A multi-source perception fusion agricultural monitoring method and system

Abstract

The present application relates to the technical field of agricultural information perception and decision-making, in particular to a multi-source perception fusion agricultural monitoring method and system; the method comprises the following steps: transforming multi-source heterogeneous agricultural perception signals into a space-time tensor and constructing a semantic resonance field to simulate inter-modal nonlinear coupling; using a gravitational evolution mechanism to drive multi-modal data adaptive aggregation to form a fusion semantic field; calculating a nonlinear response to generate an agricultural state emergence index, identifying potential risk areas and constructing a binary risk map accordingly; for the risk area, mapping the semantic disturbance into an agricultural variable disturbance vector through a modal decoupling matrix, combining a sparse optimization of an operation response matrix to generate a minimum intervention strategy, and outputting a feasible operation suggestion after verification by an agricultural knowledge graph; constructing a drift potential function based on strategy execution feedback, dynamically updating strategy parameters through gradient descent, and realizing closed-loop self-evolution optimization combined with trend prediction. The present application realizes full-link adaptive optimization from multi-source perception to regulation and decision-making.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural information perception and decision-making, and particularly relates to a multi-source perception fusion agricultural monitoring method and system. BACKGROUND

[0002] The current agricultural monitoring field is facing the technical evolution demand of continuous expansion of multi-source heterogeneous perception data scale and deepening of complex interaction characteristics. The dynamic coupling mechanism of multi-modal data such as meteorology, soil moisture, and crop physiology in the time and space dimensions needs to be quantitatively modeled to support accurate decision-making.

[0003] A crop growth prediction method and system based on artificial intelligence and crop growth model are disclosed in Chinese patent CN118709789B, which includes: constructing a three-dimensional perception network of farmland and collecting multi-source heterogeneous data, preprocessing, semantic alignment, fusion, generating initial feature representation, dimension reduction embedding, and obtaining comprehensive feature representation; constructing key concepts, relationships and attributes of the crop growth domain ontology, obtaining crop growth process knowledge and constructing a crop growth knowledge graph, embedding the crop growth process knowledge into the crop growth knowledge graph, and constructing a crop growth simulator. For each crop, identify the key environmental factors and update the crop growth knowledge graph; construct a crop growth prediction model, extract the hidden variable representation, perform structure optimization, add the comprehensive feature representation to the crop growth prediction model, obtain the yield prediction probability distribution, and generate a causal exploration path diagram.

[0004] With the development of smart agriculture towards real-time and self-adaptation, the technical requirements of agricultural production systems for early identification of implicit risks, dynamic regulation of resources, and optimization of strategy closed loop have significantly improved. It is urgent to build a full-link fusion framework from multi-source perception to intelligent decision-making through innovative methods, so as to ensure system stability while achieving a breakthrough in resource utilization efficiency. SUMMARY

[0005] The present application aims to solve the problems in the background art and proposes a multi-source perception fusion agricultural monitoring method and system.

[0006] The technical solution of the present application is a multi-source perception fusion agricultural monitoring method, which includes the following specific implementation steps:

[0007] S1, by mapping multi-source heterogeneous agricultural perception signals into a spatio-temporal semantic four-order tensor, an agricultural semantic embedding field is constructed to identify potential event trigger points, and a semantic resonance field is defined to simulate the nonlinear interaction between modalities, and a coupling strength tensor is generated;

[0008] S2, construct a dynamic gravitational tensor field by modal semantic difference and resonance intensity, calculate the resultant force on each mode to drive the semantic particles to iteratively update the tensor state in the direction of multimodal consensus, define the convergence criterion index, and output the fusion result when the norm average change is less than the convergence threshold;

[0009] S3, simulate external stimulation by introducing structured disturbance in the semantic tensor field, calculate the normalized response intensity to measure the change after disturbance, convert it into an agricultural state emergence index with enhanced spatial consistency, and construct a risk binary map to identify potential mutation areas based on the threshold value;

[0010] S4, after detecting the risk area, construct the target steady-state semantic tensor, calculate the difference to form the semantic disturbance tensor, map it to the variable disturbance vector through the modal decoupling matrix, generate the minimum intervention operation vector through sparse optimization, and verify the consistency through the agricultural knowledge graph to ensure feasibility and agronomic consistency;

[0011] S5, after executing the control suggestion, collect the feedback difference tensor, build a drift potential function to drive the gradient descent update strategy parameter tensor, dynamically correct the target state by combining the memory average mechanism and trend prediction, and encapsulate the self-evolution controller to realize closed-loop iterative optimization.

[0012] Preferably, the coupling strength tensor construction process is:

[0013] Standardize and map multi-source heterogeneous agricultural perception data to generate a four-order modal tensor containing channel number, time length, spatial height and width dimensions;

[0014] Construct a three-dimensional agricultural semantic vector field by weightedly fusing the semantic feature vectors of each modal tensor slice, calculate the L2 gradient norm of the semantic state vector and find its extreme point to locate the agricultural event excitation source;

[0015] Define the full-field semantic resonance intensity field by weighting the Frobenius norm deviation of each modal tensor from its mean value through the modal response coefficient, and measure the modal semantic resonance coupling strength by the absolute value of the mixed partial derivative of the function with respect to the bimodal tensor, i.e. construct the coupling strength tensor.

