An environment intelligent monitoring and self-adaptive regulation method and system for agricultural fields based on multi-source internet of things perception

Abstract

The application discloses an environment intelligent monitoring and self-adaptive regulation and control method and system for agricultural fields based on multi-source Internet of Things sensing, and the method comprises the following steps: constructing a credible sensing and unified description layer of multi-dimensional farmland environment parameters, deploying heterogeneous sensing nodes and redundant sensing units, running a reliable data guarantee model based on sensing source trust evaluation, outputting high-reliability environment parameters, and uniformly encapsulating the environment parameters in a JSON format; designing a double-layer data processing architecture based on edge computing and cloud cooperation, adopting a redundant communication mechanism combining LoRa, ZigBee and cellular network to transmit data, and adapting the model to edge node resources through a deep balance training strategy and an auxiliary network compression method. The application can realize omnibearing and high-reliability sensing and self-adaptive precision regulation and control of farmland water, fertilizer, pesticide, light, temperature and other environmental factors, and forms a complete closed loop of sensing, analysis, decision-making, control and optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and adaptive regulation technology for farmland environment, and in particular to a method and system for intelligent monitoring and adaptive regulation of agricultural farmland environment based on multi-source Internet of Things sensing. Background Technology

[0002] Traditional methods of farmland environmental monitoring and control typically rely on regular manual inspections and single-sensor data, resulting in problems such as limited monitoring dimensions, low data reliability, poor real-time decision-making, and inefficient control. While the development of IoT and AI technologies has led to the emergence of some intelligent agricultural monitoring systems, they still face the following challenges: First, at the perception layer, sensor data is susceptible to environmental interference and equipment malfunctions, lacking mechanisms for assessing and ensuring the reliability of the data, leading to unreliable decision-making. Second, data processing largely depends on the cloud, making it difficult to meet the urgent real-time control needs of irrigation, plant protection, and other scenarios in field environments with poor network conditions. Third, complex intelligent control models require substantial computing resources, making direct deployment on resource-constrained edge devices in the field difficult. Finally, existing systems often focus on monitoring or single-factor control, lacking the ability to collaboratively perceive and adaptively control multiple environmental factors such as water, fertilizer, pesticides, light, and temperature, failing to form a complete closed loop from intelligent perception to precise execution. Therefore, there is an urgent need for an intelligent monitoring and control method and system for farmland environment that can achieve highly reliable perception, real-time intelligent decision-making at the edge, lightweight model adaptation, and multi-factor collaborative closed-loop control. Summary of the Invention

[0003] This invention provides a method and system for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things (IoT) sensing.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] A method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things (IoT) sensing, comprising the following steps:

[0006] Step S1: Construct a reliable perception and unified description layer for multi-dimensional farmland environmental parameters:

[0007] The system deploys heterogeneous IoT sensing nodes, including soil moisture sensors, crop nutrient sensors, meteorological sensors, pest and disease monitoring equipment, and high-definition cameras, to collect comprehensive, multi-parameter data on farmland soil, crop canopy, near-ground atmosphere, and field conditions. For at least one key environmental parameter, at least three identical sensors are deployed to form a redundant sensing unit, and a reliable data assurance model based on sensing source trust evaluation is run. This model calculates the direct trust value, energy trust value, and peer-recommended trust value of each sensing node, and weights them to obtain a comprehensive trust value. Based on the comprehensive trust value, multi-channel sensor data is filtered and fused to output highly reliable environmental parameters. A standardized environmental sensing data package is generated by uniformly describing and encapsulating all sensing devices and their collected real-time data, historical data, device status, and access permissions using JSON data format.

[0008] Step S2: Design a two-tier data processing architecture based on edge computing and cloud collaboration:

[0009] A system architecture consisting of a perception layer, a transmission layer, and an application layer is constructed. In the transmission layer, a redundant communication mechanism combining LoRa, ZigBee, and cellular networks is adopted to achieve low-power, highly reliable data backhaul. Edge computing nodes are deployed at the network edge near the field perception nodes. These nodes have built-in lightweight AI models and decision-making modules to receive the standardized perception data packets from step S1 and perform localized data preprocessing, caching, and lightweight analysis tasks with high real-time requirements. Simultaneously, the data is synchronously transmitted to the cloud server, where it performs in-depth mining of massive historical data, model training, and global optimization strategy calculation.

[0010] Step S3: Implement edge adaptive model compression and deployment for proactive intelligent control:

[0011] For deep learning models deployed on edge computing nodes for farmland environmental regulation decisions, an edge-structure adaptive active intelligent control algorithm integration framework is applied to optimize them. First, considering the characteristics of continuous changes in environmental parameters and unbalanced distribution of regulation decision data in farmland regulation scenarios, a deep balanced training strategy is adopted, including label distribution resampling and feature space smoothing, to enhance the model's generalization ability and parameter regularity in continuous control target spaces. Then, by introducing a quantitative model compression method that alternates between auxiliary compression networks and main networks, the main network model is adaptively compressed to the target size according to the specific storage and computing resource constraints of edge computing nodes, enabling efficient and low-latency operation of irrigation, fertilization, and plant protection regulation models on resource-constrained embedded platforms.

[0012] Step S4: Perform intelligent environmental monitoring and early warning through multi-subsystem collaboration.

[0013] At the application layer, multiple intelligent monitoring subsystems run in parallel based on data processed from the edge and cloud:

[0014] Soil moisture and nutrient monitoring subsystem: Real-time monitoring of indicators such as temperature, humidity, pH value, nitrogen, phosphorus and potassium content in different soil layers, and early warning through built-in crop growth fertilizer and water requirement model;

[0015] Farmland microclimate monitoring subsystem: Real-time monitoring of air temperature, humidity, light intensity, wind speed, wind direction, and rainfall; combined with dynamic meteorological early warning models to predict severe weather such as frost, hot and dry winds, and rainstorms.

[0016] Crop growth and plant protection intelligent monitoring subsystem: Through fixed and mobile cameras deployed in the field, image recognition algorithms are used to identify, analyze and issue alarms in real time the crop growth, leaf color, disease and pest symptoms, and weed distribution.

