteorological records, field sensor records, and agricultural machinery operation records can all be located to the corresponding plot and time period according to the same index rule. Spatial attribution matching is mainly used to determine the attribution relationship between external observation data and target plots. Time base conversion is mainly used to unify data under different timestamp systems to the same time series base. Scale normalization processing is mainly used to eliminate the expression differences caused by different spatial resolutions, different statistical granularities, and different sampling densities.

[0020] Taking a summer maize yield prediction scenario in a plain area as an example, the target planting area includes 326 maize plots. First, the boundaries of each plot are extracted from the rural land management map, and the planting season sequence of each plot is determined by combining sowing registration information, agricultural machinery sowing records, and the regional agricultural calendar. For the target plot numbered D128, the system generates a unified index key consisting of plot identifier D128, planting season identifier 2025 summer, time phase identifier, and source identifier. Then, multi-source agricultural observation data and agricultural management data are accessed. The multi-source agricultural observation data includes 10-meter resolution optical remote sensing images, radar remote sensing images, township-level meteorological grid data, plot soil testing data, and soil moisture sensor data deployed in the fields. The agricultural management data includes sowing operation trajectories, fertilization operation records, and irrigation execution records.

[0021] For optical and radar imagery, coordinate unification and plot cropping are performed first, then the pixel coverage area is determined based on the target plot boundary, and spatial attribution matching is completed according to the effective coverage ratio within the plot; for meteorological grid data, mapping is performed based on the plot centroid location and surrounding grid weights; for agricultural machinery operation trajectories, the plot to which they belong is determined based on the overlay relationship between the trajectory and the plot boundary and the duration of the operation.

[0022] Furthermore, the remote sensing image acquisition time, hourly meteorological recording time, minute-level sensor sampling time, and agricultural machinery operation event time are uniformly converted into a standard temporal sequence under the planting season time series. Pixel-level, grid-level, point-level, and event-level data are then unified into a plot-level representation. For example, the multispectral reflectance, radar backscatter value, cumulative rainfall over the past three days, and soil moisture content of the target plot on June 18th are mapped to the standard temporal node corresponding to plot D128, generating an index mapping result for that plot under that temporal phase. For cases where multiple observations exist within the same temporal phase, aggregation can be further performed according to source priority and coverage completeness to form a unique plot-level input record.

[0023] Through the above implementation methods, a unified data organization result oriented towards plot units can be formed, so that multi-source agricultural observation data and agricultural management data are consistent in spatial attribution, temporal location and scale expression, thereby improving the data consistency of subsequent staged modeling, enhancing the standardization of multi-source data access, and providing basic support for cross-source alignment and reliable analysis in the yield prediction process.

[0024] In some embodiments, based on the crop growth stage corresponding to the plot unit, temporal reorganization, stage slicing, and cross-source alignment are performed on the mapped multi-source data to generate a staged crop condition representation, including: Using crop growth stages as time anchors, the stage boundaries of each plot unit are determined based on the planting time, stage transition characteristics, and observed change trajectories corresponding to the plot units. The mapped multi-source data are reorganized into unified stage coordinates according to the stage boundaries, and cross-source registration and feature aggregation are performed under stage association constraints to generate staged agricultural condition representations corresponding to each land parcel unit.

[0025] Specifically, this embodiment no longer directly splices various observation data according to natural dates. Instead, it uses the crop growth stage corresponding to the plot unit as a unified time-series anchor point. Based on the planting time, stage transition characteristics, and observation change trajectory of the plot unit, the stage boundaries of each plot unit within the current planting season are determined. Subsequently, the multi-source data mapped to the plot unit are reorganized into unified stage coordinates according to the stage boundaries. Cross-source registration and feature aggregation are performed under stage association constraints. This transforms multi-source data with different sampling frequencies, misaligned observation times, and inconsistent expression granularities into comparable and fusionable staged agricultural condition representations under the same growth stage, providing a unified input basis for subsequent quality scoring, confidence propagation, and yield prediction.

[0026] The planting sequence characterizes the sowing start point, key agricultural milestones, and expected harvest progress of a plot unit within the current planting season. Stage transition characteristics reflect the changes in crop growth from one stage to the next. Observational change trajectories characterize the continuous change paths formed by multi-source observations over time. Stage boundaries are not determined solely by a pre-set agricultural calendar; rather, they are determined by a comprehensive assessment of the plot unit's actual sowing time, accumulated temperature progression, remote sensing growth changes, radar structural response changes, and field monitoring trends. This allows different plots within the same region to form differentiated stage boundaries based on their own growth processes. Stage association constraints define the continuous connection between adjacent growth stages and the consistent expression of multi-source observations within the same stage.

[0027] Taking a summer maize planting scenario in a plain area as an example, for the target plot numbered D128, after completing multi-source data mapping, the system first reads the plot's sowing record on May 28, subsequent irrigation records, and accumulated regional temperature results. Combined with the optical vegetation index change sequence, radar backscattering change sequence, and soil moisture change sequence from June to September, the system identifies the transition points from the seedling stage to the jointing stage, from the jointing stage to the tasseling stage, from the tasseling stage to the grain-filling stage, and from the grain-filling stage to the maturity stage. For example, when the system detects a continuous increase in vegetation index, enhanced radar structure response, and soil moisture fluctuations matching the timing of topdressing operations, the corresponding time period is determined as the starting boundary of the jointing stage; when it detects a slowdown in canopy growth rate, temperature conditions meeting the requirements for reproductive growth transition, and a stable transition in field sensor records, the corresponding time period is determined as the boundary of the tasseling stage.

