A method and system for unmanned aerial vehicle inspection of tobacco fields

Abstract

The present application relates to the technical field of flight control of inspection unmanned aerial vehicle, and discloses a kind of unmanned aerial vehicle inspection method and system for tobacco field, which comprises: obtaining the multi-source remote sensing image data of the preset tobacco planting area.Detecting tobacco plant in remote sensing image data, spatial clustering is carried out on the detection result, and a plurality of tobacco cluster targets are obtained.Initial inspection path is generated according to the position coordinates and vegetation index of the tobacco cluster target.In the flight process, the flight pose, environmental parameters and the offset angle between the unmanned aerial vehicle and the current path segment of the target unmanned aerial vehicle are collected in real time.When the adjustment condition is met, the improved path is automatically generated, and the flight attitude and camera angle of the target unmanned aerial vehicle are adjusted.The image data collected in the flight process is identified to determine whether there is an abnormal situation.The present application realizes the inspection method integrating the tobacco planting scene, multi-source remote sensing perception, adaptive path planning and intelligent identification.

Description

Technical Field

[0001] The present invention relates to the field of flight control technology for inspection drones, and in particular to a drone inspection method and system for tobacco fields. Background Art

[0002] With the development of digital agriculture and intelligent agricultural machinery, drones are increasingly being used for agricultural inspections, crop growth monitoring, and pest and disease identification. In tobacco cultivation, in particular, the growth status of tobacco plants directly affects yield and quality. Therefore, timely and accurate detection of abnormalities such as pests and diseases, and water stress, is crucial for maintaining healthy tobacco field management. Currently, some tobacco farm managers are experimenting with using camera-equipped drones for low-altitude inspections of fields, incorporating image recognition technology to assist in identifying crop anomalies.

[0003] However, some issues remain. For example, most systems rely solely on simple image capture and GPS for grid-based inspections, failing to integrate multi-source remote sensing images (such as multispectral and depth maps) for precise target identification and spatial clustering of tobacco plants, making targeted sampling and dynamic path planning difficult. Traditional inspection routes often use fixed raster scanning or manually set waypoints, failing to incorporate growth indicators such as tobacco plant density to generate agronomically oriented inspection routes. They also fail to account for natural factors such as high wind speeds and low light levels that affect image quality and flight stability, resulting in low mission efficiency. Furthermore, current technologies typically use general image classification models for single-frame analysis and lack the ability to jointly identify complex crop anomalies (such as lesions and water stress) using multi-task recognition. Furthermore, most inspection systems lack data upload and spatial annotation mechanisms connected to cloud-based agricultural management platforms, making effective closed-loop management difficult.

[0004] Therefore, it is necessary to design a drone inspection method and system for tobacco fields to solve the problems existing in current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a drone inspection method and system for tobacco fields, aiming to solve the current problems of inaccurate tobacco plant target positioning, path generation that does not consider environmental disturbances, and poor reliability of pest and disease identification.

[0006] In one aspect, the present invention provides a method for inspecting tobacco fields using a drone, comprising:

[0007] Acquire multi-source remote sensing image data of a preset tobacco-growing area, the remote sensing image data including RGB images, multispectral images, and depth maps; detect tobacco plants in the remote sensing image data based on a target recognition network, and spatially cluster the detection results based on tobacco plant distribution density to obtain a plurality of tobacco plant cluster targets;

[0008] generating an initial inspection path according to the position coordinates and vegetation index of the tobacco cluster target, wherein the initial inspection path connects a plurality of tobacco cluster center points and avoids areas where the wind speed exceeds a first threshold and where the light intensity is less than a second threshold in the wind speed prediction map;

[0009] Control the target UAV to fly along the initial inspection path, and collect the target UAV's flight posture, environmental parameters, and the offset angle between the UAV and the current path segment in real time during the flight;

[0010] When the offset angle is greater than the angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than the set distance threshold, an improved path from the target UAV to the next target cluster is automatically generated, and the flight attitude and camera angle of the target UAV are adjusted;

[0011] The tobacco pest and disease recognition model is used to identify image data collected during flight to identify abnormalities on the leaf surface, such as lesions, pests, or water stress. The locations of abnormal tobacco plant clusters are marked and uploaded to the cloud platform.

[0012] Furthermore, when detecting tobacco plants in the remote sensing image data based on the target recognition network, the method includes:

[0013] Performing joint input processing on the RGB image and the multispectral image based on a deep convolutional neural network model, wherein the deep convolutional neural network model includes a U-Net network with an encoding-decoding structure, and the deep convolutional neural network model integrates a spatial attention mechanism;

[0014] The RGB image provides texture and edge information, the multispectral image provides vegetation characteristic band information, and the deep convolutional neural network model performs feature splicing after dual-branch feature extraction and feeds the feature into the classifier to output the pixel-level segmentation result of the tobacco plant.

[0015] The segmentation results are subjected to morphological processing to remove noise points and interference areas, and the bounding box and spatial coordinates of each tobacco plant target are extracted through connected domain analysis.

[0016] Furthermore, the detection results are spatially clustered according to the distribution density of tobacco plants to obtain several tobacco plant cluster targets, including:

[0017] Extract the center coordinates of all tobacco plant targets and construct a two-dimensional space coordinate set;

[0018] The density-based spatial clustering algorithm DBSCAN performs cluster analysis on the coordinate set, wherein the parameters of the spatial clustering algorithm DBSCAN include a distance threshold and a minimum number of neighborhood points, which are used to determine whether the tobacco plant points belong to the same cluster;

[0019] The distance threshold is dynamically set according to the remote sensing image resolution and the tobacco plant spacing;

[0020] Each high-density connected area in the clustering result is identified as a tobacco plant cluster target, and the center coordinates of the corresponding cluster, the number of tobacco plants in the cluster, and the boundary range are output.

[0021] Furthermore, when generating an initial inspection path according to the location coordinates and vegetation index of the tobacco cluster target, the method includes:

[0022] A weighted graph is constructed based on the center coordinates of each tobacco plant cluster, wherein each node in the weighted graph corresponds to the center coordinates of a tobacco plant cluster, and the edge weight between any two nodes is determined by real-time objective factors: the real-time objective factors include the Euclidean distance between nodes, the difference in vegetation index between the areas corresponding to the nodes, and the average wind speed and average light intensity values ​​of the overlapping areas on the edge paths;

[0023] Applying an improved path planning algorithm to the weighted graph to perform path optimization, where the optimization objective is to minimize the weighted sum of the total path length and an inspection risk weight, where the inspection risk weight is a penalty factor assigned based on whether the wind speed on the edge path is greater than a preset first threshold and whether the light intensity is less than a second threshold;

[0024] An inspection path connecting the center coordinates of all tobacco plant clusters is generated to obtain the initial inspection path.

[0025] Furthermore, during the flight, the real-time acquisition of the target UAV's flight posture, environmental parameters, and the offset angle between the UAV and the current path segment includes:

[0026] The flight posture of the UAV is calculated based on the inertial measurement unit, GPS module and visual inertial SLAM. The flight posture includes position coordinates and flight heading vector.

[0027] Synchronously collect environmental data, including wind speed, light, and temperature and humidity, and construct an angle between the flight posture of the drone and the fitting vector of the current path segment to obtain the offset angle;

[0028] The offset angle is calculated using the cosine similarity formula, and whether the angle exceeds the angle threshold is determined based on the angle between the fitting vector of the current path segment and the flight heading vector.

[0029] Furthermore, when the offset angle is greater than the angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than the set distance threshold, the judgment process includes:

[0030] Comparing the Euclidean space distance between the position coordinates of the UAV and the center coordinates of the current target cluster with a distance threshold, and comparing the offset angle with an angle threshold;

[0031] If the offset angle is greater than the angle threshold and the spatial distance is less than the distance threshold, the circumferential fixed-point track of the current cluster is skipped and the improved path is generated by directly turning to the center coordinates of the next cluster;

[0032] If the offset angle is greater than the angle threshold and the spatial distance is greater than the distance threshold, the heading angle and flight speed are adjusted to return to the current path segment, and the image acquisition of the current target cluster is suspended. The target observation is reactivated after the attitude stabilizes;

[0033] If the offset angle is less than or equal to the angle threshold, the yaw angle and the flight side thrust are adjusted to return the UAV to the center line of the path.

[0034] Furthermore, the improved path from the target UAV to the next target cluster is automatically generated, and the flight attitude and camera angle of the target UAV are adjusted, including:

[0035] The line connecting the current UAV's position coordinates and the center coordinates of the next target cluster is used as the main path. A smooth steering path is generated by cubic Bezier curve fitting to obtain the improved path. At the same time, the vector direction between the current UAV and the center point of the skipped cluster is calculated to adjust the camera angle.

[0036] Furthermore, the image data collected during the flight is identified based on the tobacco pest and disease recognition model to identify whether there are any abnormalities on the leaf surface, including:

[0037] Preprocessing the image data, wherein the preprocessing includes image denoising, color enhancement, and illumination normalization;

[0038] The preprocessed image is input into the tobacco pest and disease recognition model, which includes a backbone extraction network and several branch subtask heads. The backbone extraction network is used to output a semantic segmentation map of the diseased spot area, and the branch subtask heads are used to output a multi-class classification result of the pest type and a hot spot location map of the leaf water stress area.

