Air-ground integrated intelligent agricultural inspection and monitoring system

Abstract

The invention relates to the technical field of intelligent agriculture and remote sensing monitoring, and particularly discloses an air-ground integrated intelligent agriculture inspection and monitoring system. The system comprises a sky end aircraft, a ground end mobile platform and a cloud server which are interconnected through a multi-stage communication module. The aircraft carries a multi-mode remote sensing acquisition module and has the takeoff and landing and ad hoc network data return capability; the ground platform is provided with various soil environment sensors and an edge calculation module, and environment information collection and preprocessing are achieved. The cloud server has the functions of data analysis, large model training and agricultural decision making. According to the system, an agricultural large model is trained through a self-built multi-modal data set, remote sensing and ground information are fused, and crop state monitoring, anomaly detection and accurate decision making are achieved. The system has the advantages of high precision, high timeliness, multi-source cooperation and the like, can be widely applied to agricultural inspection, disaster early warning and fine management scenes, and remarkably improves the intelligent level of agricultural monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart agriculture, in particular to an air-ground integrated smart agriculture inspection and monitoring system. BACKGROUND

[0002] With the increasing demand for production efficiency and precision in modern agriculture, agricultural inspection and monitoring have become an important link to ensure the efficient and stable operation of agricultural production. Traditional agricultural inspection relies on manual inspection and manual recording, which often faces problems such as untimely information collection, delayed data analysis, fixed inspection path, etc., and is difficult to meet the needs of real-time, accuracy and flexibility in modern agricultural production.

[0003] In the prior art, although some soil monitoring and remote sensing image technologies have been introduced, most agricultural monitoring systems still have many defects, such as isolated analysis of soil data and remote sensing image data, lack of effective data fusion and intelligent algorithm support, and inability to comprehensively evaluate and warn of abnormal changes in the complex environment of farmland. In addition, the existing agricultural inspection strategy still relies on fixed path inspection and cannot be optimized and adjusted according to dynamic factors such as crop health, soil moisture and climate change, resulting in waste of resources and delayed response. The architecture of the current agricultural monitoring and inspection system is mostly based on a centralized computing platform. This architecture has bandwidth bottlenecks and inference delays when facing multi-source data from ground sensors, unmanned aerial vehicles, etc., and cannot meet the needs of real-time data processing and intelligent decision-making. At the same time, the multi-modal data fusion method in the prior art is relatively simple and cannot deeply mine the potential relationship between remote sensing images and soil data, thereby affecting the accuracy and efficiency of agricultural inspection and monitoring. SUMMARY

[0004] The purpose of the present application is to provide an air-ground integrated smart agriculture inspection and monitoring system that can combine multi-modal sensor data, remote sensing images and artificial intelligence technology to achieve accurate monitoring, dynamic inspection and intelligent decision-making of the agricultural environment and crop health status.

[0005] To achieve the above purpose, the present application provides an air-ground integrated smart agriculture inspection and monitoring system, comprising:

[0006] a sky-end aircraft, a ground-end mobile platform and a cloud-end server;

[0007] The sky-end aircraft comprises:

[0008] An autonomous flight control module for controlling takeoff and landing of the sky-side aerial vehicle; a takeoff and landing docking module for docking with a ground-side mobile platform storage compartment to achieve physical parking; a multi-modal remote sensing acquisition module including a multi-spectral camera, a remote sensing imaging unit, and a spectral information acquisition unit for acquiring ground surface remote sensing information; a first communication module for transmitting collected data to the ground-side mobile platform through an ad hoc network;

[0009] The ground-side mobile platform comprises:

[0010] A sensor module including a soil humidity sensor, a soil conductivity sensor, a soil pH sensor, and a soil temperature sensor;

[0011] An aerial vehicle storage compartment for accommodating the sky-side aerial vehicle;

[0012] A storage module for storing measurement data;

[0013] An edge computing module for synchronous preprocessing of sky-side remote sensing information and ground-side environmental information;

[0014] A second communication module for transmitting data with the sky-side aerial vehicle and a cloud server;

[0015] The cloud server comprises:

[0016] A third communication module for transmitting data with the ground-side mobile platform;

[0017] A data analysis module for structuring, organizing, and analyzing received data;

[0018] An agricultural large model module for realizing cross-regional remote sensing comparison, agricultural knowledge graph construction, and anomaly detection;

[0019] A decision module for generating response decisions based on analysis results and outputting a field patrol report;

[0020] An alarm and forecast module for implementing risk alerts and agricultural early warning strategy formulation, respectively.