[0028] Furthermore, after determining the stage boundaries, the system uniformly projects the optical remote sensing data collected daily, radar data acquired according to orbital cycles, meteorological data recorded hourly, field sensor data sampled minutely, and agricultural machinery operation data recorded by event onto the stage coordinates corresponding to the seedling emergence, jointing, tasseling, grain filling, and maturity stages. Subsequently, under stage association constraints, time alignment, spatial consistency verification, and intra-stage feature aggregation are performed on the multi-source data within each stage to generate stage-specific agricultural condition representations, such as soil moisture representations for the seedling emergence stage, growth representations for the jointing stage, stress representations for the tasseling stage, and accumulation representations for the grain filling stage. For cases where multiple observations exist within the same stage, weighted aggregation can be further performed based on the proximity of the observations to the stage center location and the degree of cross-source consistency, forming a unique stage-specific input result.

[0029] Through the above implementation methods, multi-source data with different sampling rhythms and different expression forms can be uniformly converted into a staged agricultural condition representation for crop growth process, enhancing the correspondence of multi-source data at the growth process level, improving the accuracy of cross-source alignment, and providing a consistent and stable data foundation for subsequent yield prediction processing based on quality scores and confidence propagation.

[0030] In some embodiments, crop growth stages are used as time-series anchors, and the stage boundaries corresponding to each plot unit are determined based on the planting time, stage transition characteristics, and observed change trajectories of the plot unit, including: Based on planting time series, a priori information on the corresponding reproductive process is constructed, and the stage transition probability corresponding to each time series position is determined by combining stage transition characteristics and observed change trajectories. Based on the stage transition probability, boundary positioning is performed under the constraints of stage continuity and temporal consistency of land parcel units to generate the stage boundaries corresponding to each land parcel unit.

[0031] Specifically, this embodiment targets target plots where spatiotemporal indexes for plot units have been constructed and multi-source agricultural observation data and agricultural management data have been mapped. Based on planting timelines, a priori information on the corresponding reproductive process is constructed. Then, combined with stage transition characteristics and observed change trajectories, the stage transition probability corresponding to each timeline position is determined. Subsequently, based on the stage transition probabilities, boundary positioning is performed under constraints of stage continuity and plot unit temporal consistency to generate stage boundaries for each plot unit. Through this process, the stage division of different plots no longer depends on a fixed agricultural calendar, but rather is adaptively determined based on the plot's own planting progress and multi-source observation changes, thereby providing a reliable stage benchmark for subsequent stage slicing and cross-source alignment.

[0032] Among them, planting time sequence is mainly used to reflect the sowing time, replanting time, key agricultural nodes and expected growth process of plot units in the current planting season; growth process prior is used to characterize the reasonable duration range and sequential connection relationship of each growth stage under the current crop type, planting season conditions and regional agricultural time background; stage transition characteristics are used to reflect the change pattern of growth, structure, water or heat conditions when adjacent growth stages switch; observation change trajectory is used to characterize the continuous change path formed by the evolution of multi-source agricultural observation data over time.

[0033] The stage transition probability is not a simple threshold determination result, but a measure of the likelihood of stage transition obtained by comprehensively considering planting time series constraints, multi-source change amplitude, change direction, and correlation with adjacent time series locations. Stage continuity constraints are used to restrict stage boundaries to conform to the progressive relationship between growth stages, preventing stage jumps or regressions; plot unit time series consistency constraints are used to ensure that the boundary positioning results are consistent with the plot's own previous and subsequent observation evolution, avoiding misjudgments of stages directly caused by local abnormal observations.

[0034] Taking a winter wheat yield prediction scenario in a plain area as an example, the system first obtains the sowing record on October 12th, the seedling inspection record in early November, and the relevant agricultural records for the following year's greening, jointing, heading, and grain filling stages to construct the planting sequence for this plot during the current planting season. Then, combining the regional accumulated temperature progression, soil moisture changes, and the conventional growth rhythm of winter wheat, a priori understanding of the corresponding growth process for this plot is established. Based on this, the system extracts the optical vegetation index change trajectory, radar backscattering change trajectory, near-soil temperature and humidity change trajectory, and soil moisture change trajectory for this plot, and identifies stage transition characteristics such as the extent of growth recovery before and after greening, the canopy growth rate during jointing, structural response changes before and after heading, and hydrothermal coupling changes during the grain filling stage.

[0035] Subsequently, the system calculates the probability of stage transition at each location throughout the entire planting season according to a preset time step. For example, during the period from mid-March to late March, if the vegetation index changes from a gradual increase to a continuous increase, the radar scattering intensity increases synchronously, the average daily accumulated temperature crosses the greening threshold, and the soil moisture in the field is within a suitable range, then the probability of transitioning from the overwintering period to the greening period increases during this period; around mid-April, if the canopy growth slope increases significantly, the stem structure response increases, and the topdressing operation records and observed changes corroborate each other, then the probability of transitioning from the greening period to the jointing period increases.

[0036] Finally, based on the stage transition probability distribution over continuous time, the system performs boundary positioning under stage continuity constraints and plot unit time consistency constraints, determining that the starting point of the greening period of plot W052 is March 18, the starting point of the jointing period is April 16, the starting point of the heading period is May 3, and the starting point of the grain filling period is May 17, thereby generating the corresponding stage boundary results for the plot.

[0037] Through the above implementation methods, the boundaries of the growth stages of each plot unit can be adaptively located by combining the actual planting progress of the plot and the evolution process of multi-source observation. This improves the accuracy and consistency of stage division, enhances the reliability of subsequent multi-source data stage reorganization and cross-source registration, and provides a stable stage benchmark for quality scoring, confidence propagation and interval prediction in the yield prediction process.

[0038] In some embodiments, the mapped multi-source data are reorganized to a unified stage coordinate according to the stage boundary, and cross-source registration and feature aggregation are performed under stage association constraints to generate staged agricultural condition representations corresponding to each land parcel unit, including: Based on the stage boundaries, the mapped multi-source data is projected onto a unified stage coordinate system to determine the stage position relationship of each source data within the corresponding crop growth stage. Based on the stage location relationship, cross-source spatiotemporal registration, stage consistency verification, and representation aggregation are performed under stage association constraints to generate staged agricultural condition representations corresponding to each land parcel unit.

