Intelligent monitoring method and system for whole growth period of rape based on multi-modal AI

Abstract

The present application relates to the technical field of agricultural data monitoring and analysis, in particular to a rapeseed whole growth period intelligent monitoring method and system based on multi-modal AI; the method comprises the following steps: a spatial unit with consistent growth semantics and a nonlinear growth time axis are constructed, a cross-modal consistency dynamic fusion mechanism is established, and a continuous evolution benchmark field is formed; a rapeseed growth period continuous hidden variable space is constructed, a difficult-to-observe physiological state is quantified into a continuous variable, and whole growth period dynamic modeling is realized through a cross-modal coupling evolution equation; an environment, physiology and phenotype three-layer dynamic causal structure is constructed, a continuous trajectory consistency test and causal path offset decomposition are used to accurately locate the abnormal source, and a structured abnormality explanation is generated; through risk priority dynamic scheduling and personalized intervention strategy generation, real-time feedback and causal weight iterative optimization are combined. The present application forms a whole-process closed-loop adaptive management of monitoring, decision-making, execution, feedback and strategy optimization.

Description

Technical Field

[0001] This invention relates to the field of agricultural data monitoring and analysis technology, specifically to a method and system for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI. Background Technology

[0002] With the deepening development of smart agriculture, multimodal data acquisition technologies, represented by UAV remote sensing, IoT sensors, and image recognition, are becoming increasingly mature, providing rich information sources for crop growth monitoring. As an important oilseed crop, rapeseed's full growth period monitoring is of great significance for precise management and yield assurance. However, its growth process is affected by a combination of factors such as environment, nutrients, and variety, exhibiting complex dynamic changes. Currently, research on multimodal data fusion and crop growth modeling is constantly advancing, accumulating a relatively rich technical foundation in the acquisition and processing of heterogeneous data such as images, spectra, and environments.

[0003] Chinese invention patent application CN121479271A discloses an intelligent monitoring system for the growth status of afforestation seedlings based on deep learning. This system collects multimodal growth data of seedlings using multispectral imaging sensors, 3D laser scanning sensors, and environmental monitoring sensors. The data is then processed through spatiotemporal alignment to form a standardized dataset. Morphological structure and spectral response features are extracted through parallel spatial and frequency domain analysis, and multidimensional feature representations are generated through cross-modal fusion. These features are then converted into growth state parameters using feature recombination and spatial mapping techniques. Temporal dynamic analysis identifies abnormal growth patterns. Finally, sensor operating parameters are adaptively adjusted according to the type of abnormality to form a closed-loop monitoring system.

[0004] However, effectively integrating multi-source heterogeneous observation data to form a continuous, unified, and biologically meaningful expression of the growth process remains an important technical direction in this field. Meanwhile, constructing a modeling method capable of depicting the dynamic evolution of rapeseed throughout its entire growth period, and on this basis, achieving accurate identification and interpretable diagnosis of growth anomalies, is of great significance for improving the intelligence level of monitoring systems. Facing the development needs of precision agriculture, there is an urgent need to construct an intelligent monitoring method for the entire rapeseed growth period that can integrate multimodal data, continuously depict the growth process, possess anomaly diagnosis capabilities, and support adaptive optimization, so as to further promote the development of crop monitoring technology towards systematization and intelligence. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI.

[0006] The technical solution of this invention: A method for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI, comprising the following specific implementation steps: S1. To meet the monitoring needs of rapeseed throughout its entire growth period, we construct spatial units with semantic consistency of growth, introduce a nonlinear growth time axis driven by multiple factors, establish a dynamic fusion mechanism with cross-modal consistency constraints, and further form a continuous growth evolution benchmark field. S2. Construct a continuous latent variable space for the rapeseed growth period, quantify the difficult-to-observe physiological states into continuous variables, realize dynamic modeling of the entire growth period through cross-modal coupled evolution equations, and calibrate the latent variables with the closed loop of observable features through adaptive mapping, and finally output continuous growth period indicators. S3. Construct a three-layer dynamic causal structure of environment-physiology-phenotype, identify anomalies through continuous trajectory consistency test, accurately locate the source of anomalies by using causal path offset decomposition, generate structured anomaly explanations by combining the trend of latent variable changes, and feed back the anomaly results to optimize the causal model, so as to realize the closed loop of anomaly detection, mechanism analysis and model adaptive update. S4. Closed-loop adaptive control is implemented for monitoring, intervention, and optimization management of rapeseed throughout its entire growth period. Based on continuous latent variable states, anomaly interpretations, and causal weights, closed-loop management of the entire process of monitoring, decision-making, execution, feedback, and strategy optimization is achieved through dynamic scheduling of risk priorities, generation of personalized intervention strategies, real-time feedback execution, and iterative optimization of causal weights.

[0007] Preferably, the construction of spatial units with growth semantic consistency in step S1 specifically includes: Edge computing nodes are deployed on the edge of the field. Each node is responsible for covering a monitoring area within a radius of 500 meters. The monitoring area is divided into regular grids as the basic data carrying unit. The aerial images taken by UAVs, multispectral remote sensing images, and environmental data collected by wireless sensor networks are uniformly mapped to the corresponding grid cells using geographic information system spatial registration technology; A unit internal consistency metric is introduced. When the unit internal consistency metric exceeds a preset threshold, the grid is further subdivided into sub-grids until the coefficient of variation of the normalized vegetation index value in each sub-unit is less than 10%. All modal data were normalized and scale-aligned to ensure that the growth status of rapeseed was basically uniform in each spatial unit.

[0008] Preferably, the introduction of a multi-factor driven nonlinear growth time axis in step S1 specifically includes: Based on the biological characteristics of rapeseed, a growth-driving time function was constructed, which includes temperature function, light function, humidity or soil moisture function, and soil nutrient function. A nonlinear response gating mechanism is introduced to further define the normalized form of growth time, converting natural time into a reference time that reflects the actual growth progress; Among them, the temperature function is calculated based on the daily average air temperature from field weather station data and incorporates the lower limit temperature for rapeseed growth; the light function is calculated based on the photosynthetically active radiation sensor data to simulate the promoting effect of photoperiod on rapeseed flowering induction and photosynthetic product accumulation; the humidity or soil moisture function is calculated based on the average volumetric water content of the root zone monitored by the soil moisture sensor; and the soil nutrient function is estimated by combining the nitrogen content obtained from the soil conductivity sensor and near-ground remote sensing spectral inversion, so that the growth process under different environmental conditions has a unified and comparable scale.