[0017] The early warning information and key monitoring data generated by each subsystem are displayed in an integrated visual manner through the Web / App interface and pushed to the management terminal;

[0018] Step S5: Complete closed-loop adaptive control based on trusted perception and edge intelligence;

[0019] Edge computing nodes receive real-time monitoring data and early warning information from step S4 and input them into the compressed and optimized local control model in step S3. This model performs real-time reasoning based on the current multi-dimensional environmental status, crop growth stage, historical trends, and early warning levels to generate precise control commands. The control commands are directly sent to the corresponding intelligent execution devices to drive drip irrigation / sprinkler valves, fertilizer pumps, plant protection drones, shade nets, supplemental lighting, and other equipment to perform precise operations, achieving adaptive optimization management of environmental factors such as water, fertilizer, pesticides, light, and temperature in farmland. At the same time, the control commands and execution results are fed back to the cloud for optimization and updating of the global model, forming a complete closed loop of perception, analysis, decision-making, control, and optimization.

[0020] An intelligent environmental monitoring and adaptive control system for agricultural fields based on multi-source Internet of Things (IoT) sensing, the system comprising:

[0021] The trusted sensing and data encapsulation module, deployed in farmland, includes multiple heterogeneous IoT sensing nodes and at least one redundant sensing unit, used to collect multi-dimensional farmland environmental parameters; the sensing nodes are equipped with or connected to a microprocessor, used to run the reliable data assurance model based on the trust evaluation of sensing sources, to filter and fuse multi-channel sensor data, and to encapsulate the sensing data and device status in a unified manner according to a preset JSON data format, generating a standardized environmental sensing data package;

[0022] The edge-cloud collaborative computing module includes edge computing nodes deployed in the field and a remote cloud server. The edge computing nodes are connected to the trusted perception and data encapsulation module and the cloud server through a communication gateway containing LoRa, ZigBee, and cellular network modules. They are used to receive standardized environmental perception data packets, perform local data preprocessing, lightweight AI analysis, and real-time decision-making, and synchronize data with the cloud server through the redundant communication mechanism. The cloud server is used to receive and store data from the edge computing nodes, perform massive data mining, complex model training, and global strategy optimization, and distribute the optimized model or strategy to the edge computing nodes.

[0023] The intelligent monitoring and early warning module runs at the application layer and includes a soil moisture and nutrient monitoring subunit, a farmland microclimate monitoring subunit, and a crop growth and plant protection intelligent monitoring subunit. Each subunit runs in parallel and performs analysis and early warning tasks in its respective field based on the data processed by the edge-cloud collaborative computing module. The early warning information and key data are displayed and pushed through a visual interface.

[0024] The closed-loop control and execution module includes a compressed and optimized local control model deployed in the edge computing node, and various intelligent execution devices distributed in the farmland. The local control model receives the output of the intelligent monitoring and early warning module, performs real-time inference, and generates control instructions. The control instructions are sent to the corresponding intelligent execution devices through the communication gateway to drive them to complete precise operations. The operation results of the intelligent execution devices are uploaded to the cloud server via the edge computing node as feedback information for model optimization.

[0025] The trusted perception and data encapsulation module, the edge-cloud collaborative computing module, the intelligent monitoring and early warning module, and the closed-loop control execution module are connected in sequence to form a closed-loop adaptive control system for perception, analysis, decision-making, control, and optimization.

[0026] A computing device, comprising:

[0027] One or more processors;

[0028] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0029] A computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0030] The above-described solution of the present invention has at least the following beneficial effects:

[0031] By deploying redundant sensing units and running a reliable data assurance model based on sensing source trust evaluation, the direct trust value, energy trust value, and peer-recommended trust value of nodes are comprehensively calculated. Based on these, multi-channel sensor data is filtered and weighted for fusion. This method effectively identifies and eliminates the impact of faulty, low-power, or data-anomaly nodes, outputting highly reliable environmental parameters and providing a solid and trustworthy data foundation for subsequent intelligent decision-making.

[0032] This architecture achieves synergy between low-latency real-time control and cloud-based global optimization: It constructs a two-layer data processing architecture with edge-cloud collaboration, employing a redundant communication mechanism combining LoRa, ZigBee, and cellular networks at the transmission layer. High real-time tasks (such as real-time image recognition and emergency alarms) and lightweight decision-making are completed at the edge, ensuring immediate control; massive data storage, complex model training, and global policy optimization are performed in the cloud. This architecture not only guarantees the need for real-time response in the field but also leverages the powerful computing capabilities of the cloud, ensuring the reliability of data transmission links through redundant communication.

[0033] This approach solves the deployment challenge of complex intelligent models on resource-constrained edge nodes: Through an edge-structure adaptive active intelligent control algorithm integration framework, it first employs a deep balanced training strategy (label distribution resampling and feature space smoothing) to enhance the model's generalization ability in continuous control scenarios. Then, by introducing an auxiliary compression network and alternating training with the main network, it adaptively compresses the model to the target size based on the actual storage and computing resource constraints of the edge nodes. This enables advanced deep learning control models to run efficiently and with low latency on embedded edge devices, realizing the practical application of edge intelligence.

[0034] A comprehensive and integrated intelligent monitoring and early warning system has been established: by operating multiple subsystems in parallel, including soil moisture and nutrient monitoring, farmland microclimate monitoring, and intelligent monitoring of crop growth and plant protection, the system provides full-coverage monitoring of the farmland environment, from underground to aerial, and from physical parameters to visual information. Each subsystem uses professional models for analysis and early warning, and provides integrated visualization and information push through a unified Web / App interface, helping managers to comprehensively, intuitively, and promptly grasp the field conditions.