[0039] Specifically, after determining the stage boundaries corresponding to each plot unit, this embodiment first projects the multi-source agricultural observation data and agricultural management data mapped to the plot units onto a unified stage coordinate system based on the stage boundaries. This ensures that data from different sampling periods, recording granularities, and source formats all obtain stage positional relationships corresponding to the crop growth stages. Subsequently, based on these stage positional relationships, cross-source spatiotemporal registration, stage consistency verification, and representation aggregation are performed under stage association constraints to generate staged agricultural representations corresponding to each plot unit. This transforms the original data distributed according to natural time into staged expressions oriented towards the crop growth process.

[0040] The unified stage coordinates are used to characterize the unified stage reference system established by each plot unit within the current planting season, centered around the processes of emergence, vegetative growth, reproductive growth, and maturity. The stage positional relationship is used to reflect the attribution, adjacency, and relative sequence of each source data within the corresponding crop growth stage. Cross-source spatiotemporal registration is not simply splicing data according to the collection date, but rather, based on the positional relationship of each source data in the stage coordinates, it jointly corrects the spatial coverage, stage attribution interval, and stage center offset, ensuring that multi-source data within the same plot unit and the same stage maintain consistency in terms of the object and time of representation.

[0041] Phase consistency verification is used to check whether the direction, magnitude, and phase response of data from different sources are consistent within the same phase. If there is a significant deviation, the abnormal data is downweighted by combining phase boundaries, historical evolution trajectories, and agricultural management records. Characteristic aggregation is used to aggregate the multi-source data that has passed the verification into a phased agricultural condition characterization that can reflect the growth status, hydrothermal status, management response status, and stress status of the corresponding phase.

[0042] Taking a summer maize planting scenario in a plain area as an example, for the target plot numbered D128, the system has determined the stage boundaries corresponding to its emergence, jointing, tasseling, grain-filling, and maturity stages. Subsequently, the corresponding optical remote sensing observations, radar remote sensing observations, hourly meteorological records, soil moisture sensor records, and records of sowing, topdressing, and irrigation operations are uniformly projected onto the above-mentioned stage coordinates. For the multispectral vegetation index obtained in early June and the radar backscattering characteristics of adjacent time periods within the same stage, the system performs cross-source spatiotemporal registration based on their relative positional relationship and plot cover relationship during the emergence stage to determine their common corresponding growth status in the later stages of the emergence stage.

[0043] For the continuous rainfall records, soil moisture increase records, and irrigation gap records from early July, the system first determined that these records were located in the later stages of the jointing stage based on their stage positional relationships. Then, through stage consistency verification, it confirmed that the soil moisture changes during this period were mainly caused by natural rainfall, rather than triggered by irrigation operations. For the optical growth data and radar structure response data acquired multiple times during the tasseling stage, the system performed characterization aggregation based on their proximity to the center of the tasseling stage, the consistency of cross-source responses, and the correlation with topdressing records, generating canopy structure representations and nutrient response representations for the tasseling stage. Ultimately, plot D128 developed corresponding staged agronomical representations for each growth stage, which served as standardized inputs for subsequent quality scoring, confidence propagation, and yield prediction.

[0044] Through the above implementation methods, the mapped multi-source data can be orderly reorganized and consistently expressed under a unified stage coordinate system, which enhances the correspondence between data from different sources within the same growth stage, improves the accuracy of cross-source spatiotemporal registration and stage aggregation, and reduces information offset caused by sampling misalignment and expression differences. This provides a stable, comparable, and growth stage-based agricultural condition representation for the subsequent yield prediction process.

[0045] In some embodiments, a quality score is determined based on the data completeness, source stability, spatiotemporal consistency, and stage matching degree of each stage of agricultural condition characterization, including: Extract the data coverage status, source continuity status, spatiotemporal registration status and stage adaptation status corresponding to each stage of agricultural condition characterization, and construct quality assessment parameters corresponding to each stage of agricultural condition characterization; Based on quality assessment parameters, a comprehensive quality assessment is performed under cross-source consistency constraints and quality correlation constraints to generate quality scores corresponding to agricultural condition representations at each stage.

[0046] Specifically, in this embodiment, after completing the phased reorganization of multi-source data and forming phased agricultural condition representations corresponding to each plot unit, the system does not directly input the phased agricultural condition representations into subsequent confidence propagation processing. Instead, it first extracts quality assessment information related to data reliability from each phased agricultural condition representation, constructs corresponding quality assessment parameters, and then performs a comprehensive quality assessment under cross-source consistency constraints and quality correlation constraints to generate quality scores corresponding to each phased agricultural condition representation. Through this processing, the differences in data quality across different stages, sources, and observation conditions can be explicitly quantified, providing a unified quality basis for subsequent low-confidence data identification, cross-source compensation, and yield prediction.

[0047] Among them, the data coverage status is used to characterize the effective coverage ratio, effective observation times, and missing data distribution of the observation data corresponding to a certain stage of agronomic representation within that stage; the source continuity status is used to characterize the continuity of data supply, breakpoints, and stability of the corresponding data sources in the preceding and following time series; the spatiotemporal registration status is used to characterize the spatial consistency of multi-source data within the target plot, the alignment of stage coordinates, and the coordination of responses within the same stage; and the stage adaptation status is used to characterize the degree of matching between the stage of agronomic representation and the current crop growth stage, including whether the observed changes conform to the stage evolution law and whether the management events are consistent with the stage requirements.

