Agricultural field environment dynamic detection and intelligent control system and method based on multi-modal data

Abstract

This invention relates to a system and method for dynamic detection and intelligent control of farmland environment based on multimodal data, including a data acquisition and preprocessing module, a multimodal data fusion module, a dynamic detection and prediction module, an intelligent control decision-making module, and a closed-loop feedback and optimization module. By collecting data from satellite remote sensing, ground sensors, UAVs, and meteorological data, and employing an innovative temporal conditional topological persistence fusion technology, deep spatiotemporal features of the multimodal data are extracted. A ZZTop-LSTM model incorporating topological features is used for real-time environmental state detection and forward-looking prediction, generating early warnings. A hybrid decision-making mechanism combining a large language model (LLM) and a numerical optimizer is introduced to transform semantic suggestions into executable, constraint-compliant control commands. Finally, closed-loop feedback enables self-iterative optimization of system parameters. This invention significantly improves the perception depth and predictive warning capabilities of farmland environment, achieving precise on-demand control and realizing the goals of water conservation, cost reduction, and efficiency improvement.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a system and method for dynamic detection and intelligent control of farmland environment based on multimodal data. Background Technology

[0002] With global population growth and intensified climate change, agricultural production faces multiple challenges, including increasing yields, conserving resources, and adapting to environmental fluctuations. Smart agriculture, by introducing the Internet of Things, remote sensing technology, and artificial intelligence, aims to achieve precise monitoring and automated control of the farmland environment. In this context, multimodal data fusion and intelligent decision-making technologies have become key to improving the perception and control capabilities of agricultural systems. At the level of farmland environmental perception, existing technologies mainly rely on single or limited data sources. For example, they collect data such as soil moisture and temperature in real time by deploying field sensor networks. However, the perception range is limited, and it is difficult to obtain spatially heterogeneous data information at the field scale. Existing technologies also use satellite or UAV remote sensing images for large-scale monitoring, but this method is often limited in temporal and spatial resolution and has weak perception capabilities for environmental parameters below the crop canopy (such as root moisture). Although there have been attempts to fuse sensor and remote sensing data, the fusion methods mostly stop at data layer splicing or feature layer simple weighting, without in-depth mining of the spatiotemporal correlation and topological structure between multimodal data. This results in insufficient perception depth for complex events such as farmland drought, flood, and pests and diseases, weak feature representation robustness, and limited ability to characterize deep environmental states and abnormal events. As a result, early warning of farmland conditions is delayed, making it difficult to provide sufficient time to support proactive intervention. Summary of the Invention

[0003] In view of this, the present invention provides a system and method for dynamic detection and intelligent control of farmland environment based on multimodal data, which can deeply integrate multimodal spatiotemporal data and realize forward-looking and accurate prediction, generate interpretable and executable intelligent decisions, and improve the self-optimization ability of farmland environment.

[0004] To achieve the above objectives, the present invention provides a dynamic monitoring and intelligent control system for farmland environment based on multimodal data, comprising a data acquisition and preprocessing module, a multimodal data fusion module, a dynamic monitoring and prediction module, an intelligent control decision-making module, and a closed-loop feedback and optimization module. The data acquisition and preprocessing module acquires satellite remote sensing images. Ground sensor data Aerial images taken by drones and meteorological data Construct time series multimodal data vectors For multimodal data vectors Perform noise removal and standardization preprocessing, and output the preprocessed data. ; The multimodal data fusion module is connected to the data acquisition and preprocessing module, and receives the preprocessed data. And extract the features of each modality data. A time-conditional topological persistence fusion method is adopted, which combines the topological persistence operator ZZTop with attention weights. Feature fusion is performed to generate unified spatiotemporal fusion features. ; The dynamic detection and prediction module is connected to the multimodal data fusion module and receives the spatiotemporal fusion features. The current environmental state is analyzed using the topological persistent long short-term memory network model ZZTop-LSTM. Conduct real-time monitoring and predict future environmental conditions Based on predictions of future environmental conditions Generate early warning signals based on preset environmental thresholds ; The intelligent control and decision-making module is connected to the dynamic detection and prediction module, and receives future environmental conditions. and warning signals Generate semantic-level control instructions through large language model LLM. Through numerical regression head Quantized into executable control instructions by the constraint optimizer ; The closed-loop feedback and optimization module is connected to the intelligent control and decision-making module and the data acquisition and preprocessing module, respectively, to collect the actual environmental state after intelligent control. Calculate the prediction error The prediction error The attention weights in the multimodal data fusion module are optimized and updated using the gradient descent method. .