[0021] The agricultural large model module is trained using a self-built remote sensing and ground multi-modal fusion dataset; the dataset includes: a multi-spectral image sequence, a remote sensing reflectance map, a soil humidity and temperature time sequence, crop growth stage annotation information, and the agricultural management history record of the corresponding plot.

[0022] The above training using a self-built remote sensing and ground multi-modal fusion dataset specifically includes the following steps:

[0023] S01: Perform spatial registration and time alignment processing on the collected raw data to construct a multi-modal synchronous sample pair;

[0024] S02: Extracting spatial-temporal features using a self-attention model based on the Transformer structure;

[0025] S03: Joint optimization of classification, regression and temporal tasks based on multi-task loss functions;

[0026] S04: Use knowledge distillation techniques to encode data into rule-based knowledge vectors to guide the training of the main model;

[0027] S05: Use a contrastive learning mechanism to enhance the feature separability between different plots.

[0028] The input data for the data analysis module includes:

[0029] The remote sensing image data packet is acquired by a multimodal remote sensing acquisition module and specifically includes: a unique image identifier, a capture timestamp, GPS geographic coordinates, sensor altitude, a multi-band image matrix, and corresponding wavelength information.

[0030] The environmental parameter data of the edge computing module specifically includes: sensor point coordinates, acquisition time, soil moisture, conductivity, pH value, temperature, and corresponding data confidence fields;

[0031] The remote sensing image data packets and the environmental parameter data of the edge computing module are transmitted to the cloud server through the second communication module of the ground mobile platform.

[0032] The agricultural large-scale model module includes a distributed image stitching module and a regional statistical analysis module, wherein:

[0033] The distributed image stitching module is used to spatially align remote sensing images acquired from multiple platforms and fuse them across multiple time periods within the same block.

[0034] The regional statistical analysis module calculates the normalized vegetation index, vegetation coverage, and regional drought index based on the fusion results, and compares the indicators between regions.

[0035] The regional statistical analysis module employs a multi-scale window comparison algorithm based on regional division, including the following steps:

[0036] S01: Determine the boundaries of the comparison unit plots based on the remote sensing image segmentation results;

[0037] S02: Extract the changing trends of multi-temporal remote sensing indicators on the same block;

[0038] S03: Combine historical databases and information from surrounding plots to conduct deviation analysis and determine whether there are any abnormal growth patterns or potential disaster signals;

[0039] S04: Input the analysis results and the results of the agricultural big data model module into the decision-making module to generate multi-level field inspection strategy suggestions.

[0040] Anomaly detection is based on a support vector machine (SVM) algorithm for model training, and the training method includes:

[0041] S01: Obtain multispectral images and remote sensing reflectance data collected by the multimodal remote sensing acquisition module of the sky-end flight platform, and construct them as a first type of feature vector; obtain environmental parameters generated by the edge computing module of the ground-end mobile platform, and construct them as a second type of feature vector;

[0042] S02: Synchronize the first type of features with the second type of features in time and then concatenate them to form a fused input sample;

[0043] S03: Construct training sample labels, which are manually labeled based on historical anomalies such as disease, drought or nutrient imbalance.

[0044] S04: Use radial basis kernel functions to train support vector machine models and construct nonlinear decision boundaries to distinguish between normal and abnormal samples;

[0045] S05: Deploy the trained model to the agricultural large model module in the cloud server for online identification of abnormal states of new input samples.

[0046] A multispectral camera acquires multispectral remote sensing images of farmland crops in real time and transmits them back to the ground in real time through an ad hoc network; the ad hoc network adopts an on-demand distance vector routing protocol to achieve network topology adaptation.

[0047] After receiving the multispectral remote sensing images transmitted back by the spacecraft, the edge computing module uses the edge computing unit to perform image registration algorithms to achieve image fusion, and uses the SIFT algorithm to achieve image correction and geolocation.