[0048] Quality assessment parameters are not simply given by individual indicators, but rather the aforementioned state information is transformed into a comparable set of evaluation parameters to reflect the comprehensive quality characteristics of agricultural condition representations at different stages in terms of completeness, stability, consistency, and adaptability. Cross-source consistency constraints are used to limit the range of deviation between representations from different sources within the same plot and the same stage, while quality correlation constraints are used to maintain reasonable continuity of quality changes between adjacent stages and related sources, avoiding distortion of quality scores caused by local anomalies.

[0049] Taking a summer maize planting scenario in a plain area as an example, for the target plot numbered D128, the system has generated staged agronomical characteristics corresponding to its emergence, jointing, tasseling, and grain-filling stages. For the staged agronomical characteristics of the jointing stage, the system first extracts its data coverage status, confirming that there are 4 valid optical remote sensing observations and 1 failure due to cloud obstruction during this stage, 2 valid radar remote sensing observations, and the field soil moisture sensor record integrity rate reaches 95%, thus generating a coverage parameter; then it extracts the source continuity status, confirming that the radar remote sensing and sensor data are continuous in time during this stage, while the optical remote sensing has one interruption, thus generating a source stability parameter; subsequently, it extracts the spatiotemporal registration status, determining that all source data have been assigned to plot D128, and that the remote sensing growth changes and soil moisture changes during the jointing stage are basically consistent, thus generating a registration consistency parameter; further, it extracts the stage adaptation status, identifying that the topdressing operation records and growth enhancement trajectories during this stage match each other, indicating that the characteristics match the growth features of the jointing stage, thus generating a stage adaptation parameter.

[0050] Subsequently, the system inputs the aforementioned quality assessment parameters into the comprehensive quality evaluation module. Under cross-source consistency constraints, it verifies the response differences between optical and radar representations. Under quality correlation constraints, it compares whether the quality changes during the jointing stage are smooth compared to the preceding seedling emergence stage and the subsequent tasseling stage. Finally, it generates corresponding quality scores for the staged agronomic characterization of plot D128 during the jointing stage. If there is large-area cloud cover, significant source interruption, or a stage response that is significantly inconsistent with the agricultural records, the corresponding quality score will be reduced accordingly and used as the basis for subsequent low-confidence identification.

[0051] Through the above implementation methods, the data quality of agricultural condition representation at each stage can be uniformly, refined, and correlated, improving the ability to distinguish the quality of multi-source stage representations, enhancing the pertinence of subsequent confidence propagation and cross-source compensation processing, and providing a reliable quality foundation for stable modeling and credible output in the yield prediction process.

[0052] In some embodiments, a confidence propagation relationship consisting of original observations, stage representations, compensated representations, and fused representations is established based on quality scores. Cross-source compensation, conflict resolution, and anomaly verification are performed on low-confidence data, including: Based on the quality scores corresponding to the agricultural condition representations at each stage, a hierarchical propagation node and inter-node propagation constraints are constructed to link the original observations, stage representations, compensation representations, and fusion representations, thereby generating confidence propagation relationships. Based on the confidence propagation relationship, low-confidence data is identified and the corresponding compensation path is determined, and cross-source compensation is performed on the low-confidence data; Based on the compensated propagation state and cross-source consistency results, conflict resolution and anomaly verification are performed on low-confidence data to generate confidence correction results.

[0053] Specifically, in this embodiment, after forming staged agricultural condition representations corresponding to each plot unit and obtaining quality scores, the system organizes the original observations, stage representations, compensated representations, and fused representations into a hierarchical propagation structure, constructs propagation constraints between nodes, and generates corresponding confidence propagation relationships. Subsequently, based on the confidence propagation relationships, low-confidence data is identified, corresponding compensation paths are determined, and cross-source associated data is invoked to perform compensation processing. For cases where source conflicts or response anomalies still exist after compensation, conflict adjudication and anomaly verification are performed by combining the compensated propagation status and cross-source consistency results to generate confidence correction results, which serve as inputs for subsequent yield prediction.

[0054] Among them, the original observation is used to characterize the observation records at the original acquisition level from sources such as remote sensing, meteorology, soil, field sensing and agricultural management; the stage representation is used to characterize the stage expression results of the original observation after recombination and aggregation within the corresponding crop growth stage; the compensation representation is used to characterize the alternative expression results formed after cross-source compensation of low confidence data; and the fusion representation is used to characterize the unified input result formed after comprehensive processing of the original observation, stage representation and compensation representation under propagation constraints.

[0055] Propagation constraints include stage-based propagation constraints, cross-source consistency constraints, source inheritance constraints, and spatial association constraints, used to limit the propagation direction, propagation weights, and attenuation relationships between nodes at different levels. Low-confidence data is not determined solely by a single quality score threshold, but rather by a comprehensive assessment combining the node's own quality level, the propagation status of adjacent stages, and the degree of support from associated sources. The compensation path characterizes the propagation link from low-confidence nodes to candidate compensation source nodes and its weight configuration relationships.

[0056] Taking plot D128 in a summer maize planting scenario in a plain area as an example, during the tasseling stage, the system found that the optical remote sensing observations of this plot were severely affected by cloud cover, resulting in a significantly lower quality score for the corresponding stage-specific agronomic characterization compared to radar observations, soil moisture sensor records, and near-surface meteorological records at the same stage. Based on this, the system identified the optical stage characterization of the tasseling stage as low-confidence data and established a hierarchical propagation chain in the confidence propagation relationship, consisting of the original optical observation node, the tasseling stage characterization node, the radar compensation node, the soil moisture compensation node, and the tasseling stage fusion characterization node. Subsequently, based on stage propagation constraints and cross-source consistency constraints, the system determined that the radar backscattering changes at the same stage, the optical change trends before and after the tasseling stage, the soil moisture change trajectory, and the topdressing records during the tasseling stage would be candidate compensation sources, and generated corresponding compensation paths.