[0005] Preferably, a temporal conditional topological persistence fusion method is used for feature fusion to generate spatiotemporal fused features. Specifically, it includes: The topological persistence operator ZZTop is used to calculate the characteristics of each modal data. The topological signature is expressed as: in, Indicates a point in time. Indicates the adjustment parameter. Indicates time Modal eigenvectors, Represents the spacetime potential energy field. Indicates modal features at the filtering threshold The Betti number, i.e., topological invariant; The fusion weights for each modality are calculated using a context-based attention mechanism, expressed as follows: in, Representing modes At any moment Attention weights Indicates attention score; Generate spatiotemporal fusion features The expression is: in, This represents the total number of modes.

[0006] Preferably, the topological persistence operator ZZTop is implemented through a differentiable approximation, specifically including: Persistent graphs (PDs) can be converted into multi-scale persistent images (PIs) or multi-scale persistent landscapes (PLs) through differentiable transformations. Through trainable mapping function Mapping topological characteristics at different scale levels into vectors The expression is: in, Representing modes At any moment Scale level Persistent image features on Representing modes At the scale level Topological feature vectors after mapping; Normalization is performed using softmax, and fusion is achieved using a context-based multi-scale attention mechanism, expressed as: in, Indicates the number of scale levels. Representing modes At any moment Scale level Attention weights on the surface.

[0007] Preferably, the dynamic detection and prediction module performs predictions using the ZZTop-LSTM topological persistent long short-term memory network model, and mainly includes: Based on the spatiotemporal fusion features Update hidden status The expression is: in, Indicates a hidden state; , These represent the weights and biases, respectively. Calculate and output the current environment state vector. and future Predicted state at time ; When the prediction deviation exceeds the preset tolerance Generate early warning signal The expression is: in, This indicates the environmental state threshold.

[0008] Preferably, the intelligent control decision-making module includes a semantic decision-making submodule, a numerical quantization submodule, and a constraint optimization submodule; The semantic decision submodule receives structured decision input based on the Large Language Model (LLM). ,in The constraints are represented, and the output includes semantic-level control instructions such as action type, suggestion quantity, duration, reason, and confidence level. ; The numerical quantization submodule includes a trainable numerical regression head. The semantic-level control instructions Mapping to input data to generate preliminary numerical estimates ; The constrained optimization submodule performs preliminary numerical estimation. Optimization is performed by projecting executable control variables through a constraint optimizer. The expression is: in, This indicates the upper limit of the control quantity. Indicates the cost trade-off coefficient. This represents the execution cost function. This represents a control variable.

[0009] Preferably, the closed-loop feedback and optimization module updates system parameters by: Calculate the actual environmental state after regulation With feedback status Between Norm error, expressed as: Update the fusion weights using gradient descent. : in, This indicates the updated fusion weights. Indicates the learning rate; The ZZTop-LSTM model and numerical regression head are updated based on error backpropagation. The parameters.

[0010] This invention also provides a method for dynamic detection and intelligent control of farmland environment based on multimodal data, comprising the following steps: S1. Collect multimodal farmland environmental data and perform preprocessing; S2. Extract the features of each modality data and generate unified spatiotemporal fusion features based on the time-conditional topological persistence fusion method; S201. Extracting multi-scale topological features using differentiable topological persistence operators; S202. Employ an attention-based mechanism to dynamically calculate the fusion weights for each modality and scale level. S203. Introduce a spatiotemporal potential energy field to physically modulate topological features; S3. The ZZTop-LSTM model is used to perform real-time detection and prediction of the environmental state of the spatiotemporal fusion features, and an early warning signal is generated. S4. Process the prediction results using the big data model LLM and numerical regression methods to generate semantic suggestions and quantify them into control instructions; S5. Execute the control command and collect feedback data, calculate the prediction error, and update the system parameters to achieve closed-loop optimization.

[0011] Preferably, step S4 further includes the following steps: S401. Input the prediction results, real-time detection results and constraints into the big data model LLM to generate semantic suggestions including action type, suggestion quantity, duration, reason and confidence level; S402. Map the semantic suggestions into preliminary control quantities using a numerical regression head; S403. Combining uncertainty perception and risk-sensitive optimization, the final control command is calculated under physical and equipment constraints.

