Agricultural environment collaborative perception method and system based on multi-source data fusion

Abstract

This invention relates to the field of data fusion, and provides a method and system for collaborative perception of the agricultural environment based on multi-source data fusion. The method involves collecting data from sensors, drones, and satellites, and performing spatiotemporal registration to obtain a spatiotemporally aligned dataset. Next, based on the spatiotemporally aligned dataset, the method calculates the difference values ​​of multi-source data at the same spatiotemporal location to identify candidate anomaly regions. Then, it performs structural consistency discrimination on the candidate anomaly regions to identify uncorrectable and correctable anomalies. For correctable anomalies, a Bayesian inference algorithm is used to cross-validate and calibrate the spatiotemporally aligned dataset to obtain a high-confidence dataset. Finally, using the high-confidence dataset as input and the uncorrectable anomalies as constraints, a machine learning model and spatial interpolation method are used to predict and complete the data in unmonitored areas, generating comprehensive agricultural environmental data for the entire agricultural region. This achieves complete and highly reliable perception of environmental information for the entire agricultural region.

Description

Technical Field

[0001] This invention relates to the field of data fusion, and in particular to a method and system for collaborative sensing of the agricultural environment based on multi-source data fusion. Background Technology

[0002] Agricultural environmental monitoring is a crucial foundation for precision agriculture and smart agriculture, and its results directly impact the accuracy of crop growth regulation, production management decisions, and disaster early warning. Current agricultural environmental monitoring typically relies on multiple data sources, including ground sensors, UAV remote sensing, and satellite remote sensing. Each data source has its own advantages in terms of spatial resolution, temporal continuity, coverage, and monitoring accuracy.

[0003] However, due to differences in acquisition methods, observation scales, and imaging mechanisms among multi-source data, numerical deviations or abnormal data fluctuations often exist at the same spatiotemporal location. Existing technologies mostly employ data stitching, weighted averaging, or empirical rules for fusion, lacking in-depth analysis of the causes and structural characteristics of multi-source data differences. This makes it difficult to effectively distinguish between anomalies caused by real environmental factors and those caused by sensor errors, remote sensing noise, or spatiotemporal mismatches. Regarding anomaly data processing, existing methods typically uniformly remove or correct anomalies without a mechanism to differentiate anomaly types, leading to the erroneous deletion of valid anomaly information and reducing the reliability of overall monitoring results. Furthermore, in the verification and calibration stages of anomaly data, using only simple weighted calibration methods treats real environmental anomalies alongside ordinary noise, forcibly weakening the numerical characteristics of real anomalies and masking key environmental anomaly information, thus affecting the accuracy of agricultural environmental collaborative sensing.

[0004] Therefore, it is necessary to propose a method that enables collaborative sensing of multi-source agricultural environmental data within a unified spatiotemporal framework, in order to improve the accuracy and reliability of agricultural environmental monitoring and analysis. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a collaborative sensing method and system for agricultural environment based on multi-source data fusion. This method can perform numerical prediction and supplementation of agricultural environmental data in unmonitored areas based on sensor, drone and satellite monitoring data, generate agricultural environmental data covering the entire area, realize complete and highly reliable perception of environmental information of the entire agricultural area, and provide comprehensive data support for agricultural production management, environmental monitoring and decision-making.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A collaborative sensing method for the agricultural environment based on multi-source data fusion, the method comprising:

[0008] Collect sensor, drone and satellite monitoring data, and perform spatiotemporal registration on the sensor, drone and satellite monitoring data to obtain a spatiotemporal location aligned dataset under the same spatiotemporal framework;

[0009] Based on the spatiotemporal location aligned dataset, the difference value is calculated for multi-source data at the same spatiotemporal location. When the difference value exceeds a preset threshold, the spatiotemporal location is marked as a candidate abnormal region.

[0010] The candidate abnormal regions are subjected to structural consistency discrimination. Abnormal regions with structural consistency are marked as uncorrectable abnormalities, and abnormal regions without structural consistency are marked as correctable abnormalities. For the correctable abnormalities, the spatiotemporal location alignment dataset is cross-validated and calibrated using a Bayesian inference algorithm to obtain a high-confidence dataset.

[0011] Using the high-confidence dataset as input and the uncorrectable anomaly as a condition constraint, the data of the unmonitored area is predicted and completed using machine learning models and spatial interpolation methods to generate agricultural environmental data for the entire region.

[0012] Spatiotemporal registration is performed on the sensor, UAV, and satellite monitoring data to obtain a spatiotemporally aligned dataset within the same spatiotemporal framework, including:

[0013] Based on the needs of agricultural environmental monitoring, a geographic coordinate reference system and time recording standard are set up, and a multi-source data spatial benchmark framework and time benchmark axis are constructed.

[0014] By using grid mapping and spatial resampling techniques, the sensor, UAV, and satellite monitoring data are mapped to a spatial reference frame to obtain spatially consistent multi-source data;

[0015] By using time window mapping and time interpolation algorithms, the spatially consistent multi-source data is mapped to a time reference axis to obtain spatiotemporally consistent multi-source data;

[0016] Data loss repair and spatiotemporal consistency verification are performed on the spatiotemporally consistent multi-source data to generate a spatiotemporally aligned dataset under the same spatiotemporal framework.

[0017] Based on the spatiotemporal location aligned dataset, the difference values ​​of multi-source data at the same spatiotemporal location are calculated, including:

[0018] Based on the spatiotemporal location alignment dataset, obtain sensor, drone, and satellite monitoring data under the same spatial location and the same time window;

[0019] Select sensor monitoring data or historical statistical benchmark data as reference data;

[0020] Based on the reference data, the parameter values ​​of the UAV monitoring data and satellite monitoring data are compared with the reference data to obtain the UAV monitoring difference value and the satellite monitoring difference value;

[0021] The difference values ​​are associated with their corresponding spatial location identifiers, time window identifiers, and data source identifiers to form a difference value data set.

[0022] Structural consistency determination is performed on the candidate abnormal regions, including:

[0023] Extract the multidimensional structural features of the candidate anomaly region. The multidimensional structural features include spatial morphological features, neighborhood distribution features, and temporal variation features.

[0024] Based on historical high-reliability area samples, a structural reference model for monitoring indicators is constructed as a benchmark for judging the structure of candidate abnormal areas.

[0025] The multidimensional structural features of the candidate anomaly region are matched with the structural reference model of the monitoring indicators to calculate the comprehensive structural consistency score.

[0026] The overall structural consistency score is compared with the preset consistency threshold. Candidate abnormal regions with an overall structural consistency score greater than or equal to the preset consistency threshold are marked as uncorrectable abnormalities, while candidate abnormal regions with an overall structural consistency score less than the preset consistency threshold are marked as correctable abnormalities.

[0027] The multidimensional structural features of the candidate anomaly regions are matched with the structural reference models of the monitoring indicators to calculate a comprehensive structural consistency score, including:

[0028] Based on the spatial morphological feature components of the structural reference model, the spatial morphological deviation of the candidate anomaly region is calculated using a morphological distance metric method.

[0029] Based on the neighborhood distribution feature components of the structural reference model, the neighborhood distribution deviation of the candidate anomaly region is calculated using a distribution difference measurement method.

[0030] Based on the temporal variation characteristic components of the structural reference model, the temporal variation deviation of the candidate anomaly region is calculated using a dynamic time warping algorithm;

[0031] The deviation of the multidimensional structural features is normalized and mapped to a consistency sub-score.

[0032] Based on the importance of multidimensional structural features in agricultural environment anomaly detection, corresponding weights are set, and the consistency sub-scores are weighted according to the weights to obtain the comprehensive structural consistency score.