[0048] Furthermore, the input of the PID control unit is bidirectionally connected to the decision module of the cloud server via an edge computing module. The edge computing module provides filtered and normalized environmental parameter data in real time, including soil moisture, pH value, temperature, and conductivity. The cloud server decision module periodically or event-triggeredly sends target control parameters, such as target irrigation water volume, flight path correction parameters, or environmental regulation thresholds. The PID control unit performs high-speed control calculations on the edge computing node, with a calculation cycle of no more than 50ms. It responds to deviations instantly through the proportional (P) term, eliminates accumulated errors through the integral (I) term, and suppresses overshoot caused by parameter mutations through the derivative (D) term. The calculation results form a control signal, which is directly sent to the actuator via the 5G communication unit of the communication module or a remote communication module.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] To address the shortcomings of existing agricultural inspection and monitoring systems, such as low air-ground coordination, limited data sources, insufficient cross-regional remote sensing information processing capabilities, and low anomaly identification accuracy, this invention achieves high-timeliness and high-precision acquisition of crop growth status and environmental parameters through multimodal data fusion between an aerial flight platform and a ground-based mobile platform. Combined with an agricultural large-scale model module on a cloud server, it enables deep structured processing and intelligent analysis of data from different sources and time scales, effectively improving the accuracy of anomaly detection and trend prediction. Simultaneously, the deployment of PID control units within edge computing nodes forms a closed-loop control feedback, enhancing the system's operational stability and execution accuracy in complex farmland environments. Therefore, this invention comprehensively overcomes the deficiencies of existing technologies in data fusion, intelligent decision-making, and stable control, thereby improving the automation and intelligence level of agricultural inspection and monitoring. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the accompanying drawings used in the following description of the embodiments or examples will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the technical solutions shown in these drawings without creative effort.

[0052] Figure 1 This is a system architecture diagram of an integrated air-ground smart agriculture inspection and monitoring system.

[0053] Figure 2 Flowchart of the system startup and data acquisition procedure for the integrated air-ground smart agriculture inspection and monitoring system;

[0054] Figure 3 Flowchart of cloud fusion and analysis procedures for an integrated air-ground smart agriculture inspection and monitoring system;

[0055] Figure 4 The flowchart of the feedback control procedure for the air-ground integrated smart agriculture inspection and monitoring system;

[0056] Figure 5 A diagram showing the relationship between soil moisture and remote sensing signals in an integrated air-ground smart agriculture inspection and monitoring system.

[0057] Figure 6 A diagram showing the relationship between soil moisture and remote sensing signals in an integrated air-ground smart agriculture inspection and monitoring system.

[0058] Figure 7 This is a clustering analysis result diagram of the data set of the air-ground integrated smart agriculture inspection and monitoring system. Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0060] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0061] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0062] In some embodiments, see Figures 1-5 and Figure 7 The integrated air-ground smart agriculture inspection and monitoring system can perform a complete inspection and disease monitoring task covering a target area of ​​farmland in a single operation. The system consists of an aerial vehicle, a ground-based mobile platform, and a cloud server. These modules operate collaboratively via self-organizing network communication and a 5G remote transmission mechanism. Remote communication between the cloud server and the ground-based mobile platform is based on a two-way data synchronization mechanism, employing a hybrid architecture of Transmission Control Protocol (TCP) and Message Queuing Telemetry Transport Protocol (MQTT). TCP is used for high-throughput batch data transmission of remote sensing images and sensors, while MQTT is used for low-latency alarm, control command, and parameter synchronization. Data transmission is encrypted using TLS 1.3 and incorporates an adaptive retransmission mechanism to ensure data integrity and timeliness in environments with network jitter or weak network conditions.

[0063] In some embodiments, at the start of the mission, the ground-based mobile platform is activated, and the aircraft storage compartment is launched. The airborne aircraft within the storage compartment is a quadcopter ducted structure connected to an autonomous flight control module for receiving takeoff commands from the ground. After takeoff, the aircraft plans its path using a preset route and the GNSS module, and the multispectral camera in the multimodal remote sensing acquisition module acquires remote sensing image data of the land parcel. Each frame of image data includes a unique image identifier, a capture timestamp, aircraft coordinates, camera attitude parameters, a band image matrix, and corresponding wavelength information.