[0057] Subsequently, the system assigns propagation weights to each candidate source according to the compensation path, uses radar observations to reflect the canopy structure status, soil moisture records to reflect the water stress status, and utilizes growth changes before and after each stage to maintain stage continuity, generating a compensated representation corresponding to the low-confidence data for the tasseling stage. If the compensation result shows that the radar structure response indicates enhanced growth, but the field sensor records show continuous water loss and are inconsistent with the topdressing records, the system further performs conflict resolution based on the compensated propagation status and cross-source consistency results, downweighting nodes with poor source continuity and significant deviations from the stage change pattern; simultaneously, it performs anomaly checks on compensation results showing sudden increases, decreases, or stage jumps, eliminating abnormal propagation branches that do not conform to the growth characteristics of the tasseling stage. Finally, the system generates confidence-corrected results for the tasseling stage of plot D128 and forms a fusion representation that can be used by subsequent graph-time series fusion models.

[0058] Through the above implementation methods, data quality information can be continuously transmitted along the processing chain of original observation, stage representation, compensation representation and fusion representation, realizing targeted identification and reliable compensation of low confidence data, improving data availability in missing and abnormal scenarios, enhancing the consistency and reliability of cross-source fusion process, and thus providing a more stable and reliable input basis for subsequent yield prediction.

[0059] In some embodiments, based on the confidence propagation relationship, low-confidence data is identified and the corresponding compensation path is determined. Cross-source compensation is then performed on the low-confidence data, including: Based on the confidence propagation relationship, combined with the associated propagation state and stage constraint state of the propagation node where the low confidence data is located, the candidate compensation source corresponding to the low confidence data is determined, and the corresponding compensation path is generated. According to the compensation path, propagation weight allocation, confidence-weighted fusion and stage consistency reconstruction are performed on the candidate compensation sources to generate compensation representations corresponding to low-confidence data.

[0060] Specifically, in this embodiment, after constructing the confidence propagation relationship, the system does not directly perform compensation on all staged agricultural condition representations. Instead, it first checks each propagation node layer by layer according to the confidence propagation relationship, identifying low-confidence data that lacks confidence and will affect the stability of subsequent fusion representations. Then, combining the associated propagation state and stage constraint state of the propagation node where the low-confidence data is located, it determines the corresponding candidate compensation source and generates the corresponding compensation path. Subsequently, the system performs propagation weight allocation, confidence-weighted fusion, and stage consistency reconstruction on the candidate compensation sources according to the compensation path, generating compensation representations corresponding to the low-confidence data. This ensures that missing, abnormal, or severely interfered data are reliably replaced while maintaining stage semantic consistency.

[0061] The propagation state is used to characterize the degree of connectivity, strength, and stability of propagation between the propagation node containing low-confidence data and nodes in preceding stages, related nodes in the same stage, and adjacent source nodes. The stage constraint state is used to characterize whether the current low-confidence data meets the evolutionary laws, management response laws, and stage continuity requirements corresponding to the crop growth stage. The candidate compensation sources are not arbitrarily selected related data, but are selected from sources that have a valid propagation relationship with the current propagation node and can provide semantic support for the stage, such as other observation sources in the same stage, high-confidence representations of adjacent stages, and process records related to management activities in the current stage.

[0062] The compensation path is used to characterize the propagation link from low-confidence data nodes to candidate compensation source nodes, including path priority and propagation weight. When assigning propagation weights, not only the confidence level of the candidate compensation source itself is considered, but also its matching degree with the target node in terms of stage position, change trend, and management response. Confidence-weighted fusion is used to integrate candidate compensation sources from multiple compensation paths according to their weights. Stage consistency reconstruction is used to correct deviations between the fusion result and the current stage boundary, the evolution of adjacent stages, and agricultural time series, ensuring that the generated compensation representation conforms to the actual agricultural conditions of the target stage.

[0063] Taking plot D128 in a summer maize planting scenario in a plain area as an example, during the tasseling stage, the system found that the optical remote sensing stage representation corresponding to this plot had insufficient effective coverage due to continuous cloud cover, and there was a significant quality difference between its propagation nodes and radar representation nodes at the same stage. Therefore, the optical remote sensing stage representation was identified as low-confidence data.

[0064] Subsequently, the system reads the associated propagation status of the propagation node where the low-confidence data is located and finds that it has usable propagation relationships with the radar backscattering node during the tasseling stage, the soil moisture node during the tasseling stage, the meteorological stress node during the tasseling stage, and the high-confidence growth node of the previous stage. Combined with the stage constraint status, it is confirmed that the tasseling stage should focus on reflecting the formation of canopy structure, water and heat fluctuations during flowering, and continuous growth characteristics before and after the stage. Based on this, radar observations, soil moisture changes, high temperature records, and growth changes of the previous stage are identified as candidate compensation sources, and a compensation path is generated with radar structure response as the main path, moisture and meteorological changes as auxiliary paths, and the growth characteristics of the previous stage as the continuous correction path.

[0065] Subsequently, the system performs propagation weight allocation based on the quality score, stage fit, and consistency of change of each candidate compensation source. Specifically, radar observations during the tasseling stage are assigned a higher propagation weight, the previous stage's growth characterization is assigned a continuous correction weight, and meteorological and soil moisture records are assigned a stress constraint weight. After weight allocation, the system performs confidence-weighted fusion on each candidate compensation source to generate initial compensation results for low-confidence optical characterization during the tasseling stage. Then, combining the tasseling stage boundary, the slope of growth change before and after the previous stage, and the current topdressing record, stage consistency reconstruction is performed to correct overestimation or underestimation caused by single-source bias. Finally, the compensated characterization corresponding to the low-confidence data during the tasseling stage is formed and written back to the subsequent characterization generation process.