[0035] The system achieves closed-loop adaptive regulation encompassing perception, analysis, decision-making, control, and optimization. Based on highly reliable sensing data, it utilizes lightweight intelligent models deployed at the edge for real-time inference, generating precise regulation commands that directly drive irrigation, fertilization, and plant protection equipment. Simultaneously, the regulation commands and execution results are fed back to the cloud for optimization and updating of the global model. This closed loop enables adaptive, precise, and collaborative management of multiple environmental factors in farmland, including water, fertilizer, pesticides, light, and temperature, significantly improving the intelligence level of agricultural production and resource utilization efficiency. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating an intelligent environmental monitoring and adaptive control method for agricultural fields based on multi-source Internet of Things sensing, provided by an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram of an intelligent environmental monitoring and adaptive control system for agricultural fields based on multi-source Internet of Things sensing, provided by an embodiment of the present invention. Detailed Implementation

[0038] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0039] like Figure 1 As shown, embodiments of the present invention propose a method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things (IoT) sensing. This method includes the following steps:

[0040] Step S1: Construct a reliable perception and unified description layer for multi-dimensional farmland environmental parameters:

[0041] The system deploys heterogeneous IoT sensing nodes, including soil moisture sensors, crop nutrient sensors, meteorological sensors, pest and disease monitoring equipment, and high-definition cameras, to collect comprehensive, multi-parameter data on farmland soil, crop canopy, near-ground atmosphere, and field conditions. For at least one key environmental parameter, at least three identical sensors are deployed to form a redundant sensing unit, and a reliable data assurance model based on sensing source trust evaluation is run. This model calculates the direct trust value, energy trust value, and peer-recommended trust value of each sensing node, and weights them to obtain a comprehensive trust value. Based on the comprehensive trust value, multi-channel sensor data is filtered and fused to output highly reliable environmental parameters. A standardized environmental sensing data package is generated by uniformly describing and encapsulating all sensing devices and their collected real-time data, historical data, device status, and access permissions using JSON data format.

[0042] Step S2: Design a two-tier data processing architecture based on edge computing and cloud collaboration:

[0043] A system architecture consisting of a perception layer, a transmission layer, and an application layer is constructed. In the transmission layer, a redundant communication mechanism combining LoRa, ZigBee, and cellular networks is adopted to achieve low-power, highly reliable data backhaul. Edge computing nodes are deployed at the network edge near the field perception nodes. These nodes have built-in lightweight AI models and decision-making modules to receive the standardized perception data packets from step S1 and perform localized data preprocessing, caching, and lightweight analysis tasks with high real-time requirements. Simultaneously, the data is synchronously transmitted to the cloud server, where it performs in-depth mining of massive historical data, model training, and global optimization strategy calculation.

[0044] Step S3: Implement edge adaptive model compression and deployment for proactive intelligent control:

[0045] For deep learning models deployed on edge computing nodes for farmland environmental regulation decisions, an edge-structure adaptive active intelligent control algorithm integration framework is applied to optimize them. First, considering the characteristics of continuous changes in environmental parameters and unbalanced distribution of regulation decision data in farmland regulation scenarios, a deep balanced training strategy is adopted, including label distribution resampling and feature space smoothing, to enhance the model's generalization ability and parameter regularity in continuous control target spaces. Then, by introducing a quantitative model compression method that alternates between auxiliary compression networks and main networks, the main network model is adaptively compressed to the target size according to the specific storage and computing resource constraints of edge computing nodes, enabling efficient and low-latency operation of irrigation, fertilization, and plant protection regulation models on resource-constrained embedded platforms.

[0046] Step S4: Perform intelligent environmental monitoring and early warning through multi-subsystem collaboration.

[0047] At the application layer, multiple intelligent monitoring subsystems run in parallel based on data processed from the edge and cloud:

[0048] Soil moisture and nutrient monitoring subsystem: Real-time monitoring of indicators such as temperature, humidity, pH value, nitrogen, phosphorus and potassium content in different soil layers, and early warning through built-in crop growth fertilizer and water requirement model;

[0049] Farmland microclimate monitoring subsystem: Real-time monitoring of air temperature, humidity, light intensity, wind speed, wind direction, and rainfall; combined with dynamic meteorological early warning models to predict severe weather such as frost, hot and dry winds, and rainstorms.

[0050] Crop growth and plant protection intelligent monitoring subsystem: Through fixed and mobile cameras deployed in the field, image recognition algorithms are used to identify, analyze and issue alarms in real time the crop growth, leaf color, disease and pest symptoms, and weed distribution.

[0051] The early warning information and key monitoring data generated by each subsystem are displayed in an integrated visual manner through the Web / App interface and pushed to the management terminal;

[0052] Step S5: Complete closed-loop adaptive control based on trusted perception and edge intelligence;

[0053] Edge computing nodes receive real-time monitoring data and early warning information from step S4 and input them into the compressed and optimized local control model in step S3. This model performs real-time reasoning based on the current multi-dimensional environmental status, crop growth stage, historical trends, and early warning levels to generate precise control commands. The control commands are directly sent to the corresponding intelligent execution devices to drive drip irrigation / sprinkler valves, fertilizer pumps, plant protection drones, shade nets, supplemental lighting, and other equipment to perform precise operations, achieving adaptive optimization management of environmental factors such as water, fertilizer, pesticides, light, and temperature in farmland. At the same time, the control commands and execution results are fed back to the cloud for optimization and updating of the global model, forming a complete closed loop of perception, analysis, decision-making, control, and optimization.

[0054] In this embodiment of the invention, redundant sensing units are deployed and a reliable data assurance model based on sensing source trust evaluation is run. The direct trust value, energy trust value, and peer-recommended trust value of nodes are comprehensively calculated, and multi-channel sensor data are then filtered and weighted for fusion. This method effectively identifies and eliminates the impact of faulty, low-power, or data-abnormal nodes, outputting highly reliable environmental parameters and providing a solid and trustworthy data foundation for subsequent intelligent decision-making.

[0055] This architecture achieves synergy between low-latency real-time control and cloud-based global optimization: It constructs a two-layer data processing architecture with edge-cloud collaboration, employing a redundant communication mechanism combining LoRa, ZigBee, and cellular networks at the transmission layer. High real-time tasks (such as real-time image recognition and emergency alarms) and lightweight decision-making are completed at the edge, ensuring immediate control; massive data storage, complex model training, and global policy optimization are performed in the cloud. This architecture not only guarantees the need for real-time response in the field but also leverages the powerful computing capabilities of the cloud, ensuring the reliability of data transmission links through redundant communication.

[0056] This approach solves the deployment challenge of complex intelligent models on resource-constrained edge nodes: Through an edge-structure adaptive active intelligent control algorithm integration framework, it first employs a deep balanced training strategy (label distribution resampling and feature space smoothing) to enhance the model's generalization ability in continuous control scenarios. Then, by introducing an auxiliary compression network and alternating training with the main network, it adaptively compresses the model to the target size based on the actual storage and computing resource constraints of the edge nodes. This enables advanced deep learning control models to run efficiently and with low latency on embedded edge devices, realizing the practical application of edge intelligence.