[0012] Preferably, the system parameters are updated using an online learning method, combining historical errors with current feedback for incremental optimization, and incorporating a learning rate scheduling and parameter update protection mechanism.

[0013] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention overcomes the limitations of simple splicing or weighted averaging of traditional multimodal data by using the innovative "Temporal Conditional Topological Persistence Fusion (ZZTop)" method. It can effectively extract and fuse the inherent spatiotemporal topological structure and evolution pattern from multi-source heterogeneous data such as satellites, sensors, UAVs, and meteorology. This makes the system more capable of representing complex events such as crop stress (such as drought and waterlogging) and the spread of pests and diseases. The feature representation is more robust to interference such as data noise and occlusion, which significantly improves the overall perception depth of the system. (2) The present invention combines the ZZTop-LSTM prediction model with topological features, which can more accurately capture the nonlinear spatiotemporal dependence between environmental parameters, thus winning a valuable time window for proactive intervention and reducing disaster losses; (3) This invention introduces the semantic reasoning capability of large language model (LLM) into the closed loop of agricultural regulation decision-making; LLM can generate decision suggestions that conform to agricultural semantic logic, and then quantify and calibrate them through a learnable numerical regression head and an optimizer that integrates physical constraints and economic costs. This not only makes the decision-making process more reliable, but also greatly improves the safety and reliability of decision-making in scenarios with data noise or sudden environmental changes by introducing uncertainty perception and risk-sensitive optimization. Attached Figure Description

[0014] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0015] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0016] Example 1 This embodiment proposes a dynamic monitoring and intelligent control system for farmland environment based on multimodal data, including a data acquisition and preprocessing module, a multimodal data fusion module, a dynamic monitoring and prediction module, an intelligent control decision-making module, and a closed-loop feedback and optimization module; The data acquisition and preprocessing module acquires satellite remote sensing images. Ground sensor data (including soil moisture) ,temperature Aerial images taken by drones and meteorological data Construct time series multimodal data vectors For multimodal data vectors Perform noise removal and standardization preprocessing, and output the preprocessed data. ; The dimension is , These represent the height and width, respectively. Indicates the number of channels; The dimension is , Indicates the number of sensors; For UAV image matrix; For meteorological vectors; Data collection Gaussian filtering was applied to remove noise and then normalized to obtain... The expression is: in, This represents the mean. Indicates standard deviation, Indicates the length of the time series; The data acquisition and preprocessing module collects and preprocesses multimodal data to ensure data quality and consistency, providing reliable input for subsequent fusion and avoiding noise interference with depth perception.

[0017] The multimodal data fusion module is connected to the data acquisition and preprocessing module to receive the preprocessed data. And extract the features of each modality data. We use CNN to extract the rejection features, expressed as follows: in, Represents mode, Represents a modality-specific feature vector (dimension) , (where is the feature dimension), CNN is the convolutional neural network function; A time-conditional topological persistence fusion method is adopted, which combines the topological persistence operator ZZTop with attention weights. Feature fusion is performed to generate unified spatiotemporal fusion features. This includes the following steps: The topological persistence operator ZZTop is used to calculate the characteristics of each modal data. The topological signature is expressed as: in, Indicates a point in time. Indicates the adjustment parameter. Indicates time Modal eigenvectors, Represents the spacetime potential energy field. Indicates modal features at the filtering threshold The Betti number, i.e., topological invariant; The topological persistence operator ZZTop is implemented through differentiable approximation, specifically including: Persistent graphs (PDs) can be converted into multi-scale persistent images (PIs) or multi-scale persistent landscapes (PLs) through differentiable transformations. Through trainable mapping function Mapping topological characteristics at different scale levels into vectors The expression is: in, Representing modes At any moment Scale level Persistent image features on Representing modes At the scale level Topological feature vectors after mapping; Normalization is performed using softmax, and fusion is achieved using a context-based multi-scale attention mechanism, expressed as: in, Indicates the number of scale levels. Representing modes At any moment Scale level Attention weights on; The fusion weights for each modality are calculated using a context-based attention mechanism, expressed as follows: in, Representing modes At any moment Attention weights Indicates attention score; Generate spatiotemporal fusion features The expression is: in, Indicates the total number of modes; Spacetime potential energy field The scale and influence range of the modulated topological signature are calculated using the ZZTop operator, which is expressed by the persistence cohomology theory as the Betti number varying with the threshold. The cumulative component is introduced, and the field potential energy correction is also introduced; The multimodal data fusion module enhances the perception depth of multimodal data through topological fusion, captures spatiotemporal dependencies, provides comprehensive feature representation for real-time monitoring, and enables deep perception of the rice environment (such as changes in soil moisture).