[0064] The collected data is transmitted in real time to the second ground-based communication module via the communication module. This module includes a 5G communication unit and a remote communication module. The image data is first sent to the edge computing module for preprocessing, performing the following operations: radiometric and geometric correction; image registration and multi-temporal stitching; feature point extraction for geolocation based on the SIFT algorithm; extraction of common vegetation indices and construction of a candidate feature set. The key steps of the SIFT algorithm specifically include: S01: Scale-space extremum detection, which involves convolving the image I(x,y) at different scales to construct the scale space L(x,y,σ)=G(x,y,σ)*I(x,y);

[0065] Where G(x,y,σ) is a Gaussian function, σ is a scale parameter, and * represents a convolution operation. The extreme points in the space L(x,y,σ) at different scales are used to find potential feature points.

[0066] S02: Gaussian difference calculation, using Gaussian functions G(x,y,σ) at different scales to perform image convolution and calculate the difference between images at different scales;

[0067] S03: Feature descriptor generation. Local gradient calculation is performed on the region R(x,y) surrounding each feature point. This is achieved by calculating the gradient value within each region. And generate descriptor d i ={d1,d2,...,d n}, where d i Descriptors for points, and encode features based on gradient histograms.

[0068] The common vegetation index datasets in this embodiment are shown in the table below.

[0069] Table 1 Vegetation Index Dataset

[0070]

[0071]

[0072] In the table, R NIR It is the near-infrared reflectance, R GREEN R RED R BLUE These represent the green band, red band, and blue band, respectively. This indicates the reflectivity corresponding to the wavenumber.

[0073] The edge computing module initiates a feature selection algorithm to optimize the aforementioned vegetation indices, selecting several features most sensitive to crops. These features form the final remote sensing input vector. Simultaneously, soil parameter data of the target area, including humidity, electrical conductivity, temperature, and pH, are collected on the ground, with each data point tagged with the collection time and location.

[0074] Image feature vectors and environmental parameters are integrated into a unified multimodal sample through a temporal alignment mechanism, and then transmitted back to the communication module of the cloud server through the second communication module.

[0075] After receiving the data, the cloud server inputs it into the data analysis module. This module first performs normalization and cleaning operations, and then proceeds to the following processing flow: using a convolutional neural network to extract multimodal joint features; the neural network model classifies the fused features and outputs the disease level of each plot: healthy, mild, or severe;

[0076] The data then enters the agricultural big data model module, where the distributed image stitching module further combines historical plot data, crop stage annotations, management records, and other information to perform time-series modeling and trend prediction. The regional statistical analysis module calls the image stitching module to generate a complete regional map and calculates indicators such as DVI and drought index through a regional division multi-scale window algorithm for comparison. If a certain block differs significantly from neighboring areas or deviates from historical trends by more than a threshold, intervention logic is triggered in the alarm and forecast module.

[0077] All results are aggregated into the decision-making module to generate a field inspection report, including regional level, disease risk, drought probability, and inspection recommendations.

[0078] After the report is generated, it is sent to the ground-based mobile platform via the communication module. The PID control unit receives the priority location information from this report. Deployed within the edge computing node, the PID control unit performs proportional-integral-derivative (PID) operations to generate a control signal based on environmental parameter data output from the edge computing module and target control parameters sent from the cloud server. The PID control algorithm output formula is:

[0079]

[0080] Where u(t) is the current error term; Adjust the direction and speed of movement; e(t) = T c (t)-T set e(t) represents the historical bias; K represents the rate of change of the error. i K d These are the proportional, integral, and differential gains, respectively.

[0081] The PID algorithm calculates the current path error in real time and outputs a control signal. This control signal is used to adjust the robot's current steering, speed, and other control variables to ensure accurate inspection. The robot can acquire high-resolution images on-site or receive manual confirmation before transmitting the data back.