[0033] For the correctable anomalies, the spatiotemporal alignment dataset is cross-validated and calibrated using a Bayesian inference algorithm to obtain a highly reliable dataset, including:

[0034] Based on the spatiotemporal location and multi-source data corresponding to the correctable anomaly, an anomaly cross-validation data unit is constructed.

[0035] Based on the data source type and historical monitoring accuracy, a prior confidence level is set for the abnormal cross-validation data unit, and a Bayesian prior probability model is constructed.

[0036] Based on the aforementioned cross-validation data unit and the Bayesian prior probability model, the posterior estimation results of multi-source data at the same spatiotemporal location are calculated through Bayesian inference, and the correctable anomalies are calibrated and corrected based on the posterior estimation results.

[0037] The calibrated data and data from non-abnormal areas are integrated to form a highly reliable dataset.

[0038] By using machine learning models and spatial interpolation methods to predict and complete data from unmonitored areas, agricultural environmental data for the entire region is generated, including:

[0039] Using the high-confidence dataset as input, a nonlinear mapping relationship between agricultural environmental parameters and spatial distribution characteristics is established through a machine learning model, generating prediction results for environmental parameters in unmonitored areas;

[0040] Based on the spatial distribution characteristics of the high-confidence dataset, the environmental parameters of the unmonitored area are estimated by spatial interpolation method to obtain spatial interpolation results;

[0041] The prediction results of the machine learning model and the spatial interpolation results are weighted and fused to generate agricultural environmental data for the entire region.

[0042] Using the uncorrectable anomaly as a conditional constraint includes:

[0043] When predicting environmental parameters using machine learning models, the range of prediction results for environmental parameters in uncorrectable anomaly regions is limited by introducing the spatiotemporal location of the uncorrectable anomaly region and multi-source data.

[0044] When estimating environmental parameters using spatial interpolation methods, the spatiotemporal location and multi-source data corresponding to the uncorrectable anomaly region are used as interpolation boundaries or interpolation weights.

[0045] An agricultural environment collaborative sensing system based on multi-source data fusion, the system includes a spatiotemporal registration module, a candidate anomaly identification module, a cross-validation and calibration module, and a data prediction and completion module;

[0046] The spatiotemporal registration module is used to map multi-source monitoring data to the same spatiotemporal framework and generate a spatiotemporal location aligned dataset.

[0047] The candidate anomaly identification module is used to calculate the difference value of multi-source data at the same spatiotemporal location, and identify candidate anomaly regions by comparing the difference value with a preset threshold.

[0048] The cross-validation and calibration module is used to perform cross-validation and calibration on the spatiotemporally aligned dataset to generate a highly reliable dataset.

[0049] The data prediction and completion module is used to generate agricultural environmental data for the entire region based on a highly reliable dataset.

[0050] The cross-validation and calibration module includes a structural consistency discrimination unit, an anomaly calibration unit, and a high-confidence dataset generation unit.

[0051] The structural consistency discrimination unit is used to determine whether a candidate abnormal region is a correctable abnormality;

[0052] Anomaly calibration unit is used to perform cross-validation and probabilistic calibration on correctable anomalies;

[0053] The high-confidence dataset generation unit is used to integrate calibrated data and non-abnormal region data to generate a high-confidence dataset.

[0054] Compared with the prior art, the beneficial effects of this application are:

[0055] This application unifies sensor, UAV, and satellite monitoring data into a single spatiotemporal framework, enabling collaborative sensing of multi-source data under comparable conditions. By introducing a structural consistency discrimination mechanism, it identifies correctable anomalies in candidate anomaly regions from multiple dimensions, including spatial morphology, neighborhood distribution, and temporal changes, effectively distinguishing between anomalies caused by real agricultural environmental changes and those caused by sensor errors, remote sensing noise, or spatiotemporal mismatches. For correctable anomalies, Bayesian inference is introduced for cross-validation and probability calibration, quantitatively correcting correctable anomaly data based on historical reliable data to reduce error accumulation and generate a high-reliability dataset. By using the high-reliability dataset as input and uncorrectable anomalies as constraints, machine learning models and spatial interpolation methods are combined to predict and complete agricultural environmental data in unmonitored areas. This method improves the accuracy and reliability of multi-source agricultural environmental data fusion without weakening real environmental anomalies, achieving continuous monitoring and reliable sensing of the agricultural environment across the entire region. Attached Figure Description

[0056] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0057] Figure 1 This is a flowchart illustrating the collaborative sensing method for the agricultural environment based on multi-source data fusion, as described in an embodiment of the present invention.

[0058] Figure 2 A schematic diagram illustrating the process of generating a set of difference value data in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the process for generating a highly reliable dataset according to an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the process for calculating the overall structural consistency score according to an embodiment of the present invention;

[0061] Figure 5 This is a block diagram of an agricultural environment collaborative sensing system based on multi-source data fusion, according to an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0063] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0064] In further explaining the application scenarios and technical concepts of this application, it is necessary to point out that the agricultural environment collaborative sensing method based on multi-source data fusion is not a completely new creation of existing technologies, but rather a precise improvement and optimization upgrade of traditional methods based on agricultural monitoring systems and data fusion technologies. Traditional agricultural environment multi-source data fusion methods mostly rely on simple means such as data splicing and weighted averaging to integrate data. They lack in-depth analysis of the causes of differences in multi-source data and have not established an effective anomaly classification mechanism, making it difficult to distinguish between real environmental anomalies and false anomalies such as sensor errors and remote sensing noise, thus affecting the reliability of monitoring results.

[0065] The agricultural environment collaborative sensing method involved in this application can be widely applied to multiple core scenarios in the entire smart agriculture industry chain, covering large-scale agricultural monitoring, refined management and control of facility agriculture, early warning and prevention of agricultural disasters, and dynamic management of agricultural resources. These scenarios all take precise monitoring of the agricultural environment as their core requirement, but due to differences in monitoring scale, spatial characteristics, or core objectives, they face problems that traditional data fusion technologies cannot solve.

[0066] For example, in large-scale field planting scenarios, due to the wide monitoring range and complex environmental variables, problems such as spatiotemporal asynchrony and large numerical deviations in multi-source monitoring data often occur. Traditional methods lack effective spatiotemporal unification mechanisms and difference analysis capabilities, making it difficult to distinguish between real environmental changes and data anomalies such as sensor errors and remote sensing noise. At the same time, monitoring blind spots under wide coverage lead to incomplete data coverage, directly affecting the efficiency of accurate decision-making in large-scale production. In the scenario of refined management and control of facility agriculture, which involves environmental control in closed or semi-closed spaces, the requirements for data accuracy and real-time performance are extremely high. However, traditional data fusion methods often use simple weighting or splicing methods, which easily confuse sensor errors with real environmental anomalies, leading to inaccurate calculation of control parameters and affecting crop growth stability.

[0067] For example, the core requirement of agricultural disaster early warning and prevention scenarios is to quickly identify environmental anomalies caused by disasters, thus buying time for disaster prevention and mitigation. However, traditional methods lack a structured mechanism for identifying anomaly types and rely solely on single threshold judgments, which can easily lead to delayed early warnings or misjudgments, missing the best opportunity for prevention and control. Agricultural resource dynamic management scenarios, on the other hand, face the pain points of low reliability of resource distribution monitoring data and missing data for unmonitored areas. Traditional methods struggle to achieve comprehensive perception and dynamic updating of resource data, making it difficult to support the efficient and rational allocation of water resources, soil resources, and other resources.