[0009] Preferably, the dynamic fusion mechanism for establishing cross-modal consistency constraints in step S1, and further forming a continuous growth and evolution reference field, specifically includes: Under a unified time scale, a dynamic weight allocation mechanism based on the degree of difference between modalities is constructed to address the potential conflicts between multimodal data. By evaluating the consistency of each modal feature and adaptively adjusting its fusion weight, while introducing time smoothing constraints to suppress the impact of short-term fluctuations, a stable, reliable and continuous multimodal fusion expression result is formed. A continuous evolutionary baseline field covering the entire growth process is constructed based on the fused multimodal expression results; Furthermore, the growth change rate is introduced to characterize the dynamic trend, and the current growth state is output through the state mapping function, realizing the transformation from discrete observation data to the expression of a continuous dynamic growth process.

[0010] Preferably, the construction of the continuous latent variable space for the rapeseed growth period in step S2 specifically includes: The latent variable space of each field unit is constructed at the edge node. The latent variable space of each grid unit is constructed based on the continuous fusion features and evolution rate. The latent variables are initialized as fusion feature mappings and physiological rationality boundary constraints are applied to ensure that the latent variables are quantifiable, interpretable and closely coupled with the observation data. Define a dynamic evolution equation for latent variables with normalized growth time, combine fusion characteristics and evolution rate to drive evolution, and introduce nonlinear evolution constraints to ensure that the changes of latent variables are continuous and conform to physiological laws. A recursive solution strategy is used for the evolution of latent variables to ensure that the evolution results are updated in real time; Edge nodes use long short-term memory network model parameters distributed from the cloud and local real-time data to recursively calculate the evolution trajectory of latent variables with a set step size.

[0011] Preferably, step S2, which involves aligning the latent variables with the closed loop of observable features through adaptive mapping, specifically includes: The cloud is responsible for model training and calibration, while edge nodes perform forward inference, establish a mapping function from latent variables to observable fused features, and define the reconstruction loss for calibration. The cloud aggregates historical fusion features and latent variable data uploaded by each edge node, periodically performs model retraining, updates mapping function parameters, and distributes the updated model to each edge node through a secure channel. The edge nodes load the new model for subsequent forward inference. By minimizing the reconstruction loss, the state of the latent variables is continuously calibrated so that it conforms to physiological laws and can accurately predict observable features, thus achieving closed-loop alignment between continuous latent variables and observable features.

[0012] Preferably, the construction of the three-layer dynamic causal structure of environment-physiology-phenotype in step S3 specifically includes: A three-layer variable system is defined, consisting of a set of variables in the environment-driven layer, latent variables in the physiological state layer, and fused features in the phenotypic response layer. A stage-adaptive causal function is constructed to define the causal relationship under different growth stages. The physiological to phenotypic response function is constructed synchronously. The cloud learns from historical data to form a causal template library and automatically selects the most matching template according to the current growth progress. The template parameters are sent to the edge nodes, and the edge nodes execute local causal inference after loading the template. The environmental driving variables include temperature, humidity, soil nutrients, and light, while the physiological state layer latent variables include crop potential biomass, nitrogen nutrient index, water stress coefficient, stress resistance level, and reproductive growth intensity.

[0013] Preferably, step S3, which identifies anomalies through continuous trajectory consistency testing and precisely locates the source of the anomalies using causal path offset decomposition, specifically includes: Based on the current environmental variables and causal functions, the trajectory of latent variables is predicted, the cumulative amount of trajectory bias is calculated, and the phenotypic consistency bias is constructed. Edge nodes calculate the cumulative trajectory deviation and phenotypic consistency deviation locally in real time. When any deviation continues to accumulate and exceeds the threshold, an anomaly is determined and an anomaly event is generated. After an anomaly is detected, the offset from the environment to the physiological path and the offset from the physiological path to the phenotype are calculated to construct an offset ratio coefficient. Based on the ratio coefficient, the anomaly type is determined to be an environment-dominated anomaly, a physiological anomaly, or a composite anomaly. The specific anomaly type is further identified by combining the direction of change of latent variables, thus realizing the transformation from anomaly detection to mechanism inversion.

[0014] Preferably, step S4, through dynamic scheduling of risk priorities, generation of personalized intervention strategies, real-time feedback execution, and iterative optimization of causal weights, achieves closed-loop management of the entire process of monitoring, decision-making, execution, feedback, and strategy optimization, specifically including: Edge nodes construct a comprehensive risk index based on the anomaly explanation and continuous latent variable state output in step S3. Areas with a risk index greater than 0.7 are identified as high-risk areas, 0.3 to 0.7 are medium-risk areas, and below 0.3 are low-risk areas. The cloud-based system integrates risk data from all edge nodes, combined with drone patrol capabilities, ground sensor distribution, and network bandwidth limitations, to generate a global monitoring task plan. The cloud-based strategy engine generates intervention measures based on the type of anomaly. When there is an environmental anomaly, it automatically generates an irrigation prescription map. When there is a physiological anomaly, it generates a drone variable fertilization prescription map. When there is a combined anomaly, it prioritizes irrigation and evaluates the root function recovery after irrigation. Edge nodes receive job plans from the cloud and execute real-time closed-loop control, comparing real-time observation data with predicted status to calculate deviation indicators: when the deviation exceeds the threshold, intervention parameters are adjusted immediately. The cloud aggregates the feedback deviations reported by each edge node, uses the feedback deviations to iteratively update the causal weights, recalculates the risk level and intervention parameters after updating the causal weights, forms the adaptive scheduling and intervention plan for the next monitoring cycle, and distributes it to each edge node.