[0016] Preferably, the iterative update process of the iterative update tensor state is:

[0017] Based on the modal semantic difference and resonance intensity, construct a dynamic gravitational field inversely proportional to the square of the modal tensor difference to drive the energy flow between modes for adaptive aggregation;

[0018] By calculating the resultant force tensor of any mode under the influence of other modes, drive its semantic tensor to iteratively update in the direction of multimodal consensus with an update amplitude, and perform cross-modal collaborative fusion and semantic consistency convergence:

[0019] ;

[0020] ;

[0021] where, represents the resultant force of all other modal gravitational forces that modal j is subjected to at (x, y, t), i.e., the sum of all semantic guiding effects from other modalities; represents the semantic tensor of modal j at this spatiotemporal position; represents the magnitude of the first semantic tensor update; represents the semantic gravitational force vector exerted by the i-th modality on the j-th modality at the three-dimensional spatiotemporal coordinate (x, y, t); (x, y) is the spatial position; t is the time stamp.

[0022] Preferably, the convergence determination index is:

[0023] ;

[0024] If the following condition is met: it is considered that the modal aggregation has reached a stable state, and the final fused tensor is output;

[0025] where, represents the average change magnitude of modal j in the k-th evolution process; represents the tensor state of modal j after the k-th iteration; represents the tensor state at the previous iteration; N is the total number of all (x, y, t) spatiotemporal positions participating in fusion; represents the set convergence threshold.

[0026] Preferably, the risk binary map construction process is:

[0027] By superimposing a structured disturbance term simulating external stimulation on the fused semantic tensor field, a disturbed modal semantic tensor is constructed: ;

[0028] where, represents the semantic tensor of modal j after adding the disturbance; represents the original fused semantic tensor of modal j at position (x, y) and time t; represents the disturbance term;

[0029] By calculating the Frobenius norm difference of the semantic tensor before and after disturbance and performing unit disturbance normalization, the structural response strength is quantified: ;

[0030] where, represents the response amplitude tensor, i.e., the structural response strength of modal j at point (x, y, t) to the disturbance; represents the strength of the disturbance term itself; denotes a stability factor;

[0031] Each response amplitude component is weighted by a modal weight, and the spatial Laplacian operator is combined to enhance the aggregation feature, to generate an agricultural state emergence index:

[0032] wherein, denotes an agricultural state emergence index; denotes a modal weight; denotes a regulation coefficient of the Laplacian term; denotes a spatial Laplacian operator;

[0033] According to the emergence index map, a risk binary map is constructed

[0034]

[0035] wherein, denotes an agricultural state risk marker map, indicating whether a point (x, y, t) is a potential key emergence area, 1 indicating detection of an anomaly, and 0 indicating normality; denotes an emergence index threshold value.

[0036] Preferably, the variable perturbation vector generation process is:

[0037] When it is detected that a state emergence exists in a certain area, the expected values of the semantic tensors of each mode in the normal area with a risk marker of zero are calculated by a conditional expectation operator, to construct a target steady-state semantic tensor to define an ideal agricultural state:

[0038] wherein, denotes an expected agricultural system semantic tensor, i.e., at a spatial position (x, y) and a time point t; denotes a conditional expectation operator; denotes a semantic tensor of the jth mode at a spatial position and a time point t; denotes a binary value of the agricultural state risk marker map;

[0039] A modal semantic perturbation tensor is constructed by calculating the difference between the target state tensor and the current actual state tensor, and is mapped to a variable perturbation vector by a modal decoupling matrix trained based on agricultural domain knowledge and historical data:

[0040]

[0041]

[0042] wherein, denotes a semantic perturbation tensor under a mode j;​​​​​​ denotes the modal semantic tensor, which represents the semantic information of the agricultural field, and is trained based on the agricultural field knowledge and historical data, and represents the corresponding relationship between each dimension of the modal semantic tensor and the agricultural variable; denotes the agricultural variable disturbance vector, i.e., the agricultural variable value that the modal j should adjust at the point (x, y, t), i.e., the disturbance response value of the p agricultural variables associated with the modal j.

[0043] Preferably, the verification process of consistency verification through the agricultural knowledge graph is as follows:

[0044] The operation variable vector is mapped through the agricultural operation response matrix, and an optimization problem of minimizing the square sum of response error and regularization of the operation vector L1 norm is solved to generate a sparse constraint minimum intervention strategy operation vector:

[0045] , wherein, denotes the agricultural operation vector; denotes the agricultural operation response matrix, which describes the influence strength of each operation on the agricultural variable; denotes the operation variable vector; denotes the regularization parameter, which controls the weight of the sparse constraint term; denotes the 1-norm of the operation vector;

[0046] Consistency verification is performed through the agricultural knowledge graph, and a judgment function is used to screen whether the operation scheme conforms to the agricultural specification:

[0047] , wherein, denotes the agricultural knowledge graph; denotes the final agricultural operation suggestion after consistency verification and correction; is used to judge whether the operation scheme conforms to the agricultural knowledge and specification; denotes adjustment of the unreasonable strategy.

[0048] Preferably, the updating process of the strategy parameter tensor is as follows:

[0049] According to the regulation and control suggestion , after a delay time , a new perception tensor is collected Comparison, construct a feedback difference tensor:

[0050] , wherein, denotes the multi-source perception tensor re-collected after the strategy execution ; denotes the feedback difference tensor;

[0051] Construct a state drift potential function:​​​​

[0052] ;

[0053] ;

[0054] wherein, represents the stability score of the agricultural system at spatial location x, y, time t; represents the tensor fluctuation amplitude of the specified region in the past K2 time units; K2 represents the length of the time window for calculating the historical fluctuation; represents the multi-source perception semantic tensor at spatial location (x, y) at time point t-k; represents the average tensor state within the time window [t-K2, t-1]; represents the state drift potential energy function value; represents the stability adjustment coefficient;

[0055] The potential energy calculated in the previous step is used to adjust the output tensor of the next round of strategy, and the strategy parameter is represented as tensor , then the following response gradient descent is adopted:

[0056] ;

[0057] wherein, represents the strategy parameter tensor, i.e. the control model weight or feature mapping parameter used to generate the suggestion at the tth round of decision-making; represents the control model weight used to generate the suggestion at the t+1th round of decision-making; represents the learning rate; represents the gradient of the drift potential energy with respect to the strategy parameter.