[0082] After the mission, the spacecraft returned to the ground platform's storage compartment, where it was automatically docked by the takeoff and landing module. All data was backed up in a large-capacity storage device on the ground, while the agricultural knowledge graph and risk model were updated in the cloud.

[0083] In some embodiments, the system is deployed in an agricultural planting base in a high-temperature, low-rainfall area, undertaking the tasks of drought risk inspection, identification, and remote early warning for five consecutive crop planting areas. The system acquires multi-temporal remote sensing images via an aerial vehicle following a predetermined path, and transmits multispectral images and remote sensing reflectance data back to a ground-based mobile platform via a multimodal remote sensing acquisition module. The image data carries the acquisition timestamp, GPS location, sensor flight altitude, and band matrix information.

[0084] Meanwhile, the ground-based sensor module samples the soil environment of the deployment area, acquiring humidity, conductivity, pH, and temperature values ​​at each sensor location, and labeling the acquisition time, geographical location, and confidence level. Remote sensing images and soil parameters are simultaneously transmitted to the edge computing module. The image data first undergoes radiometric and geometric correction, followed by spatial registration and multi-scale fusion operations to extract the vegetation sensitivity index dataset. Simultaneously, the sensor data is standardized and noise-filtered, and aligned temporally with the image data to form one-to-one data pairs.

[0085] Table 2 Standardized Dataset

[0086]

[0087]

[0088] The system integrates the above features to construct sample vectors, which are then input into an agricultural big data model module deployed in the cloud. A drought detection model based on the support vector machine algorithm, trained with a radial basis function kernel, is used. The distributed image stitching module determines the current state as moderate drought with a confidence level of 92.4%. Subsequently, the regional statistics module performs a horizontal comparative analysis of similar indicators between the current plot and neighboring plots, confirming that its GNDVI decrease is greater than twice the standard deviation of the neighboring area's average. Combining this with the agricultural big data model's recorded data on the current crop growth stage, variety, and water requirement for the region, the system automatically generates a drought early warning report and transmits it to a remote user interface via the communication module, recommending that irrigation tasks be initiated within 24 hours and manually reviewed. This processing strategy is automatically registered by the alarm and forecast module and included in the system management log as part of the early warning response plan.

[0089] In some embodiments, this is a typical agricultural monitoring task. The system is deployed in a long strip of farmland with a depth of 600 meters to identify the trend of disease risk changing with distance. The aerial platform flies along the main axis of the area to collect multispectral images, while the ground-based mobile platform simultaneously collects data such as soil moisture, electrical conductivity, and pH value.

[0090] The system employs artificial intelligence algorithms and a vector machine model, and its construction steps include the following:

[0091] S01: Find an optimal hyperplane that maximizes the margin;

[0092] S02: Filtering based on constraints; the optimization objective is to maximize the interval.

[0093] S03: Introducing the Lagrange multiplier α i The obtained support vector solutions are used for classification prediction, and the constructed classification function is:

[0094]

[0095] Where f(x) is the final output of the model; sign(·) is the sign function; n is the number of samples in the training set; α i These are Lagrange multipliers, corresponding to the weights of the i-th training sample, and α corresponding to the support vectors. i y is 0; i x is the class label of the i-th training sample; i Let be the feature vector of the i-th training sample; b is the bias term, which adjusts the position of the hyperplane in the feature space.

[0096] Feature prediction is performed on the fused data at each 200-meter interval, and the output is a "target feature value" that comprehensively reflects the intensity of disease risk, which can be understood as a comprehensive disease score.

[0097] Meanwhile, to enhance the model's discriminative ability, the system executes the K-means clustering algorithm on historical data to cluster the time-series data of several plots within the region, forming representative feature cluster centers. The "feature values ​​after clustering" in the figure represent the feature expressions corresponding to these cluster centers.

[0098] Compare the two sets of data—the predicted values ​​and the cluster center values—at various distance points. (See [reference]) Figure 6 The trend shown in the graph indicates that:

[0099] In the initial stage (0–200m), the predicted values ​​are not much different from the cluster center values, indicating that the areas identified by the system are highly consistent with existing historical patterns. After 300m, the predicted values ​​are significantly higher than the cluster center values, which may indicate that the system has detected new high-risk disease areas that are not covered in the clusters. This deviation trend can serve as proof of the model's sensitivity, showing that the system can actively identify "potentially emerging abnormal areas" in actual deployment.