[0068] To address the aforementioned challenges, this application establishes a unified spatiotemporal benchmark to eliminate differences in collection scale and time dimension among multi-source data, enabling data from different sources to have a comparable basis. It overcomes the limitations of traditional single-threshold judgment by distinguishing between real-environmental anomalies and data error-related anomalies at the structural feature level, avoiding the accidental deletion of valid information or the retention of false anomalies. For correctable error-related anomalies, a probabilistic calibration mechanism is introduced, combining data credibility and historical experience to achieve precise correction and improve data reliability. Based on highly reliable data and combined with spatial and model prediction capabilities, information on unmonitored areas is supplemented, achieving complete and highly reliable perception of the agricultural environment across the entire region, providing comprehensive data support for agricultural production decisions.

[0069] Next, with reference to the accompanying drawings, the agricultural environment collaborative sensing method based on multi-source data fusion provided in the embodiments of this application will be further described. Please refer to the accompanying drawings. Figure 1 , Figure 1This is a flowchart illustrating the agricultural environment collaborative sensing method based on multi-source data fusion according to an embodiment of the present invention. The specific steps of the method are as follows:

[0070] S1: Collect sensor, drone and satellite monitoring data, perform spatiotemporal registration on the sensor, drone and satellite monitoring data to obtain a spatiotemporal location aligned dataset under the same spatiotemporal framework;

[0071] In this embodiment, agricultural environmental monitoring data of the target agricultural area is collected by ground sensors, drones and satellite remote sensing systems. The sensor, drone and satellite monitoring data are uniformly spatiotemporally registered. Through spatial coordinate transformation, spatial grid mapping and time window mapping, the sensor, drone and satellite monitoring data are mapped to the same geographic coordinate reference system and unified time base to obtain a spatiotemporally aligned dataset under the same spatiotemporal framework.

[0072] S2: Based on the spatiotemporal location aligned dataset, calculate the difference value of multi-source data at the same spatiotemporal location. When the difference value exceeds a preset threshold, mark the spatiotemporal location as a candidate abnormal region.

[0073] In this embodiment, sensor monitoring data or historical statistical benchmark data are used as a reference to compare the UAV monitoring data and satellite monitoring data numerically, calculate the difference value between different data sources, and mark the corresponding spatiotemporal location as a candidate abnormal area when the difference value exceeds a preset threshold.

[0074] S3: Perform structural consistency discrimination on the candidate abnormal regions, mark the abnormal regions with structural consistency as uncorrectable abnormalities, and mark the abnormal regions without structural consistency as correctable abnormalities. For the correctable abnormalities, perform cross-validation and calibration on the spatiotemporal location alignment dataset using a Bayesian inference algorithm to obtain a high-confidence dataset.

[0075] In this embodiment, spatial morphological features, neighborhood distribution features, and temporal variation features of candidate abnormal regions are extracted and matched with a structural reference model constructed based on historical high-confidence region samples to calculate a comprehensive structural consistency score. The comprehensive structural consistency score is compared with a preset consistency threshold. Based on the comparison results, abnormal regions with a comprehensive structural consistency score greater than or equal to the preset consistency threshold are determined as uncorrectable abnormalities, and abnormal regions with a comprehensive structural consistency score less than the preset consistency threshold are determined as correctable abnormalities. For the correctable abnormalities, an abnormal cross-validation data unit is constructed, and a Bayesian inference algorithm is introduced. The prior confidence level is set in combination with the data source type and historical monitoring accuracy. Cross-validation and probability calibration are performed on multi-source data at the same spatiotemporal location to finally obtain a calibrated high-confidence dataset.

[0076] S4: Using the high-confidence dataset as input and the uncorrectable anomaly as a condition constraint, predict and complete the data of the unmonitored area through machine learning models and spatial interpolation methods to generate agricultural environmental data for the entire region.

[0077] In this embodiment, a nonlinear mapping relationship between agricultural environmental parameters and spatial distribution characteristics is established through a machine learning model to predict environmental parameters in unmonitored areas. Simultaneously, based on the spatial distribution characteristics of a high-confidence dataset, a spatial interpolation method is used to estimate environmental parameters in unmonitored areas. The prediction results of the machine learning model and the spatial interpolation results are weighted and fused to generate environmental data covering the entire agricultural area.

[0078] Next, we will further elaborate on the technical aspects of the method in this application regarding the generation of spatiotemporally aligned datasets within the same spatiotemporal framework.

[0079] S1.1: Based on the needs of agricultural environmental monitoring, establish a geographic coordinate reference system and time recording standards, and construct a multi-source data spatial reference framework and time reference axis;

[0080] In this step, the agricultural environment is simultaneously affected by regional-scale climate change, plot-scale crop growth, and changes in microenvironmental conditions. A single data source is insufficient to comprehensively and accurately reflect the true state of the agricultural environment, necessitating the introduction of multiple monitoring methods for complementary sensing. Satellite remote sensing data is mainly used to provide periodically acquired wide-area, multispectral agricultural environmental information, while high-definition remote sensing image data acquired by UAVs is used to supplement the shortcomings of satellite data in terms of spatial resolution and observation timeliness. Internet of Things (IoT) sensors, as a near-ground monitoring method, are used to continuously and in real-time collect agricultural environmental parameters such as soil temperature, soil moisture, soil pH, and air temperature and humidity.

[0081] Spatially, based on the size of the agricultural monitoring area, the accuracy of plot division, and the sensitivity of monitoring parameters to spatial resolution, a unified geographic coordinate reference system is determined. This system can use geodetic coordinates or unified projected coordinates, defining the coordinate datum, coordinate axis directions, coordinate units, and spatial accuracy level. Temporally, based on the variation cycle of agricultural environmental parameters, monitoring response speed requirements, and differences in sampling frequencies from different data sources, a time recording standard is selected, determining the timestamp format, time accuracy, and time synchronization principles.

[0082] S1.2: By using grid mapping and spatial resampling techniques, the sensor, UAV and satellite monitoring data are mapped to a spatial reference frame to obtain spatially consistent multi-source data;

[0083] In this step, the spatial reference frame is first defined with a regularized grid, defining the spatial range, resolution, and spatial index identifier of each spatial grid. Secondly, the data collected by the ground sensors is mapped. Specifically, for monitoring data collected from fixed sensor locations, the spatial grid cell to which the sensor belongs is determined based on the sensor's geographic coordinates. The corresponding monitoring value is then mapped to that spatial grid cell. When multiple sensor locations are mapped to the same spatial grid cell, their monitoring values ​​are aggregated to form representative sensor data for that spatial grid cell, thus realizing the mapping of point data to the spatial reference frame.

[0084] Then, the UAV remote sensing data is mapped. Specifically, the original pixel coordinate system of the UAV image is transformed to the geographic coordinate reference system adopted by the spatial reference frame. Through image georegistration technology, the exterior orientation elements of the image and the mapping model between pixels and geographic coordinates are determined. By comparing the pixel coordinates and the spatial range boundaries of each spatial grid unit, the target grid corresponding to a single pixel is determined. If a pixel spans multiple grid units, its main grid or split allocation needs to be determined according to the pixel center attribution principle or area weighted allocation principle. An index association between the pixel and the corresponding spatial grid unit is established, and the monitoring data of the UAV pixel is accurately allocated to the spatial reference grid unit to which it belongs.

[0085] Finally, spatial resampling processing is performed on the satellite remote sensing data. Specifically, due to the scale difference between the spatial resolution of the satellite remote sensing data and the spatial reference grid, coordinate alignment processing is performed on the satellite remote sensing data. According to the resolution requirements of the spatial reference frame, the spatial scale of the satellite remote sensing data is adjusted, and a data representation form that matches the spatial reference grid unit is generated through resampling technology. Furthermore, during the resampling process, based on the requirements of agricultural environmental monitoring for data accuracy and continuity, the target resolution is determined, the resampling strategy is selected, and the pixel value is reconstructed. The resampling strategy can include nearest neighbor resampling, bilinear resampling, or area-weighted resampling. If data timeliness is prioritized and local detail accuracy is less critical, nearest neighbor resampling can be used, directly selecting the satellite pixel value closest to the center of the target grid as the grid data. If a balance between data smoothness and accuracy is required, bilinear resampling can be used, determining the grid data by calculating the weighted average of the four neighboring pixels surrounding the target grid. If the scale difference between the satellite data and the spatial reference grid is significant, and the overall regional data distribution characteristics need to be considered, an area-weighted resampling method can be used, allocating weights based on the overlap area ratio between the satellite pixels and the target grid, and calculating the weighted summed grid data.