[0066] Through the above implementation methods, low-confidence data can be specifically identified and path-based compensated according to the confidence propagation relationship, thereby improving the accuracy of candidate compensation source screening, enhancing the matching degree between compensation results and the current crop growth stage, reducing the interference of missing and abnormal data on subsequent fusion modeling, and thus improving the integrity and reliability of staged agricultural condition representation.

[0067] In some embodiments, the compensated staged agricultural condition representation and confidence propagation results are jointly input into the graph-time series fusion model to generate, under uncertainty constraints, the predicted yield value, prediction interval, and ranking of influencing factors corresponding to the plot unit, including: Based on the spatial and management transmission relationships between land parcel units, a land parcel association graph corresponding to the graph-temporal fusion model is constructed. The compensated staged agricultural condition representation and confidence propagation results are mapped to the time-series node representations corresponding to the plot association graph. Time-series association modeling and graph association modeling are performed to generate the yield distribution representations corresponding to each plot unit. Based on the output distribution characterization, the output forecast value and forecast interval corresponding to each plot unit are determined under uncertainty constraints, and the contribution ranking of the influencing factors is generated according to the model response results.

[0068] Specifically, after completing the construction of phased agricultural condition representations, quality scoring, confidence propagation, and cross-source compensation for each plot unit, this embodiment further utilizes the spatial and management transmission relationships between plots to construct a plot association graph. Then, the compensated phased agricultural condition representations and confidence propagation results are mapped to the temporal node representations corresponding to the plot association graph. Through temporal association modeling and graph association modeling, the yield distribution representations corresponding to each plot unit are generated. Finally, under uncertainty constraints, the predicted yield values ​​and prediction intervals corresponding to each plot unit are determined, and the contribution ranking of influencing factors is generated based on the model response results.

[0069] Spatial association relationships characterize the degree of association between plot units in terms of spatial adjacency, irrigation and drainage connectivity, soil similarity, and topographic proximity. Management transmission relationships characterize the synergistic and transmission effects of adjacent plots on management activities such as sowing, fertilization, irrigation, plant protection, and harvesting. Nodes in the plot association graph correspond to different plot units, and edges characterize the strength of spatial and management associations between these plot units. Temporal node representations characterize the staged crop condition, compensation state, and confidence state of a plot unit at each crop growth stage. Temporal association modeling extracts the evolutionary association features of the same plot unit across different growth stages, while graph association modeling extracts the spatial propagation and management coupling features between different plot units.

[0070] Yield distribution characterization reflects the central estimation trend and fluctuation distribution of the final yield of each plot unit. Uncertainty constraints constrain the model output based on confidence propagation results, compensation ratios, cross-source conflict levels, and sample distribution bias, ensuring that the prediction results not only output a single value but also simultaneously form interval results that match the current data credibility. The contribution ranking of influencing factors characterizes the order of contribution of different growth stages, different agricultural conditions, and different management factors to the final yield prediction results.

[0071] Taking a summer maize planting scenario in a plain area as an example, the system constructs a plot association graph for adjacent plots numbered D128, D129, and D130. D128 and D129 are adjacent by boundary and share the same irrigation branch. Although D128 and D130 are not directly adjacent, they are also linked by a management transfer edge due to using the same fertilization route and having similar soil types. For plot D128, the system maps the staged agronomical characteristics corresponding to the seedling, jointing, tasseling, and grain-filling stages, as well as the compensation status and confidence propagation results for each stage, to the temporal node representation of that plot. For example, the tasseling stage node includes canopy structure representation, soil moisture representation, high-temperature stress representation, and corresponding compensation confidence; the grain-filling stage node includes accumulated temperature representation, soil moisture stability representation, and irrigation response representation.

[0072] Subsequently, the system first performs temporal correlation modeling on the nodes of each stage of plot D128 along the time dimension, learning its growth evolution and management response trajectory from the seedling stage to the grain-filling stage. Then, combined with the plot correlation graph, it performs graph correlation modeling on the correlation edges between D128 and D129 and D130, extracting the correlation features formed by the common influence of rainfall processes, shared irrigation conditions, and similar management measures among neighboring plots. Based on the temporal modeling results and graph modeling results, the system generates a yield distribution representation corresponding to plot D128.

[0073] Furthermore, the system applies uncertainty constraints based on the compensation ratios at each stage, cross-source consistency levels, and the degree of response consistency of related plots for plot D128, determining its predicted yield value to be 742 kg / mu, with a prediction range of 718 kg / mu to 768 kg / mu. Simultaneously, based on model response intensity, stage sensitivity, and node contribution propagation results, the system generates a ranking of influencing factors, identifying high-temperature stress during the tasseling stage, soil moisture stability during the grain-filling stage, topdressing response during the jointing stage, and shared irrigation conditions among neighboring plots as the main contributing factors affecting the yield prediction results for plot D128.

[0074] Through the above implementation methods, spatiotemporal collaborative modeling can be carried out by jointly utilizing staged agricultural condition representation and confidence propagation results at the plot unit level, thereby improving the utilization of inter-plot correlation information and stage evolution information, enhancing the adaptability of yield prediction results to the state of missing data compensation and data credibility, and improving the stability and interpretability of prediction interval output and influence factor ranking results.

[0075] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0076] Figure 2 This is a schematic diagram of the structure of the agricultural product yield prediction system based on confidence propagation provided in an embodiment of this application. Figure 2 As shown, the system includes: Module 201 is used to construct a spatiotemporal index of plot units corresponding to the target agricultural products and to map multi-source agricultural observation data and agricultural management data to plot units; The generation module 202 is used to perform temporal reorganization, stage slicing and cross-source alignment on the mapped multi-source data based on the crop growth stage corresponding to the plot unit, and generate staged crop condition representations. The determination module 203 is used to determine the quality score based on the data integrity, source stability, spatiotemporal consistency and stage matching degree of each stage of agricultural condition representation, and to establish a confidence propagation relationship consisting of original observations, stage representations, compensation representations and fusion representations based on the quality score, and to perform cross-source compensation, conflict resolution and anomaly verification on low confidence data. Output module 204 is used to combine the spatial relationship and management transmission relationship between plot units, input the compensated staged agricultural condition representation and confidence propagation results into the graph time series fusion model, generate the yield prediction value, prediction interval and influence factor contribution ranking corresponding to the plot unit under uncertainty constraints, and output the yield prediction value, prediction interval and influence factor contribution ranking as the yield prediction result of the target plot.