[0058] Preferably, the correction process for dynamically correcting the target state is:

[0059] By a memory average mechanism, the new strategy weight and the historical weight are fused with a decay factor to avoid oscillation, and a trend prediction module is combined to model the multi-channel coupling characteristics based on the historical observation tensor sequence, and the future target state is dynamically decoded and predicted to realize dynamic correction:

[0060] ;

[0061] ;

[0062] wherein, represents the strategy memory decay factor; represents the time trend prediction function; represents the observation sequence within the time window; represents the target agricultural state of the next round.

[0063] The technical scheme of the present application: a multi-source perception fusion agricultural monitoring system for executing the above-mentioned multi-source perception fusion agricultural monitoring method, comprising:

[0064] A multi-source perception semantic tensor construction module is responsible for receiving and fusing multi-source heterogeneous data, and constructing a high-dimensional semantic tensor through spatial and temporal fusion and resonance field initialization processing, to complete deep semantic expression of agricultural environment and crop growth information;

[0065] A multi-modal aggregation module is used for coupling analysis of multi-modal semantic tensors, and different modalities and time series data are organically aggregated by constructing a gravitational tensor model between devices;

[0066] An agricultural state emergence detection module is used for dynamic analysis of the aggregated agricultural state data, to identify potential abnormalities, emergence behaviors and nonlinear evolution characteristics, and to provide a scientific basis for risk early warning and state diagnosis;

[0067] An agricultural regulation suggestion generation module is used for automatically generating targeted agricultural regulation strategies and operation suggestions in combination with detected agricultural state changes and emergence characteristics;

[0068] An agricultural strategy dynamic feedback and self-evolution module is used for combining multi-source perception data after the implementation of regulation measures, constructing a feedback difference tensor and state drift potential, and dynamically adjusting and optimizing the agricultural strategy in combination with strategy gradient update and trend prediction.

[0069] Compared with the prior art, the above technical scheme of the present application has the following beneficial technical effects:

[0070] The present application designs a multi-source perception fusion agricultural monitoring method and system, which significantly improves the processing efficiency and accuracy of agricultural perception data by constructing a physical model driven multi-modal data fusion mechanism. Firstly, the nonlinear coupling strength between modalities is quantified by using spatiotemporal tensor modeling and semantic resonance field, and multi-source data is adaptively aggregated by combining gravitational evolution, to overcome the shortcomings of traditional methods in interactive modeling of heterogeneous data. Secondly, an agricultural state emergence index is generated through a structured disturbance response mechanism, to accurately locate potential risk areas such as disease outbreak or drought center, and to significantly improve the sensitivity and accuracy of mutation early warning. Thirdly, the semantic disturbance is mapped to a variable disturbance vector based on a modal decoupling matrix, and the minimum intervention strategy is generated through sparse optimization of an operation response matrix, and the agricultural feasibility and safety of the regulation suggestion are ensured by combining agricultural knowledge graph verification. Finally, a closed-loop self-evolution system is constructed, the drift potential is calculated based on the feedback difference tensor to drive the strategy parameter gradient update, the target state is dynamically adjusted by fusing historical memory and trend prediction, and the whole-link adaptive optimization from perception to decision is realized, so as to comprehensively improve the real-time performance, stability and resource regulation efficiency of the agricultural system. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 A method flow chart of a multi-source perception fusion agricultural monitoring method is proposed in the present application.

[0072] Figure 2 A system architecture diagram of a multi-source perception fusion agricultural monitoring system is proposed in the present application. DETAILED DESCRIPTION

[0073] Embodiment one, as shown in the figure, a multi-source perception fusion agricultural monitoring method is proposed in the present application, including the following specific implementation steps: Figure 1

[0074] S1, extract the tensor representation of the embeddable semantic structure from the heterogeneous agricultural perception signals from different sources, and construct a dynamic driving data fusion preparation state by introducing the physical abstraction method of semantic resonance field, to provide a unified tensor representation and semantic interaction field for the subsequent coupling perception mechanism, and the specific implementation process is:

[0075] S11, standardize the heterogeneous agricultural perception signals into a four-order tensor with space-time and semantic distribution attributes: let the original data source set be: , each data source is preliminarily nested and encoded by a modal perception device to generate a modal tensor :

[0076] ;

[0077] Among them, represents the high-order tensor representation of the mth modal perception data after embedding mapping; represents the semantic embedding function corresponding to the mth modal, which is used to convert the original data into a standard tensor form; represents the original mth agricultural perception data; M is the total number of modalities of agricultural perception data; C m represents the channel number of the mth modal tensor; T represents the length of the time dimension; H is the height dimension of the spatial dimension in the tensor, corresponding to the vertical direction division unit in the agricultural area grid map; W is the width dimension of the spatial dimension in the tensor, corresponding to the horizontal direction division unit in the grid map;

[0078] S12, based on the tensor mapping result, construct an agricultural semantic embedding field , which describes the semantic expression state at different times and geographical positions, and identifies the excitation source point of the potential important agricultural event through the gradient evolution of the field:

[0079] ;