[0057] A comprehensive and integrated intelligent monitoring and early warning system has been established: by operating multiple subsystems in parallel, including soil moisture and nutrient monitoring, farmland microclimate monitoring, and intelligent monitoring of crop growth and plant protection, the system provides full-coverage monitoring of the farmland environment, from underground to aerial, and from physical parameters to visual information. Each subsystem uses professional models for analysis and early warning, and provides integrated visualization and information push through a unified Web / App interface, helping managers to comprehensively, intuitively, and promptly grasp the field conditions.

[0058] The system achieves closed-loop adaptive regulation encompassing perception, analysis, decision-making, control, and optimization. Based on highly reliable sensing data, it utilizes lightweight intelligent models deployed at the edge for real-time inference, generating precise regulation commands that directly drive irrigation, fertilization, and plant protection equipment. Simultaneously, the regulation commands and execution results are fed back to the cloud for optimization and updating of the global model. This closed loop enables adaptive, precise, and collaborative management of multiple environmental factors in farmland, including water, fertilizer, pesticides, light, and temperature, significantly improving the intelligence level of agricultural production and resource utilization efficiency.

[0059] In a preferred embodiment of the present invention, step 1 above may include:

[0060] Step 1.1: Deploy heterogeneous sensing nodes and redundant sensing units. In the farmland area to be monitored, based on the terrain, crop planting layout, and monitoring needs, plan and deploy a heterogeneous IoT sensing node network including soil moisture sensors, crop nutrient sensors, meteorological sensors, pest spore traps, and high-definition PTZ cameras. For key environmental parameters such as soil temperature and humidity, and air temperature and humidity, deploy at least three sensors of the same type within each standard monitoring unit (e.g., per acre) to form a redundant sensing unit, providing a basis for data backup and cross-validation. Step 1.2: Run a reliable data assurance model to calculate the comprehensive trust value of the nodes. Each sensing node within the redundant sensing unit periodically (e.g., every 5 minutes) collects and reports environmental data. The system runs a reliable data assurance model based on sensing source trust evaluation to calculate a comprehensive trust value T for each node. total ;

[0061] Calculate the direct trust value T direct : The statistical node reports N data points {x1, x2, ..., x} within the most recent time window (e.g., 1 hour). N} Calculate the median M of all node data for the redundant unit in the same time period. Then, the direct trust value of this node is defined as the negative correlation function of the mean absolute deviation:

[0062]

[0063] Meanwhile, if a node experiences a data reporting interruption within W, a penalty discount will be applied based on the duration of the interruption.

[0064] Calculate the energy trust value T energy Let the current remaining power of the node be E. current The preset alarm charge level is E. threshold The recent average energy consumption per unit time is P avg The energy trust value can then be calculated as follows:

[0065]

[0066] Here, k is an adjustment coefficient that reflects the degree to which energy consumption affects trust. This value ensures that the trust level of nodes with high energy consumption or low power consumption is reduced.

[0067] Calculate the peer recommendation trust value T recommend The system maintains a network topology of sensing nodes. For a target node i, it finds its K nearest neighbors that are physically closest or monitor similar parameters. Each neighbor node j is determined based on its spatiotemporal correlation with node i in historical data (e.g., Pearson correlation coefficient ρ). ij Based on the data quality and its own characteristics, a recommendation score R is given. ji The peer recommendation trust value of node i is the weighted average of all its recommendation scores:

[0068]

[0069] Calculate the overall trust value: Based on the application scenario's emphasis on data accuracy, node lifespan, and network consistency, set adjustable weights α, β, and γ (satisfying α + β + γ = 1), and calculate the final overall trust value:

[0070] Step 1.3, Data filtering and fusion based on trust values: For redundant sensing units, the T values ​​of all nodes are... total Compare with a preset threshold; remove T. total <θ low Node data is considered untrusted data; for T total ≥θ high The node data is weighted and averaged according to its trust value, and the result is used as the final high-reliability environmental parameter value X output by this unit. fused :

[0071] Step 1.4, Unified Resource Description and Encapsulation: Each sensing device and its data are uniformly described using the standard JSON-LD (JSON for Linked Data) format. Each data packet contains:

[0072] Device description block: Device ID, type, geographic location, manufacturer, status (normal / alarm), access permission list;

[0073] Data description block: timestamp, list of collected parameters (each parameter includes name, value, unit, and confidence score T) total );

[0074] Historical data index: URL links pointing to historical data segments stored in the cloud;

[0075] This JSON data packet serves as a standardized environment-aware data packet, which can be used by subsequent layers of the system.

[0076] In a preferred embodiment of the present invention, step 2 above may include:

[0077] Step 2.1: Construct a sensing network and a self-organizing network. Configure ZigBee communication modules on heterogeneous sensing nodes (such as sensors and cameras) in the field. Also configure ZigBee modules on each edge computing node (such as a ruggedized gateway device deployed on the field ridge) to serve as the aggregation point for sensing nodes in the area. The sensing nodes and edge computing nodes automatically form a ZigBee self-organizing network to aggregate the collected raw data to the edge computing nodes in a low-power, multi-hop manner.

[0078] Step 2.2, Lightweight processing and caching at the edge: After receiving the sensing data from the ZigBee network through the edge computing node, localized processing is performed, including preprocessing: filtering invalid values, simple threshold alarms (such as instantaneous excessive temperature), and local data fusion based on the method described in Example 1 (if computing resources allow).

[0079] By running lightweight AI models, we can perform analysis tasks with extremely high real-time requirements, such as calling lightweight image recognition models to perform real-time analysis of camera video streams and make preliminary identification of pests and diseases.

[0080] The preprocessed data and lightweight analysis results are stored in a time series on a local SD card or Flash memory to form a recent (e.g., 24-hour) historical dataset.

[0081] Step 2.3, using a redundant communication mechanism to transmit data back to the cloud, includes:

[0082] The edge computing node is configured with dual-path wide area network communication modules: a LoRa module and a 4G / 5G cellular network module;

[0083] Under normal circumstances, edge nodes periodically (e.g., every 15 minutes) upload cached standardized data packets to the remote LoRa gateway via the LoRa network, and then the gateway forwards them to the cloud server. This mode consumes very little power.

[0084] Edge nodes continuously monitor the signal strength (RSSI) and round-trip time (RTT) of the previous packet data in the LoRa link. When the RSSI is less than the threshold or the RTT is greater than the tolerance limit, the LoRa link quality is deemed substandard.