[0039] According to the pixel position of the abnormal area in the recognition result and the flight posture when the corresponding image was collected, combined with the coordinates of the center point of the current tobacco cluster, spatial back-calculation and annotation were performed to generate structured abnormality information including abnormality type, spatial location and image timestamp.

[0040] Furthermore, the drone inspection method for tobacco fields also includes:

[0041] The remaining power of the drone is collected in real time, and the power demand is obtained based on the remaining path. The remaining power is compared with the power demand. If the difference between the remaining power of the drone and the power demand is less than the power threshold, the drone is controlled to return to the nest for charging.

[0042] Compared with the existing technology, the beneficial effects of the present invention are: by combining multi-source remote sensing image acquisition with deep learning recognition, accurate identification and spatial clustering of tobacco plants in tobacco-growing areas are achieved, breaking through the accuracy limitations of traditional inspections that only rely on RGB images and GPS grids; integrating tobacco plant distribution density and vegetation index information to generate an initial inspection path, and introducing wind speed and light factors to optimize the path obstacle avoidance, thereby improving the environmental adaptability and operational safety of path planning; by real-time acquisition of the UAV's flight posture and its offset angle from the inspection path, a dynamic inspection path reconstruction mechanism is constructed, enabling the UAV to have the ability of jump-point flight and posture-linked shooting, thereby improving flight efficiency and image acquisition stability; utilizing a multi-task tobacco disease and pest recognition model to jointly identify multi-dimensional abnormal conditions such as lesions, pests, and water stress in the collected images, and performing spatial back-labeling based on the flight posture and synchronously uploading to the cloud, forming an identification-positioning-feedback closed loop, effectively solving the problems of single identification dimension, fixed path, and broken feedback chain in the existing technology.

[0043] On the other hand, the present application also provides a drone inspection system for tobacco fields, which is used to apply the above-mentioned drone inspection method for tobacco fields, including:

[0044] The acquisition unit is configured to acquire multi-source remote sensing image data of a preset tobacco planting area; detect tobacco plants in the remote sensing image data based on a target recognition network, and spatially cluster the detection results according to tobacco plant distribution density to obtain a plurality of tobacco plant cluster targets;

[0045] a processing unit configured to generate an initial inspection path based on the position coordinates of the tobacco cluster target and the vegetation index, wherein the initial inspection path connects a plurality of tobacco cluster center points and avoids areas where the wind speed exceeds a first threshold and where the light intensity is less than a second threshold in the wind speed prediction map;

[0046] a judgment unit configured to control the target UAV to fly along the initial inspection path and collect the flight posture, environmental parameters, and offset angle between the target UAV and the current path segment in real time during the flight;

[0047] The judgment unit is further configured to, when the offset angle is greater than an angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than a set distance threshold, automatically generate an improved path for the target UAV to the next target cluster and adjust the flight attitude and camera angle of the target UAV;

[0048] The recognition unit is configured to recognize the image data collected during flight based on a tobacco pest recognition model, recognize whether there is an abnormal situation on the leaf surface, the abnormal situation including a disease spot, a pest, or water stress, and upload the location of the abnormal situation of the tobacco plant cluster to the cloud platform after labeling.

[0049] It can be understood that the unmanned aerial vehicle inspection method and system for tobacco fields have the same beneficial effects, and will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0050] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments and are not meant to limit the present application. Furthermore, the same reference numerals in the attached drawings are intended to refer to the same components throughout. In the drawings:

[0051] Figure 1 a flowchart of the unmanned aerial vehicle inspection method for tobacco fields provided by the embodiments of the present application;

[0052] Figure 2 a functional block diagram of the unmanned aerial vehicle inspection system for tobacco fields provided by the embodiments of the present application. DETAILED DESCRIPTION

[0053] Exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and the scope of the present disclosure can be accurately conveyed to those skilled in the art. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0054] In traditional drone inspection methods, insufficient fusion of multi-source remote sensing data limits tobacco plant detection accuracy. Plant density and spatial clustering results cannot support dynamic path planning, and the impact of environmental factors on flight stability and image acquisition quality is not effectively mitigated. For example, when deploying drone inspection systems in tobacco-growing areas, RGB and multispectral images are processed independently, without a joint feature extraction mechanism. This results in unacceptable plant boundary segmentation errors. The path planning module uses a fixed waypoint sequence generation strategy and fails to incorporate wind speed prediction data and light intensity monitoring values ​​into path weight calculations. When the drone enters an area with sudden wind speed changes, the flight attitude control system, lacking a pre-set obstacle avoidance strategy, experiences heading deviations, causing the image acquisition area to become spatially misaligned with the pre-set tobacco plant cluster center. The image recognition module uses a single-task classification model for leaf anomaly detection, failing to simultaneously output a segmentation map of the lesion area and the pest type classification results. This results in inaccurate identification of spatial overlap between water-stressed and pest-infested areas.

[0055] If these issues are not addressed, spatial positioning errors in tobacco cluster targets will cause inspection routes to deviate from high-density growth areas, resulting in missed detection of critical abnormal samples. High wind speed areas not avoided in the flight path will cause the drone's posture to oscillate, reducing the image acquisition frame rate and resolution, and affecting the accuracy of feature extraction for subsequent anomaly identification. Single-task recognition models are unable to establish correlations between lesion morphology and pest activity, resulting in reduced diagnostic accuracy for complex anomaly scenarios. This ultimately leads to incomplete basis for tobacco field management decisions, increasing the risk of pest and disease spread and errors in resource allocation.

[0056] For this, see Figure 1 As shown, this application proposes a method for drone inspection of tobacco fields, comprising:

[0057] S100: Acquire multi-source remote sensing image data of a preset tobacco planting area, where the remote sensing image data includes RGB images, multispectral images, and depth maps.

[0058] S200: Detecting tobacco plants in the remote sensing image data based on a target recognition network, and spatially clustering the detection results according to tobacco plant distribution density to obtain a number of tobacco plant cluster targets.

[0059] S300: Generate an initial inspection path based on the location coordinates and vegetation index of the tobacco cluster target. The initial inspection path connects several tobacco cluster center points, and avoids areas where the wind speed exceeds the first threshold and areas where the light intensity is less than the second threshold in the wind speed prediction map.

[0060] S400: Control the target UAV to fly along the initial inspection path, and collect the target UAV's flight posture, environmental parameters, and the offset angle between the UAV and the current path segment in real time during flight. When the offset angle exceeds an angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than a set distance threshold, an improved path is automatically generated for the target UAV to the next target cluster, and the target UAV's flight posture and camera angle are adjusted.

[0061] S500: Based on the tobacco pest and disease recognition model, the image data collected during the flight is identified to identify whether there are any abnormalities on the leaf surface, such as lesions, pests or water stress. The locations of the abnormal tobacco plant clusters are marked and uploaded to the cloud platform.

[0062] Specifically, multi-source remote sensing image data refers to image data that contains RGB images, multispectral images, and depth maps. This can be achieved by using a drone equipped with a multispectral camera, RGB camera, and depth sensor to simultaneously collect data. By fusing information from different spectral bands with three-dimensional depth information, the accuracy of tobacco plant detection and spatial positioning can be improved. Among them, the target recognition network refers to a deep learning model used to detect tobacco plants in remote sensing images. Specifically, it can be implemented using a U-Net network with an encoding-decoding structure combined with a spatial attention mechanism. Through dual-branch feature extraction and fusion, the recognition ability of tobacco plant texture, edges, and vegetation features is enhanced, solving the problem that traditional methods cannot accurately segment tobacco plants. Among them, spatial clustering refers to dividing the detection results into multiple tobacco plant cluster targets based on the distribution density of tobacco plants. Specifically, it can be implemented using the density-based DBSCAN algorithm. By dynamically adjusting the distance threshold to adapt to different planting row spacing scenarios, high-density connected areas are generated as inspection target clusters, providing a basis for dynamic path planning. Among them, the initial inspection path refers to the flight path that connects the center points of the tobacco clusters and avoids high wind speed and low light areas. Specifically, it can be achieved by constructing a weighted graph and applying an improved path planning algorithm. The path is optimized by combining Euclidean distance, vegetation index difference and environmental risk factors to solve the problem of low efficiency of traditional fixed grid paths. Among them, the offset angle refers to the angle between the UAV flight heading vector and the fitting vector of the current path segment. Specifically, it can be achieved through the joint measurement of inertial measurement unit, GPS and visual SLAM. The path is dynamically adjusted by real-time monitoring of angle deviation to ensure that the UAV can still stably perform inspection tasks under environmental interference. Among them, the improved path refers to the smooth turning path generated when the UAV deviates from the original path. Specifically, it can be achieved by fitting the current UAV position and the center point of the next target cluster with a cubic Bezier curve. Combined with the camera angle adjustment, it compensates for the observation needs of the skipped clusters to solve the problem of inspection interruption caused by sudden environmental interference. The tobacco pest and disease identification model is a multi-task deep learning model used to detect leaf surface lesions, pests, and water stress. It can be implemented using a backbone network combined with a multi-branch subtask head architecture. It combines semantic segmentation, classification, and hotspot location to output anomaly types and locations, improving recognition accuracy in complex agricultural scenarios. The cloud platform is a remote server that receives and stores anomaly annotation information. It can upload the spatial location, timestamp, and type of abnormal tobacco plant clusters via wireless communication modules, providing structured data support for subsequent precision agriculture decision-making and forming a closed-loop management process.