[0015] The technical solution of this invention: A smart monitoring system for the entire growth period of rapeseed based on multimodal AI, comprising: Memory; processor; A computer program stored in the memory and capable of running on the processor; When the processor executes the computer program, it implements the aforementioned intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a multimodal AI-based intelligent monitoring method and system for the entire growth period of rapeseed. Employing a cloud-edge architecture, it constructs spatial units with semantically consistent growth patterns and a nonlinear growth time axis, unifying multimodal heterogeneous observation data into a continuous evolutionary expression. This significantly improves the spatiotemporal consistency and biological interpretability of multi-source data fusion, providing a stable and reliable data foundation for subsequent growth modeling and causal analysis. The continuous latent variable space established on this basis can quantify crop physiological states that are difficult to observe directly into computable continuous variables. Combined with cross-modal coupled evolutionary equations, it achieves dynamic modeling of the entire growth period, making the description of the growth process more continuous and physiologically reasonable. Further, it constructs environmental, physiological, and phenotypic three-dimensional... The layered dynamic causal structure breaks through the limitations of traditional anomaly detection, which can only output alarm signals. Through continuous trajectory consistency testing and causal path offset decomposition, it can accurately locate the source of anomalies and distinguish between environmentally driven anomalies, internal physiological anomalies, or complex anomalies, generating structured and interpretable diagnostic results. This achieves a technological leap from anomaly detection to mechanism inversion. Finally, the closed-loop adaptive management mechanism can dynamically adjust monitoring priorities, intervention strategies, and causal model weights based on real-time feedback. This makes monitoring, decision-making, execution, feedback, and strategy optimization an organic whole, significantly improving the operability, interpretability, and adaptability of intelligent monitoring and control throughout the entire growth period, and providing scientific and efficient technical support for precision agricultural management. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes an intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI, which includes the following specific implementation steps: S1. To meet the monitoring needs of rapeseed throughout its entire growth period, this study constructs spatial units with semantically consistent growth characteristics, introduces a multi-factor-driven nonlinear growth time axis, establishes a dynamic fusion mechanism with cross-modal consistency constraints, and further forms a continuous growth evolution benchmark field. This transforms multimodal data from heterogeneous discrete observations to a unified continuous evolutionary expression, providing a unified, stable, and biologically meaningful data foundation for subsequent growth modeling and causal analysis. The specific implementation process is as follows: S11. Based on a cloud-edge collaborative architecture, the monitoring area is initially divided into grids at the edge, and the grids are adaptively subdivided based on a multimodal feature consistency index. This ensures that the growth state within each spatial unit remains relatively uniform. Simultaneously, multi-source data such as images, spectra, and environmental data are uniformly mapped to this unit and normalized, thereby constructing a multimodal structured representation unit with spatial semantic consistency. Specifically: Edge computing nodes (such as smart gateways or edge servers) are deployed on the edge of the field. Each node is responsible for covering a monitoring area within a radius of 500 meters. Considering that the width of the planting beds and the spacing between rows in rapeseed planting are usually 2-3 meters, in order to balance monitoring accuracy and computing efficiency, the monitoring area is divided into regular grids with a side length of 20m×20m as the basic data carrying unit. Environmental data collected by UAV aerial images, multispectral remote sensing images, and wireless sensor networks (WSN) are uniformly mapped to the corresponding grid units through geographic information system (GIS) spatial registration technology. Define the multimodal observation set of the i-th field grid cell at time t as: ; Introduce an internal consistency metric: ; If the internal consistency metric of the unit When the threshold is exceeded (set to 0.15 in this embodiment), it indicates that the rapeseed growth status in the unit is significantly different. The 20m×20m grid is further subdivided into 10m×10m subgrids until the NDVI value variation coefficient in each sub-unit is less than 10%, ensuring that the rapeseed growth status (such as seedling condition and density) in each spatial unit is basically uniform. Then, all modal data were normalized and scaled: ; in, This represents the multimodal integrated feature vector of the i-th field unit at time t; The image feature subvectors are derived from the visible light camera on the drone and are extracted using a convolutional neural network (CNN) to extract rapeseed canopy coverage, leaf color index (such as green leaf index), and texture features. The spectral feature subvectors are derived from multispectral sensors (bands covering 450nm, 550nm, 670nm, 750nm, and 850nm), and extract Normalized Difference Vegetation Index (NDVI), Red Edge Chlorophyll Index (RECI), etc. The environmental feature subvectors are derived from soil temperature and humidity sensors, electrical conductivity sensors, and small weather stations deployed in the field, which collect data on soil moisture content, soil temperature, air temperature and humidity, and light intensity. This represents the consistency index within the i-th unit; This indicates the number of sampling points within the i-th unit; This represents the feature of the p-th sampling point in the i-th unit; This represents the mean value of features within a cell, calculated by averaging across all sampling points. This represents the normalized feature vector, which is the result of standardizing the original features. D represents the vector of historical data means; D represents the diagonal matrix of standard deviations. S12. By introducing multiple environmental factors such as temperature, light, humidity, and nutrients, a growth driving function with a nonlinear gating mechanism is constructed. Natural time is converted into a reference time reflecting the actual growth progress, and this time is normalized to ensure that the growth process under different environmental conditions has a uniform and comparable scale. Specifically: Based on the biological characteristics of rapeseed, its growth rate is mainly affected by temperature (accumulated temperature), photoperiod, soil moisture, and nitrogen, phosphorus, and potassium nutrients. A growth-driving time function is constructed, and a nonlinear response gating mechanism is further introduced: ; Further define the normalized form of growth time: ; in, Indicates the reference growth time (growth progress); Indicates the start time (sowing or emergence time). Indicates the integral variable (intermediate time variable); The temperature function is represented by data from field weather stations, using daily average air temperature and incorporating the lower limit temperature for rapeseed growth (base temperature, usually 3℃) for calculation. The illumination function is represented by the photoperiod function, which simulates the promoting effect of photoperiod on flowering induction and photosynthetic product accumulation in rapeseed based on photosynthetically active radiation (PAR) sensor data. It represents the humidity or soil moisture function, based on the average volumetric water content in the root zone (10-30cm depth) monitored by a soil moisture sensor. When the water content is lower than the wilting coefficient or higher than the field capacity, its contribution to growth will decrease significantly. The soil nutrient function is represented by nitrogen content obtained through soil conductivity sensors and near-ground remote sensing spectral inversion. , , and Indicates the weight of environmental factors; This represents a nonlinear gated function, in the form of a Sigmoid function, used to simulate how the growth-promoting effect of environmental factors will tend to saturate or even inhibit growth when environmental factors exceed the suitable range for crops (such as temperatures exceeding 30°C). Indicates the normalized growth rate; and These represent the minimum and maximum growth rates, respectively. It should be noted that the temperature function The data used to characterize the thermal environment conditions of rapeseed during its growth process are derived from real-time monitoring results of meteorological sensors deployed in the field or regional meteorological stations, and are corrected for different grid cells using spatial interpolation methods. This function not only reflects the instantaneous temperature level, but also comprehensively reflects the diurnal temperature range and the cumulative effect of temperature in stages, in order to characterize the driving effect of temperature on crop physiological metabolism, enzyme activity and growth rate. In this embodiment, the temperature function, as one of the key growth driving factors, works together with light, water and nutrients in the construction process of the growth reference time, thereby realizing the dynamic mapping of the actual growth progress. It should be noted that the illumination function This function is used to characterize the effective light intensity and dynamic changes received by rapeseed during its growth process. It is derived from field light sensors or meteorological data acquisition systems and takes into account factors such as solar radiation, sunshine duration, and shading conditions. This function not only reflects the instantaneous light level but also implies the regulatory effect of photoperiod on photosynthetic efficiency. In practical applications, it is usually combined with time cumulative effect for smoothing to avoid short-term fluctuations from interfering with growth judgment, thereby providing stable and reliable light energy driven information for growth progress modeling. It should be noted that the humidity or soil moisture function This method is used to characterize the dynamic impact of soil moisture and environmental humidity on rapeseed growth during the growth process. By collecting information such as soil moisture content, surface humidity, and root zone water distribution through real-time sensors, the water status at different depths and regions is integrated into a single continuous function to reflect the water supply situation changing over time. This function not only reflects the absolute amount of soil moisture but also implies the influence of the rate of water change and humidity fluctuations, thereby providing an environmental driving force indicator for growth progress calculation and ensuring that the growth reference time under different environmental conditions can accurately reflect the actual development status of the crop. It should be noted that soil nutrient function This method is used to characterize the dynamic contribution of key soil nutrients to rapeseed growth over time. It comprehensively considers the content and availability of nitrogen, phosphorus, potassium, and trace elements, and forms a continuous-time function through real-time monitoring by sensors or estimation from historical data. During the rapeseed growth process, It not only reflects the level of soil fertility, but also reflects the impact of factors such as fertilization, rainfall and nutrient loss on crop absorption efficiency, thus providing a scientific basis for growth-driving time and multimodal fusion, enabling the system to accurately measure the real-time effect of soil nutrients on crop development rate and growth status. S13. Under a unified time scale, to address potential conflicts between multimodal data, a dynamic weight allocation mechanism based on intermodal differences is constructed. This mechanism assesses the consistency of each modal feature and adaptively adjusts its fusion weights. Simultaneously, a time smoothing constraint is introduced to suppress the impact of short-term fluctuations, thereby forming a stable, reliable, and continuous multimodal fusion representation. Specifically: To avoid information conflicts between multiple modalities under a unified timeline, a consistent feedback weight update mechanism is introduced. Define fusion expression: ; The dynamic weight update rule is as follows: ; The degree of difference is defined as: ; To further improve robustness, a time smoothing term is introduced: ; in, This represents the fused multimodal feature vector; M represents the number of modes. Represents the m-th modal feature; Represents the weight of the m-th mode; This represents the difference between the m-th mode and other modes. The greater the difference, the lower the weight of the mode in the next time step, thereby effectively suppressing "data conflicts" caused by sensor failure or local environmental influences. This represents the smoothed fusion features; This represents the current observation weight, which is set based on experience or learned. The empirical value is set to 0.7 to suppress instantaneous fluctuations caused by sudden weather changes, making the growth trajectory more continuous. Indicates the time step; S14. Based on the fused multimodal expression results, a continuous evolutionary reference field covering the complete growth process is constructed, and the growth change rate is further introduced to characterize the dynamic trend. The current growth state is output through a state mapping function, realizing the transformation from discrete observation data to the expression of a continuous dynamic growth process. Specifically: After obtaining stable fusion expressions, a continuous evolutionary baseline field is constructed: ; To characterize the trend of evolutionary change, the evolutionary rate is defined: ; And define the overall growth state: ; in, Represents the growth and evolution reference field of the i-th unit; Indicates the normalized growth range. ; It represents the rate of growth and evolution, including continuous estimates of key agronomic indicators such as the current growth stage (e.g., five-leaf stage, budding stage, full bloom stage), leaf area index (LAI), and aboveground biomass. Represents the final growth state vector; The state mapping function is represented by a pre-trained multilayer perceptron (MLP) regression model. Its training data comes from a pairing dataset of agronomic indicators obtained through manual observation and field sampling in the past three rapeseed growing seasons, along with the fusion features and rates at corresponding times. This realizes a structured mapping from high-dimensional perceptual data to low-dimensional interpretable agronomic states. It should be noted that the state mapping function This tool is used to convert the fused multimodal feature vectors and their evolution rates into interpretable rapeseed growth state vectors. By learning the nonlinear mapping relationship between multimodal features and growth stages, phenotypic indicators, and environmental responses, it achieves a structured mapping from high-dimensional perception data to low-dimensional growth states, while preserving continuous evolution information and dynamic trends. It can accurately reflect the growth potential, health status, and potential anomalies of different grid units at different growth stages, providing directly usable state inputs for subsequent causal analysis and monitoring scheduling.