[0085] If the LoRa link fails, the system automatically switches to cellular network for data transmission. In cellular network mode, it can transmit larger amounts of data and higher-resolution image or video clips. Once the LoRa link quality is restored, it can switch back to low-power mode.

[0086] Step 2.4: Through cloud-based deep analysis and model training, the cloud server receives data from all edge nodes and performs global processing.

[0087] The data is stored in a spatiotemporal database, and big data analytics techniques (such as clustering and time-series prediction) are used to uncover patterns of regional environmental change and crop growth correlation models.

[0088] Using massive amounts of historical data, we train and optimize complex AI models (such as large-scale pest and disease identification models, global yield prediction models, and optimal control strategy models).

[0089] The trained new model or optimized control strategy parameters are distributed in batches to the corresponding edge computing nodes through the cellular network to achieve cloud-based collaborative evolution of the model and strategy.

[0090] In a preferred embodiment of the present invention, step 3 above may include:

[0091] Step 3.1: Enhance the model's generalization ability through deep balanced training. Prepare the dataset D={(s) in the cloud for training farmland regulation models (such as deciding when to irrigate and how much to fertilize). i a i )}, where s i For the environmental state, a i For continuous control commands (such as irrigation volume 0-100 cubic meters);

[0092] By adaptively discretizing the continuous action space A, assuming the action value range is [a min ,a max First, the distribution of all action values ​​in the training set is statistically analyzed. Finer boundaries are set in densely distributed intervals, and coarser boundaries are set in sparsely distributed intervals, forming B non-uniform partitions (bins): A1, A2, ..., A... B ;

[0093] By statistically analyzing each partition A b The number of samples N bFor partitions with fewer samples than the average number of samples N (i.e., partitions with fewer samples), a generative adversarial network (GAN) is used to learn the joint distribution of state and action pairs (s,a) within the partition, and new samples (s′,a′) are synthesized and added to the training set until the sample size of each partition is relatively balanced.

[0094] During model training, a feature contrast loss L is added to the standard loss function (such as mean squared error). con For a sample pair (s) i s j If their action labels are similar (i.e., belong to the same or adjacent partitions), then they are forced to be in a certain hidden layer f( of the model). The characteristic expressions are also similar:

[0095]

[0096] Where I( ) is the indicator function, and δ is the action similarity threshold. This loss term smooths the feature space and improves the stability of the model in predicting continuous actions.

[0097] Step 3.2: Construct and jointly train the main network and the auxiliary network by constructing a main control network M with more parameters (such as a multi-layer fully connected network) and an auxiliary network A with a more compact structure and fewer layers;

[0098] During the training phase, the feature map F of a certain layer in the middle of the main network is... M As input to the auxiliary network, the total loss function is:

[0099]

[0100] Where y is the actual control command. and These are the predicted outputs of the main network and the auxiliary network, respectively, L task λ is the task loss (such as the smoothed MSE), and λ is the balancing weight. In this way, the auxiliary network learns "knowledge" from the main network.

[0101] Step 3.3: When it is necessary to deploy the model to a specific edge computing node, the system first evaluates the real-time available resources (available memory Mem) of that node. avl CPU / Peak FLOPS max );

[0102] Determine the size of the target model based on resource constraints. target and computational load (FLOPS) target ;

[0103] If the main network M itself has redundancy, a compression strategy is selected. If the importance of neurons is evaluated, the weights of M are pruned to remove unimportant connections until the Size is satisfied. target Require;

[0104] If the extreme resource constraints cannot be met after pruning, then all or part of the layers of the main network M will be directly replaced with the pre-trained auxiliary network A.

[0105] For the compressed model, a short-term fine-tuning is performed on a portion of the locally cached data to recover the accuracy loss that may have been caused by compression.

[0106] Step 3.4, Lightweight Model Deployment and Update: Deploy the compressed and optimized lightweight control model to the inference engine of the edge computing node. The cloud will periodically (e.g., weekly) distribute new versions of the model or compression strategies trained on global data. The edge node can automatically perform incremental updates or replacements of the model during idle periods (e.g., at night).

[0107] In a preferred embodiment of the present invention, step 4 above may include:

[0108] Step 4.1, Soil moisture and nutrient monitoring and early warning: The soil moisture and nutrient monitoring subsystem continuously receives data from soil sensors;

[0109] Data on temperature, humidity, pH, nitrogen, phosphorus, and potassium content of each soil layer were integrated according to soil depth (e.g., 0-20cm, 20-40cm).

[0110] The subsystem has a built-in model for the water and fertilizer requirements of the currently planted crop. This model defines the appropriate water and nutrient ranges [lower limit i, upper limit i] for each soil layer at different growth stages of the crop (seedling stage, jointing stage, grain filling stage, etc.).

[0111] Real-time monitoring value v i Compare with the appropriate range defined in the model. For any parameter i, if v i <lower limit i or v i If the upper limit i is reached, an alert will be triggered. The alert level can be dynamically divided according to the degree of deviation (e.g., attention, minor, major).

[0112] Generate early warning information that includes warning parameters, degree of deviation, location (plot, depth) of occurrence, and time.

[0113] Step 4.2, Farmland Microclimate and Severe Weather Prediction: The farmland microclimate monitoring subsystem integrates meteorological sensor data;

[0114] Continuously monitor air temperature and humidity, light intensity, wind speed and direction, and rainfall, and calculate derived indicators such as dew point temperature and perceived temperature.

[0115] It has a built-in short-term early warning model trained based on local historical meteorological data. For example, by combining the current temperature and humidity, wind speed and nighttime radiative cooling patterns, it can predict whether frost is likely to occur in the next 3-6 hours (minimum ground temperature ≤0°C).

[0116] By calling on regional meteorological forecast data distributed from the cloud and combining it with local real-time monitoring, the system uses early warning models to predict the probability and possible intensity of severe weather events such as hot and dry winds (high temperature, low humidity, and strong winds) and regional rainstorms within the next 12-24 hours.

[0117] Generate early warning reports that include disaster type, expected time of occurrence, potential intensity, scope of impact, and recommended measures.