[0086] S1.3: By using time window mapping and time interpolation algorithms, the spatially consistent multi-source data is mapped to the time reference axis to obtain spatiotemporally consistent multi-source data;

[0087] Specifically, firstly, the time reference axis is discretized according to a fixed time resolution or an adaptive time step to form a continuous and ordered set of standard time windows. Each standard time window corresponds to a start time and an end time, used to constrain the time allocation range of multi-source data. Secondly, the original timestamp information of the spatially consistent multi-source data is parsed. Based on the correspondence between the timestamps and the time reference axis, the spatially consistent multi-source data is initially assigned to the target time window closest to its time position. For spatially consistent multi-source data spanning time windows, the data is split according to the time coverage ratio or sampling density. Then, the time series from different data sources within each time window are aligned to identify situations where there are inconsistent time sampling intervals, inconsistent sampling frequencies, or missing time. Based on preset time interpolation rules, missing or sparse time nodes are interpolated for compensation. The time interpolation rules may include linear interpolation, piecewise polynomial interpolation, or interpolation methods based on historical sequence trends, generating continuous time series data that correspond one-to-one with the nodes of the time reference axis. Finally, the multi-source data, after being processed by time window mapping and time interpolation, is reorganized and indexed according to the time reference axis order to form spatiotemporally consistent multi-source data.

[0088] S1.4: Perform data missing repair and spatiotemporal consistency verification on the spatiotemporally consistent multi-source data to generate a spatiotemporally aligned dataset under the same spatiotemporal framework.

[0089] In this step, the spatiotemporally consistent multi-source data is subjected to an integrity scan. According to the combination relationship between spatial grid cells and time reference axes, a spatiotemporal index matrix is ​​constructed. The data occupancy status corresponding to each spatiotemporal position is detected in the spatiotemporal index matrix, and spatiotemporal nodes with data gaps, abnormal interruptions, or no valid observations are marked to form a missing data identifier set.

[0090] Secondly, for the set of missing data identifiers, missing data repair constraints are constructed based on the spatial location, time domain, and data source type of the missing data nodes. Multi-source observation data that are spatially adjacent and temporally continuous with the missing data nodes are selected as reference samples. Repair algorithms based on spatial neighborhood weighting, time series extrapolation, or multi-source correlation constraints are used to reconstruct the data values ​​of the missing data nodes and generate candidate datasets.

[0091] Then, the candidate dataset is subjected to spatiotemporal consistency verification. According to the preset spatial consistency rules and temporal consistency rules, the continuity of data at the same spatial location at different time nodes and the coordination of data at the same time node at different spatial locations are verified, including spatial adjacency difference detection, time rate of change constraint detection and cross-data source numerical interval consistency detection.

[0092] Finally, for spatiotemporal nodes that fail the spatiotemporal consistency check, the corresponding check deviation information is recorded. Based on the deviation type, secondary repair or data backtracking is triggered, and the spatiotemporal nodes are reintroduced into the missing data repair process until all spatiotemporal nodes meet the preset spatiotemporal consistency rules. The multi-source data that passes the spatiotemporal consistency check are rearranged and encapsulated according to spatial coordinate identifiers and time window identifiers to form a spatiotemporally aligned dataset within the same spatiotemporal framework.

[0093] Next, we will further elaborate on the technical content of the method in this application regarding the generation of a set of difference value data.

[0094] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the process of generating a set of difference value data in an embodiment of the present invention.

[0095] S2.1: Based on the spatiotemporal location alignment dataset, obtain sensor, drone, and satellite monitoring data under the same spatial location and the same time window;

[0096] S2.2: Select sensor monitoring data or historical statistical benchmark data as reference data;

[0097] Specifically, using the spatial coordinate identifiers and time window identifiers in the spatiotemporal location alignment dataset as joint index conditions, a multidimensional search is performed on the spatiotemporal location alignment dataset to extract sensor monitoring data, UAV monitoring data, and satellite monitoring data within the same spatial grid cell and the same time window. Among them, the spatial dimension is used to characterize the specific location of farmland or monitoring area in a unified coordinate system, the temporal dimension is used to characterize the synchronous sampling interval of multi-source data on the time reference axis, and the data source dimension is used to distinguish data records formed by different acquisition platforms.

[0098] Then, the multi-source monitoring data from the same spatial location and time window are structured and organized, constructing multi-source data units according to a three-dimensional data organization method of "spatial location-time window-data source type". Within each multi-source data unit, the monitoring indicators of each data source are analyzed for attributes and labeled with dimensions, classifying the monitoring indicators into environmental state indicators, growth state indicators, and background condition indicators. Environmental state indicators are used to describe meteorological and soil-related monitoring parameters, growth state indicators are used to describe observation parameters of crop canopy or surface characteristics, and background condition indicators are used to describe auxiliary parameters related to acquisition conditions or imaging status.

[0099] Finally, sensor monitoring data or historical statistical benchmark data are selected from the multi-source data units as reference data. Specifically, when the sensor monitoring data has a complete timestamp, stable sampling frequency, and clear calibration relationship within the spatial location and time window, the sensor monitoring data is marked as reference data. When the sensor monitoring data is missing, abnormal, or does not meet the reference conditions, historical statistical data matching the spatial location, crop type, and time window are retrieved from the historical statistical benchmark dataset and used as reference data. The selected reference data is standardized by adding spatial location identifiers, time window identifiers, and indicator dimension labels.

[0100] S2.3: Based on the reference data, compare the parameter values ​​of the UAV monitoring data and satellite monitoring data with the reference data to obtain the UAV monitoring difference value and the satellite monitoring difference value;

[0101] In this step, based on the reference data, parameter dimension mapping is performed on the UAV monitoring data and satellite monitoring data respectively, establishing a correspondence between the UAV monitoring data, satellite monitoring data and the reference data. Specifically, the parameter dimensions may include environmental parameter dimensions, growth state parameter dimensions, and surface feature parameter dimensions.

[0102] Secondly, for each parameter dimension, the corresponding UAV monitoring parameter value and satellite monitoring parameter value are extracted within the time window, and compared with the corresponding parameter value in the reference data. The comparison process can be performed independently for each parameter dimension by calculating the absolute difference, relative difference, or based on the standardized difference method, to obtain the UAV monitoring difference value and the satellite monitoring difference value.

[0103] S2.4: Associate the difference value with the corresponding spatial location identifier, time window identifier and data source identifier to form a difference value data set. When the difference value exceeds a preset threshold, mark the spatiotemporal location as a candidate abnormal region.