[0019] S2. Based on a cloud-edge collaborative architecture, edge nodes perform real-time latent variable evolution calculations, while the cloud is responsible for global model training and calibration. Based on the fused features and evolution rate from step S1, a continuous latent variable space for the rapeseed growth period is constructed, quantifying difficult-to-observe physiological states into continuous variables. Dynamic modeling of the entire growth period is achieved through cross-modal coupled evolutionary equations, and latent variables are aligned with observable features through adaptive mapping. Finally, continuous growth period indicators are output for dynamic monitoring, visualization, and subsequent anomaly analysis. The specific implementation process is as follows: S21. Construct the latent variable space for each field unit at the edge nodes, and map the fused features from step S1 to the initial latent variables. Simultaneously, apply physiological rationality boundary constraints to provide core state variables for continuous evolution modeling, ensuring that the latent variables are quantifiable, interpretable, and tightly coupled with observational data. Specifically: Based on the continuous fusion features output in step S1 and evolution rate Construct a latent variable space for each grid cell i: ; Latent variables are initialized to fused feature maps: ; To ensure the physiological rationality of the latent variables, boundary constraints are applied: ; Edge nodes load pre-trained model parameters (including but not limited to boundary constraint values) from the cloud and perform local latent variable initialization calculations; in, This indicates the growth progress of the i-th plot unit. The hidden variable state vector; denoted as the k-th latent variable in the i-th unit; K represents the number of dimensions of the latent variable, which is set according to agronomic mechanisms. In this embodiment, K represents the potential biomass of crops, nitrogen nutrition index, water stress coefficient, stress resistance level, and reproductive growth intensity, respectively. Represents the fused feature vector; The latent variable initialization mapping function is a linear mapping function obtained by regressing the fusion features in historical data with measured physiological indicators (such as plant nitrogen content and leaf water potential). Indicates the state of the hidden variables at the initial moment; and These represent the minimum and maximum values ​​of the k-th latent variable, respectively. For example, the reasonable range for the nitrogen nutrition index is 0.8-1.2. Exceeding this range indicates severe nitrogen deficiency or nitrogen excess. S22. Define a dynamic evolution equation for latent variables at the edge nodes as a function of normalized growth time. Combine this with the fusion features and evolution rate driven by step S1, and limit the magnitude of change through nonlinear constraints to form a continuous and smooth growth trajectory. This achieves a combination of dynamic modeling of latent variables and physiological rationality. Specifically: Define hidden variables as growth time The evolution equation: ; Introducing nonlinear evolution constraints ensures that the changes in latent variables are continuous and conform to physiological laws: ; A recursive solution strategy is used for the evolution of latent variables, allowing the evolution results to be updated in real time: ; Edge nodes utilize LSTM model parameters distributed from the cloud, combined with local real-time data, to... The evolution trajectory of hidden variables is calculated recursively for the step size; the calculated trajectory is... Stored locally on edge nodes in, This represents the rate of change of latent variables over growth time; The latent variable evolution function is obtained through data-driven learning and is implemented using a Long Short-Term Memory (LSTM) network. Its parameter θ is obtained by training with time-series data (including environment, fusion features, final yield, etc.) from multiple rapeseed growing seasons. Represents the set of parameters for the evolution function; Indicates the time step; This represents the constraint function for the range of change, which is set or learned based on experience of the reproductive stage. For example, during the budding stage, the upper limit of the daily range of change of the latent variable of biomass is significantly higher than that during the maturity stage. The vector norm is represented (in this embodiment, it is set to the Euclidean norm). Indicates the hidden variable state at the next moment; S23. The cloud is responsible for model training and calibration, while edge nodes perform forward inference, establishing a mapping function from latent variables to fused phenotypic features. Adaptive calibration is performed by minimizing reconstruction error, making latent variables both interpretable and predictable. This achieves closed-loop alignment between continuous latent variables and observable features, improving model accuracy and operability. Specifically: Establish a mapping function from latent variables to observable fused features: ; Define the reconstruction loss for calibration: ; The cloud aggregates historical fusion features and latent variable data uploaded by each edge node, and performs model retraining periodically (e.g., daily) to update the data. Parameters; The updated model is distributed to each edge node via a secure channel, and the edge nodes load the new model for subsequent forward inference; After variable mapping calibration, it can be used to predict fusion features at future moments; Where N represents the number of field units; This represents the fusion feature predicted by latent variables; The mapping function representing latent variables to phenotypic features is a fully connected network used to map the latent variable space back to the observable fused feature space. This represents the reconstruction loss function. By minimizing the reconstruction loss, the state of the latent variables can be continuously calibrated so that it conforms to physiological laws and can accurately predict directly observable features such as canopy coverage and NDVI at the next moment, thus achieving closed-loop alignment. S24. Generate comprehensive continuous reproductive period indicators, including latent variables, growth rates, and fusion characteristics, and visualize them to form a dynamic representation of the entire reproductive period. This provides continuous and operable state input for subsequent causal analysis, anomaly identification, and decision support. Specifically: Define the comprehensive growth status index: ; Edge nodes push comprehensive growth status indicators to local monitoring terminals (such as farmers' mobile apps or field screens) in real time, and report key indicators (early warnings of hidden variables exceeding thresholds, changes in growth stages) to the cloud on an hourly basis. The cloud aggregates data from all edge nodes to generate a continuous growth period curve at the whole field scale. On the GIS platform, the changing trend of latent variables in each grid cell is plotted with growth progress as the horizontal axis and dynamically displayed on the map in the form of a heat map. At the same time, the system automatically calculates and outputs key dynamic indicators, such as the value of the maximum growth rate (corresponding to the peak growth period) and the first time that each latent variable exceeds the stress threshold (corresponding to the stress sensitive period), to provide quantitative basis for precision management. in, Indicators representing continuous growth status.