[0018] The dynamic detection and prediction module is connected to the multimodal data fusion module to receive spatiotemporal fusion features. The current environmental state is analyzed using the topological persistent long short-term memory network model ZZTop-LSTM. Conduct real-time monitoring and predict future environmental conditions : Based on the spatiotemporal fusion features Update hidden status The expression is: in, Indicates a hidden state; , These represent the weights and biases, respectively. Calculate and output the current environment state vector. and future Predicted state at time The expression is: in, Indicates the forecast time window; Based on predictions of future environmental conditions Generate early warning signals based on preset environmental thresholds When the prediction deviation exceeds the preset tolerance Generate early warning signal The expression is: in, This represents the environmental condition threshold (e.g., a humidity threshold of 0.5), and in this embodiment, the tolerance. Set to 0.1; The dynamic detection and prediction module enables real-time monitoring and forward-looking prediction, generating early warning signals. This provides drought early warnings for rice paddies, enabling proactive responses.

[0019] The intelligent control and decision-making module is connected to the dynamic detection and prediction module. The intelligent control and decision-making module includes a semantic decision-making submodule, a numerical quantization submodule, and a constraint optimization submodule; the intelligent control and decision-making module receives future environmental states. and warning signals Generate semantic-level control instructions through large language model LLM. Through numerical regression head Quantized into executable control instructions by the constraint optimizer ; The semantic decision submodule receives structured decision input based on the Large Language Model (LLM). ,in The constraints are represented, and the output includes semantic-level control instructions such as action type, suggestion quantity, duration, reason, and confidence level. ; The numerical quantization submodule includes a trainable numerical regression head. The semantic-level control instructions Mapping to input data to generate preliminary numerical estimates ; The constrained optimization submodule performs preliminary numerical estimation. Optimization is performed by projecting executable control variables through a constraint optimizer. The expression is: in, This indicates the upper limit of the control quantity. Indicates the cost trade-off coefficient. This represents the execution cost function. Indicates control variables; The intelligent control and decision-making module generates intelligent control instructions on demand, thereby reducing costs and increasing efficiency, such as optimizing rice irrigation to reduce waste.

[0020] The closed-loop feedback and optimization module is connected to the intelligent control and decision-making module and the data acquisition and preprocessing module, respectively, to collect the actual environmental state after intelligent control. Calculate the actual environmental state after regulation. With feedback status Between Norm error, prediction error The expression is: Prediction error The attention weights in the multimodal data fusion module are optimized and updated using the gradient descent method. The expression is: in, This indicates the updated fusion weights. In this embodiment, the learning rate is represented. Set to 0.01, update the ZZTop-LSTM model and numerical regression head based on error backpropagation. The parameters.

[0021] Example 2 This embodiment provides a method for dynamic detection and intelligent control of farmland environment based on multimodal data, including the following steps: S1. Collect multimodal farmland environmental data and perform preprocessing; S2. Extract the features of each modality data and generate unified spatiotemporal fusion features based on the time-conditional topological persistence fusion method; S201. Extracting multi-scale topological features using differentiable topological persistence operators; S202. Employ an attention-based mechanism to dynamically calculate the fusion weights for each modality and scale level. S203. Introduce a spatiotemporal potential energy field to physically modulate topological features; S3. The ZZTop-LSTM model is used to perform real-time detection and prediction of the environmental state of the spatiotemporal fusion features, and an early warning signal is generated. S4. Process the prediction results using the big data model LLM and numerical regression methods to generate semantic suggestions and quantify them into control instructions; S401. Input the prediction results, real-time detection results and constraints into the big data model LLM to generate semantic suggestions including action type, suggestion quantity, duration, reason and confidence level; S402. Map the semantic suggestions into preliminary control quantities using a numerical regression head; S403. Combining uncertainty perception and risk-sensitive optimization, the final control command is calculated under physical and equipment constraints.

[0022] S5. Execute the control command and collect feedback data, calculate the prediction error, and update the system parameters to achieve closed-loop optimization. The system parameter update adopts an online learning method, combines historical errors with current feedback for incremental optimization, and introduces a learning rate scheduling and parameter update protection mechanism.