[0080] ;​

[0081] wherein, represents represents the semantic state vector at coordinate point (x, y, t) in the three-dimensional semantic vector field, reflecting the agricultural meaning expression (such as "humidity is too high" "crop growth is good") of the point at the current space-time position; represents a function of mapping a slice of the m-th modal tensor (i.e. the perception data at point (x, y, t)) to a semantic feature vector; represents the tensor At the slice of coordinate point (x, y, t), that is, all channel data (dimension: C m ) of the point in this modal; represents the semantic fusion weight matrix of the m-th modal, which is adaptively estimated by mutual information in this embodiment; represents the spatial position and time point of the semantic excitation source point, that is, the point with the most drastic change in the semantic field in the agricultural area, which may represent a disease outbreak point, a drought development center point, etc. represents the gradient vector of the semantic vector field, that is, the change rate (i.e. semantic volatility) of the semantic in the local space and time; represents the vector norm, and L2 norm (Euclidean length) is adopted in this embodiment to measure the size of the semantic change rate; represents the value of the variable that maximizes the expression; (x, y) is the spatial position; t is the time stamp;

[0082] S13, define a semantic resonance field to simulate the nonlinear interaction behavior (similar to the phase resonance phenomenon in physics) between perception data due to semantic proximity, which provides a driving force for subsequent multi-modal tensor self-aggregation, specifically:

[0083] Define the resonance field: ;

[0084] To measure the semantic resonance coupling strength of modal i and j, a coupling strength tensor is constructed:

[0085] ;

[0086] wherein, represents the semantic resonance strength field value generated by all modal perception data at point (x, y, t), which is used to represent the excitation energy of the coupling resonance behavior between perceptions; represents the modal response coefficient, that is, the participation degree (response strength) of the m-th modal to the semantic resonance behavior, which is learned by the modal volatility and coupling empirical entropy in the historical data in this embodiment; represents the Frobenius norm, that is, the channel norm of the tensor at the slice position, which is used to measure the difference size from the mean; denotes the average tensor value of all modalities at this point, used as the resonance reference center; denotes the semantic resonance coupling strength between modalities i and j at point (x, y, t);

[0087] Accordingly, the inter-modal response relationship tensor network is constructed by tensor partial derivation, and the coupled physical modeling is introduced into the agricultural semantic perception system.

[0088] S2, after obtaining the multi-source perception semantic tensor and the semantic coupling strength therebetween, a set of dynamic evolution mechanisms is constructed, and the analogy model of physical gravitational evolution is used to drive the adaptive aggregation between modalities, so as to construct the coupled state semantic structure, so as to realize the efficient integrated modeling of agricultural multi-modal perception information, and the specific implementation process is as follows:

[0089] S21, a tensor gravitational action determined by the modal semantic difference and the resonance strength is constructed, and an evolvable coupling gravitational tensor field is constructed in the space-time-semantic domain. The gravitational field acts on the tensor space position and the internal semantic distribution structure, so as to form the dynamic energy flow between modalities, and the specific implementation process is as follows:

[0090] ;

[0091] wherein, denotes the semantic gravitational vector (tensor) exerted by the i-th modality on the j-th modality at the three-dimensional space-time coordinates (x, y, t), and whether the semantic tensor of the j-th modality should be attracted to the i-th modality, the greater the gravitational value, the stronger the fusion willingness; denotes the inter-modal semantic gravitational coefficient, which is a global adjustment factor controlling the size of the semantic gravitational force; denotes a stability factor for preventing numerical instability caused by 0 or excessively small values, which is set to 1e-6 in this embodiment;

[0092] S22, a tensor behavior system driven by gravitational force is constructed, each modality is modeled as a semantic particle, and the fusion behavior of the semantic particle is dynamically regulated by the global coupling tensor field, and an evolutionary cross-modal collaboration mechanism is provided:

[0093] The total gravitational field force tensor of the j-th modality is defined as: ;

[0094] The state update formula is defined as: ;

[0095] wherein, denotes the resultant force of all other modalities on the j-th modality at (x, y, t), that is, the sum of all semantic guiding actions from other modalities, which is used to guide the j-th modality to approach the multi-modal consensus; represents the semantic tensor of modality j at this spatiotemporal position, which gradually converges to the consensus direction of other modalities after each update, realizing the iterative convergence of semantic consistency. represents the amplitude of a semantic tensor update, controlling the influence of the gravitational tensor on the modal tensor;

[0096] S23, define the convergence criterion index: ;

[0097] If the following condition is met: , it is considered that the modal aggregation has reached a stable state, and the final fusion tensor is output.

[0098] wherein, represents the average change amplitude of modality j in the kth evolution process, which is a key measure of whether the fusion is stable. If the change is very small, it means that the modal state has basically stabilized, and the fusion is considered complete. represents the tensor state of modality j after the kth iteration; represents the tensor state of the previous iteration, and the difference between the two is used to determine whether there is enough change; N is the total number of all (x, y, t) spatiotemporal positions participating in the fusion; represents the set convergence threshold.