[0104] Specifically, the differences between UAV and satellite monitoring are structurally encapsulated and associated in the form of "spatial location-time window-data source type-difference value" to obtain difference value data units. These difference value data units are then aggregated according to parameter dimensions to form a difference value data set. Based on parameter dimension type, data source type, and historical statistical distribution, difference value thresholds are set. For each spatiotemporal location, the difference value is compared against these thresholds to determine if at least one difference record in the difference value data set exceeds its corresponding preset threshold. The preset thresholds are set for three dimensions: environmental parameters, growth status parameters, and surface characteristic parameters. Environmental parameters include soil temperature, air temperature, and soil pH, and are quantified using absolute thresholds. For example, for soil temperature, the difference value threshold for UAV monitoring is ±1.5℃, and for satellite monitoring, it is ±3℃. Growth status parameters include vegetation index, canopy coverage, and crop biomass. Parameters such as leaf area index are quantified using relative thresholds. For example, for vegetation index parameters, the threshold for difference values ​​monitored by drones is ±4%, and the threshold for difference values ​​monitored by satellites is ±5%. Surface feature parameters, including surface reflectance, surface roughness, and imaging brightness, are quantified using standardized thresholds. For example, for surface reflectance parameters, the threshold for difference values ​​monitored by drones is ±1.96, and the threshold for difference values ​​monitored by satellites is ±2.58. When the difference record corresponding to a certain spatiotemporal location exceeds the preset threshold, a candidate anomaly identifier is added to that spatiotemporal location. This identifier is then associated and stored with the corresponding spatial location coordinates, time window number, and parameter dimension information that triggered the threshold exceedance. Finally, according to spatial adjacency and temporal continuity, the spatiotemporal locations of adjacent or consecutive anomaly identifiers are merged to obtain candidate anomaly regions, and each candidate anomaly region is assigned a unique region number.

[0105] Next, we will further elaborate on the technical aspects of the method in this application regarding the generation of highly reliable datasets.

[0106] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the process of generating a highly reliable dataset according to an embodiment of the present invention.

[0107] S3.1: Extract the multidimensional structural features of the candidate anomaly region, including spatial morphological features, neighborhood distribution features, and temporal variation features;

[0108] Specifically, based on the spatial location identifier of the candidate anomaly region, the spatial boundary corresponding to the candidate anomaly region is determined under the spatial reference framework. The spatial boundary is discretized based on rasterization rules or vector contour rules to form a set of anomaly region spatial units composed of multiple basic spatial units.

[0109] Secondly, based on the set of spatial units in the abnormal region, spatial morphological features are extracted. These features may include region area, perimeter, shape compactness, aspect ratio, boundary curvature distribution, and connectivity parameters. The connectivity parameters are obtained by calculating the geometric relationships, boundary point distribution, and connectivity structure of the spatial units in the abnormal region.

[0110] Then, a spatial neighborhood of a preset scale is constructed with the abnormal area as the center. Normal area spatial units within the neighborhood are selected, and the distribution relationship between the abnormal area and its spatial neighborhood is statistically analyzed. Neighborhood distribution features are extracted, which may include the mean distribution, variance distribution, and gradient change rate of monitoring indicators within the neighborhood.

[0111] Next, based on the time reference axis, multi-time data corresponding to the candidate abnormal region are extracted along the time dimension to construct a time series data set of the abnormal region. Temporal change features are extracted from the time series data set, which may include parameter change amplitude, change rate, consistency of change direction, periodicity, and change point location.

[0112] Finally, the extracted spatial morphological features, neighborhood distribution features, and temporal change features are uniformly encoded and structurally encapsulated, and each type of feature is given a corresponding spatial identifier, neighborhood identifier, and temporal identifier to form a multi-dimensional structural feature set corresponding to the candidate abnormal region.

[0113] S3.2: Based on historical high-reliability area samples, construct a structural reference model for monitoring indicators as a benchmark for judging the structure of candidate abnormal areas;

[0114] Specifically, when constructing the structural reference model, spatial morphological features, neighborhood distribution features, and temporal change features of the same type as the candidate anomaly region are extracted from the historical high-confidence region samples. Statistical modeling is performed on various structural features to obtain descriptions of various structural features under normal agricultural environmental conditions.

[0115] S3.3: Match the multidimensional structural features of the candidate abnormal region with the structural reference model of the monitoring indicators, and calculate the comprehensive structural consistency score;

[0116] Please see Figure 4 , Figure 4 The flowchart for calculating the overall structural consistency score in an embodiment of the present invention is shown in section S3.3. The specific steps are as follows:

[0117] S3.3.1: Based on the spatial morphological feature components of the structural reference model, calculate the spatial morphological deviation of the candidate anomaly region using a morphological distance metric method;

[0118] In this step, spatial morphological feature components are first extracted from the structural reference model, and a morphological distance metric function is selected. Then, the distance value of each morphological parameter in the spatial morphological features is calculated using the morphological distance metric function. Finally, a morphological parameter weighting coefficient is introduced to weight the distance value of each morphological parameter to obtain the spatial morphological distance value between the candidate anomaly region and the structural reference model, that is, the spatial morphological deviation of the candidate anomaly region.

[0119] Specifically, the distance values ​​of each morphological parameter in the spatial morphological features are calculated using morphological distance metrics. These morphological distance metrics may include Euclidean distance-based metrics, weighted Manhattan distance-based metrics, and morphological parameter similarity-based metrics. The selection of the morphological distance metric is based on a comprehensive assessment of the parameter attributes of the spatial morphological feature components in the structural reference model and the actual needs of agricultural environmental monitoring. First, the spatial morphological feature components in the structural reference model are classified into continuous numerical parameters, discrete feature parameters, and proportional similarity parameters. Region area, perimeter, and boundary curvature distribution are continuous numerical parameters; connectivity parameters are discrete feature parameters; and shape compactness and aspect ratio are proportional similarity parameters. Then, different metric functions are matched for different types of parameters. For continuous numerical parameters, Euclidean distance-based metrics are preferred; for discrete feature parameters, weighted Manhattan distance-based metrics are selected; and for proportional similarity parameters, morphological parameter similarity-based metrics are selected.

[0120] Furthermore, morphological parameter weighting coefficients are introduced to constrain the participation ratio of different spatial morphological parameters. The distance values ​​of each morphological parameter are weighted and summed according to the parameter dimension order to obtain the spatial morphological distance value between the candidate anomaly region and the structural reference model. Finally, this spatial morphological distance value is used as the spatial morphological deviation degree of the candidate anomaly region.

[0121] S3.3.2: Based on the neighborhood distribution feature components of the structural reference model, calculate the neighborhood distribution deviation of the candidate anomaly region using a distribution difference measurement method;

[0122] In this step, firstly, neighborhood distribution feature components are extracted from the structural reference model, and the neighborhood distribution features of the candidate anomaly region are formally aligned with the neighborhood distribution feature components of the structural reference model. Then, the differences between the candidate anomaly region and the structural reference model in each neighborhood distribution parameter dimension are calculated using a neighborhood distribution difference measurement function. The distribution difference measurement function can be a relative entropy measurement function, Jensen-Shannon divergence, minimum working distance function, or a measurement method based on statistical distance. Finally, the differences in each neighborhood distribution parameter dimension are quantified, and the quantification results are integrated to form the neighborhood distribution deviation of the candidate anomaly region.

[0123] It should be noted that, for formal alignment of the neighborhood distribution features of candidate anomaly regions with the neighborhood distribution feature components of the structural reference model, the following steps are taken: First, the distribution description elements of the neighborhood distribution feature components in the structural reference model are extracted, including the distribution interval division method, statistical type, and probability or frequency expression form, as the benchmark template for alignment. Second, based on multi-source monitoring data within the neighborhood of the candidate anomaly region, the neighborhood distribution features of the candidate anomaly region are calculated. Then, the distribution intervals of the neighborhood distribution features of the candidate anomaly region are reconstructed and mapped to the benchmark template. When the neighborhood distribution features of the candidate anomaly region are inconsistent with the benchmark template, the distribution intervals are made consistent through interval splitting, merging, or resampling. Finally, the neighborhood distribution features after interval consistency are normalized, and the normalized neighborhood distribution features are aligned with the order and dimension identifiers.