[0023] Example 3 This embodiment evaluates the perception depth of multimodal data fusion, the accuracy of real-time monitoring and prediction, the cost reduction and efficiency improvement capabilities of on-demand intelligent control, and the autonomy of the fully automatic intelligent control closed loop. The experiment was conducted in a rice planting scenario, and combined with multimodal data sources such as satellite remote sensing, ground sensors, UAV aerial photography, and meteorological data, the system's performance under different environmental conditions was tested.

[0024] 1. Acquiring and preprocessing experimental data (1) Obtaining experimental data: The experimental data were obtained from the following multimodal data sources, covering the 2022-2023 growing season data of a rice-growing area (approximately 500 hectares), spanning 6 months (May to October): Satellite remote sensing images The data is from the Sentinel-2 satellite with a resolution of 10 meters, including multispectral bands (RGB, near-infrared, etc.), acquired every 5 days, and downloaded from the Copernicus open data platform. Ground sensor data 100 sensor nodes deployed in the field collect soil moisture data. Soil temperature Data such as pH value are sampled once per hour, and the data is transmitted to the cloud via LoRa network; Drone aerial data Using a DJI Phantom 4 Pro drone, high-resolution imagery (0.1-meter resolution) was collected weekly, covering the phenotypic characteristics of crops in the field. Meteorological data Data includes Growing Days (GDD), Heating Days (HDD), rainfall, wind speed, etc., sourced from local weather stations, updated daily, and obtained via API.

[0025] (2) Data preprocessing: To ensure data consistency and quality, all data underwent the following preprocessing steps: Noise removal: Gaussian filtering (kernel size 5x5) is applied to satellite images, and medium filtering (window size 3) is applied to sensor data. Standardization: Modal data are processed according to the formula. Standardization is carried out, among which and These are the mean and standard deviation of the time series, respectively. Time alignment: A linear interpolation method is used to uniformly align data with one time point per hour for data with different sampling frequencies; Data splitting: The data is divided into training, validation and test sets in an 8:1:1 ratio to ensure the continuity of the time series.

[0026] (3) Data scale statistics and labeling A total of approximately 4,320 time-series data points (6 months × 30 days × 24 hours) were collected, each containing multimodal features. Target events (such as drought and flood) were confirmed through expert annotation and historical records, with a total of 58 drought events and 32 flood events annotated, serving as monitoring signals for prediction and early warning tasks.

[0027] 2. Experiment Content (1) Intermediate data generation: The following key intermediate data were generated during the experiment for subsequent model use: Feature extraction results Features were extracted using CNNs (such as ResNet-18) for each modality of data. The feature dimension of satellite images was 512, the feature dimension of sensor data was 64, and the feature dimension of UAV images was 256. Topology fusion features Based on temporal conditional topological persistence fusion (ZZTop), a unified spatiotemporal feature representation with a dimension of 128 is generated; fusion weights are used. The formula is dynamically calculated using an attention mechanism. ; Predicted state The ZZTop-LSTM model is used to predict the environmental conditions in the next 6 hours, and the output includes indicators such as soil moisture and temperature. Control Orders Generate semantic-level decision instructions through LLM and quantify them into irrigation amounts. ,in , ; Feedback error : The actual state after regulation and the predicted state Norm error, the formula is: Used for closed-loop optimization; (4) Experimental operation process: The experiment is carried out in sequence according to the five core modules of the technical solution. The specific operation is as follows: Data acquisition and preprocessing module: Downloads data from various data sources periodically via scripts, calls preprocessing functions (such as Gaussian filtering) to process noise, and saves standardized data to HDF5 format; Multimodal data fusion module: CNN feature extraction and ZZTop topology fusion are implemented using the PyTorch framework. The fused features are computed using GPU acceleration (NVIDIA RTX 3090), taking approximately 0.2 seconds per time point. Dynamic detection and prediction module: Real-time monitoring and prediction based on the ZZTop-LSTM model, prediction window. Hourly warning threshold Output early warning signal ; The intelligent regulation and decision-making module (LLM, based on GPT-3.5 fine-tuning) receives prediction results, generates structured instructions, and quantifies irrigation amounts using regression analysis. The command was sent to the simulated irrigation equipment; Closed-loop feedback and optimization module: Collects the actual state after regulation and calculates the error. Update the fusion weights using gradient descent Learning rate ; 3. Experimental results and comparison with other methods (1) The experimental results were evaluated from four aspects: perception depth, prediction accuracy, regulation efficiency, and system autonomy. The specific indicators are as follows: Perceptual depth: through multimodal feature fusion Classification performance evaluation, The score was 0.89, which is better than single-modality (satellite imagery). ,sensor ); Forecast accuracy: Drought warning lead time Within 1 hour, the recall rate was 0.87, the precision rate was 0.83, and the AUC was 0.91. Regulation efficiency: Irrigation water consumption decreased by 12.3%, crop stress reduction rate ΔStress = 15.6%, and return on investment (ROI) increased by 8.7%; System autonomy: closed-loop error The convergence rate is 5 iterations, and the oscillation amplitude is less than 0.05. (2) The method provided in this embodiment is compared with the following baseline methods. The experimental results are shown in Table 1: Single-modal LSTM: Prediction and control based solely on sensor data; Traditional multimodal fusion (Concat): simply concatenates multimodal features without topological fusion; Traditional control strategies: irrigation rules based on fixed thresholds, without prediction and intelligent decision-making; Table 1 As can be seen from Table 1, the method provided in this embodiment is significantly better than the baseline method in all indicators, especially in terms of early warning time and water consumption reduction rate, which demonstrates the advantages of multimodal topology fusion and intelligent regulation.