[0099] S3, by introducing a structured perturbation term in the semantic tensor field, calculate the nonlinear response of the tensor on the spatiotemporal structure before and after the perturbation, so as to identify the latent areas that are extremely sensitive to the perturbation. These areas are potential agricultural situation mutation points or critical evolution sources, and the specific implementation process is as follows:

[0100] S31, perturb the fusion semantic tensor field (modality j) output by step S2, simulate the weak external stimuli that the agricultural ecosystem may face, such as microclimate change, water fluctuation, and pathogen signal penetration:

[0101] Construct the perturbation tensor: ;

[0102] wherein, represents the semantic tensor of modality j after adding the perturbation, i.e. the structural state of the modality at the spatiotemporal point (x, y, t) after being disturbed; represents the original fusion semantic tensor of modality j (such as image, weather, soil, etc.) at position (x, y) and time t; represents the perturbation term, corresponding to the structural disturbance of modality j at this point, simulating small changes in climate, humidity, and pathogens;

[0103] S32, after introducing the perturbation, pay attention to the response intensity of the system state in the semantic space, which is specifically measured as the unit perturbation normalization result of the difference between the tensors before and after the perturbation:

[0104] ;

[0105] wherein, represents the response amplitude tensor, i.e. the response intensity of the modal j at the point (x, y, t) to the perturbation, i.e. the degree of change after the semantic structure is disturbed; represents the intensity of the perturbation item itself, as a normalization item, used to eliminate the influence of different perturbation amplitudes on the response value; represents a stabilization factor (such as 1e-6), used to prevent instability caused by a denominator of 0 or too small;

[0106] S33, introduce spatial consistency analysis, further transform high response points in the response amplitude field into emergent indexes with evolutionary mutation possibility, for spatial aggregation evaluation:

[0107] ;

[0108] wherein, represents the agricultural state emergent index, i.e. whether the point has mutation possibility in the multimodal perturbation response field, which is a measure of the potential critical state of the system; represents the modal weight, reflecting the relative importance of the modal j in the agricultural state judgment; represents the adjustment coefficient of the Laplace term, controlling the influence of spatial aggregation on the overall emergent index; represents the spatial Laplace operator;

[0109] S34, construct a risk binary map according to the emergent index map , for subsequent decision-making of the agricultural system:

[0110] ;

[0111] wherein, represents the agricultural state risk marking map, a binary output map, marking whether the point is a potential key emergent area in the agricultural system; represents the emergent index threshold, used to judge whether there is a key state mutation.

[0112] S4, construct an agricultural regulation suggestion generation mechanism based on state emergence driving, inverse through semantic perturbation guide variables, combine modal decoupling and sparse optimization, construct a reversible mapping mechanism from agricultural state emergence to operation strategy, and through knowledge graph, correct the feasibility and agronomic consistency of the regulation suggestion, realize dynamic and self-adaptive precision regulation strategy generation in intelligent agricultural environment, and the specific implementation process is:

[0113] S41, when the system detects that a region (x, y) has state emergence at time t (i.e. =1), construct the desired state tensor field to define the target steady state: ;

[0114] in, represents the desired semantic tensor of the agricultural system, i.e., the ideal steady-state modal semantic state of the agricultural system at the spatial location (x, y) and time point t; represents the conditional expectation operator; Indicates the spatial position of the jth mode , the semantic tensor at time point t; It represents the binary value of the agricultural status risk marker map, where 1 indicates that an abnormality is detected and 0 indicates that it is normal;

[0115] S42, according to the obtained target state tensor and the current actual state tensor , construct the semantic perturbation tensor of the modal layer: ;

[0116] This tensor represents the deviation of the system's current perceived state from the desired steady state in mode j;

[0117] in, represents the semantic perturbation tensor under modality j, that is, the difference between the current state and the target steady state;

[0118] For example, in image mode, it can reflect vegetation color changes, texture density fluctuations, etc.; in meteorological mode, it may indicate humidity drop or abnormal temperature difference.

[0119] Introducing the modal decoupling matrix , mapping the semantic perturbation tensor to the variable perturbation vector:

[0120] ;

[0121] in, Representing the modal decoupling matrix, mapping the modal semantic tensor space to the specific agricultural control variable space, based on agricultural domain knowledge and historical data training, characterizing the correspondence between each dimension of modal semantics and agricultural variables; represents the disturbance vector of agricultural variables, i.e., the agricultural variable value that should be adjusted for mode j at point (x, y, t), i.e., the disturbance response value of the p agricultural variables associated with mode j;

[0122] It should be noted that the modal decoupling matrix The learning is obtained by constructing a large-scale labeled sample set between multi-source perception data and corresponding agricultural variables: firstly, the semantic tensor output and the actual measured control variable value of each modality are collected in the experimental field or greenhouse at the same time, and the prior feature correlation information given by the agricultural experts is combined, and the mapping relationship between the perception features and the variable changes is fitted by using multivariate regression or tensor decomposition and other data-driven methods; on this basis, through cross-validation and calibration of the domain knowledge graph, the matrix parameters are continuously optimized, so that they can not only accurately reflect the contribution of each modality to different control variables, but also have good generalization and anti-interference ability, and finally form a modality decoupling matrix that conforms to the physical agronomic law and can be efficiently calculated in the actual system;

[0123] S43, introducing the agricultural operation response matrix , the minimum intervention strategy is calculated by optimization:

[0124] ;

[0125] Among them, represents the agricultural operation vector, representing the agricultural intervention operation to be executed (such as fertilizer amount, irrigation time, etc.); represents the agricultural operation response matrix, which describes the influence intensity of each operation on the agricultural variable, and is constructed according to historical agricultural experimental data, physical models or expert experience, and serves as a mapping bridge between the control variable and the operation; represents the operation variable vector, which contains the intensity or parameters of all possible agricultural intervention measures; represents the regularization parameter, which controls the weight of the sparse constraint term; represents the 1-norm of the operation vector, that is, the sum of the absolute values of the operation intensity;

[0126] S44, after preliminarily obtaining the operation vector , consistency verification needs to be carried out through the agricultural knowledge graph :

[0127] ;

[0128] Among them, represents the agricultural knowledge graph; represents the final agricultural operation suggestion after consistency verification and correction; is used to judge whether the operation scheme conforms to the agricultural knowledge and specifications, that is, to screen whether there is an illegal or conflicting strategy; represents adjusting the unreasonable strategy to make it meet the agricultural specifications within the minimum change range;

[0129] It should be noted that the agricultural knowledge graph The entity set is the core element in the agricultural knowledge graph, including but not limited to crop type entity, soil type entity, climate environment entity, agricultural disease entity, agricultural operation entity and agricultural variable entity; the relationship set describes the semantic relationship between entities, for example: <rice yellow leaf, need operation, follow-up nitrogen fertilizer>; the attribute set defines the numerical or enumeration attribute label for entities and relationships.