[0124] Furthermore, the neighborhood distribution difference measurement function is selected based on the type of neighborhood distribution feature components in the structural reference model and the actual needs of agricultural environmental monitoring. When the neighborhood distribution features are expressed in the form of a probability distribution, the relative entropy measurement function is selected; when the neighborhood distribution features are expressed in the form of a symmetrical distribution, the Jensen-Shannon divergence is selected; when the neighborhood distribution features are strongly correlated with spatial location, the minimum workload distance function is selected; and when the neighborhood distribution features are expressed in the form of frequency or statistics, a measurement method based on statistical distance is selected.

[0125] S3.3.3: Based on the temporal variation characteristic components of the structural reference model, the temporal variation deviation of the candidate anomaly region is calculated using a dynamic time warping algorithm;

[0126] Specifically, firstly, temporal variation feature components are extracted from the structural reference model; secondly, the temporal variation features of candidate anomaly regions and the temporal variation feature components of the structural reference model are preprocessed, including time series length normalization, numerical scale standardization, and outlier pruning; then, a time alignment path function based on a dynamic time warping algorithm is constructed. Under the constraint of allowing nonlinear stretching and compression of the time axis, the temporal variation features of candidate anomaly regions are matched with the temporal variation feature components of the structural reference model, and the local distance between each temporal variation feature parameter is calculated to form a global warping distance matrix; finally, the global warping distance matrix is ​​normalized to obtain the temporal variation distance value of the candidate anomaly region relative to the structural reference model, and this distance value is used as the temporal variation deviation of the candidate anomaly region.

[0127] S3.3.4: Normalize the deviation of the multidimensional structural features and map the deviation of the multidimensional structural features into a consistency sub-score;

[0128] In this step, the deviation of multidimensional structural features is mapped to obtain a consistency sub-score through linear normalization. The deviation of the multidimensional structural features includes spatial morphological deviation, neighborhood distribution deviation, and temporal variation deviation. The consistency sub-score includes the consistency sub-score corresponding to the spatial morphological features, the consistency sub-score corresponding to the neighborhood distribution features, and the consistency sub-score corresponding to the temporal variation features.

[0129] S3.3.5: Set corresponding weights based on the importance of multidimensional structural features in agricultural environment anomaly detection, and perform weighted calculation on the consistency sub-scores according to the weights to obtain the comprehensive structural consistency score.

[0130] First, multidimensional structural features are classified and labeled, assigning unique feature dimension identifiers to the consistency sub-scores corresponding to spatial morphology features, neighborhood distribution features, and temporal change features. Second, based on agricultural environmental anomaly discrimination rules and structural reference model configuration parameters, weights are set for the consistency sub-scores corresponding to different structural features. These weight coefficients can be set according to the discrimination priority of feature dimensions, historical sample statistical results, or expert configuration rules. For example, in a large-scale field agricultural monitoring scenario, the agricultural environment often exhibits regional sprawl characteristics, so the discrimination priority of neighborhood distribution features is the highest, while spatial morphology features and temporal change features have the same discrimination priority. Therefore, the weight of the consistency sub-score corresponding to neighborhood distribution features is set to 0.4, and the weight of the consistency sub-scores corresponding to spatial morphology features and temporal change features is set to 0.3. In an agricultural disaster early warning monitoring scenario, disaster-related anomalies such as drought, flood, and pests have strong temporal abrupt change characteristics, so the discrimination priority of temporal change features is the highest. Therefore, the weight of the consistency sub-score corresponding to temporal change features is set to 0.6, and the weight of the consistency sub-scores corresponding to neighborhood distribution features and spatial morphology features is set to 0.2. Then, within the same candidate anomaly region, a weighted sub-score set is generated based on the weights of the consistency sub-scores corresponding to different structural features. Finally, the weighted sub-score sets are integrated, and the weighted sub-scores of different structural features are accumulated or combined according to a unified calculation order to form a comprehensive structural consistency score.

[0131] S3.4: Compare the overall structural consistency score with the preset consistency threshold, mark the candidate abnormal regions with an overall structural consistency score greater than or equal to the preset consistency threshold as uncorrectable abnormalities, and mark the candidate abnormal regions with an overall structural consistency score less than the preset consistency threshold as correctable abnormalities.

[0132] In this step, the overall structural consistency score is compared with a preset consistency threshold. When the overall structural consistency score is greater than or equal to the preset consistency threshold, an uncorrectable anomaly identifier is written into the corresponding candidate anomaly region record. When the overall structural consistency score is less than the preset consistency threshold, an correctable anomaly identifier is written into the corresponding candidate anomaly region record. The candidate anomaly regions that have completed classification and marking are then updated.

[0133] It should be noted that the consistency threshold is the boundary between correctable and uncorrectable anomalies, and the consistency threshold can be set according to different monitoring scenarios such as field planting, facility agriculture, and disaster early warning. For example, in a large-scale field agriculture monitoring scenario, the preset consistency threshold for the comprehensive structural consistency score is set to 0.6. If the comprehensive structural consistency score of a candidate anomaly area is 0.7, then the candidate anomaly area is marked as an uncorrectable anomaly; if the comprehensive structural consistency score is 0.4, then the candidate anomaly area is marked as a correctable anomaly. In an agricultural disaster early warning monitoring scenario, in order to more accurately identify uncorrectable anomalies caused by real disasters such as drought, floods, and pests and diseases, and reduce the probability of missed detection, the preset consistency threshold for the comprehensive structural consistency score is set to 0.5. If the comprehensive structural consistency score of a candidate anomaly area is 0.55, then the candidate anomaly area is marked as an uncorrectable anomaly; if the comprehensive structural consistency score is 0.48, then the candidate anomaly area is marked as a correctable anomaly.

[0134] S3.5: Construct an anomaly cross-validation data unit based on the spatiotemporal location and multi-source data corresponding to the correctable anomaly;

[0135] Specifically, firstly, regions marked as correctable anomalies are selected from the candidate anomaly region set, and their corresponding spatial location identifiers and time window identifiers are extracted. Secondly, based on the spatial location identifiers and time window identifiers, sensor monitoring data, UAV monitoring data, and satellite monitoring data corresponding to the correctable anomaly region are extracted from the spatiotemporal alignment dataset, and integrity verification and field consistency processing are performed. Then, the multi-source monitoring data is structurally reorganized and uniformly encapsulated to form a data organization structure centered on spatiotemporal location. Next, data identification information is set for the anomaly cross-validation data unit, including anomaly region number, spatiotemporal location index, data source identifier, and parameter dimension identifier. The above data identification information is associated with the corresponding multi-source data content to obtain the anomaly cross-validation data unit.

[0136] S3.6: Based on the data source type and historical monitoring accuracy, set the prior confidence level for the abnormal cross-validation data unit and construct a Bayesian prior probability model;

[0137] Specifically, historical monitoring accuracy parameters corresponding to each data source type are extracted. These historical monitoring accuracy parameters may include historical error statistics, stability indicators, data integrity rates, or calibration consistency parameters. The historical monitoring accuracy parameters are quantified and organized according to preset accuracy evaluation rules to form an accuracy parameter set. Based on the data source type and the corresponding historical monitoring accuracy parameter set, prior confidence parameters are set for each data source component in the anomaly cross-validation data unit. These prior confidence parameters are expressed in the form of probability values ​​or probability distribution parameters. Based on the prior confidence parameters, a Bayesian prior probability model for the anomaly cross-validation data unit is constructed. The Bayesian prior probability model is then bound and stored with the corresponding anomaly cross-validation data unit.