[0028] 4. Ablation test results To further verify the contribution of each key component, this embodiment also conducted an ablation experiment, removing or replacing the following modules respectively, and evaluating their impact on system performance. The experimental results are shown in Table 2: No ZZTop fusion: directly uses weighted average fusion features; No LLM decision: Only the regression head is used to generate regulatory instructions; No Feedback: No parameter optimization is performed; the initial parameters are used directly. Table 2 As shown in Table 2, topology fusion (ZZTop) contributes the most to early warning time and prediction accuracy, LLM decision-making significantly improves control efficiency, while closed-loop feedback contributes to the long-term stability and adaptability of the system.

[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A farmland environment dynamic detection and intelligent control system based on multi-modal data, characterized in that, It includes a data acquisition and preprocessing module, a multimodal data fusion module, a dynamic detection and prediction module, an intelligent control and decision-making module, and a closed-loop feedback and optimization module; The data acquisition and preprocessing module acquires satellite remote sensing images. Ground sensor data Aerial images taken by drones and meteorological data Construct time series multimodal data vectors For multimodal data vectors Perform noise removal and standardization preprocessing, and output the preprocessed data. ; The multimodal data fusion module is connected to the data acquisition and preprocessing module, and receives the preprocessed data. And extract the features of each modality data. A time-conditional topological persistence fusion method is adopted, which combines the topological persistence operator ZZTop with attention weights. Feature fusion is performed to generate unified spatiotemporal fusion features. ; The dynamic detection and prediction module is connected to the multimodal data fusion module and receives the spatiotemporal fusion features. The current environmental state is analyzed using the topological persistent long short-term memory network model ZZTop-LSTM. Conduct real-time monitoring and predict future environmental conditions Based on predictions of future environmental conditions Generate early warning signals based on preset environmental thresholds ; The intelligent control and decision-making module is connected to the dynamic detection and prediction module, and receives future environmental conditions. and warning signals Generate semantic-level control instructions through large language model LLM. Through numerical regression head Quantized into executable control instructions by the constraint optimizer ; The closed-loop feedback and optimization module is connected to the intelligent control and decision-making module and the data acquisition and preprocessing module, respectively, to collect the actual environmental state after intelligent control. Calculate the prediction error The prediction error The attention weights in the multimodal data fusion module are optimized and updated using the gradient descent method. .

2. The farmland environment dynamic detection and intelligent control system based on multimodal data according to claim 1, characterized in that, The temporal conditional topological persistence fusion method is used to perform feature fusion and generate spatiotemporal fused features. Specifically, it includes: The topological persistence operator ZZTop is used to calculate the characteristics of each modal data. The topological signature is expressed as: in, Indicates a point in time. Indicates the adjustment parameter. Indicates time Modal eigenvectors, Represents the spacetime potential energy field. Indicates modal features at the filtering threshold The Betti number, i.e., topological invariant; The fusion weights for each modality are calculated using a context-based attention mechanism, expressed as follows: in, Representing modes At any moment Attention weights Indicates attention score; Generate spatiotemporal fusion features The expression is: in, This represents the total number of modes.