[0130] S5, a dynamic self-evolution method of agricultural strategy based on strategy feedback tensor construction, drift potential energy estimation and response gradient update is constructed, through constructing a closed-loop system of perception-regulation-response, the agricultural strategy can be self-adaptively adjusted and long-term evolved according to the actual perception deviation, and the intelligence and stability of the system are improved, and the specific implementation process is:

[0131] S51, according to the regulation suggestion After execution, a new perception tensor is collected after a delay time Compare and construct a feedback difference tensor: The difference reflects whether the system responds as expected after the strategy is executed;

[0132] Wherein, Indicates the multi-source perception tensor re-collected after the strategy is executed, that is, the change of the real agricultural environment; Indicates the feedback difference tensor, that is, the difference between the actual effect after the strategy is executed and the expected target;

[0133] S52, a state drift potential energy function is constructed:

[0134] ;

[0135] ;

[0136] Wherein, Indicates the stability score of the agricultural system at spatial position x, y and time t; Indicates the tensor fluctuation amplitude (standard deviation approximation) of the specified area in the past K2 time units; K2 represents the length of the time window for calculating the historical fluctuation; Indicates the multi-source perception semantic tensor at spatial position (x, y) at time point t-k, containing crop growth, physiology, weather, soil and other information characteristics; Indicates the average tensor state in the time window [t-K2, t-1]; Indicates the state drift potential energy function value, which measures the damage degree of the current feedback difference to the stability of the system; Indicates the stability adjustment coefficient, which is used to balance the weight of feedback energy and system stability; ​​

[0137] S53, use the potential energy calculated in the previous step to adjust the next round of policy output tensor, let the policy parameter be represented as a tensor , then use the following response gradient descent:

[0138] ;

[0139] wherein, represents the policy parameter tensor, that is, the control model weight or feature mapping parameter used to generate the suggestion at the tth round of decision-making; represents the control model weight or feature mapping parameter used to generate the suggestion at the t+1th round of decision-making; represents the learning rate, which is used to control the update step of the policy weight to avoid oscillation; represents the gradient of the drift potential with respect to the policy parameter, reflecting which direction can most reduce the deviation potential;

[0140] S54, introduce a first-order memory average mechanism to avoid oscillation caused by response update:

[0141] ;

[0142] At the same time, combined with the time window sliding trend prediction module , the long-term trend deviation is corrected and the system policy target point is adjusted: ;

[0143] wherein, represents the policy memory decay factor, which controls the fusion ratio of historical weight and new weight; represents the time trend prediction function, which is used to predict the reasonable interval of the future target state according to the historical observation; represents the observation sequence in the time window, which is used for state evolution speculation by the trend prediction module; represents the new round of target agricultural state, which is dynamically updated by trend prediction;

[0144] It should be noted that the time trend prediction function is a state evolution modeling method based on sliding time series. The core idea is to use a historical observation tensor sequence of a certain length to extract the potential time series evolution trend of the agricultural system through a nonlinear state regression or embedded tensor decoder model, and then predict the optimal expected state at the next moment or in the future. The module not only considers the smooth trend of numerical change, but also considers the interactive coupling characteristics of multi-dimensional channels (such as soil moisture, air temperature, pest signals, etc.). A tensor dynamic decoding mechanism is used to model the multi-modal evolution path, ensuring that the trend prediction result has robustness and actual adaptability, thereby providing a sustainable target reference direction for dynamic adjustment of the strategy, playing a dual role of system foresight and evolution harmonization.

[0145] S55, encapsulating the above feedback, energy, gradient update, trend prediction, etc. modules into a self-evolution controller unit , and adding a time iteration formula: ;

[0146] wherein, represents the agricultural regulation and control self-evolution controller state of the current space-time point (x, y); represents the iteration function of the controller from t to t+1, using the policy network evolution function in reinforcement learning; represents the gradient of the potential energy function to the policy parameter.

[0147] Embodiment two, as shown in Figure 2 , the present application proposes a multi-source perception fusion agricultural monitoring system, which is applied to the multi-source perception fusion agricultural monitoring method proposed in embodiment one, comprising: a multi-source perception semantic tensor construction module, a multi-modal aggregation module, an agricultural state emergence detection module, an agricultural regulation and control suggestion generation module, and an agricultural strategy dynamic feedback and self-evolution module.

[0148] The multi-source perception semantic tensor construction module is responsible for receiving and fusing multi-source heterogeneous data from satellite remote sensing, ground sensors, weather stations, etc., and processing these heterogeneous data through spatial and temporal fusion and resonance field initialization to construct a high-dimensional semantic tensor, complete deep semantic expression of agricultural environment and crop growth information, and ensure the richness and consistency of the perception information;

[0149] The multi-modal aggregation module is used for coupling analysis of multi-modal semantic tensors, and through the construction of a gravitational tensor model between devices, different modal and time series data are organically aggregated to reveal the mutual influence and dynamic coupling relationship between multiple factors in the agricultural system, and provide accurate comprehensive state evaluation;

[0150] The agricultural state emergence detection module is used for dynamic analysis of the aggregated agricultural state data, identifies potential anomalies, emergence behavior and nonlinear evolution characteristics, provides scientific basis for risk warning and state diagnosis, and improves the sensitivity and accuracy of monitoring;