[0063] This application uses multi-source remote sensing data fusion and dynamic path planning technology, combined with the results of tobacco plant spatial clustering, to generate an initial inspection path that avoids environmental risks. It also monitors the offset angle in real time during flight to trigger dynamic path adjustment. At the same time, it uses a multi-task recognition model to achieve joint detection of pests and diseases and water stress, and finally uploads abnormal information annotations to the cloud platform, forming a complete closed loop from data collection to decision support.

[0064] The working process and principle of this application are as follows: multi-source remote sensing image data of a preset tobacco-growing area is obtained, including RGB images, multispectral images, and depth maps. These different types of image data provide multi-dimensional information about tobacco plants. The target recognition network is used to detect tobacco plants in the remote sensing image data. The target recognition network locates and identifies tobacco plants by analyzing image features. After the detection is completed, the detection results are spatially clustered according to the distribution density of the tobacco plants to obtain several tobacco plant cluster targets. Spatial clustering can group adjacent tobacco plants together to form tobacco plant clusters.

[0065] An initial inspection path is generated based on the location coordinates and vegetation index of the tobacco cluster targets. This initial inspection path connects several tobacco cluster centers while avoiding areas where wind speeds exceed the first threshold and where light intensity falls below the second threshold, as shown in the wind speed forecast. This ensures that the inspection path covers key tobacco plant areas while avoiding unfavorable flight conditions.

[0066] The target drone is controlled to fly along the initial inspection route. During flight, the target drone's flight posture, environmental parameters, and the offset angle between the drone and the current path segment are collected in real time. This data is used to monitor the drone's flight status and environmental conditions. When the offset angle exceeds a threshold and the spatial distance between the target drone and the center of the current target cluster is less than a set distance threshold, an improved path is automatically generated for the target drone to the next target cluster, and the target drone's flight posture and camera angle are adjusted. This dynamic adjustment mechanism optimizes the inspection route and improves inspection efficiency.

[0067] The flight-generated image data is analyzed using a tobacco pest and disease recognition model to identify leaf surface anomalies, including lesions, insect damage, and water stress. The locations of abnormal tobacco plant clusters are annotated and uploaded to a cloud platform, providing data support for subsequent tobacco field management.

[0068] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0069] Aerial photography of a pre-determined tobacco planting area was conducted using a multispectral camera and a depth camera, generating RGB images, multispectral images, and depth maps. The RGB images had a resolution of 4000 × 3000 pixels, while the multispectral images included near-infrared, red-edge, and red light bands. The depth maps had centimeter-level accuracy.

[0070] The acquired image data is fed into a pretrained deep learning network for tobacco plant detection. This network uses a U-Net architecture, fusing RGB and multispectral information for feature extraction. The detection results output a pixel-level segmentation mask and center coordinates for each tobacco plant.

[0071] The detected tobacco plants were spatially clustered using the DBSCAN algorithm. The cluster radius was set to 2 meters, and the minimum number of samples was set to 5. The clustering results yielded multiple tobacco plant clusters, each containing multiple adjacent tobacco plants.

[0072] Based on the clustering results, a weighted graph model was constructed. Nodes in the graph represent the centers of tobacco plant clusters, and edge weights take into account inter-node distance, vegetation index differences, wind speed, and light intensity. An improved A* algorithm was used to search for the optimal path on the graph and generate the initial inspection route.

[0073] The drone is controlled to fly along the initial path, at an altitude of 5 meters above the ground. During flight, onboard sensors collect real-time data such as position, attitude, and wind speed. If the deviation angle exceeds 5 degrees and the distance from the center of the current target cluster is less than 10 meters, a new Bezier curve path is automatically generated to the next target cluster.

[0074] During the inspection, the drone's high-resolution camera captures five images per second. These images are transmitted in real time to a ground station and fed into a pre-trained multi-task convolutional neural network for pest and disease identification. This network simultaneously outputs a lesion segmentation map, pest classification results, and a water stress heat map.

[0075] Finally, the identified anomalies are associated with the corresponding GPS coordinates to generate structured data containing the anomaly type, location, and timestamp, which is then uploaded to the cloud platform via the 4G network for storage and analysis.

[0076] Through the above solution, this application achieves efficient and accurate inspections of tobacco-growing areas. The integration of multi-source remote sensing data improves the accuracy of tobacco plant detection, and dynamic path planning avoids adverse environmental factors, improving inspection efficiency. The multi-task recognition model can detect multiple anomalies simultaneously, providing strong support for the timely discovery and resolution of tobacco field issues. Real-time data upload and spatial annotation provide comprehensive and timely information support for tobacco field management decisions, helping to improve the management level and yield quality of tobacco cultivation.

[0077] In some of the above-mentioned solutions of this application, the tobacco plant detection process relies on a single image data source, resulting in insufficient accuracy in the segmentation results. In particular, it is difficult to distinguish tobacco plants from interference objects in complex backgrounds, and it is impossible to effectively integrate vegetation features in multispectral information, affecting the accuracy of subsequent spatial clustering and path planning.

[0078] This application further proposes a method for joint input processing of RGB and multispectral images based on a deep convolutional neural network model. The deep convolutional neural network model includes a U-Net network with an encoding-decoding structure and incorporates a spatial attention mechanism. The RGB image provides texture and edge information, while the multispectral image provides vegetation characteristic band information. The deep convolutional neural network model performs feature splicing after dual-branch feature extraction and feeds the feature into a classifier, outputting the pixel-level segmentation results of the tobacco plants. The segmentation results are morphologically processed to remove noise and interference areas, and the bounding box and spatial coordinates of each tobacco plant target are extracted through connected domain analysis.

[0079] The U-Net network with an encoder-decoder architecture uses skip connections to preserve multi-scale features, and a spatial attention mechanism is embedded in the decoding layer to enhance the feature response of tobacco plant regions. During the dual-branch feature extraction process, the RGB image branch extracts texture and shape features, while the multispectral image branch extracts features related to vegetation indices. The feature maps output by the two branches are concatenated in the channel dimension and input to the classifier. Morphological processing uses a combination of opening and closing operations, and connected domain analysis uses the four-neighborhood or eight-neighborhood criteria to label individual tobacco plant targets.

[0080] Specifically, the RGB image and the multispectral image are input to a two-branch network, each consisting of a convolutional layer and a pooling layer for feature extraction. The intermediate feature map of the RGB branch contains the outline and edge information of the tobacco plant, while the intermediate feature map of the multispectral branch contains vegetation characteristics such as near-infrared reflectance. After spatial alignment, the feature maps of the two branches are fused into a multi-channel feature map through a concatenation operation and input to the U-Net decoding layer, which incorporates a spatial attention mechanism. The spatial attention mechanism calculates spatial weights on the feature maps to suppress interference from background regions and enhance feature activation values ​​in the tobacco plant region. The classifier outputs a probability map for each pixel belonging to a tobacco plant, which is then segmented using thresholding to obtain a binary segmentation result. The segmentation result undergoes morphological filtering to remove isolated noise points, and connected component analysis is then used to determine the minimum bounding rectangle and center coordinates of each tobacco plant. For example, the red edge band in the multispectral image is used to calculate the normalized vegetation index to enhance the contrast between the tobacco plant and the soil. The U-Net encoding layer uses a ResNet50 architecture, while the decoding layer utilizes deconvolution and skip connections to restore spatial resolution. As a result, the accuracy of tobacco plant detection is improved, providing reliable input for subsequent clustering.

[0081] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0082] The RGB and multispectral images are jointly processed based on a deep convolutional neural network model. The deep convolutional neural network model includes a U-Net network with an encoder-decoder structure and integrates a spatial attention mechanism.

[0083] RGB images provide texture and edge information, while multispectral images provide information about vegetation characteristic bands. The deep convolutional neural network model performs dual-branch feature extraction, then performs feature concatenation and feeds this into a classifier, outputting pixel-level segmentation results for tobacco plants.

[0084] The segmentation results were morphologically processed to remove noise and interference areas, and the bounding box and spatial coordinates of each tobacco plant target were extracted through connected domain analysis.

[0085] Specifically, the RGB and multispectral images first undergo independent convolutional layers for feature extraction. The RGB branch primarily extracts visual features such as the shape and texture of the tobacco plant, while the multispectral branch focuses on extracting spectral response characteristics across different wavelengths. The feature maps from both branches are then weighted and fused using a spatial attention module to highlight the features of key regions.

[0086] The fused feature map is fed into the encoder portion of the U-Net network, where high-level semantic features are gradually extracted through multiple layers of convolution and pooling. The decoder then gradually restores the spatial resolution of the feature map through upsampling and skip connections, ultimately outputting a segmentation map of the same size as the input image. Each pixel in the segmentation map is assigned a probability value for belonging to a tobacco plant or background.

