Camellia oleifera tree disease detection method and system based on unmanned aerial vehicle and YOLO algorithm

Abstract

The invention discloses a camellia oleifera disease detection method and system based on an unmanned aerial vehicle and a YOLO algorithm, and the method comprises the steps: enabling the unmanned aerial vehicle to collect the video stream data of a to-be-detected camellia oleifera in real time, inputting the obtained video stream data into an optimized and improved YOLOv11 model, and obtaining a camellia oleifera disease detection result; according to the improved YOLOv11 model, a C2PSA module in a backbone network in an original YOLOv11 model structure is replaced by a C2B module, and an adaptive spatial feature fusion module is introduced between a neck network and a head network in the original YOLOv11 model structure. The method has the advantages of being high in detection precision, high in response real-time performance and accurate in space positioning, the efficiency and precision of camellia oleifera disease control can be remarkably improved, a basis is provided for optimization of a subsequent planting management strategy, and the method has important application value and popularization prospects in the field of intelligent agriculture and forestry disease control.

Description

Technical Field

[0001] This invention belongs to the field of image or video recognition technology, specifically relating to a method and system for detecting diseases in camellia trees based on unmanned aerial vehicles and the YOLO algorithm. Background Technology

[0002] Drone technology has been gradually applied to fields such as crop protection, crop growth monitoring, and yield assessment, achieving relatively mature application results. In camellia oleifera cultivation, diseases can significantly impact the growth and yield of the trees. Traditional camellia oleifera disease detection mainly relies on manual inspections, which suffers from low efficiency, high labor intensity, strong subjectivity, and difficulty in achieving large-scale real-time monitoring, making it difficult to meet the needs of large-scale planting and refined management.

[0003] Currently, although drones are widely used in agriculture, there are still relatively few dedicated technical solutions for detecting diseases in camellia oleifera trees. In particular, there is a lack of mature and effective technical systems for acquiring high-quality image data using drones in complex scenarios and building artificial intelligence models capable of accurately identifying specific diseases in camellia oleifera trees. Therefore, there is an urgent need to propose a drone-based method and system for detecting diseases in camellia oleifera trees to improve the automation and accuracy of disease detection and provide technical support for the precise prevention and control of camellia oleifera diseases. Summary of the Invention

[0004] To address the shortcomings of existing camellia oleifera disease detection technologies, such as low efficiency, insufficient precision in prevention and control, and lack of disease prediction capabilities, this solution proposes a camellia oleifera disease detection method and system based on unmanned aerial vehicles (UAVs) and the YOLO algorithm. This method and system aim to achieve off-site, automated detection of camellia oleifera diseases. While ensuring the accuracy of disease identification and the efficiency of prevention and control, early detection provides a basis for prevention and control planning, thus providing a technical foundation for building a green and intelligent camellia oleifera industry planting and management system.

[0005] A method for detecting diseases in camellia oleifera trees based on UAVs and the YOLO algorithm includes: real-time acquisition of video stream data of camellia oleifera trees to be detected by UAVs; inputting the obtained video stream data into an optimized and improved YOLOv11 model to obtain the disease detection results of camellia oleifera trees; the improved YOLOv11 model is to replace the C2PSA module in the backbone network of the original YOLOv11 model structure with a C2B module, and to introduce an adaptive spatial feature fusion module between the neck network and the head network of the original YOLOv11 model.

[0006] Furthermore, during the detection process, the obtained video stream data of the camellia trees is input into the optimized and improved YOLOv11 model to obtain disease target frame images labeled with disease categories and corresponding bounding box information. The confidence level of the disease target frame images is then filtered to retain high-confidence disease images, generating crop disease detection results.

[0007] Furthermore, a method for detecting diseases in camellia trees based on drones and the YOLO algorithm includes the following steps:

[0008] Step 1: The drone collects video stream data of the camellia trees to be inspected in real time and uploads the collected video stream data to the cloud server;

[0009] Step 2: The cloud server uses the optimized CO_YOLOv11 model (i.e. the improved YOLOv11 model mentioned above) to detect diseases and obtain disease target frame images labeled with disease categories and corresponding bounding box information.

[0010] Step 3: Filter the identified disease target frame images by confidence level, retaining disease images with high confidence; and delete the bounding boxes of targets with low confidence level.

[0011] Preferably, the C2B module consists of parallel CBAM module branches and BiFormer module branches.

[0012] The input features of the C2B module first pass through a simplified and efficient BiFormer routing attention branch (i.e., the BiFormer module branch), thereby capturing the global semantic context and long-distance spatial interactions in the image with extremely low computational overhead, making up for the global perception shortcomings of CBAM. At the same time, the features are input in parallel into a CBAM structure branch (i.e., the CBAM module branch). This branch performs fine calibration on the channel dimension relationship of the feature map through channel attention, and then focuses on the most salient local detail regions through spatial attention, which enhances the representation of the subtle texture of lesions. Finally, the outputs of the two branches are integrated through an adaptive weighting, which fundamentally improves the model's perception and discrimination capabilities in complex scenes.