[0020] S3. Based on a cloud-edge collaborative architecture, edge nodes perform real-time anomaly detection and causal path offset calculation, while the cloud performs global causal model training and template library management. Building upon the continuous latent variable modeling output of step S2, a three-layer dynamic causal structure of environment, physiology, and phenotype is constructed. Anomalies are identified through continuous trajectory consistency checks, and the source of anomalies is precisely located using causal path offset decomposition. Structured anomaly explanations are generated by combining latent variable change trends, and the anomaly results are fed back to optimize the causal model. This achieves a closed loop of anomaly detection, mechanism analysis, and adaptive model updates, enabling interpretable, dynamic, continuous, and operable anomaly monitoring and diagnosis. The specific implementation process is as follows: S31. Construct a three-layer causal structure, including an environmental variable layer, a physiological latent variable layer, and a phenotypic expression layer. Introduce a weight function that dynamically adjusts with growth progress to transform multimodal data from parallel features into a continuous, directed, and interpretable causal chain. This provides the basic structure for subsequent anomaly analysis and ensures that the model adaptively adjusts with the growth stage. Specifically: Define a three-level variable system: Environment-driven layer: a set of environmental variables such as temperature, humidity, and nutrients. ; Physiological state layer: latent variables ; Phenotypic Response Layer: Feature Fusion ; An adaptive causal function is constructed, with different weights for causal relationships at different growth stages. For example, seedlings are sensitive to temperature, while flowering is sensitive to nutrients. Definition: ; Simultaneous construction of physiological-phenotypic response functions: ; The cloud learns from historical data to form a causal template library, and automatically selects the most matching template based on the current growth progress, and sends the template parameters (such as weight functions) to the edge nodes; after the edge nodes load the template, they perform local causal inference. in, This represents the set of environmental variables for the i-th field unit, including but not limited to temperature, humidity, soil nutrients, and light. The causal function representing the changes in latent variables driven by the environment is formed based on historical data training and stage adaptive learning, simulating the causal effect of the environment on physiological state, and adjusting the weights according to the growth stage. This represents the adaptive weighting function for the growth stage, which varies with the normalized growth time. change; It represents the causal function that drives changes in phenotypic features through latent variables. It is a non-linear mapping that can be fitted by a multilayer perceptron (MLP). Its input is latent variables and its output is phenotypic features such as NDVI and canopy temperature. It represents a function that describes the relationship between a single environmental factor (such as temperature) and a latent variable (such as the rate of biomass accumulation). It takes the form of the Mitscherlich growth function or the Beta function, and the parameters are obtained by fitting historical data. This represents an adaptive weighting function for different growth stages. For example, during the seedling stage, temperature has the largest weight; during the flowering stage, the weights of light and nutrients increase significantly; and during the pod-setting stage, the sensitivity of the water weight increases. S32. Based on the causal structure in step S31, a continuity consistency test is performed on the latent variables and phenotypic feature trajectories. Anomalies are identified by accumulating trajectory deviations, avoiding transient noise interference. This achieves cross-time and cross-level anomaly detection and ensures more stable and reliable anomaly judgment. Specifically: Based on current environmental variables and causal functions, predict the trajectory of latent variables: ; Calculate the cumulative trajectory deviation: ; Simultaneously construct phenotypic consistency bias: ; Edge nodes perform local real-time computation and When any deviation ( or When the accumulation continues and exceeds the threshold, it is determined that there is an anomaly, rather than a fluctuation at a single time point. For example, when the predicted rapeseed biomass accumulation and the actual observed canopy coverage continue to deviate, it indicates that there is an anomaly in the growth process. in, This represents the state of latent variables predicted based on environmental variables and causal functions; and These represent the start and end times of integration, i.e., the time period for trajectory detection, which are derived from the current continuous observation window; It represents the cumulative amount of latent variable trajectory deviation, used to determine the continuity and severity of abnormal physiological states; This represents the cumulative deviation of the phenotypic trajectory, used to determine the consistency between the phenotypic and latent variable mappings; S33. At edge nodes, for identified anomalies, the source of the anomaly is precisely decomposed by calculating the offsets and proportions of the environment-physiology and physiology-phenotype paths. Combined with the trends in latent variable changes, the specific anomaly type is further identified, achieving a transformation from anomaly detection to mechanism inversion. This enables the system to locate environment-driven anomalies or internal physiological anomalies, providing interpretable diagnostic evidence. Specifically: After an anomaly is detected, the source of the anomaly is defined by decomposing the offsets of each segment in the causal chain. Define path offset metrics: Environmental → Physiological Shift This reflects environmentally driven anomalies; Physiological → Phenotypic Shift: This reflects an abnormal internal state; Construct offset scaling factor: ; According to the proportional relationship Determine the type of exception: when When this occurs, it is determined to be an environmentally dominant anomaly (such as a decrease in fruit set rate due to low temperatures during the flowering period). when When this occurs, it is determined to be a physiological abnormality (such as a decline in root vitality due to disease, but with normal environmental conditions). when The median value (i.e.) When this occurs, it is classified as a complex anomaly; If the direction of change of latent variables is combined, such as the nitrogen nutrition index continuously decreasing while the water stress coefficient is normal, the abnormality can be further diagnosed as "leaf yellowing caused by decreased nitrogen absorption efficiency". After the edge node completes the anomaly type determination, it packages the anomaly event, offset, and determination basis and reports them to the cloud, while generating local early warning information. in, This represents the offset scaling factor, used to quantify the source of the anomaly; This represents the offset of the environment layer from the hidden variable layer; This represents the offset of the latent variable layer from the phenotype layer; S34. Generate structured anomaly explanations at edge nodes, aggregate anomalies in the cloud and optimize the causal model, combine anomaly types, impact paths, and key variables to generate structured anomaly explanations, and feed the results back to dynamically adjust the causal weight function to achieve adaptive optimization of the causal model. Simultaneously, output comprehensive information combined with the continuous state indicators from step S2 to form an anomaly analysis-explanation-optimization closed loop, specifically: Edge nodes construct anomaly explanation structures locally: anomaly type (environmental / physiological / complex); impact path (e.g., environmental "insufficient light" → physiological "decreased growth intensity" → phenotype "reduced number of siliques"); key variables (e.g., average daily light hours 20 days after flowering). Generate explanatory semantics, such as: "The current field is in the late flowering stage. Due to continuous rainy weather, there is insufficient effective light (environmental abnormality), which causes the latent variable of reproductive growth to be 20% lower than the normal trajectory. It is predicted that the final number of siliques will be reduced by 15%. It is recommended to carry out artificial lighting or spray growth regulators in time." Edge nodes report abnormal events and offsets to the cloud. The cloud aggregates all abnormal data from edge nodes and periodically (e.g., weekly) retrains the causal model, adjusting the weight function. Fine-tuning can be performed; for example, if the system frequently detects growth deviations caused by abnormal soil moisture, the water weight value can be increased in the next iteration to make the model more sensitive to changes in water and achieve adaptive optimization of the causal model. Output the overall result: ; in, This represents the comprehensive output result of the i-th field unit, which includes continuous growth status and anomaly explanation information.