[0087] Furthermore, thresholding is applied to the segmentation map to obtain a binary image, followed by morphological opening and closing operations to remove small noise points and smooth the boundaries. Finally, a connected domain analysis algorithm is used to extract each independent tobacco plant region, calculate its minimum enclosing rectangle as the bounding box, and record the coordinates of the center point.

[0088] Through the above technical solutions, this application achieves accurate detection and location of tobacco plants. The fusion of RGB and multispectral information improves recognition accuracy, particularly for tobacco plant segmentation against complex backgrounds. The spatial attention mechanism enhances the model's perception of key regions. The U-Net architecture ensures high-resolution segmentation results, facilitating subsequent refined analysis. Morphological processing and connected domain analysis further enhance the robustness of detection results, laying the foundation for subsequent tobacco plant clustering.

[0089] In some of the above-mentioned schemes in this application, spatial clustering is required after tobacco plant detection to generate inspection paths, but existing clustering methods may not be able to adapt to different planting densities and remote sensing image resolution differences, resulting in inaccurate clustering results, affecting the rationality of subsequent path planning and inspection efficiency.

[0090] This application further proposes to extract the central coordinates of all tobacco plant targets and construct a two-dimensional spatial coordinate set. The density-based spatial clustering algorithm DBSCAN performs cluster analysis on the coordinate set. The parameters of the spatial clustering algorithm DBSCAN include a distance threshold and a minimum number of neighborhood points, which are used to determine whether the tobacco plant points belong to the same cluster. Among them, the distance threshold is dynamically set according to the remote sensing image resolution and the tobacco plant planting row spacing. Each high-density connected area in the clustering result is identified as a tobacco plant cluster target, and the center coordinates of the corresponding cluster, the number of tobacco plants in the cluster, and the boundary range are output.

[0091] The two-dimensional spatial coordinate set is constructed by extracting the coordinates of the geometric center points of the tobacco plant segmentation results, ensuring that the location information of each tobacco plant target is accurately recorded. The DBSCAN algorithm controls the clustering density using two parameters: a distance threshold and a minimum number of neighboring points. The distance threshold is dynamically adjusted based on the remote sensing image resolution and the tobacco plant row spacing. For example, when the image resolution is 0.1 meters and the row spacing is 1.5 meters, the distance threshold can be set to 0.8 meters to avoid clustering errors caused by excessively high resolution or variable row spacing. The minimum number of neighboring points is set to 5 to ensure that low-density noise points are filtered out. The dynamically set distance threshold is calculated using the resolution parameter in the remote sensing image metadata and pre-recorded tobacco field row spacing data. For example, the formula: distance threshold = row spacing × 0.5 + resolution × 10 is used to adapt the threshold to different scenarios. The center coordinates of the output cluster are determined by calculating the geometric mean of all points within the cluster, and the boundary range is generated using the convex hull algorithm.

[0092] Specifically, after the tobacco plant center coordinate set is input into the DBSCAN algorithm, the algorithm traverses all coordinate points and searches for neighboring points using a dynamic distance threshold as the neighborhood radius. If the number of points in the neighborhood of a point reaches the minimum number of neighborhood points, the point and its neighboring points are classified into the same cluster and recursively expanded to all points where the density can be reached. Unclassified points are marked as noise. The distance threshold is dynamically adjusted by combining the remote sensing image resolution and the planting row spacing. For example, the threshold is increased in low-resolution images to avoid cluster dispersion due to coordinate errors, and the threshold is reduced in high-density planting areas to distinguish adjacent tobacco plant clusters. The final output cluster center coordinates are used to generate inspection paths. The number of tobacco plants in the cluster reflects the regional growth density, and the boundary range is used to determine the drone observation area. This process solves the problem of insufficient adaptability of fixed parameter clustering algorithms in complex tobacco field environments, ensuring that the inspection path accurately covers high-density tobacco plant areas.

[0093] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0094] The center coordinates of the detected tobacco plant targets are extracted to construct a two-dimensional spatial coordinate set, which contains the X and Y coordinate values ​​of each tobacco plant.

[0095] The DBSCAN density clustering algorithm was applied to the coordinate set for analysis. Key parameters of the DBSCAN algorithm include the distance threshold ε and the minimum number of neighborhood points, MinPts. The distance threshold ε is dynamically set based on the remote sensing image resolution and the tobacco plant row spacing. For example, when the image resolution is 0.1 meters per pixel and the tobacco plant row spacing is 0.5 meters, ε can be set to 5 pixels. MinPts can be set to 3-5, indicating that a core point must have a neighborhood of at least 3-5 points within it to form a cluster.

[0096] The DBSCAN algorithm execution process is as follows:

[0097] Randomly select an unvisited point P.

[0098] Retrieve the points in the ε neighborhood of P. If the number of points contained is greater than or equal to MinPts, then P is the core point and a new cluster is formed.

[0099] Recursively perform density-reachable expansion on the points in the ε neighborhood of P, and add density-reachable points to the cluster.

[0100] If P is not a core point, mark P as a noise point.

[0101] Repeat the steps until all points have been visited.

[0102] After clustering is completed, each high-density connected area is identified as a tobacco plant cluster target. For each cluster, calculate and output:

[0103] The center coordinates of the cluster: the average value of the X and Y coordinates of all points in the cluster.

[0104] Number of plants in a cluster: the total number of points belonging to the cluster.

[0105] Cluster boundary range: the minimum enclosing rectangle of the points within the cluster.

[0106] Through the above technical solution, this application realizes adaptive clustering based on the actual distribution density of tobacco plants. Compared with fixed grid division, this method can more accurately reflect the spatial distribution characteristics of tobacco plants and provide more targeted target cluster information for subsequent inspection path planning. At the same time, by dynamically setting clustering parameters, this method can adapt to tobacco fields with different planting densities and terrain conditions, improving the accuracy and robustness of clustering results. In addition, the output cluster target information contains key data such as center coordinates, number of tobacco plants, and boundary range, which provides an important basis for subsequent precise inspections and pest and disease monitoring.

[0107] In some of the above-mentioned schemes in this application, the impact of real-time objective factors on path planning, such as the distance between nodes, vegetation index differences, wind speed and lighting conditions, is not comprehensively considered when generating the initial inspection path, resulting in the inspection path possibly passing through high wind speed or low light areas, increasing flight risks or affecting image acquisition quality.

[0108] The present application further proposes that when generating the initial inspection path based on the location coordinates and vegetation index of the tobacco cluster target, it includes: constructing a weighted graph based on the central coordinates of each tobacco cluster, each node in the weighted graph corresponds to the central coordinates of a tobacco cluster, and the edge weight between any two nodes is determined by real-time objective factors: real-time objective factors include the Euclidean distance between nodes, the difference in vegetation index between the areas corresponding to the nodes, and the average wind speed and average light intensity values ​​of the overlapping areas on the edge path. Applying an improved path planning algorithm to the weighted graph to optimize the path, the optimization goal is to minimize the weighted sum of the total path length and the inspection risk weight, and the inspection risk weight is a penalty factor assigned based on whether the wind speed on the edge path is greater than a preset first threshold and whether the light intensity is less than a second threshold. Generate an inspection path connecting the central coordinates of all tobacco clusters to obtain an initial inspection path.

[0109] The weighted graph is constructed by using the coordinates of tobacco plant cluster centers as nodes, and the edge weights between nodes are dynamically calculated using Euclidean distance, vegetation index differences, wind speed, and light intensity. The improved path planning algorithm uses a heuristic search strategy and combines a penalty factor to adjust path weights. For example, when the average wind speed on an edge path exceeds a threshold, the penalty factor increases the edge weight, causing the algorithm to prioritize low-wind-speed paths. The calculation of inspection risk weights further incorporates weighting coefficients. For example, the weighted sum of total path length and risk weight can be expressed as a linear combination of α × total path length + β × risk weight, where α and β are dynamically adjusted based on actual needs.

[0110] Specifically, when constructing a weighted graph, the center coordinates of each tobacco cluster serve as nodes, and the edge weights between nodes are calculated by integrating multiple factors. For example, the edge weight between node A and node B can be expressed as: Weight = Euclidean distance × distance weight + vegetation index difference × vegetation weight + wind speed penalty factor × wind speed weight + light penalty factor × light weight. The improved path planning algorithm uses a variant of the Dijkstra algorithm to achieve path optimization by dynamically adjusting edge weights. When the average wind speed of a path exceeds a threshold, the edge weight of that path is assigned a high penalty factor, such as a 50% increase in weight, thereby reducing the probability of the algorithm selecting that path. When generating the initial inspection path, the algorithm traverses all possible path combinations and selects the path with the smallest total weight as the final result. This ensures that the initial inspection path is short while avoiding areas of high wind speed and low light, improving flight safety and image acquisition quality.

[0111] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0112] A weighted graph is constructed based on the central coordinates of each tobacco cluster. Each node in the weighted graph corresponds to the central coordinates of a tobacco cluster. The edge weight between any two nodes is determined by real-time objective factors: real-time objective factors include the Euclidean distance between nodes, the difference in vegetation index values ​​in the areas corresponding to the nodes, and the average wind speed and average light intensity values ​​in the overlapping areas on the edge paths.