[0118] Step 4.3, Crop growth and plant protection monitoring based on lightweight image recognition: The intelligent monitoring subsystem for crop growth and plant protection processes video streams from cameras;

[0119] Deploy a lightweight convolutional neural network model based on improved YOLO-v5s within edge computing nodes or high-performance camera terminals. This model is pre-trained in the cloud using massive amounts of labeled images of pests, weeds, and crop growth, and is periodically updated through incremental learning.

[0120] Real-time inference is performed on the input video frames. The model outputs the bounding box (x) of the identified target. center ,y center (w,h), category (e.g., "rice blast", "barnyard grass", "nitrogen deficiency yellowing") and confidence score;

[0121] For identified pests, diseases, or weeds, a field severity index (Severity) is estimated based on their pixel area ratio in the image, distribution density, and the severity level of the category itself.

[0122]

[0123] Where Atarget is the area or number of target areas, Atotal is the total area or total amount, and wrisk is the risk weight coefficient;

[0124] An alarm is generated when a target is identified and the confidence level is higher than a threshold, or the severity index exceeds a preset level. The alarm information includes the target category, confidence level, location (image coordinates, which can be mapped to the approximate location in the field), severity level, and a snapshot of the scene.

[0125] Step 4.4, Integrated Visualization and Information Push: The monitoring data, early warning information, and identification results (such as pest and disease annotation maps) of the above subsystems are pushed to the central processing module of the application layer in real time via WebSocket or MQTT protocol;

[0126] All farmland plots are overlaid on the GIS map on the web or mobile app. Clicking on any plot allows you to view its real-time soil data, microclimate data, crop growth assessment, and all current activity alerts in a single interface, either in layers or columns.

[0127] For high-level (severe) warnings, in addition to highlighting them on the visual interface, the system also sends key alarm information (such as "Severe rice blast warning has occurred in plot A3, please handle it in time!") to the mobile terminals of farm managers through App push, SMS and other means.

[0128] In a preferred embodiment of the present invention, step 5 above may include:

[0129] Step 5.1, constructing the reinforcement learning regulation model includes:

[0130] State Space: The real-time monitoring and early warning status after the fusion of all subsystems in Example 4 is defined as the state space. It includes: soil parameter vectors, meteorological parameter vectors, crop growth score, pest / weed alarm vectors (including category, severity, and location codes), crop growth period codes, and historical parameter trends (such as the soil moisture change rate over the past 24 hours).

[0131] Action space: Defined as the combination of instructions for all controllable execution devices. For example, an action vector could be: [drip irrigation valve on / off state (0 / 1), fertilizer pump flow rate (L / h), coordinates of point 1 on the agricultural drone's operating path, shade net opening / closing degree (%), supplemental lighting brightness (%)]. The action space can be discrete (selected from preset combinations) or continuous (directly outputting control quantities).

[0132] In the cloud, using historical data or constructing a farmland simulation environment, an initial global control network is trained using deep reinforcement learning algorithms (such as DDPG or PPO), with crop growth indicators (such as estimated yield and quality) and resource consumption (water, fertilizer, and electricity) as reward functions.

[0133] Step 5.2, Edge Model Inference and Instruction Generation includes:

[0134] Input current state: Edge computing nodes collect data from various subsystems in real time and integrate it into a standardized state vector s. t ;

[0135] Model inference: s t The input is fed into a local control model deployed at the edge after compression and optimization as described in Example 3. This model (such as a deep Q-network or policy network) outputs the current optimal action vector a. t ;

[0136] Security and rule verification: Before outputting, the system will verify a against built-in agricultural rules and security constraints. t Perform verification. For example, prohibit irrigation during rainfall forecast periods, or set a safe upper limit for fertilizer concentration. If a t If a rule is violated, the action will be corrected to the closest action that conforms to the rule.

[0137] Step 5.3, instruction distribution and device collaborative execution includes:

[0138] Protocol parsing: The protocol conversion module within the edge computing node parses the abstract action vector 'at' into a set of underlying control instructions that conform to the communication protocols of the specific execution device (such as Modbus, CAN, MQTT);

[0139] Redundant communication delivery: For fixed equipment such as irrigation valves, fertilizer pumps, and shade net controllers, control commands are delivered through a low-latency, highly reliable ZigBee self-organizing network;

[0140] For mobile agricultural drones, control commands (such as operation paths and spraying switch commands) are sent to the drone control station via cellular networks (4G / 5G).

[0141] Collaborative operations: Instructions can include collaborative logic between devices. For example, first activate the shade net to cool down, then activate the sprinkler system 10 minutes later to increase humidity; or control a drone to perform variable-rate spraying within a specified coordinate area.

[0142] Step 5.4, closed-loop feedback and model optimization include:

[0143] Execution feedback: After completing the instructions, the intelligent execution device will feed back the execution results (such as "valve has been opened", "actual amount of fertilizer applied", "drone operation completed") to the edge computing node.

[0144] Data closure: Edge nodes will record the current "state s" t -command a t -Execution result-Subsequent statuses t "+1" is stored as a complete decision-execution-feedback tuple and uploaded to the cloud.

[0145] Model iteration: The cloud server collects feedback tuples from all farmlands to form new training data. This new data is periodically (e.g., monthly) used to retrain or fine-tune the global control model, and the optimized model parameters or structure are then updated and distributed to edge nodes. This forms a continuous improvement loop of perception, analysis, decision-making, control, and optimization.

[0146] like Figure 2As shown, embodiments of the present invention also provide an intelligent environmental monitoring and adaptive control system for agricultural fields based on multi-source Internet of Things sensing, characterized in that the system includes:

[0147] The trusted sensing and data encapsulation module, deployed in farmland, includes multiple heterogeneous IoT sensing nodes and at least one redundant sensing unit, used to collect multi-dimensional farmland environmental parameters; the sensing nodes are equipped with or connected to a microprocessor, used to run the reliable data assurance model based on the trust evaluation of sensing sources, to filter and fuse multi-channel sensor data, and to encapsulate the sensing data and device status in a unified manner according to a preset JSON data format, generating a standardized environmental sensing data package;

[0148] The edge-cloud collaborative computing module includes edge computing nodes deployed in the field and a remote cloud server. The edge computing nodes are connected to the trusted perception and data encapsulation module and the cloud server through a communication gateway containing LoRa, ZigBee, and cellular network modules. They are used to receive standardized environmental perception data packets, perform local data preprocessing, lightweight AI analysis, and real-time decision-making, and synchronize data with the cloud server through the redundant communication mechanism. The cloud server is used to receive and store data from the edge computing nodes, perform massive data mining, complex model training, and global strategy optimization, and distribute the optimized model or strategy to the edge computing nodes.