[0077] In some embodiments, Figure 2The construction module 201 extracts the plot boundaries, planting season sequence and data source identifiers corresponding to the target planting area, constructs a unified index key for associated plot identifiers, time phase identifiers and source identifiers, and forms a plot unit spatiotemporal index; based on the plot unit spatiotemporal index, it performs spatial attribution matching, time benchmark conversion and scale normalization processing on multi-source agricultural observation data and agricultural management data, and generates index mapping results corresponding to each plot unit.

[0078] In some embodiments, Figure 2 The generation module 202 uses the crop growth stage as the time anchor point. Based on the planting time, stage transition characteristics and observation change trajectory of the plot unit, it determines the stage boundary of each plot unit. According to the stage boundary, the mapped multi-source data is reorganized into a unified stage coordinate, and cross-source registration and feature aggregation are performed under the stage association constraint to generate the staged agricultural condition representation of each plot unit.

[0079] In some embodiments, Figure 2 The generation module 202 constructs the corresponding fertility process prior based on the planting time sequence, and determines the stage transition probability corresponding to each time sequence position by combining stage transition characteristics and observed change trajectory; according to the stage transition probability, boundary positioning is performed under the constraints of stage continuity and plot unit time sequence consistency to generate the stage boundary corresponding to each plot unit.

[0080] In some embodiments, Figure 2 The generation module 202 projects the mapped multi-source data onto a unified stage coordinate based on the stage boundary, and determines the stage position relationship of each source data within the corresponding crop growth stage. Based on the stage position relationship, cross-source spatiotemporal registration, stage consistency verification and representation aggregation are performed under stage association constraints to generate staged agricultural condition representations corresponding to each plot unit.

[0081] In some embodiments, Figure 2 The determination module 203 extracts the data coverage status, source continuity status, spatiotemporal registration status and stage adaptation status corresponding to each stage of agricultural condition representation, and constructs quality assessment parameters corresponding to each stage of agricultural condition representation; based on the quality assessment parameters, a comprehensive quality assessment is performed under cross-source consistency constraints and quality correlation constraints to generate quality scores corresponding to each stage of agricultural condition representation.

[0082] In some embodiments, Figure 2The determination module 203 constructs hierarchical propagation nodes and inter-node propagation constraints based on the quality scores corresponding to the agricultural condition representations at each stage, and generates confidence propagation relationships. Based on the confidence propagation relationships, it identifies low-confidence data and determines the corresponding compensation paths, and performs cross-source compensation on the low-confidence data. Based on the compensated propagation status and cross-source consistency results, it performs conflict adjudication and anomaly verification on the low-confidence data, and generates confidence correction results.

[0083] In some embodiments, Figure 2 The determination module 203, based on the confidence propagation relationship and combined with the associated propagation state and stage constraint state of the propagation node where the low confidence data is located, determines the candidate compensation source corresponding to the low confidence data and generates the corresponding compensation path; according to the compensation path, it performs propagation weight allocation, confidence weighted fusion and stage consistency reconstruction on the candidate compensation source to generate the compensation representation corresponding to the low confidence data.

[0084] In some embodiments, Figure 2 The output module 204 constructs a plot association map corresponding to the graph-time series fusion model based on the spatial and management transmission relationships between plot units; it maps the compensated staged agricultural condition representation and confidence propagation results to the time series node representations corresponding to the plot association map, performs time series association modeling and graph association modeling, and generates the yield distribution representations corresponding to each plot unit; based on the yield distribution representations, it determines the yield prediction values ​​and prediction intervals corresponding to each plot unit under uncertainty constraints, and generates the impact factor contribution ranking based on the model response results.

[0085] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0086] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0087] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0088] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0089] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0090] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0091] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0092] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting agricultural product yield based on confidence propagation, characterized in that, include: Construct a spatiotemporal index for plot units corresponding to the target agricultural products, and map multi-source agricultural observation data and agricultural management data to the plot units; Based on the crop growth stage corresponding to the plot unit, temporal reorganization, stage slicing and cross-source alignment are performed on the mapped multi-source data to generate a staged agricultural condition characterization. A quality score is determined based on the data integrity, source stability, spatiotemporal consistency, and stage matching degree of each stage of agricultural condition representation. A confidence propagation relationship consisting of original observations, stage representations, compensation representations, and fusion representations is established based on the quality score. Cross-source compensation, conflict resolution, and anomaly verification are performed on low-confidence data. Combining the spatial relationships and management transmission relationships between plot units, the compensated staged agricultural condition representation and confidence propagation results are jointly input into the graph time series fusion model. Under uncertainty constraints, the predicted yield value, prediction interval, and ranking of influencing factors corresponding to the plot unit are generated, and the predicted yield value, prediction interval, and ranking of influencing factors are output as the yield prediction result of the target plot. Specifically, the step of performing temporal reorganization, stage slicing, and cross-source alignment on the mapped multi-source data based on the crop growth stage corresponding to the plot unit to generate a staged agricultural condition representation includes: Using the crop growth stage as a time-series anchor point, the stage boundaries corresponding to each plot unit are determined based on the planting time, stage transition characteristics, and observed change trajectory of the plot unit. According to the stage boundary, the mapped multi-source data is reorganized to a unified stage coordinate, and cross-source registration and feature aggregation are performed under stage association constraints to generate staged agricultural condition representations corresponding to each plot unit. The step of jointly inputting the compensated staged agricultural condition representation and confidence propagation results into the time-series fusion model to generate the yield forecast, prediction interval, and ranking of influencing factors corresponding to the plot unit under uncertainty constraints includes: Based on the spatial and management relationships between the land parcel units, a land parcel association map corresponding to the graph-time series fusion model is constructed. The compensated staged agricultural condition representation and the confidence propagation result are mapped to the time-series node representation corresponding to the plot association graph. Time-series association modeling and graph association modeling are performed to generate the yield distribution representation corresponding to each plot unit. Based on the aforementioned output distribution characterization, the predicted output values ​​and prediction intervals for each plot unit are determined under the aforementioned uncertainty constraints, and the contribution ranking of influencing factors is generated based on the model response results.