[0138] S3.7: Based on the aforementioned cross-validation data unit and the Bayesian prior probability model, calculate the posterior estimation results of multi-source data at the same spatiotemporal location through Bayesian inference, and calibrate and correct the correctable anomaly based on the posterior estimation results;

[0139] In this step, the spatial location identifier and time window identifier associated with the anomaly cross-validation data unit are read, and multi-source monitoring data from sensors, UAVs, and satellites at that spatiotemporal location are extracted. The Bayesian prior probability model corresponding to the anomaly cross-validation data unit is then extracted. Next, based on the multi-source monitoring data, an observation likelihood function matching the Bayesian prior probability model is constructed. This observation likelihood function describes the probability relationship of observations from each data source under a given true-state assumption, and the observation data is modeled according to data source type and parameter dimension. Then, based on the Bayesian prior probability model and the observation likelihood function, joint probability calculation is performed on the multi-source data at the same spatiotemporal location to obtain the corresponding posterior probability distribution. Next, posterior estimation results are extracted from the posterior probability distribution. These posterior estimation results may include the posterior expected value, the maximum posterior probability value, or the posterior distribution parameter. The posterior estimation results are then associated with the corresponding spatial location identifier, time window identifier, and parameter dimension identifier. Finally, based on the posterior estimation results, the multi-source monitoring data marked as correctable anomalies are calibrated and corrected. Specifically, the original observation values ​​are replaced with the data values ​​corresponding to the posterior estimation results to complete the correction operation of the correctable anomalies at the corresponding spatiotemporal locations, and the corrected data is updated to the spatiotemporally aligned dataset.

[0140] S3.8: Integrate the calibrated data and data from non-abnormal regions to form a highly reliable dataset.

[0141] Specifically, the process involves first filtering correctable anomalous data records from the spatiotemporally aligned dataset after calibration correction, and then attaching calibration status identifiers, corresponding spatial location identifiers, and time window identifiers to these records. Next, non-anomalous region data not marked as anomalous areas is extracted from the spatiotemporally aligned dataset, and this data undergoes field integrity checks and format consistency processing. Then, following a unified spatiotemporal indexing rule, the correctable anomalous data and non-anomalous region data are integrated, and conflict resolution is performed on multiple data records potentially existing at the same spatiotemporal location. Duplicate or overlapping data are filtered, replaced, or rearranged based on data source identifiers, calibration status identifiers, or time sequence identifiers. Finally, the integrated data is encapsulated to generate a highly reliable dataset containing calibrated data and non-anomalous region data.

[0142] Next, we will further elaborate on the technical content of the method in this application regarding the generation of agricultural environmental data for the entire region.

[0143] S4.1: Using the high-confidence dataset as input, a nonlinear mapping relationship between agricultural environmental parameters and spatial distribution characteristics is established through a machine learning model to generate prediction results for environmental parameters in unmonitored areas;

[0144] Specifically, the high-confidence dataset undergoes feature parsing and sample construction. Data is organized based on spatial location and time window identifiers, and environmental parameters are correlated with corresponding spatial distribution features. The environmental parameters serve as the target variables for the machine learning model, while the spatial distribution features serve as the input features. These spatial distribution features include at least spatial coordinate features, neighborhood statistical features, and spatial structure description features. Next, the constructed sample data undergoes feature preprocessing operations, including numerical normalization, feature scaling, outlier pruning, and missing feature imputation. Samples are filtered or weighted according to spatial continuity or temporal consistency rules to form a training sample set. Then, based on the correlation between agricultural environmental parameters and spatial distribution features, the input, hidden, and output layers of the machine learning model are configured. The machine learning model can employ support vector regression, random forest regression, gradient boosting, or neural network models, using spatial distribution feature vectors as model input. Next, based on the training sample set, the machine learning model is trained. An iterative optimization algorithm is used to adjust the internal parameters of the machine learning model, enabling it to gradually fit the nonlinear mapping relationship between spatial distribution features and environmental parameters during training, thus completing the machine learning model training. Finally, the trained machine learning model is applied to the spatial distribution features corresponding to unmonitored areas, and model inference calculations are performed on the spatial distribution features of these unmonitored areas to output the corresponding predicted values ​​of environmental parameters.

[0145] S4.2: Based on the spatial distribution characteristics of the high-confidence dataset, the environmental parameters of the unmonitored area are estimated by spatial interpolation method to obtain the spatial interpolation result;

[0146] Specifically, environmental parameter data and spatial coordinate information corresponding to the monitored spatial locations are read from the high-confidence dataset, and the spatial coordinate information is subjected to unified coordinate system verification and scale standardization to construct a set of spatial sample points containing spatial locations and environmental parameter values.

[0147] Furthermore, based on the set of spatial sample points, spatial distribution features are extracted, including spatial distance relationships between sample points, spatial density distribution, and spatial adjacency structure, and the spatial neighborhood range of the location to be interpolated in the unmonitored area is determined.

[0148] Then, according to the preset spatial interpolation method, valid sample points for interpolation calculation are selected within the spatial neighborhood to form an interpolation input data structure. The spatial interpolation method may include inverse distance weighted interpolation, kriging interpolation, or spline interpolation.

[0149] Next, using the calculation rules corresponding to the selected spatial interpolation method, interpolation is performed on the target spatial location in the unmonitored area. By weighted combination of parameter values ​​of effective sample points within the spatial neighborhood or calculation using spatial correlation functions, the environmental parameter value corresponding to the target spatial location is obtained.

[0150] Finally, the spatial location of the unmonitored area is correlated with the corresponding environmental parameter values ​​to obtain the spatial interpolation results.

[0151] S4.3: The prediction results of the machine learning model and the spatial interpolation results are weighted and fused to generate agricultural environmental data for the entire region.

[0152] Specifically, the machine learning model predictions and spatial interpolation results are aligned in terms of spatial location, time window, parameter dimensions, and numerical scale to construct fusionable data pairs. Then, based on preset fusion weight configuration rules, weight coefficients are assigned to the machine learning model predictions and spatial interpolation results, and weighted fusion calculations are performed at spatial locations in unmonitored areas. Finally, the fused environmental parameter values ​​are correlated with their corresponding spatial locations and time windows, and integrated with data from monitored areas to generate comprehensive agricultural environmental data for the entire region.

[0153] It should be noted that when predicting environmental parameters using machine learning models, the range of prediction results for environmental parameters in uncorrectable anomalous regions is limited by introducing the spatiotemporal location and multi-source data corresponding to the uncorrectable anomalous regions; when estimating environmental parameters using spatial interpolation methods, the spatiotemporal location and multi-source data corresponding to the uncorrectable anomalous regions are used as interpolation boundaries or interpolation weights.

[0154] Please see Figure 5 , Figure 5 This is a block diagram of an agricultural environment collaborative sensing system based on multi-source data fusion according to an embodiment of the present invention. The system includes a spatiotemporal registration module, a candidate anomaly identification module, a cross-validation and calibration module, and a data prediction and completion module.

[0155] The spatiotemporal registration module is used to map multi-source monitoring data to the same spatiotemporal framework and generate a spatiotemporal location aligned dataset.

[0156] The candidate anomaly identification module is used to calculate the difference value of multi-source data at the same spatiotemporal location, and identify candidate anomaly regions by comparing the difference value with a preset threshold.

[0157] The cross-validation and calibration module is used to perform cross-validation and calibration on the spatiotemporally aligned dataset to generate a highly reliable dataset.

[0158] The data prediction and completion module is used to generate agricultural environmental data for the entire region based on a highly reliable dataset.

[0159] The cross-validation and calibration module includes a structural consistency discrimination unit, an anomaly calibration unit, and a high-confidence dataset generation unit.

[0160] The structural consistency discrimination unit is used to determine whether a candidate abnormal region is a correctable abnormality;

[0161] Anomaly calibration unit is used to perform cross-validation and probabilistic calibration on correctable anomalies;

[0162] The high-confidence dataset generation unit is used to integrate calibrated data and non-abnormal region data to generate a high-confidence dataset.