[0113] An improved path planning algorithm is applied to the weighted graph for path optimization. The optimization goal is to minimize the weighted sum of the total path length and the inspection risk weight. The inspection risk weight is a penalty factor assigned according to whether the wind speed on the edge path is greater than a preset first threshold and whether the light intensity is less than a second threshold.

[0114] Generate an inspection path connecting the center coordinates of all tobacco plant clusters to obtain an initial inspection path.

[0115] Specifically, first construct a weighted graph based on the center coordinates of the tobacco clusters. For example, suppose there are 5 tobacco clusters, whose center coordinates are A(1,2), B(3,4), C(5,6), D(7,8), and E(9,10). Calculate the Euclidean distance between any two nodes, such as the AB distance is 2.83, the AC distance is 5.66, and so on. Then obtain the vegetation index of the area corresponding to each node, such as 0.6 for area A, 0.7 for area B, and so on, and calculate the difference value. Then obtain the average wind speed and light intensity on the edge path, such as the average wind speed on the AB path is 3m / s, and the average light intensity is 800lux.

[0116] Furthermore, an improved ant colony algorithm was applied to the weighted graph for path optimization. The total path length was weighted at 0.6, and the inspection risk weight was set at 0.4. The first threshold was assumed to be a wind speed threshold of 5 m / s, and the second threshold was assumed to be a light intensity threshold of 500 lux. A penalty factor of 1.5 was applied to path segments where the wind speed exceeded 5 m / s or the light intensity was below 500 lux. After multiple iterations of optimization, the optimal inspection path was ultimately determined to be ACEDB.

[0117] Thus, an initial inspection path ACEDB connecting the center coordinates of all tobacco plant clusters is generated as the flight path of the UAV.

[0118] Through the above technical solution, this application can generate an optimized inspection route based on the distribution of tobacco clusters and real-time environmental factors. By considering inter-node distances, vegetation index differences, wind speed, and lighting conditions, unfavorable flight areas are avoided, improving inspection efficiency and data collection quality. Furthermore, a path optimization algorithm balances total path length and inspection risk, achieving overall optimization of the inspection route and providing more reasonable and efficient flight planning for drones.

[0119] In some of the above-mentioned solutions of this application, the flight path of the drone may deviate due to environmental interference or positioning errors during the inspection process, affecting the image acquisition quality and inspection efficiency, and it is necessary to monitor the flight status in real time and make timely adjustments.

[0120] This application further proposes a solution for collecting the flight posture, environmental parameters and offset angle between the target drone and the current path segment in real time during flight, including the joint measurement of the drone's flight posture based on an inertial measurement unit, a GPS module and visual inertial SLAM. The flight posture includes position coordinates and a flight heading vector. Environmental data is collected synchronously, including wind speed, light, temperature and humidity, and the flight posture of the drone is constructed at an angle with the fitting vector of the current path segment to obtain the offset angle. The offset angle is calculated using the cosine similarity formula, and the angle between the fitting vector of the current path segment and the flight heading vector is used to determine whether it exceeds the angle threshold.

[0121] The flight pose is calculated using a multi-sensor fusion approach. The inertial measurement unit provides acceleration and angular velocity data, the GPS module provides global coordinate information, and visual-inertial SLAM uses the camera and inertial data to construct a local map and correct positioning errors. Environmental parameters are collected using onboard meteorological sensors, including a three-cup anemometer, a photosensor, and a temperature and humidity probe. The deviation angle is calculated using vector space relationship analysis, defining the fitted vector of the current path segment as the direction vector from the start point to the end point. The flight heading vector is determined by the yaw angle output by the attitude sensor.

[0122] Specifically, the joint measurement process of the flight posture uses an extended Kalman filter to fuse multi-source data and output high-frequency position coordinates and heading angle data. Environmental parameters are recorded synchronously with the flight posture data at a fixed sampling frequency to form a time-space aligned data stream. The cosine similarity calculation of the offset angle is implemented using the vector dot product formula, specifically the dot product of two vectors divided by the product of their modulus lengths. When the calculation result is lower than the preset threshold, the angle offset is judged to be out of limit. For example, the path segment fitting vector is (0.8, 0.6), the flight heading vector is (0.6, 0.8), and the cosine similarity calculation result is 0.96, corresponding to an angle of approximately 15 degrees. When it is lower than the 30-degree threshold, it is judged to be normal. When the ambient wind speed exceeds 5m / s or the light intensity is lower than 10,000 lux, the system automatically increases the sampling frequency of the posture data to 50Hz to enhance control accuracy.

[0123] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0124] The following steps are included when collecting the target UAV's flight posture, environmental parameters, and the offset angle between the UAV and the current path segment in real time during flight:

[0125] First, the drone's flight posture is calculated using a combination of an inertial measurement unit (IMU), a GPS module, and visual-inertial SLAM. Specifically, the IMU provides acceleration and angular velocity data, the GPS module provides global positioning information, and the visual-inertial SLAM system performs feature matching and tracking using image sequences captured by the camera. This data is fused using a Kalman filter to produce highly accurate position coordinates and flight heading vectors.

[0126] Secondly, environmental data is collected simultaneously. This includes wind speed, light intensity, and temperature and humidity. Wind speed is measured by an onboard anemometer, light intensity is collected by a light sensor, and temperature and humidity are acquired by a temperature and humidity sensor. These environmental parameters have a significant impact on the drone's flight stability and image acquisition quality.

[0127] Finally, the angle between the UAV's flight posture and the fitting vector of the current path segment is constructed to obtain the offset angle. The specific steps are as follows:

[0128] Calculates the direction vector of the path segment based on the start and end point coordinates of the current path segment.

[0129] Extract the flight heading vector of the drone.

[0130] The cosine similarity formula is used to calculate the angle between the two vectors, which is the offset angle.

[0131] The calculated offset angle is compared with the preset angle threshold to determine whether it exceeds the allowable range.

[0132] Through the above technical solution, this application realizes real-time and accurate monitoring of the UAV flight status. This can promptly detect the situation where the UAV deviates from the planned path, providing a basis for subsequent path adjustment and flight attitude control. Furthermore, the synchronous collection of environmental parameters helps to analyze the impact of external factors on flight stability, improving the reliability of inspection tasks and the quality of data collection. In addition, the real-time calculation of the offset angle enables the system to quickly respond to deviations in the flight trajectory, thereby ensuring the accurate execution of the inspection path and comprehensive coverage of the target area.

[0133] In some of the above-mentioned schemes in this application, the drone collects flight posture, environmental parameters and offset angle in real time during the inspection process, but it is not clear how to dynamically adjust the path and posture when the offset angle exceeds the threshold and the distance is close to the target cluster. As a result, the drone may be unable to stably perform the inspection task due to sudden environmental interference, affecting the image acquisition quality and task efficiency.

[0134] The present application further proposes that when the offset angle is greater than the angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than the set distance threshold, the judgment process includes: comparing the Euclidean spatial distance between the position coordinates of the UAV and the center coordinates of the current target cluster with the distance threshold, and comparing the offset angle with the angle threshold. If the offset angle is greater than the angle threshold and the spatial distance is less than the distance threshold, the circular fixed-point track of the current cluster is skipped and the center coordinates of the next cluster are directly turned to generate an improved path. If the offset angle is greater than the angle threshold and the spatial distance is greater than the distance threshold, the heading angle and flight speed are adjusted to return to the current path segment, and the image acquisition of the current target cluster is suspended, and the target observation is reactivated after the attitude stabilizes. If the offset angle is less than or equal to the angle threshold, the yaw angle and flight side thrust are adjusted to return the UAV to the center line of the path.

[0135] The distance threshold is set based on the tobacco cluster's boundary and the drone's turning radius, while the angle threshold is determined by the drone's flight attitude adjustment capabilities and the influence of ambient wind speed. When the drone approaches the target cluster, if the offset angle is too large and the distance is close, it indicates that the drone has deviated from the preset path and cannot adjust to the surrounding observation state in time. In this case, skipping the current cluster can avoid ineffective hovering or repeated adjustments. If the distance is far, the original path is prioritized to maintain overall inspection continuity. Pausing image acquisition prevents image blur caused by unstable attitude. Once the heading stabilizes, reactivating observation ensures data validity.

[0136] Specifically, the drone calculates the offset angle from the current path segment in real time during flight, and obtains precise spatial coordinates through GPS and SLAM fusion positioning. When it is detected that the offset angle exceeds the preset threshold, the adjustment strategy is determined in combination with the distance from the target cluster. For example, when the wind speed suddenly changes and the drone deviates from the path and is only 5 meters away from the center of the target cluster, an improved path to the next cluster is directly generated, and a cubic Bezier curve is used for smooth steering. At the same time, the camera angle is adjusted to compensate for the observation angle deviation. If there is still 20 meters away from the center of the target cluster and the offset angle is large, the flight speed is reduced and the heading angle is adjusted to gradually return to the original path, and image acquisition is suspended until the posture is stable. As a result, the drone can still efficiently complete the inspection task under complex environmental interference, reduce path duplication and data redundancy, and improve the accuracy and timeliness of anomaly detection.