[0021] S4. Based on a cloud-edge collaborative architecture, the cloud is responsible for global resource scheduling and strategy optimization, while edge nodes are responsible for local real-time control and feedback execution. Based on the multimodal causal anomaly analysis output from step S3, closed-loop adaptive control is implemented for the monitoring, intervention, and optimization management of rapeseed throughout its entire growth period. Using continuous latent variable states, anomaly interpretations, and causal weights as a foundation, through dynamic scheduling based on risk priorities, generation of personalized intervention strategies, real-time feedback execution, and iterative optimization of causal weights, a closed-loop management system encompassing monitoring, decision-making, execution, feedback, and strategy optimization is achieved. This makes intelligent monitoring and regulation throughout the entire growth period operable, interpretable, and adaptive. The specific implementation process is as follows: S41. Edge nodes calculate local risk levels, and the cloud performs global resource scheduling. Based on the anomaly explanation and continuous status output in step S3, the risk level of the field unit is calculated. Dynamic monitoring priorities are generated according to risk level, and combined with the constraints of UAVs, ground sensors, and network resources, a multimodal monitoring task plan executable throughout the entire growth period is formed to achieve on-demand data collection and key monitoring of high-risk areas. Specifically: Explanation of the anomaly output in step S3 and continuous hidden variable states Construct comprehensive risk indicators: ; Based on the set threshold, Areas with a value greater than 0.7 are identified as high-risk areas, those between 0.3 and 0.7 are medium-risk areas, and those below 0.3 are low-risk areas. Edge nodes report risk levels to the cloud. The cloud then integrates the risk data from all edge nodes, taking into account drone patrol capabilities, ground sensor distribution, and network bandwidth limitations, to generate a global monitoring task plan. For example, considering the practical operational constraints such as a single drone flight covering a maximum of 500 acres and a battery life of 30 minutes, high-risk areas are allocated for daily multimodal monitoring (visible light + multispectral + thermal infrared), medium-risk areas for routine monitoring every 3 days, and low-risk areas for monitoring weekly; based on growth progress... Dynamic adjustments can be made, such as increasing the monitoring frequency during the budding stage; The cloud will distribute the monitoring task plan to the edge nodes and the drone flight control system, and adjust the monitoring frequency according to the progress of the growth stage. For example, the seedling stage focuses on growth rate and environmental sensitivity, the flowering stage focuses on flower organs and photosynthesis, and the maturity stage focuses on grain development and leaf senescence. in, This represents the risk level index of the i-th field unit, used to assess the risk of abnormal growth and to prioritize monitoring. The risk assessment function, designed based on the anomaly identification and continuous state output in step S3, transforms multimodal growth information into an actionable risk level. S42. Cloud-based intervention strategies are generated, and edge nodes execute local control. For different types of anomalies (environmental, physiological, or combined anomalies), personalized intervention strategies are generated by combining causal paths and continuous states. Anomaly types are mapped to executable operational parameters (such as irrigation volume, fertilizer application, and light duration), and the operation sequence and resource consumption are optimized to maximize intervention effectiveness, minimize costs, and achieve precise control. Specifically: For each anomaly type, the cloud-based policy engine generates targeted intervention measures: Environmental anomalies (such as) (And for water stress): Automatically generate irrigation prescription maps, including but not limited to irrigation volume (e.g., 30m³ / acre), irrigation area and optimal operation time (e.g., early morning). Physiological abnormalities (such as) And the nitrogen latent variable is low): Generate a drone variable fertilizer prescription map and determine the fertilizer application amount based on the degree of nitrogen deficit (e.g., apply 5-10 kg of urea per mu). Complex anomalies: For example, the simultaneous presence of water stress and decreased root activity. The system will prioritize irrigation, assess root function recovery after irrigation, and then decide whether to supplement with root-promoting biostimulants. Further, the intervention measures are transformed into a sequence of executable parameters: ; The cloud uses optimization algorithms (such as greedy algorithms) to optimize the order of operations, ensuring that the areas with the highest risk and the greatest impact on yield are processed first under limited resources (such as a drone or an irrigation system); S43. Edge nodes execute real-time closed-loop control, with cloud monitoring of the execution effect. Based on the generated monitoring tasks and intervention strategies, they implement drone inspections, ground sensor data collection, and automated control. Real-time observation data is compared with predicted states to calculate deviation indicators, which are used to adjust monitoring frequency and intervention operations in real time. This ensures accurate strategy execution and forms continuous closed-loop feedback. Specifically: Edge nodes receive work plans from the cloud and automatically schedule drones to take off and perform inspection tasks according to the plan. At the same time, they issue instructions through the Internet of Things (IoT) platform to automatically start the drip irrigation system or start the fertilizer drone. Edge nodes collect real-time data to form a continuous data stream. And upload it to the cloud or local edge computing node; By comparing the predicted state of edge nodes with real-time observations, the deviation index is calculated. ; When the deviation exceeds the threshold, the edge node can adjust the intervention parameters in real time. For example, if the soil moisture sensor shows that the moisture content has increased by less than 80% of the expected value 2 hours after irrigation, the intervention parameters can be adjusted immediately, such as extending the irrigation time or checking for pipe blockage, to form a local closed-loop correction. At the same time, the deviation event is reported to the cloud for subsequent strategy optimization. The monitoring-intervention-feedback cycle is formed, and the operation sequence, intervention amount and collection frequency are adaptively adjusted according to the real-time status. S44. Cloud-based aggregation of edge node feedback data enables global model iterative optimization. Real-time deviation data from step S3 is used to update causal weights, achieving adaptive model optimization. Simultaneously, risk levels and intervention strategies are recalculated to form an adaptive scheduling plan for the next cycle. Through multi-cycle iterative optimization, continuous adaptive management throughout the entire reproductive period is achieved, and a final comprehensive monitoring and intervention plan is output for decision-making. Specifically: The cloud aggregates the feedback deviations reported by each edge node and uses the causal weights output in step S3. Feedback deviation Perform iterative updates: ; After updating the causal weights, the risk level is calculated in the cloud. With intervention operation parameters The system generates an adaptive scheduling and intervention plan for the next monitoring cycle and distributes it to each edge node. The edge nodes receive the updated model parameters and continue to perform local real-time monitoring and control. Through continuous growth stage monitoring and anomaly feedback, the strategy is continuously optimized throughout the entire growth period, gradually adapting to environmental differences between different fields and years, and ultimately outputting a complete intelligent management solution for the entire growth period: ; in, This represents the updated causal weight vector, used for intervention decisions in the next monitoring period; The learning rate for weight updates controls the iteration amplitude and is set by the system or adjusted adaptively. In this embodiment, 0.05 is used to balance the update speed and stability, ensuring that the strategy converges and does not diverge. The final output represents a smart management solution for the entire reproductive cycle, including dynamic monitoring tasks, intervention strategies, and optimization records, providing operable, traceable, and optimizable full-cycle decision support for agricultural management.