[0149] The intelligent monitoring and early warning module runs at the application layer and includes a soil moisture and nutrient monitoring subunit, a farmland microclimate monitoring subunit, and a crop growth and plant protection intelligent monitoring subunit. Each subunit runs in parallel and performs analysis and early warning tasks in its respective field based on the data processed by the edge-cloud collaborative computing module. The early warning information and key data are displayed and pushed through a visual interface.

[0150] The closed-loop control and execution module includes a compressed and optimized local control model deployed in the edge computing node, and various intelligent execution devices distributed in the farmland. The local control model receives the output of the intelligent monitoring and early warning module, performs real-time inference, and generates control instructions. The control instructions are sent to the corresponding intelligent execution devices through the communication gateway to drive them to complete precise operations. The operation results of the intelligent execution devices are uploaded to the cloud server via the edge computing node as feedback information for model optimization.

[0151] The trusted perception and data encapsulation module, the edge-cloud collaborative computing module, the intelligent monitoring and early warning module, and the closed-loop control execution module are connected in sequence to form a closed-loop adaptive control system for perception, analysis, decision-making, control, and optimization.

[0152] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0153] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0154] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0155] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things (IoT) sensing, characterized in that, The method includes the following steps: Step S1: Construct a reliable perception and unified description layer for multi-dimensional farmland environmental parameters: The system deploys heterogeneous IoT sensing nodes, including soil moisture sensors, crop nutrient sensors, meteorological sensors, pest and disease monitoring equipment, and high-definition cameras, to collect comprehensive, multi-parameter data on farmland soil, crop canopy, near-ground atmosphere, and field conditions. For at least one key environmental parameter, at least three identical sensors are deployed to form a redundant sensing unit, and a reliable data assurance model based on sensing source trust evaluation is run. This model calculates the direct trust value, energy trust value, and peer-recommended trust value of each sensing node, and weights them to obtain a comprehensive trust value. Based on the comprehensive trust value, multi-channel sensor data is filtered and fused to output highly reliable environmental parameters. A standardized environmental sensing data package is generated by uniformly describing and encapsulating all sensing devices and their collected real-time data, historical data, device status, and access permissions using JSON data format. Step S2: Design a two-tier data processing architecture based on edge computing and cloud collaboration: A system architecture consisting of a perception layer, a transmission layer, and an application layer is constructed. In the transmission layer, a redundant communication mechanism combining LoRa, ZigBee, and cellular networks is adopted to achieve low-power, highly reliable data backhaul. Edge computing nodes are deployed at the network edge near the field perception nodes. These nodes have built-in lightweight AI models and decision-making modules to receive the standardized perception data packets from step S1 and perform localized data preprocessing, caching, and lightweight analysis tasks with high real-time requirements. Simultaneously, the data is synchronously transmitted to the cloud server, where it performs in-depth mining of massive historical data, model training, and global optimization strategy calculation. Step S3: Implement edge adaptive model compression and deployment for proactive intelligent control: For deep learning models deployed on edge computing nodes for farmland environmental regulation decisions, an edge-structure adaptive active intelligent control algorithm integration framework is applied to optimize them. First, considering the characteristics of continuous changes in environmental parameters and unbalanced distribution of regulation decision data in farmland regulation scenarios, a deep balanced training strategy is adopted, including label distribution resampling and feature space smoothing, to enhance the model's generalization ability and parameter regularity in continuous control target spaces. Then, by introducing a quantitative model compression method that alternates between auxiliary compression networks and main networks, the main network model is adaptively compressed to the target size according to the specific storage and computing resource constraints of edge computing nodes, enabling efficient and low-latency operation of irrigation, fertilization, and plant protection regulation models on resource-constrained embedded platforms. Step S4: Perform intelligent environmental monitoring and early warning through multi-subsystem collaboration. At the application layer, multiple intelligent monitoring subsystems run in parallel based on data processed from the edge and cloud: Soil moisture and nutrient monitoring subsystem: Real-time monitoring of indicators such as temperature, humidity, pH value, nitrogen, phosphorus and potassium content in different soil layers, and early warning through built-in crop growth fertilizer and water requirement model; Farmland microclimate monitoring subsystem: Real-time monitoring of air temperature, humidity, light intensity, wind speed, wind direction, and rainfall; combined with dynamic meteorological early warning models to predict severe weather such as frost, hot and dry winds, and rainstorms; Crop growth and plant protection intelligent monitoring subsystem: Through fixed and mobile cameras deployed in the field, image recognition algorithms are used to identify, analyze and issue alarms in real time the crop growth, leaf color, disease and pest symptoms and weed distribution. The early warning information and key monitoring data generated by each subsystem are displayed in an integrated visual manner through the Web / App interface and pushed to the management terminal; Step S5: Complete closed-loop adaptive control based on trusted perception and edge intelligence; Edge computing nodes receive real-time monitoring data and early warning information from step S4 and input them into the compressed and optimized local control model in step S3. This model performs real-time reasoning based on the current multi-dimensional environmental status, crop growth stage, historical trends, and early warning levels to generate precise control commands. The control commands are directly sent to the corresponding intelligent execution devices to drive drip irrigation / sprinkler valves, fertilizer pumps, plant protection drones, shade nets, supplemental lighting, and other equipment to perform precise operations, achieving adaptive optimization management of environmental factors such as water, fertilizer, pesticides, light, and temperature in farmland. At the same time, the control commands and execution results are fed back to the cloud for optimization and updating of the global model, forming a complete closed loop of perception, analysis, decision-making, control, and optimization.

2. The method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things sensing according to claim 1, characterized in that, In step S1, the reliable data assurance model based on perception source trust evaluation specifically includes: The direct trust value is determined based on the average deviation between the data reported by the same type of sensor nodes within a preset time window and the bit value in the data of the redundant sensing unit, as well as a function of the continuity of its historical data. The energy trust value is determined based on the ratio of the current remaining power of the sensor node to a preset power threshold, and the negative correlation function of the node's recent average energy consumption per unit time. The peer recommendation trust value is an indirect evaluation score given by other sensor nodes in the perception network that are geographically adjacent or monitor the same parameters, based on the spatiotemporal correlation of the data. The comprehensive trust value is calculated using the weighted sum formula T. total =α*T direct +β*T energy +γ*T recommend The calculation is performed, where α, β, and γ are adjustable weight coefficients, and α+β+γ=1; based on the comprehensive trust value, the multi-channel sensor data is filtered and fused, specifically: sensor data with a comprehensive trust value lower than the first threshold is removed, and sensor data with a trust value higher than the second threshold are fused by weighted average according to their trust value weights, and the highly reliable environmental parameters are output.