[0151] The agricultural regulation and control suggestion generation module is used to automatically generate targeted agricultural regulation and control strategies and operation suggestions in combination with the detected agricultural state changes and emergence characteristics, covering irrigation, fertilization, pest control, etc., effectively guiding agricultural production practice, promoting healthy growth of crops and optimizing resource utilization;

[0152] The agricultural strategy dynamic feedback and self-evolution module is used for combining multi-source sensing data after the implementation of the regulation measures, constructing a feedback difference tensor and a state drift potential, and combining strategy gradient updating and trend prediction to dynamically adjust and optimize the agricultural strategy, realize the closed-loop self-adaptation and continuous evolution of the system, and enhance the intelligent level and long-term stability of the agricultural monitoring and regulation system.

[0153] The embodiments of the application are described in detail above with reference to the drawings, but the application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A multi-source perception fusion agricultural monitoring method, characterized in that: The specific implementation steps include the following: S1. By normalizing and mapping multi-source heterogeneous agricultural perception signals into a spatiotemporal semantic fourth-order tensor, an agricultural semantic embedding field is constructed to identify potential event excitation sources. A semantic resonance field is also defined to simulate nonlinear interactions between modalities and drive the generation of a coupling strength tensor. The coupling strength tensor construction process is: The multi-source heterogeneous agricultural sensing data is standardized and mapped to generate a fourth-order modal tensor containing the dimensions of channel number, time length, spatial height and width; By weighted fusion of semantic feature vectors of each modal tensor slice, a three-dimensional agricultural semantic vector field is constructed. The L2 gradient norm of the semantic state vector is calculated and its extreme value is found to locate the triggering source of agricultural events. The full-field semantic resonance intensity field is defined by weighting the Frobenius norm deviation of each modal tensor from its mean through the modal response coefficient. The absolute value of the mixed partial derivative of the function with respect to the bimodal tensor measures the semantic resonance coupling strength between the modalities, that is, constructing the coupling strength tensor. S2. Construct a dynamic gravitational tensor field based on modal semantic differences and resonance strength, calculate the net force acting on each modality, and iteratively update the tensor state of the semantic particles in the direction of multimodal consensus. Define a convergence judgment indicator, and output the fusion result when the average change in the norm is lower than the convergence threshold. S3. By introducing structured perturbations into the semantic tensor field to simulate external stimuli, the normalized response intensity is calculated to measure the degree of change after the perturbation, which is converted into an agricultural state emergence index with enhanced spatial consistency. Based on the threshold, a binary risk map is constructed to identify potential mutation areas. The process of constructing the risk binary map is as follows: By superimposing a structured perturbation term simulating external stimulation on the fused semantic tensor field, the perturbed modal semantic tensor is constructed: ； in, Represents the semantic tensor of modality j after adding perturbation; Represents the original fused semantic tensor of modality j at position (x, y) and time t; represents the disturbance term; The structural response strength is quantified by calculating the Frobenius norm difference of the semantic tensor before and after perturbation and performing unit perturbation normalization: ； in, represents the response amplitude tensor, that is, the structural response intensity of mode j to the disturbance at point (x, y, t); represents the intensity of the disturbance term itself; represents the stability factor; The modal weights are used to weight the amplitude components of each response, and the spatial Laplace operator is used to enhance the aggregation characteristics to generate the agricultural state emergence index: ； in, represents the agricultural status emergence index; w j represents the modal weight; represents the adjustment coefficient of the Laplace term; represents the spatial Laplacian operator; Construct a binary risk map based on the emergence index map : ； in, Represents the agricultural status risk marker map, marking whether the point (x, y, t) is a potential key emergence area, 1 means an abnormality is detected, and 0 means normal; represents the emergence index threshold; S4. After detecting the risk area, the target steady-state semantic tensor is constructed, the difference is calculated to form a semantic disturbance tensor, which is mapped into a variable disturbance vector through the modal decoupling matrix. Sparse optimization is used to generate the minimum intervention operation vector, and consistency verification is performed through the agricultural knowledge graph to ensure feasibility and agronomic consistency. S5. After executing the control suggestions, the feedback difference tensor is collected, and the drift potential function is constructed to drive the gradient descent to update the strategy parameter tensor. The target state is dynamically corrected by combining the memory averaging mechanism and trend prediction, and the self-evolution controller is encapsulated to realize closed-loop iterative optimization.

2. The agricultural monitoring method based on multi-source perception fusion according to claim 1 is characterized in that: The iterative update process of tensor state is: A dynamic gravitational field is constructed based on modal semantic differences and resonance strength, which is inversely proportional to the square of the modal tensor difference, driving the energy flow between modalities for adaptive aggregation; By calculating the net force tensor of the gravitational forces exerted on any modality by other modalities, its semantic tensor is driven to iteratively update in the direction of multimodal consensus with an update amplitude, thus achieving cross-modal collaborative fusion and semantic consistency convergence: ； ； in, represents the resultant force of all other modal attractions on modal j at (x, y, t), that is, the sum of all semantic guidance effects from other modalities; Represents the semantic tensor of modality j at this spatiotemporal position; Indicates the magnitude of a semantic tensor update; Represents the semantic gravitational vector exerted by the i-th modality on the j-th modality at the three-dimensional space-time coordinate (x, y, t); (x, y) is the spatial position; t is the timestamp.