[0137] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0138] The Euclidean distance between the drone's position and the center of the current target cluster is compared to a distance threshold, and the offset angle is compared to an angle threshold. The distance threshold can be set to 5 meters, and the angle threshold can be set to 15 degrees.

[0139] If the deviation angle is greater than 15 degrees and the spatial distance is less than 5 meters, the current cluster's circular point track is skipped, and the next cluster's center coordinate is directly turned to generate an improved path. Specifically, a Bezier curve can be used to generate a smooth turning path, and the camera angle can be adjusted to compensate for the observation of the skipped area.

[0140] If the deviation angle is greater than 15 degrees and the spatial distance is greater than 5 meters, the heading angle and flight speed are adjusted to return to the current path segment, and the image collection of the current target cluster is paused. After the attitude is stabilized, the target observation is reactivated. For example, the flight speed can be reduced to 2 meters / second, and the heading angle can be adjusted to be less than 5 degrees with the predetermined path.

[0141] If the deviation angle is less than or equal to 15 degrees, the yaw angle and lateral thrust are adjusted to return the UAV to the path centerline. The yaw angle and lateral thrust can be adjusted in real time by a PID control algorithm to control the lateral deviation of the UAV from the path centerline to within 0.5 meters.

[0142] Through the above technical solutions, the present application can dynamically adjust the inspection path and observation strategy according to the actual flight state of the UAV, improving the adaptability and completion of the inspection task. At the same time, by setting reasonable angle and distance thresholds, the flight trajectory is optimized under the premise of ensuring observation quality, reducing unnecessary track adjustment, and improving inspection efficiency. In addition, different handling strategies are adopted for different deviations, which not only ensures the observation coverage of key areas, but also avoids the increase in energy consumption and flight instability caused by frequent adjustments, thereby achieving efficient completion of the inspection task and reliable collection of observation data.

[0143] In some of the above schemes of the present application, when the UAV deviates from the threshold due to environmental interference during the inspection process and approaches the current target cluster, the traditional straight turning path can cause sudden changes in flight attitude and deviation in image collection angle, and cannot effectively maintain the observation coverage of the skipped cluster.

[0144] The present application further proposes a main path connecting the current UAV position coordinates and the center coordinates of the next target cluster, a smooth turning path is generated by fitting a cubic Bezier curve, an improved path is obtained, and the camera angle is adjusted by calculating the vector direction of the current UAV and the center point of the skipped cluster.

[0145] The fitting of the cubic Bezier curve realizes the gradual change of the path curvature by setting two control points, and the control point position is dynamically adjusted based on the current speed and turning angle of the UAV. The camera angle adjustment calculates the azimuth and pitch angle of the center point of the skipped cluster to drive the gimbal to rotate to the target direction.

[0146] Specifically, when the path improvement condition is triggered, the path planning module generates a straight main path based on the current UAV coordinates and the next cluster center coordinates, and introduces cubic Bezier curve control points based on this straight line. The position of the control point is calculated according to the current flight speed of the UAV and the steering requirements. For example, when the speed is 3m / s, the first control point is set 1.5m behind the starting end of the straight path, and the second control point is set at 30% of the length of the front path to form a smooth transition curve. The flight control system solves the heading angle and roll angle instructions in real time according to the curve equation, so that the UAV can fly at a speed of 0.5m / s 2 At the same time, the vision module obtains the geographic coordinates of the center point of the skipped cluster, calculates its azimuth and pitch angle relative to the drone through coordinate conversion, and drives the camera gimbal to rotate to the target direction at an angular velocity of 15 degrees per second, ensuring continuous collection of image data of the skipped cluster during the path switching process.

[0147] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0148] When the drone's offset angle exceeds a threshold and the spatial distance between it and the center of the current target cluster is less than a set distance threshold, an improved path is automatically generated for the drone to the next target cluster, and the drone's flight attitude and camera angle are adjusted. Specifically, a smooth steering path is generated using a cubic Bezier curve fit, using the line connecting the current drone's position coordinates and the center coordinates of the next target cluster as the primary path. This improved path is then obtained by simultaneously calculating the vector direction between the current drone and the center of the skipped cluster, and adjusting the camera angle.

[0149] For example, the current position of the drone is (x1, y1, z1), and the center coordinates of the next target cluster are (x2, y2, z2). First, construct the main path vector V = (x2-x1, y2-y1, z2-z1). Then, select two control points P1 and P2, located at 1 / 3 and 2 / 3 of the main path vector, respectively. Using these four points (current position, P1, P2, and target position), a cubic Bezier curve is constructed to obtain a smooth, improved path.

[0150] Furthermore, the vector direction W = (x0-x1, y0-y1, z0-z1) between the current position of the drone and the center point of the skipped cluster (x0, y0, z0) is calculated. According to the direction of W, the pitch and yaw angles of the camera are adjusted to point to the center of the skipped cluster. For example, the pitch angle of the camera can be set to arctan((z0-z1) / sqrt((x0-x1) 2 +(y0-y1) 2 )), the yaw angle is set to arctan2(y0-y1,x0-x1).

[0151] Thus, the unmanned aerial vehicle can maintain observation of the skipped area while skipping the current cluster, ensuring the continuity and integrity of the inspection.

[0152] Through the above technical solutions, the present application realizes dynamic optimization of the unmanned aerial vehicle inspection path and intelligent adjustment of the camera angle. When the unmanned aerial vehicle deviates from the predetermined path, the system can quickly generate a smooth improved path, avoiding the impact of sharp turns on flight stability. At the same time, by adjusting the camera angle, continuous observation of the skipped area is ensured, improving the efficiency and coverage of the inspection. This method not only optimizes the flight trajectory, but also ensures the continuity of data collection, effectively improving the flexibility and adaptability of the unmanned aerial vehicle inspection of tobacco fields.

[0153] In some of the above schemes of the present application, when identifying tobacco pests and diseases based on image data collected during flight, existing methods usually use general image classification models for single-frame analysis, lacking multi-task joint identification capability for complex crop abnormalities, such as being unable to handle different types of abnormalities such as disease spots, pests, and water stress at the same time. In addition, traditional methods do not accurately associate the identification results with the spatial position of the tobacco plant cluster, making it difficult to form an effective closed-loop management with the cloud-based agricultural management platform.

[0154] The present application further proposes to identify the image data collected during flight based on a tobacco pest and disease identification model, identify whether there is an abnormal situation on the leaf surface, the abnormal situation including disease spots, pests, or water stress, and upload the location of the abnormal situation of the tobacco plant cluster to the cloud platform after labeling.

[0155] Among them, image data preprocessing includes image denoising, color enhancement, and illumination normalization to eliminate the influence of environmental factors on image quality. The tobacco pest and disease identification model uses a main extraction network and several branch sub-task heads. The main extraction network outputs a semantic segmentation map of the disease spot area, and the branch sub-task heads respectively output multi-class classification results of pest types and heat spot positioning maps of leaf water stress areas. According to the pixel position of the abnormal area in the identification result and the flight pose when the image is collected, combined with the center point coordinates of the current tobacco plant cluster, spatial inverse calculation labeling is performed to generate structured abnormal information containing abnormal type, spatial position, and image timestamp.

[0156] Specifically, the preprocessing stage eliminates sensor noise by image denoising, enhances the contrast of leaf surface details by color enhancement, and eliminates color difference interference under different light conditions by light normalization. The preprocessed image is input into the backbone extraction network, which extracts multi-scale features through convolution layers and outputs a pixel-level semantic segmentation map of the lesion area. The branch sub-task head is connected to different levels of the main network: the pest classification sub-task head classifies pest types through a fully connected layer, and the water stress sub-task head locates the leaf thermal spot area through a heat map regression. After identification, according to the pixel coordinates of the abnormal area in the image, combined with the latitude, longitude, height and camera parameters in the flight pose of the unmanned aerial vehicle, the actual geographic coordinates of the abnormal area in the tobacco field are calculated through spatial geometric transformation. The coordinates are superimposed with the corresponding tobacco cluster center point coordinates to generate structured data containing abnormal type, geographic coordinates and collection time, which is directly uploaded to the cloud platform for visual labeling and agricultural decision support.

[0157] As a preferred embodiment, the scheme of the application is implemented as follows: when preprocessing the tobacco leaf images collected by the unmanned aerial vehicle, first, Gaussian filtering is used for denoising to eliminate image noise interference, then histogram equalization algorithm is used to enhance the color contrast of the image, and finally, the Retinex theory is used for normalization processing on the unevenly illuminated area. The preprocessed image is input into the pest and disease identification model, the backbone extraction network of the model uses ResNet-50 architecture to extract multi-scale features, the semantic segmentation branch uses U-Net structure to output a binary mask of the lesion area, the classification branch outputs a pest category probability distribution through a fully connected layer, and the thermal spot positioning branch generates a heat map of the leaf water stress area through a deconvolution layer. In the spatial inverse calculation labeling process, the RTK-GPS module carried by the unmanned aerial vehicle is used to obtain the latitude and longitude coordinates and height data at the image collection time, combined with the camera intrinsic parameter matrix to calculate the actual coordinates of the lesion pixel points in three-dimensional space, and finally generate a structured data package containing the lesion area ratio, pest type code, water stress heat map path and corresponding geographic coordinates, which is uploaded to the cloud platform through the 4G communication module.