2. The method according to claim 1, characterized in that, The construction of a spatiotemporal index for plot units corresponding to the target agricultural product, and the mapping of multi-source agricultural observation data and agricultural management data to the plot units, includes: Extract the plot boundaries, planting season sequence, and data source identifiers corresponding to the target planting area, construct a unified index key for the associated plot identifier, time phase identifier, and source identifier, and form a spatiotemporal index of plot units; Based on the spatiotemporal index of the plot unit, spatial attribution matching, time reference conversion and scale normalization are performed on the multi-source agricultural observation data and the agricultural management data to generate index mapping results corresponding to each plot unit.

3. The method according to claim 1, characterized in that, The process of using the crop growth stage as a time-series anchor point, and determining the stage boundaries corresponding to each plot unit based on the planting time, stage transition characteristics, and observed change trajectories of the plot unit, includes: Based on the planting time sequence, a corresponding reproductive process prior is constructed, and combined with the stage transition characteristics and the observed change trajectory, the stage transition probability corresponding to each time sequence position is determined. Based on the stage transition probability, boundary positioning is performed under stage continuity constraints and land parcel unit temporal consistency constraints to generate stage boundaries corresponding to each land parcel unit.

4. The method according to claim 1, characterized in that, The process of reorganizing the mapped multi-source data to a unified stage coordinate according to the stage boundary, and performing cross-source registration and feature aggregation under stage association constraints to generate staged agricultural condition representations corresponding to each land parcel unit includes: Based on the stage boundaries, the mapped multi-source data is projected onto a unified stage coordinate system to determine the stage position relationship of each source data within the corresponding crop growth stage. Based on the stage position relationship, cross-source spatiotemporal registration, stage consistency verification, and representation aggregation are performed under the stage association constraints to generate staged agricultural condition representations corresponding to each land parcel unit.

5. The method according to claim 1, characterized in that, The quality score is determined based on the data completeness, source stability, spatiotemporal consistency, and stage matching degree of agricultural condition representation at each stage, including: Extract the data coverage status, source continuity status, spatiotemporal registration status and stage adaptation status corresponding to each stage of agricultural condition characterization, and construct quality assessment parameters corresponding to each stage of agricultural condition characterization; Based on the aforementioned quality assessment parameters, a comprehensive quality assessment is performed under cross-source consistency constraints and quality correlation constraints to generate quality scores corresponding to the agricultural condition representations at each stage.

6. The method according to claim 1, characterized in that, The process of establishing a confidence propagation relationship based on the quality score, consisting of original observations, stage representations, compensation representations, and fusion representations, and performing cross-source compensation, conflict resolution, and anomaly verification on low-confidence data, includes: Based on the quality scores corresponding to the agricultural condition representations at each stage, a hierarchical propagation node and inter-node propagation constraints are constructed to link the original observations, stage representations, compensation representations, and fusion representations, thereby generating confidence propagation relationships. Based on the confidence propagation relationship, low-confidence data is identified and the corresponding compensation path is determined, and cross-source compensation is performed on the low-confidence data; Based on the compensated propagation state and cross-source consistency results, conflict resolution and anomaly verification are performed on the low-confidence data to generate confidence correction results.

7. The method according to claim 6, characterized in that, The step of identifying low-confidence data and determining corresponding compensation paths based on the confidence propagation relationship, and performing cross-source compensation on the low-confidence data, includes: Based on the confidence propagation relationship, combined with the associated propagation state and stage constraint state of the propagation node where the low confidence data is located, the candidate compensation source corresponding to the low confidence data is determined, and the corresponding compensation path is generated. According to the compensation path, propagation weight allocation, confidence-weighted fusion, and phase consistency reconstruction are performed on the candidate compensation sources to generate the compensation representation corresponding to the low-confidence data.

8. A system for predicting agricultural product yield based on confidence propagation, according to any one of claims 1 to 7, characterized in that, include: The construction module is used to construct a spatiotemporal index of the plot unit corresponding to the target agricultural product, and to map multi-source agricultural observation data and agricultural management data to the plot unit; The generation module is used to perform temporal reorganization, stage slicing and cross-source alignment on the mapped multi-source data based on the crop growth stage corresponding to the plot unit, and generate staged agricultural condition representations. The determination module is used to determine the quality score based on the data integrity, source stability, spatiotemporal consistency and stage matching degree of each stage of agricultural condition representation, and to establish a confidence propagation relationship consisting of original observations, stage representations, compensation representations and fusion representations based on the quality score, and to perform cross-source compensation, conflict resolution and anomaly verification on low confidence data; The output module is used to combine the spatial relationship and management transmission relationship between plot units, input the compensated staged agricultural condition representation and confidence propagation results into the graph time series fusion model, generate the yield prediction value, prediction interval and influence factor contribution ranking corresponding to the plot unit under uncertainty constraints, and output the yield prediction value, prediction interval and influence factor contribution ranking as the yield prediction result of the target plot.

Images