3. The agricultural monitoring method based on multi-source perception fusion according to claim 2 is characterized in that: The convergence judgment index is: ； If satisfied: , it is considered that the modal aggregation has reached a stable state and the final fusion tensor is output; in, represents the average change amplitude of mode j during the k-th evolution process; Represents the tensor state of mode j after the kth iteration; Represents the tensor state at the previous iteration; N is the total number of all (x, y, t) spatiotemporal positions involved in the fusion; Indicates the set convergence threshold.

4. The agricultural monitoring method based on multi-source perception fusion according to claim 3 is characterized in that: The variable disturbance vector generation process is: When a state emergence is detected in a certain area, the expected value of each modal semantic tensor in the normal area with a risk mark of zero is calculated through the conditional expectation operator, and the target steady-state semantic tensor is constructed to define the ideal agricultural state: ； in, The semantic tensor representing the desired agricultural system, i.e., at the spatial location (x, y) and time point t; represents the conditional expectation operator; Indicates the spatial position of the jth mode , the semantic tensor at time point t; A binary map representing the agricultural status risk marker map; The difference between the target state tensor and the current actual state tensor is calculated to construct the modal semantic disturbance tensor, which is mapped into a variable disturbance vector through the modal decoupling matrix trained based on agricultural domain knowledge and historical data: ； ； in, represents the semantic perturbation tensor under modality j; Representing the modal decoupling matrix, mapping the modal semantic tensor space to the specific agricultural control variable space, based on agricultural domain knowledge and historical data training, characterizing the correspondence between each dimension of modal semantics and agricultural variables; represents the disturbance vector of agricultural variables, that is, the agricultural variable value that should be adjusted for mode j at point (x, y, t), that is, the disturbance response value of the p agricultural variables associated with mode j.

5. The agricultural monitoring method based on multi-source perception fusion according to claim 4 is characterized in that: The verification process for consistency verification through the agricultural knowledge graph is as follows: By mapping the operation variable vector through the agricultural operation response matrix, the optimization problem of minimizing the sum of squared response errors and the L1 norm regularization of the operation vector is solved to generate the minimum intervention strategy operation vector with sparsity constraints: ； in, represents the agricultural operation vector; A (j) represents the agricultural operation response matrix, which describes the impact intensity of each operation on agricultural variables; O represents the operation variable vector; represents the regularization parameter, which controls the weight of the sparse constraint term; ||O||1 represents the 1-norm of the operation vector; Consistency verification is performed through the agricultural knowledge graph, and judgment functions are used to screen whether the operation plan complies with agricultural regulations: ； in, Represents the agricultural knowledge graph; Indicates final agricultural practice recommendations after consistency verification and revision; Used to determine whether the operation plan complies with agricultural knowledge and standards; Indicates adjustments to unreasonable strategies.

6. The agricultural monitoring method of multi-source perception fusion according to claim 5 is characterized in that: The update process of the policy parameter tensor is: According to regulatory recommendations After execution, during the delay Then collect new perception tensors Compare and build the feedback difference tensor: ; in, Indicates that during policy execution The multi-source perception tensor is then recaptured; represents the feedback difference tensor; Construct the state drift potential energy function: ； ； Among them, S x,y,t represents the stability score of the agricultural system at spatial location x, y and time t; Indicates the tensor fluctuation amplitude of the specified area in the past K2 time units; K2 represents the time window length used to calculate historical fluctuations; T x,y,t-k Represents the multi-source perception semantic tensor at the spatial position (x, y) at the time point tk; represents the average tensor state within the time window [t-K2, t-1]; Represents the state drift potential energy function value; represents the stability adjustment coefficient; The potential energy calculated in the previous step is used to adjust the next round of strategy output tensor, and the strategy parameters are expressed as tensors , the following response gradient descent is used: ； in, represents the policy parameter tensor, i.e., the control model weights or feature map parameters used to generate recommendations at the tth round of decision making; represents the control model weight used to generate recommendations in the t+1th round of decision making; represents the learning rate; represents the gradient of the drift potential with respect to the policy parameters.

7. The agricultural monitoring method of multi-source perception fusion according to claim 6 is characterized in that: The correction process of dynamically correcting the target state is: The memory averaging mechanism uses a decay factor to fuse the new strategy weights with historical weights to avoid oscillation. At the same time, the trend prediction module uses historical observation tensor sequence modeling to model multi-channel coupling characteristics, dynamically decode and predict future target states to achieve dynamic correction: ； ； in, represents the strategy memory decay factor; represents the time trend prediction function; represents the observation sequence within the time window; Indicates the target agricultural status of a new round.

8. An agricultural monitoring system with multi-source perception fusion, which is used to execute the agricultural monitoring method with multi-source perception fusion according to any one of claims 1 to 7, characterized in that: include: The multi-source perception semantic tensor construction module is responsible for receiving and fusing multi-source heterogeneous data, and then processing it through spatial and temporal fusion and resonance field initialization to construct a high-dimensional semantic tensor, completing the deep semantic expression of agricultural environment and crop growth information. The multimodal aggregation module is used to analyze the coupling of multimodal semantic tensors. By building a gravitational tensor model between devices, it organically aggregates different modal and time series data. The agricultural status emergence detection module is used to dynamically analyze aggregated agricultural status data, identify potential anomalies, emergent behaviors, and nonlinear evolution characteristics, and provide a scientific basis for risk warning and status diagnosis; The agricultural regulation suggestion generation module is used to automatically generate targeted agricultural regulation strategies and operational suggestions based on the detected agricultural status changes and emergent features; The agricultural strategy dynamic feedback and self-evolution module is used to combine multi-source perception data after the execution of control measures to construct feedback difference tensors and state drift potential energy, and dynamically adjust and optimize agricultural strategies by combining policy gradient updates and trend predictions.

Images