[0158] Through the above technical scheme, the application effectively solves the problem that the traditional single-task model cannot identify multiple abnormal types simultaneously, and through multi-branch joint analysis, realizes the synchronous processing of lesion segmentation, pest classification and water stress detection, and improves the integrity and efficiency of abnormal identification. At the same time, combined with the flight pose data and spatial coordinate inverse calculation technology, the abnormal area is accurately mapped to the geographic coordinates of the tobacco cluster, forming traceable structured data, providing a directly callable data basis for subsequent precise pesticide application and farmland management.

[0159] In some of the above-mentioned solutions of this application, there is a risk of mission interruption due to insufficient power management during the drone inspection process. The existing technology has not established a real-time power monitoring and path demand matching mechanism, which may cause a crash or data collection failure due to insufficient remaining power to complete the remaining inspection path.

[0160] This application further proposes to collect the remaining power of the drone in real time, obtain the power demand based on the remaining path, compare the remaining power with the power demand, and if the difference between the remaining power of the drone and the power demand is less than the power threshold, control the drone to return to the nest autonomously for charging.

[0161] The remaining battery power is calculated in real time using voltage sensors and a current integration algorithm. Power requirements are dynamically predicted based on the length of each remaining route segment and a preset flight power consumption model. The battery threshold is set based on the minimum required return power and a safety margin factor to prevent battery depletion during the return journey. When an autonomous return is triggered, the flight control system replans the shortest safe route to the landing zone, suspends the current inspection mission, and saves the breakpoint data.

[0162] Specifically, the battery voltage and discharge current are periodically collected during flight, and the remaining power percentage is calculated after eliminating noise interference through Kalman filtering. The total length of the remaining path is calculated by adding the Euclidean distance of the coordinate sequence of the center points of the uninspected clusters. The flight power consumption is corrected by combining the headwind coefficient and the climb altitude to generate a power demand forecast. When the difference between the remaining power and the required power is lower than the preset threshold, the inspection mission is immediately terminated, and the A* algorithm is called to generate an obstacle avoidance return path. The drone is controlled to return to the nest along the path for charging. At the same time, the unfinished path segments and abnormal annotation data are uploaded to the cloud platform. After charging is completed, the task breakpoint is reloaded to continue execution.

[0163] As a preferred embodiment, the solution of the present application is specifically implemented as follows: During the flight of the drone, the percentage value of the current remaining power is obtained in real time through the built-in battery management system, and at the same time, the path planning module estimates the power consumption value required to complete the remaining inspection tasks based on the remaining flight length of the initial inspection path and the flight speed. When the difference between the remaining power and the required power is lower than the preset critical threshold, the navigation control module of the drone immediately terminates the current inspection task, starts the autonomous return program, generates the shortest path from the current position back to the machine nest, and adjusts the flight altitude and speed to reduce energy consumption during the return process. Among them, a wireless charging device is configured inside the machine nest. After the drone lands on the charging platform, the charging process is automatically triggered until the battery is restored to the preset power level.

[0164] Through the above technical solution, this application effectively avoids the problem of drone inspection missions being interrupted due to insufficient remaining power. By dynamically monitoring the matching relationship between the remaining power and the mission requirements, the return charging mechanism is triggered in time to ensure that the drone completes the inspection operation within a safe power range. At the same time, through the closed-loop management of autonomous return and charging, the continuity and reliability of the system operation are improved.

[0165] In some of the above-mentioned schemes in this application, a method of realizing drone inspection through multi-source remote sensing image data and dynamic path planning is proposed. However, at the system implementation level, the existing technology lacks modular design, resulting in low coordination efficiency between data acquisition, path planning, flight control and anomaly recognition, making it difficult to achieve real-time environmental adaptation and closed-loop management. At the same time, an autonomous power management mechanism has not been established, affecting the continuity of the inspection task.

[0166] In the above embodiment, by combining multi-source remote sensing image acquisition with deep learning recognition, accurate identification and spatial clustering of tobacco plants in tobacco-growing areas are achieved, breaking through the accuracy limitations of traditional inspections that rely solely on RGB images and GPS grids. The tobacco plant distribution density and vegetation index information are integrated to generate an initial inspection path, and wind speed and light factors are introduced to optimize the path obstacle avoidance, thereby improving the environmental adaptability and operational safety of path planning. By real-time acquisition of the UAV's flight posture and its offset angle from the inspection path, a dynamic inspection path reconstruction mechanism is constructed, enabling the UAV to have the ability of point-hopping flight and posture-linked shooting, thereby improving flight efficiency and image acquisition stability. A multi-task tobacco pest and disease recognition model is used to jointly identify multi-dimensional abnormalities such as lesions, pests, and water stress in the collected images, and spatial back-labeling is performed based on the flight posture and synchronously uploaded to the cloud, forming an identification-positioning-feedback closed loop, which effectively solves the problems of single identification dimension, fixed path, and feedback chain disconnection in the existing technology.

[0167] In another preferred embodiment based on the above embodiment, refer to Figure 2 As shown, this embodiment provides a drone inspection system for tobacco fields, which is used to apply the above-mentioned drone inspection method for tobacco fields, including:

[0168] The acquisition unit is configured to acquire multi-source remote sensing image data of a preset tobacco planting area, detect tobacco plants in the remote sensing image data based on a target recognition network, and spatially cluster the detection results based on tobacco plant distribution density to obtain a plurality of tobacco plant cluster targets.

[0169] The processing unit is configured to generate an initial patrol path based on the location coordinates and vegetation index of the tobacco cluster target. The initial patrol path connects several tobacco cluster center points, and the initial patrol path avoids areas where the wind speed exceeds the first threshold and areas where the light intensity is less than the second threshold in the wind speed prediction map.

[0170] The judgment unit is configured to control the target UAV to fly along the initial inspection path, and collect the flight posture, environmental parameters and offset angle between the target UAV and the current path segment in real time during the flight.

[0171] The judgment unit is also configured to automatically generate an improved path from the target UAV to the next target cluster and adjust the flight posture and camera angle of the target UAV when the offset angle is greater than the angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than the set distance threshold.

[0172] The recognition unit is configured to identify the image data collected during the flight based on the tobacco pest and disease recognition model, identify whether there are abnormalities on the leaf surface, such as lesions, pests or water stress, and mark the locations of the abnormal tobacco plant clusters and upload them to the cloud platform.

[0173] The acquisition unit improves tobacco plant detection accuracy by fusing multi-source remote sensing data. The processing unit optimizes the route by combining vegetation indices and environmental parameters. The judgment unit dynamically corrects the trajectory using real-time pose and offset angles. The recognition unit employs a multi-task model for joint anomaly identification. The power management module ensures mission continuity by predicting remaining battery power and route demand. Each unit interacts with control signals through data streams. For example, the route generated by the processing unit is dynamically adjusted by the judgment unit, and the output of the recognition unit triggers cloud platform annotation and subsequent agricultural decision-making.

[0174] Specifically, during system execution, the acquisition unit acquires multispectral and depth map data, extracts features and segments tobacco plants using a two-branch convolutional network, and outputs tobacco plant cluster targets through morphological processing and connected domain analysis. The processing unit constructs a weighted graph, using Euclidean distance, vegetation index difference, wind speed, and light intensity as edge weights, and applies an improved path planning algorithm to generate an initial path. The judgment unit calculates the offset angle in real time based on the fusion of inertial measurement unit and visual SLAM data. If the offset exceeds the limit and approaches the target cluster, it generates a smooth steering path using cubic Bezier curves, while simultaneously adjusting the camera angle to compensate for observational gaps. The recognition unit performs lesion segmentation, pest classification, and moisture hotspot location on the preprocessed images. Based on the flight posture, it inversely calculates the location of anomalies and uploads structured data. The power management module continuously monitors the remaining battery power and triggers a return command when insufficient to complete the remaining path, ensuring the drone's safe return to the nest for recharging. These units collaborate to achieve environmentally adaptive inspections, precise anomaly identification, and closed-loop data management, improving the efficiency and reliability of tobacco field inspections.

[0175] As a preferred embodiment, the solution of the present application is specifically implemented as follows: the drone inspection system includes an acquisition unit, a processing unit, a judgment unit, and an identification unit. The acquisition unit is configured to obtain RGB images, multispectral images, and depth maps of the tobacco planting area, and perform joint feature extraction on the images through a dual-branch deep convolutional neural network, wherein the RGB image and the multispectral image are respectively input into the convolution layer to extract texture features and vegetation features, and the pixel-level segmentation results of the tobacco plants are output after weighted fusion by the spatial attention mechanism, and the coordinates of the tobacco plant bounding box are extracted after removing noise through morphological filtering. The processing unit receives the tobacco plant coordinate set and uses the DBSCAN algorithm for spatial clustering, and constructs a weighted graph based on the clustering results, wherein the node weights are comprehensively calculated by Euclidean distance, vegetation index difference, and environmental parameters, and generates an initial inspection path connecting all cluster center points through an improved path planning algorithm, which automatically avoids areas where the wind speed exceeds a preset value. The judgment unit calculates the drone's position in real time by fusing GPS and visual SLAM data. When it detects that the deviation angle between the flight heading and the path segment exceeds a threshold and approaches the current target cluster, it uses cubic Bezier curves to generate an improved path around high-wind speed areas. It also adjusts the gimbal pitch angle to align the camera with the skipped tobacco cluster. The recognition unit performs multi-task analysis on images collected during flight. The backbone network uses ResNet-50 to extract features, and the three branches output a lesion segmentation map, pest classification results, and moisture hotspot location maps, respectively. Abnormal data is mapped to the geographic coordinates of tobacco clusters through spatial inverse calculation and uploaded to a cloud database.