3. The method for intelligent monitoring and adaptive control of farmland environment based on multi-source Internet of Things sensing according to claim 2, characterized in that, In step S2, the redundant communication mechanism specifically refers to: Each edge computing node is configured with a LoRa module, a ZigBee module, and a cellular network module; the edge computing node forms an ad-hoc network with the heterogeneous IoT sensing nodes through the ZigBee network to aggregate sensing data; When transmitting data to the cloud server, LoRa network is used first for low-power data transmission; when the LoRa network signal strength is detected to be lower than the preset value or the transmission delay exceeds the tolerance range, the system automatically switches to cellular network for data backhaul to ensure the reliability of the data transmission link.

4. The method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things sensing according to claim 3, characterized in that, In step S3, the edge structure adaptive active intelligent control algorithm integration framework specifically includes: In the deep balanced training strategy, the label distribution resampling is to discretize the continuous control target space by using an adaptive bounding box algorithm for continuously changing control command values, and to perform oversampling of synthetic samples based on generative adversarial networks for control command partitions with a small number of samples. The feature space smoothing is achieved by adding a feature contrast loss term to the model training loss function, which forces samples with similar input environment state features to have similar representations in the hidden layer feature space of the model, so as to smooth the decision boundary. The quantitative model compression method that introduces alternating training of the auxiliary compression network and the main network specifically includes: constructing an auxiliary network with a smaller parameter size than the main network; during the training phase, using the feature maps of the intermediate layers of the main network as inputs to the auxiliary network and the final output of the main network as a supervision signal to jointly train the main network and the auxiliary network; during the compression phase, based on the real-time available memory and computing power evaluation results of the edge computing nodes, adaptively selecting to prune some redundant layers from the main network, or partially or completely replacing them with the trained auxiliary network, to obtain a compressed and controlled model that meets resource constraints.

5. The method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things sensing according to claim 4, characterized in that, In step S4, the intelligent monitoring subsystem for crop growth and plant protection specifically includes: The image recognition algorithm adopts a lightweight convolutional neural network model based on the improved YOLO architecture. This model is pre-trained in the cloud using a large-scale labeled farmland image dataset and is incrementally learned and fine-tuned periodically using new samples uploaded by edge computing nodes. The real-time identification and analysis alarm is specifically as follows: the lightweight convolutional neural network model is deployed on field edge computing nodes or high-performance embedded modules integrated with cameras to perform real-time analysis of video streams. After identifying specific pest or disease symptoms or weeds, it not only generates alarm information, but also marks the type, confidence level, and position coordinates in the image, and estimates the range of influence or severity level.

6. The method for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things sensing according to claim 5, characterized in that, In step S5, generating precise control commands and sending them to the intelligent execution device specifically includes: The local control model in the edge computing node is built based on a reinforcement learning framework. Its state space is the multi-dimensional environmental monitoring and early warning state after the fusion of each subsystem in step S4, and its action space is the combination of control commands for each intelligent execution device (drip irrigation / sprinkler valve, fertilizer pump, plant protection drone, shade net, supplemental lighting). During model inference, an action vector is output based on the current state. This vector is then parsed by the protocol conversion module into a set of specific control instructions that conform to the communication protocols of each execution device. The control command set is issued through the redundant communication mechanism of the transmission layer. Specifically, fixed equipment such as irrigation valves and fertilizer pumps are issued through the ZigBee network, while mobile devices such as plant protection drones are issued through the cellular network, forming a precise and coordinated control operation.

7. A system for intelligent environmental monitoring and adaptive control of agricultural fields based on multi-source Internet of Things sensing as described in any one of claims 1-6, characterized in that, The system includes: The trusted sensing and data encapsulation module, deployed in farmland, includes multiple heterogeneous IoT sensing nodes and at least one redundant sensing unit, used to collect multi-dimensional farmland environmental parameters; the sensing nodes are equipped with or connected to a microprocessor, used to run the reliable data assurance model based on the trust evaluation of sensing sources, to filter and fuse multi-channel sensor data, and to encapsulate the sensing data and device status in a unified manner according to a preset JSON data format, generating a standardized environmental sensing data package; The edge-cloud collaborative computing module includes edge computing nodes deployed in the field and a remote cloud server. The edge computing nodes are connected to the trusted perception and data encapsulation module and the cloud server through a communication gateway containing LoRa, ZigBee, and cellular network modules. They are used to receive standardized environmental perception data packets, perform local data preprocessing, lightweight AI analysis, and real-time decision-making, and synchronize data with the cloud server through the redundant communication mechanism. The cloud server is used to receive and store data from the edge computing nodes, perform massive data mining, complex model training, and global strategy optimization, and distribute the optimized model or strategy to the edge computing nodes. The intelligent monitoring and early warning module runs at the application layer and includes a soil moisture and nutrient monitoring subunit, a farmland microclimate monitoring subunit, and a crop growth and plant protection intelligent monitoring subunit. Each subunit runs in parallel and performs analysis and early warning tasks in its respective field based on the data processed by the edge-cloud collaborative computing module. The early warning information and key data are displayed and pushed through a visual interface. The closed-loop control and execution module includes a compressed and optimized local control model deployed in the edge computing node, and various intelligent execution devices distributed in the farmland. The local control model receives the output of the intelligent monitoring and early warning module, performs real-time inference, and generates control commands. The control commands are sent to the corresponding intelligent execution devices through the communication gateway to drive them to complete precise operations. The operation results of the intelligent execution devices are uploaded to the cloud server via the edge computing node as feedback information for model optimization. The trusted perception and data encapsulation module, the edge-cloud collaborative computing module, the intelligent monitoring and early warning module, and the closed-loop control execution module are connected in sequence to form a closed-loop adaptive control system for perception, analysis, decision-making, control, and optimization.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images