[0100] Areas with final predicted values ​​higher than historical cluster values ​​will be marked as priority inspection areas by the system, triggering a pre-control mechanism to generate emergency tasks.

[0101] In other embodiments, the system is applied to real-time detection and response to disease occurrence in a restricted tomato growing area. A flight platform collects multispectral images of the target area daily at fixed points. An image stitching module performs spatial registration and multi-temporal fusion on images from the past three days, forming a fused image raster. The edge computing module calculates plant pigment correlation indices, including ARI and EVI, from these images. The system identifies a specific area where the ARI has consistently increased to 0.21, significantly higher than the average of 0.11 for neighboring areas. Simultaneously, ground-based sensors report that the soil pH in this area has decreased to 5.1, and the electrical conductivity has decreased by more than 10% for two consecutive days. The fused multimodal data is fed as input samples into an agricultural large-scale model module, which contains a pre-trained BPNN neural network structure. A three-layer fully connected network is used to output the disease risk level. The analysis result for this area is "Disease Risk: Severe," with an output confidence level of 86.7%. In response to the analysis results, the system automatically generates task response suggestions, including task review, application area planning, and impact estimation reports, which are then pushed to the agricultural dispatch center via a remote communication module. If the system is connected to automatic spraying equipment, the task can also automatically generate control commands and trigger the spraying execution process via downlink, achieving a closed loop from system self-diagnosis to treatment.

[0102] In some embodiments, the system identifies a nutrient stress area in a rice experimental field and outputs targeted fertilization recommendations. Analysis of multi-temporal remote sensing images acquired from the sky shows that the DVI index is consistently below 0.18, while the GNDVI is also less than 0.35. Ground sensor feedback data shows an electrical conductivity of 0.28 mS / cm, far below the normal nutrient zone range of 0.45–0.55, a pH value of 5.7, and a soil temperature of 27.9℃. This data is aggregated into a fusion vector and input into a support vector machine training model. The model outputs a label of "mild nitrogen deficiency," with a system confidence level of 91.6%. Subsequently, the system retrieves information on the crop variety and growth stage of the area, identifying it as late tillering stage rice. Based on the corresponding nutrient requirement standards, the system recommends a nitrogen fertilizer application rate of 5.5 kg / mu and marks the coordinates of the fertilization area. This recommendation is uploaded to the scheduling system via a communication module for user reception and confirmation.

Claims

1. An integrated air-ground smart agriculture inspection and monitoring system, characterized in that, This includes aerial vehicles, ground-based mobile platforms, and cloud servers; The sky-end aircraft includes: The autonomous flight control module is used to control the takeoff and landing of the aerial vehicle; the takeoff and landing docking module is used to dock with the ground mobile platform storage compartment to achieve physical docking; the multimodal remote sensing acquisition module includes a multispectral camera, a remote sensing imaging unit and a spectral information acquisition unit, used to acquire surface remote sensing information; and the first communication module is used to transmit the acquired data back to the ground mobile platform through a self-organizing network. The ground-based mobile platform includes: The sensor module includes a soil moisture sensor, a soil conductivity sensor, a soil pH sensor, and a soil temperature sensor. The aircraft storage compartment is used to house the aforementioned airborne aircraft; Storage module, used to save measurement data; The edge computing module is used for synchronous preprocessing of remote sensing information from the sky and environmental information from the ground. The second communication module is used to transmit data with the airborne aircraft and the cloud server. The cloud server includes: The third communication module is used to transmit data with the ground-based mobile platform; The data analysis module is used to structure and analyze the received data. The agricultural big data model module is used to realize cross-regional remote sensing comparison, agricultural knowledge graph construction, and anomaly detection. The decision-making module is used to generate response decisions based on the analysis results and output field inspection reports; The alarm and forecast module is used to implement risk alerts and agricultural early warning strategies, respectively.