[0163] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for collaborative perception of agricultural environment based on multi-source data fusion, characterized in that, The method includes: Collect sensor, drone and satellite monitoring data, and perform spatiotemporal registration on the sensor, drone and satellite monitoring data to obtain a spatiotemporal location aligned dataset under the same spatiotemporal framework; Based on the spatiotemporal location aligned dataset, the difference value is calculated for multi-source data at the same spatiotemporal location. When the difference value exceeds a preset threshold, the spatiotemporal location is marked as a candidate abnormal region. The candidate abnormal regions are subjected to structural consistency discrimination. Abnormal regions with structural consistency are marked as uncorrectable abnormalities, and abnormal regions without structural consistency are marked as correctable abnormalities. For the correctable abnormalities, the spatiotemporal location alignment dataset is cross-validated and calibrated using a Bayesian inference algorithm to obtain a high-confidence dataset. Using the high-confidence dataset as input and the uncorrectable anomaly as a condition constraint, the data of the unmonitored area is predicted and completed using machine learning models and spatial interpolation methods to generate agricultural environmental data for the entire region.

2. The method of claim 1, wherein, Spatiotemporal registration is performed on the sensor, UAV, and satellite monitoring data to obtain a spatiotemporally aligned dataset within the same spatiotemporal framework, including: Based on the needs of agricultural environmental monitoring, a geographic coordinate reference system and time recording standard are set up, and a multi-source data spatial benchmark framework and time benchmark axis are constructed. By using grid mapping and spatial resampling techniques, the sensor, UAV, and satellite monitoring data are mapped to a spatial reference frame to obtain spatially consistent multi-source data; By using time window mapping and time interpolation algorithms, the spatially consistent multi-source data is mapped to a time reference axis to obtain spatiotemporally consistent multi-source data; Data loss repair and spatiotemporal consistency verification are performed on the spatiotemporally consistent multi-source data to generate a spatiotemporally aligned dataset under the same spatiotemporal framework.

3. The method of claim 1, wherein, Based on the spatiotemporal location aligned dataset, the difference values ​​of multi-source data at the same spatiotemporal location are calculated, including: Based on the spatiotemporal location alignment dataset, obtain sensor, drone, and satellite monitoring data under the same spatial location and the same time window; Select sensor monitoring data or historical statistical benchmark data as reference data; Based on the reference data, the parameter values ​​of the UAV monitoring data and satellite monitoring data are compared with the reference data to obtain the UAV monitoring difference value and the satellite monitoring difference value; The difference values ​​are associated with their corresponding spatial location identifiers, time window identifiers, and data source identifiers to form a difference value data set.

4. The method of claim 1, wherein, Structural consistency determination is performed on the candidate abnormal regions, including: Extract the multidimensional structural features of the candidate anomaly region. The multidimensional structural features include spatial morphological features, neighborhood distribution features, and temporal variation features. Based on historical high-reliability area samples, a structural reference model for monitoring indicators is constructed as a benchmark for judging the structure of candidate abnormal areas. The multidimensional structural features of the candidate anomaly region are matched with the structural reference model of the monitoring indicators to calculate the comprehensive structural consistency score. The overall structural consistency score is compared with the preset consistency threshold. Candidate abnormal regions with an overall structural consistency score greater than or equal to the preset consistency threshold are marked as uncorrectable abnormalities, while candidate abnormal regions with an overall structural consistency score less than the preset consistency threshold are marked as correctable abnormalities.

5. The method of claim 4, wherein, The multidimensional structural features of the candidate anomaly regions are matched with the structural reference models of the monitoring indicators to calculate a comprehensive structural consistency score, including: Based on the spatial morphological feature components of the structural reference model, the spatial morphological deviation of the candidate anomaly region is calculated using a morphological distance metric method. Based on the neighborhood distribution feature components of the structural reference model, the neighborhood distribution deviation of the candidate anomaly region is calculated using a distribution difference measurement method. Based on the temporal variation characteristic components of the structural reference model, the temporal variation deviation of the candidate anomaly region is calculated using a dynamic time warping algorithm; The deviation of the multidimensional structural features is normalized and mapped to a consistency sub-score. Based on the importance of multidimensional structural features in agricultural environment anomaly detection, corresponding weights are set, and the consistency sub-scores are weighted according to the weights to obtain the comprehensive structural consistency score.

6. The agricultural environment collaborative sensing method based on multi-source data fusion according to claim 1, characterized in that, For the correctable anomalies, the spatiotemporal alignment dataset is cross-validated and calibrated using a Bayesian inference algorithm to obtain a highly reliable dataset, including: Based on the spatiotemporal location and multi-source data corresponding to the correctable anomaly, an anomaly cross-validation data unit is constructed. Based on the data source type and historical monitoring accuracy, a prior confidence level is set for the abnormal cross-validation data unit, and a Bayesian prior probability model is constructed. Based on the aforementioned cross-validation data unit and the Bayesian prior probability model, the posterior estimation results of multi-source data at the same spatiotemporal location are calculated through Bayesian inference, and the correctable anomalies are calibrated and corrected based on the posterior estimation results. The calibrated data and data from non-abnormal areas are integrated to form a highly reliable dataset.

7. The agricultural environment collaborative sensing method based on multi-source data fusion according to claim 1, characterized in that, By using machine learning models and spatial interpolation methods to predict and complete data from unmonitored areas, agricultural environmental data for the entire region is generated, including: Using the high-confidence dataset as input, a nonlinear mapping relationship between agricultural environmental parameters and spatial distribution characteristics is established through a machine learning model, generating prediction results for environmental parameters in unmonitored areas; Based on the spatial distribution characteristics of the high-confidence dataset, the environmental parameters of the unmonitored area are estimated by spatial interpolation method to obtain spatial interpolation results; The prediction results of the machine learning model and the spatial interpolation results are weighted and fused to generate agricultural environmental data for the entire region.

8. The agricultural environment collaborative sensing method based on multi-source data fusion according to claim 7, characterized in that, Using the uncorrectable anomaly as a conditional constraint includes: When predicting environmental parameters using machine learning models, the range of prediction results for environmental parameters in uncorrectable anomaly regions is limited by introducing the spatiotemporal location of the uncorrectable anomaly region and multi-source data. When estimating environmental parameters using spatial interpolation methods, the spatiotemporal location and multi-source data corresponding to the uncorrectable anomaly region are used as interpolation boundaries or interpolation weights.

9. An agricultural environment collaborative sensing system based on multi-source data fusion, characterized in that, The system includes a spatiotemporal registration module, a candidate anomaly identification module, a cross-validation and calibration module, and a data prediction and completion module; The spatiotemporal registration module is used to map multi-source monitoring data to the same spatiotemporal framework and generate a spatiotemporal location aligned dataset. The candidate anomaly identification module is used to calculate the difference value of multi-source data at the same spatiotemporal location, and identify candidate anomaly regions by comparing the difference value with a preset threshold. The cross-validation and calibration module is used to perform cross-validation and calibration on the spatiotemporally aligned dataset to generate a highly reliable dataset. The data prediction and completion module is used to generate agricultural environmental data for the entire region based on a highly reliable dataset.

10. The agricultural environment collaborative sensing system based on multi-source data fusion according to claim 9, characterized in that, The cross-validation and calibration module includes a structural consistency discrimination unit, an anomaly calibration unit, and a high-confidence dataset generation unit. The structural consistency discrimination unit is used to determine whether a candidate abnormal region is a correctable abnormality; Anomaly calibration unit is used to perform cross-validation and probabilistic calibration on correctable anomalies; The high-confidence dataset generation unit is used to integrate calibrated data and non-abnormal region data to generate a high-confidence dataset.

Images