[0022] Example 2: The present invention proposes a multimodal AI-based intelligent monitoring system for the entire growth period of rapeseed, which is used to execute the multimodal AI-based intelligent monitoring method for the entire growth period of rapeseed proposed in Example 1, comprising: Memory; processor; A computer program stored in the memory and capable of running on the processor; The processor executes a computer program to implement the intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI in the first embodiment described above.

[0023] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI, characterized in that, The specific implementation steps include the following: S1. To meet the monitoring needs of rapeseed throughout its entire growth period, we construct spatial units with semantic consistency of growth, introduce a nonlinear growth time axis driven by multiple factors, establish a dynamic fusion mechanism with cross-modal consistency constraints, and further form a continuous growth evolution benchmark field. S2. Construct a continuous latent variable space for the rapeseed growth period, quantify the difficult-to-observe physiological states into continuous variables, realize dynamic modeling of the entire growth period through cross-modal coupled evolution equations, and calibrate the latent variables with the closed loop of observable features through adaptive mapping, and finally output continuous growth period indicators. S3. Construct a three-layer dynamic causal structure of environment-physiology-phenotype, identify anomalies through continuous trajectory consistency test, accurately locate the source of anomalies by using causal path offset decomposition, generate structured anomaly explanations by combining the trend of latent variable changes, and feed back the anomaly results to optimize the causal model, so as to realize the closed loop of anomaly detection, mechanism analysis and model adaptive update. S4. Closed-loop adaptive control is implemented for monitoring, intervention, and optimization management of rapeseed throughout its entire growth period. Based on continuous latent variable states, anomaly interpretations, and causal weights, closed-loop management of the entire process of monitoring, decision-making, execution, feedback, and strategy optimization is achieved through dynamic scheduling of risk priorities, generation of personalized intervention strategies, real-time feedback execution, and iterative optimization of causal weights.

2. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 1, characterized in that, The specific steps in step S1 for constructing spatial units with growth semantic consistency include: Edge computing nodes are deployed on the edge of the field. Each node is responsible for covering a monitoring area within a radius of 500 meters. The monitoring area is divided into regular grids as the basic data carrying unit. The aerial images taken by UAVs, multispectral remote sensing images, and environmental data collected by wireless sensor networks are uniformly mapped to the corresponding grid cells using geographic information system spatial registration technology; A unit-internal consistency metric is introduced. When the unit-internal consistency metric exceeds a preset threshold, the grid is further subdivided into sub-grids until the coefficient of variation of the normalized vegetation index value in each sub-unit is less than 10%. All modal data were normalized and scale-aligned to ensure that the growth status of rapeseed was basically uniform in each spatial unit.

3. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 2, characterized in that, The introduction of a multi-factor driven nonlinear growth time axis in step S1 specifically includes: Based on the biological characteristics of rapeseed, a growth-driving time function was constructed, which includes temperature function, light function, humidity or soil moisture function, and soil nutrient function. A nonlinear response gating mechanism is introduced to further define the normalized form of growth time, converting natural time into a reference time that reflects the actual growth progress; Among them, the temperature function is calculated based on the daily average air temperature from field weather station data and incorporates the lower limit temperature for rapeseed growth; the light function is calculated based on the photosynthetically active radiation sensor data to simulate the promoting effect of photoperiod on rapeseed flowering induction and photosynthetic product accumulation; the humidity or soil moisture function is calculated based on the average volumetric water content of the root zone monitored by the soil moisture sensor; and the soil nutrient function is estimated by combining the nitrogen content obtained from the soil conductivity sensor and near-ground remote sensing spectral inversion, so that the growth process under different environmental conditions has a unified and comparable scale.

4. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 3, characterized in that, Step S1 establishes a dynamic fusion mechanism for cross-modal consistency constraints and further forms a continuous growth and evolution reference field, specifically including: Under a unified time scale, a dynamic weight allocation mechanism based on the degree of difference between modalities is constructed to address the potential conflicts between multimodal data. By evaluating the consistency of each modal feature and adaptively adjusting its fusion weight, while introducing time smoothing constraints to suppress the impact of short-term fluctuations, a stable, reliable and continuous multimodal fusion expression result is formed. A continuous evolutionary baseline field covering the entire growth process is constructed based on the fused multimodal expression results; Furthermore, the growth change rate is introduced to characterize the dynamic trend, and the current growth state is output through the state mapping function, realizing the transformation from discrete observation data to the expression of a continuous dynamic growth process.

5. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 4, characterized in that, The construction of the continuous latent variable space for the rapeseed growth period in step S2 specifically includes: The latent variable space of each field unit is constructed at the edge node. The latent variable space of each grid unit is constructed based on the continuous fusion features and evolution rate. The latent variables are initialized as fusion feature mappings and physiological rationality boundary constraints are applied to ensure that the latent variables are quantifiable, interpretable and closely coupled with the observation data. Define a dynamic evolution equation for latent variables with normalized growth time, combine fusion characteristics and evolution rate to drive evolution, and introduce nonlinear evolution constraints to ensure that the changes of latent variables are continuous and conform to physiological laws. A recursive solution strategy is used for the evolution of latent variables to ensure that the evolution results are updated in real time; Edge nodes use long short-term memory network model parameters distributed from the cloud and local real-time data to recursively calculate the evolution trajectory of latent variables with a set step size.

6. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 5, characterized in that, Step S2, which involves aligning the latent variables with the closed loop of observable features through adaptive mapping, specifically includes: The cloud is responsible for model training and calibration, while edge nodes perform forward inference, establish a mapping function from latent variables to observable fused features, and define the reconstruction loss for calibration. The cloud aggregates historical fusion features and latent variable data uploaded by each edge node, periodically performs model retraining, updates mapping function parameters, and distributes the updated model to each edge node through a secure channel. The edge nodes load the new model for subsequent forward inference. By minimizing the reconstruction loss, the state of the latent variables is continuously calibrated so that it conforms to physiological laws and can accurately predict observable features, thus achieving closed-loop alignment between continuous latent variables and observable features.

7. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 6, characterized in that, Step S3, which involves constructing a three-tiered dynamic causal structure of environment, physiology, and phenotype, specifically includes: A three-layer variable system is defined, consisting of a set of variables in the environment-driven layer, latent variables in the physiological state layer, and fused features in the phenotypic response layer. A stage-adaptive causal function is constructed, and the causal relationship weights under different growth stages are defined. The physiological to phenotypic response function is constructed synchronously. The cloud learns from historical data to form a causal template library and automatically selects the most matching template according to the current growth progress. The template parameters are sent to the edge nodes, and the edge nodes execute local causal inference after loading the template. The environmental driving variables include temperature, humidity, soil nutrients, and light, while the physiological state layer latent variables include crop potential biomass, nitrogen nutrient index, water stress coefficient, stress resistance level, and reproductive growth intensity.

8. The intelligent monitoring method for the entire growth period of rapeseed based on multimodal AI according to claim 7, characterized in that, Step S3 involves identifying anomalies through continuous trajectory consistency testing and precisely locating the source of the anomalies using causal path offset decomposition. Specifically, this includes: Based on the current environmental variables and causal functions, the trajectory of latent variables is predicted, the cumulative amount of trajectory bias is calculated, and the phenotypic consistency bias is constructed. Edge nodes calculate the cumulative trajectory deviation and phenotypic consistency deviation locally in real time. When any deviation continues to accumulate and exceeds the threshold, an anomaly is determined and an anomaly event is generated. After an anomaly is detected, the offset from the environment to the physiological path and the offset from the physiological path to the phenotype are calculated to construct an offset ratio coefficient. Based on the ratio coefficient, the anomaly type is determined to be an environment-dominated anomaly, a physiological anomaly, or a composite anomaly. The specific anomaly type is further identified by combining the direction of change of latent variables, thus realizing the transformation from anomaly detection to mechanism inversion.

9. A method for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI according to claim 8, characterized in that, Step S4 achieves closed-loop management of the entire process of monitoring, decision-making, execution, feedback, and strategy optimization through dynamic scheduling of risk priorities, generation of personalized intervention strategies, real-time feedback execution, and iterative optimization of causal weights. Specifically, this includes: Edge nodes construct a comprehensive risk index based on the anomaly explanation and continuous latent variable state output in step S3. Areas with a risk index greater than 0.7 are identified as high-risk areas, 0.3 to 0.7 are medium-risk areas, and below 0.3 are low-risk areas. The cloud-based system integrates risk data from all edge nodes, combined with drone patrol capabilities, ground sensor distribution, and network bandwidth limitations, to generate a global monitoring task plan. The cloud-based strategy engine generates intervention measures based on the type of anomaly. When there is an environmental anomaly, it automatically generates an irrigation prescription map. When there is a physiological anomaly, it generates a drone variable fertilization prescription map. When there is a combined anomaly, it prioritizes irrigation and evaluates the root function recovery after irrigation. Edge nodes receive job plans from the cloud and execute real-time closed-loop control, comparing real-time observation data with predicted status to calculate deviation indicators: when the deviation exceeds the threshold, intervention parameters are adjusted immediately. The cloud aggregates the feedback deviations reported by each edge node, uses the feedback deviations to iteratively update the causal weights, recalculates the risk level and intervention parameters after updating the causal weights, forms the adaptive scheduling and intervention plan for the next monitoring cycle, and distributes it to each edge node.

10. A smart monitoring system for the entire growth period of rapeseed based on multimodal AI, characterized in that, include: Memory; processor; A computer program stored in the memory and capable of running on the processor; When the processor executes the computer program, it implements a method for intelligent monitoring of the entire growth period of rapeseed based on multimodal AI as described in any one of claims 1 to 9.

Images