2. The integrated air-ground smart agriculture inspection and monitoring system according to claim 1, characterized in that, The agricultural large model module is trained using a self-built remote sensing and ground multimodal fusion dataset. The dataset includes: multispectral image sequences, remote sensing reflectance maps, soil moisture and temperature time series, crop growth stage annotation information, and corresponding agricultural management history records of the plots.

3. The integrated air-ground smart agriculture inspection and monitoring system according to claim 2, characterized in that, The training process using a self-built remote sensing and ground-based multimodal fusion dataset specifically includes the following steps: S01: Perform spatial registration and temporal alignment on the collected raw data to construct multimodal synchronous sample pairs; S02: Extracting spatial-temporal features using a self-attention model based on the Transformer structure; S03: Joint optimization of classification, regression and temporal tasks based on multi-task loss functions; S04: Use knowledge distillation techniques to encode data into rule-based knowledge vectors to guide the training of the main model; S05: Use a contrastive learning mechanism to enhance the feature separability between different plots.

4. The integrated air-ground smart agriculture inspection and monitoring system according to claim 1, characterized in that, The input data for the data analysis module includes: The remote sensing image data packet is acquired by a multimodal remote sensing acquisition module and specifically includes: a unique image identifier, a capture timestamp, GPS geographic coordinates, sensor altitude, a multi-band image matrix, and corresponding wavelength information. The environmental parameter data of the edge computing module specifically includes: sensor point coordinates, acquisition time, soil moisture, conductivity, pH value, temperature, and corresponding data confidence fields; The remote sensing image data packets and the environmental parameter data of the edge computing module are transmitted to the cloud server through the second communication module of the ground mobile platform.

5. The integrated air-ground smart agriculture inspection and monitoring system according to claim 1, characterized in that, The large-scale agricultural model module includes a distributed image stitching module and a regional statistical analysis module, wherein: The distributed image stitching module is used to spatially align remote sensing images acquired from multiple platforms and fuse them across multiple time periods within the same block. The regional statistical analysis module calculates the normalized vegetation index, vegetation coverage, and regional drought index based on the fusion results, and compares the indicators between regions.

6. The integrated air-ground smart agriculture inspection and monitoring system according to claim 5, characterized in that, The regional statistical analysis module employs a multi-scale window comparison algorithm based on regional division, including the following steps: S01: Determine the boundaries of the comparison unit plots based on the remote sensing image segmentation results; S02: Extract the changing trends of multi-temporal remote sensing indicators on the same block; S03: Combine historical databases and information from surrounding plots to conduct deviation analysis and determine whether there are any abnormal growth patterns or potential disaster signals; S04: Input the analysis results and the results of the agricultural big data model module into the decision-making module to generate multi-level field inspection strategy suggestions.

7. The integrated air-ground smart agriculture inspection and monitoring system according to claim 1, characterized in that, The anomaly detection is based on a support vector machine algorithm for model training, and the training method includes: S01: Obtain multispectral images and remote sensing reflectance data collected by the multimodal remote sensing acquisition module of the sky-end flight platform, and construct them as a first type of feature vector; obtain environmental parameters generated by the edge computing module of the ground-end mobile platform, and construct them as a second type of feature vector; S02: Synchronize the first type of features with the second type of features in time and then concatenate them to form a fused input sample; S03: Construct training sample labels, which are manually labeled based on historical anomalies such as disease, drought or nutrient imbalance. S04: Use radial basis kernel functions to train support vector machine models and construct nonlinear decision boundaries to distinguish between normal and abnormal samples; S05: Deploy the trained model to the agricultural large model module in the cloud server for online identification of abnormal states of new input samples.

8. The integrated air-ground smart agriculture inspection and monitoring system according to claim 1, characterized in that, The multispectral camera acquires multispectral remote sensing images of farmland crops in real time and transmits them back to the ground in real time through an ad hoc network; the ad hoc network adopts an on-demand distance vector routing protocol to achieve network topology adaptation.

9. The integrated air-ground smart agriculture inspection and monitoring system according to claim 1, characterized in that, After receiving the multispectral remote sensing images transmitted back by the spacecraft, the edge computing module uses the edge computing unit to perform image registration algorithms to achieve image fusion, and uses the SIFT algorithm to achieve image correction and geolocation.

Images