3. The farmland environment dynamic detection and intelligent control system based on multimodal data according to claim 2, characterized in that, The topological persistence operator ZZTop is implemented through a differentiable approximation, specifically including: Persistent graphs (PDs) can be converted into multi-scale persistent images (PIs) or multi-scale persistent landscapes (PLs) through differentiable transformations. Through trainable mapping function Mapping topological characteristics at different scale levels into vectors The expression is: in, Representing modes At any moment Scale level Persistent image features on Representing modes At the scale level Topological feature vectors after mapping; Normalization is performed using softmax, and fusion is achieved using a context-based multi-scale attention mechanism, expressed as: in, Indicates the number of scale levels. Representing modes At any moment Scale level Attention weights on the surface.

4. The farmland environment dynamic detection and intelligent control system based on multimodal data according to claim 1, characterized in that, The dynamic detection and prediction module performs predictions using the ZZTop-LSTM topological persistent long short-term memory network model, and mainly includes: Based on the spatiotemporal fusion features Update hidden status The expression is: in, Indicates a hidden state; , These represent the weights and biases, respectively. Calculate and output the current environment state vector. and future Predicted state at time ; When the prediction deviation exceeds the preset tolerance Generate early warning signal The expression is: in, This indicates the environmental state threshold.

5. The farmland environment dynamic detection and intelligent control system based on multimodal data according to claim 1, characterized in that, The intelligent control decision-making module includes a semantic decision-making submodule, a numerical quantization submodule, and a constraint optimization submodule; The semantic decision submodule receives structured decision input based on the Large Language Model (LLM). ,in The constraints are represented, and the output includes semantic-level control instructions such as action type, suggestion quantity, duration, reason, and confidence level. ; The numerical quantization submodule includes a trainable numerical regression head. The semantic-level control instructions Mapping to input data to generate preliminary numerical estimates ; The constrained optimization submodule performs preliminary numerical estimation. Optimization is performed by projecting executable control variables through a constraint optimizer. The expression is: in, This indicates the upper limit of the control quantity. Indicates the cost trade-off coefficient. This represents the execution cost function. Indicates a control variable.

6. The farmland environment dynamic detection and intelligent control system based on multimodal data according to claim 1, characterized in that, The closed-loop feedback and optimization module updates system parameters including: Calculate the actual environmental state after regulation With feedback status Between Norm error, expressed as: Update the fusion weights using gradient descent. : in, This indicates the updated fusion weights. Indicates the learning rate; The ZZTop-LSTM model and numerical regression head are updated based on error backpropagation. The parameters.

7. A method for dynamic detection and intelligent control of farmland environment based on multimodal data, employing the control system described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Collect multimodal farmland environmental data and perform preprocessing; S2. Extract the features of each modality data and generate unified spatiotemporal fusion features based on the time-conditional topological persistence fusion method; S201. Extracting multi-scale topological features using differentiable topological persistence operators; S202. Employ an attention-based mechanism to dynamically calculate the fusion weights for each modality and scale level. S203. Introduce a spatiotemporal potential energy field to physically modulate topological features; S3. The ZZTop-LSTM model is used to perform real-time detection and prediction of the environmental state of the spatiotemporal fusion features, and an early warning signal is generated. S4. Process the prediction results using the big data model LLM and numerical regression methods to generate semantic suggestions and quantify them into control instructions; S5. Execute the control command and collect feedback data, calculate the prediction error, and update the system parameters to achieve closed-loop optimization.

8. The method for dynamic detection and intelligent control of farmland environment based on multimodal data according to claim 7, characterized in that, Step S4 also includes the following steps: S401. Input the prediction results, real-time detection results and constraints into the big data model LLM to generate semantic suggestions including action type, suggestion quantity, duration, reason and confidence level; S402. Map the semantic suggestions into preliminary control quantities using a numerical regression head; S403. Combining uncertainty perception and risk-sensitive optimization, the final control command is calculated under physical and equipment constraints.

9. The method for dynamic detection and intelligent control of farmland environment based on multimodal data according to claim 7, characterized in that, The system parameters are updated using an online learning approach, combining historical errors with current feedback for incremental optimization, and incorporating learning rate scheduling and parameter update protection mechanisms.

Images