[0176] Through the above technical solutions, this application realizes the precise positioning of tobacco plants and dynamic path planning based on multi-source remote sensing data, effectively avoids flight risk areas through real-time environmental perception and path correction mechanism, adopts multi-task recognition model to improve the accuracy of complex agricultural anomaly detection, and forms closed-loop management through structured data annotation and cloud collaboration, solving the problems of rigid paths, single recognition capabilities and data silos in traditional inspection methods.

[0177] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0178] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0179] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0180] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for inspecting tobacco fields using drones, characterized in that: include: Acquire multi-source remote sensing image data of a preset tobacco planting area, wherein the remote sensing image data includes RGB images, multispectral images, and depth maps; Detecting tobacco plants in the remote sensing image data based on a target recognition network, and spatially clustering the detection results according to tobacco plant distribution density to obtain a plurality of tobacco plant cluster targets; generating an initial inspection path according to the position coordinates and vegetation index of the tobacco cluster target, wherein the initial inspection path connects a plurality of tobacco cluster center points and avoids areas where the wind speed exceeds a first threshold and where the light intensity is less than a second threshold in the wind speed prediction map; A weighted graph is constructed based on the center coordinates of each tobacco plant cluster, wherein each node in the weighted graph corresponds to the center coordinates of a tobacco plant cluster, and the edge weight between any two nodes is determined by real-time objective factors: the real-time objective factors include the Euclidean distance between nodes, the difference in vegetation index between the areas corresponding to the nodes, and the average wind speed and average light intensity values ​​of the overlapping areas on the edge paths; Applying an improved path planning algorithm to the weighted graph to perform path optimization, where the optimization objective is to minimize the weighted sum of the total path length and an inspection risk weight, where the inspection risk weight is a penalty factor assigned based on whether the wind speed on the edge path is greater than a preset first threshold and whether the light intensity is less than a second threshold; Generating an inspection path connecting the center coordinates of all tobacco plant clusters to obtain the initial inspection path; Control the target UAV to fly along the initial inspection path, and collect the target UAV's flight posture, environmental parameters, and the offset angle between the UAV and the current path segment in real time during the flight; When the offset angle is greater than the angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than the set distance threshold, an improved path from the target UAV to the next target cluster is automatically generated, and the flight attitude and camera angle of the target UAV are adjusted; Comparing the Euclidean space distance between the position coordinates of the UAV and the center coordinates of the current target cluster with a distance threshold, and comparing the offset angle with an angle threshold; If the offset angle is greater than the angle threshold and the spatial distance is less than the distance threshold, the circumferential fixed-point track of the current cluster is skipped and the improved path is generated by directly turning to the center coordinates of the next cluster; If the offset angle is greater than the angle threshold and the spatial distance is greater than the distance threshold, the heading angle and flight speed are adjusted to return to the current path segment, and the image acquisition of the current target cluster is suspended. The target observation is reactivated after the attitude stabilizes; If the offset angle is less than or equal to the angle threshold, adjust the yaw angle and flight side thrust to return the UAV to the path centerline; Automatically generate an improved path from the target drone to the next target cluster and adjust the target drone's flight attitude and camera angle, including: Taking the line between the current UAV's position coordinates and the center coordinates of the next target cluster as the main path, a smooth steering path is generated by cubic Bezier curve fitting to obtain the improved path. At the same time, the vector direction between the current UAV and the center point of the skipped cluster is calculated to adjust the camera angle. The tobacco pest and disease recognition model is used to identify image data collected during flight to identify abnormalities on the leaf surface, such as lesions, pests, or water stress. The locations of abnormal tobacco plant clusters are marked and uploaded to the cloud platform.

2. The method for inspecting tobacco fields using a drone according to claim 1, wherein: Detecting tobacco plants in the remote sensing image data based on a target recognition network includes: Performing joint input processing on the RGB image and the multispectral image based on a deep convolutional neural network model, wherein the deep convolutional neural network model includes a U-Net network with an encoding-decoding structure, and the deep convolutional neural network model integrates a spatial attention mechanism; The RGB image provides texture and edge information, the multispectral image provides vegetation characteristic band information, and the deep convolutional neural network model performs feature splicing after dual-branch feature extraction and feeds the feature into the classifier to output the pixel-level segmentation result of the tobacco plant. The segmentation results are subjected to morphological processing to remove noise points and interference areas, and the bounding box and spatial coordinates of each tobacco plant target are extracted through connected domain analysis.

3. The method for inspecting tobacco fields using a drone according to claim 2, wherein: When the detection results are spatially clustered according to the density of tobacco plant distribution, several tobacco plant cluster targets are obtained, including: Extract the center coordinates of all tobacco plant targets and construct a two-dimensional space coordinate set; The density-based spatial clustering algorithm DBSCAN performs cluster analysis on the coordinate set, wherein the parameters of the spatial clustering algorithm DBSCAN include a distance threshold and a minimum number of neighborhood points, which are used to determine whether the tobacco plant points belong to the same cluster; The distance threshold is dynamically set according to the remote sensing image resolution and the tobacco plant spacing; Each high-density connected area in the clustering result is identified as a tobacco plant cluster target, and the center coordinates of the corresponding cluster, the number of tobacco plants in the cluster, and the boundary range are output.

4. The method for inspecting tobacco fields using a drone according to claim 1, wherein: During flight, the real-time acquisition of the target UAV's flight posture, environmental parameters, and the offset angle between the UAV and the current path segment includes: The flight posture of the UAV is calculated based on the inertial measurement unit, GPS module and visual inertial SLAM. The flight posture includes position coordinates and flight heading vector. Synchronously collect environmental data, including wind speed, light, and temperature and humidity, and construct an angle between the flight posture of the drone and the fitting vector of the current path segment to obtain the offset angle; The offset angle is calculated using the cosine similarity formula, and whether the angle exceeds the angle threshold is determined based on the angle between the fitting vector of the current path segment and the flight heading vector.

5. The method for inspecting tobacco fields using a drone according to claim 1, wherein: Based on the tobacco pest and disease recognition model, the image data collected during the flight is identified to determine whether there are any abnormalities on the leaf surface, including: Preprocessing the image data, wherein the preprocessing includes image denoising, color enhancement, and illumination normalization; The preprocessed image is input into the tobacco pest and disease recognition model, which includes a backbone extraction network and several branch subtask heads. The backbone extraction network is used to output a semantic segmentation map of the diseased spot area, and the branch subtask heads are used to output a multi-class classification result of the pest type and a hot spot location map of the leaf water stress area. According to the pixel position of the abnormal area in the recognition result and the flight posture when the corresponding image was collected, combined with the coordinates of the center point of the current tobacco cluster, spatial back-calculation and annotation were performed to generate structured abnormality information including abnormality type, spatial location and image timestamp.

6. The method for inspecting tobacco fields using a drone according to claim 1, wherein: Also includes: The remaining power of the drone is collected in real time, and the power demand is obtained based on the remaining path. The remaining power is compared with the power demand. If the difference between the remaining power of the drone and the power demand is less than the power threshold, the drone is controlled to return to the nest for charging.

7. A drone inspection system for tobacco fields, used to apply the drone inspection method for tobacco fields according to any one of claims 1 to 6, characterized in that: include: An acquisition unit configured to acquire multi-source remote sensing image data of a preset tobacco planting area; Detecting tobacco plants in the remote sensing image data based on a target recognition network, and spatially clustering the detection results according to tobacco plant distribution density to obtain a plurality of tobacco plant cluster targets; a processing unit configured to generate an initial inspection path based on the position coordinates of the tobacco cluster target and the vegetation index, wherein the initial inspection path connects a plurality of tobacco cluster center points and avoids areas where the wind speed exceeds a first threshold and where the light intensity is less than a second threshold in the wind speed prediction map; a judgment unit configured to control the target UAV to fly along the initial inspection path and collect the flight posture, environmental parameters, and offset angle between the target UAV and the current path segment in real time during the flight; The judgment unit is further configured to, when the offset angle is greater than an angle threshold and the spatial distance between the target UAV and the center point of the current target cluster is less than a set distance threshold, automatically generate an improved path for the target UAV to the next target cluster and adjust the flight attitude and camera angle of the target UAV; The recognition unit is configured to identify the image data collected during the flight based on the tobacco pest and disease recognition model, identify whether there are abnormalities on the leaf surface, such as lesions, pests or water stress, and mark the location of the abnormal tobacco plant clusters and upload them to the cloud platform.

Images