[0013] Preferably, the adaptive spatial feature fusion module is located between the neck network output and the head network input in the original YOLOv11 model structure.

[0014] This invention introduces an Adaptive Spatial Feature Fusion (ASFF) module in the detection head. Located between the Neck output and the decoupling head, this module aims to address the consistency issue between multi-scale feature layers. Specifically, ASFF adjusts the feature maps from the Neck at three different scales (P3, P4, and P5) to the spatial dimensions corresponding to the current level through upsampling or downsampling operations. Subsequently, it generates adaptive weight matrices for spatial and channel dimensions using learnable parameters to recalibrate and fuse features from different levels. The fused feature map retains multi-scale characteristics and contains rich semantic and detailed information across scales, which is then fed into the decoupling detection head for bounding box regression and class prediction. This design not only effectively alleviates semantic conflicts between multi-scale features but also significantly enhances the model's perception sensitivity and localization accuracy for small target lesions against complex leaf backgrounds.

[0015] Furthermore, the improved YOLOv11 model is trained using the SGD optimizer.

[0016] Furthermore, the dataset used to optimize the improved YOLOv11 model includes a training set and a validation set. During training, the improved YOLOv11 model is trained using the training set. Simultaneously, during each training epoch, the model's precision, recall, and mean average precision (mAP) are evaluated on the validation set. An early stopping strategy is adopted: if the validation set mAP metric does not improve for a set number of consecutive epochs, training is automatically terminated.

[0017] Furthermore, the dataset used to optimize the improved YOLOv11 model is a tea oil leaf disease dataset constructed using artificially induced disease methods.

[0018] Furthermore, the method for constructing the camellia leaf disease dataset is as follows:

[0019] (1) Pathogenic fungi were isolated and extracted from samples of diseased Camellia oleifera leaves in the natural environment and classified to obtain pathogenic fungi whose species were determined;

[0020] (2) The pathogenic fungi obtained by culture were inoculated onto the leaves of healthy Camellia oleifera trees using artificial inoculation techniques to induce disease.

[0021] (3) Collect typical image data of different types of diseases and different disease stages of diseased camellia trees;

[0022] (4) The typical image data obtained are labeled according to the disease category. The labeling information includes the disease category and the corresponding bounding box coordinates. Finally, a data set of tea tree leaf diseases containing information on various tea tree diseases is obtained.

[0023] A method for detecting diseases in camellia oleifera trees based on drones and the YOLO algorithm includes the following steps:

[0024] Step S1: Setting up the disease detection equipment, specifically including:

[0025] Step S1.1: Install a Raspberry Pi 4B embedded module on the DJI Matrice 4T drone platform to complete the Linux system environment configuration and image acquisition and preprocessing program development;

[0026] Step S1.2: Configure the cloud server as a data receiving and storage node to realize real-time data interaction and collaborative processing with the deep learning model inference module.

[0027] Step S2: Disease data collection, specifically including:

[0028] Step S2.1: Induce disease in healthy tea leaves through artificial inoculation;

[0029] Step S2.2: Take images of the diseased leaves and use the Labelimg annotation tool to annotate the disease. The annotation information includes the disease category and its corresponding bounding box coordinates.

[0030] Step S2.3: Divide all labeled camellia disease image datasets into training set, validation set and test set according to a sample size ratio of 7:2:1.

[0031] Step S3: Disease detection model training, specifically including:

[0032] Step S3.1: Based on the YOLOv11n algorithm, improve the feature fusion module of its backbone network and the detection head to enhance the ability to extract fine-grained features in the diseased area of ​​Camellia oleifera, so as to cope with the challenges of light changes, branch and leaf occlusion and small target detection from the perspective of UAV, and thus improve the accuracy of disease detection.

[0033] Step S3.2: Training and Optimizing the Camellia oleifera Disease Detection Model. The improved CO_YOLOv11 model was trained using the SGD optimizer. After training, the model was compared with other models in terms of precision, recall, and mean AP to select the optimal model structure and parameters.

[0034] Step S4: System Testing and Result Saving

[0035] Step S4.1: Model Detection: Upload the video stream information collected during the drone's flight to the cloud server in real time. Use the trained CO_YOLOv11 model to detect defects, and filter the identified defect target frames by confidence, retaining high-confidence defect images;

[0036] Step S4.2: Data archiving: Store disease-related data in a database on a cloud server to provide a complete and reliable data foundation for subsequent analysis of the occurrence patterns of camellia tree diseases, optimization of prevention and control strategies, and precise application of pesticides.

[0037] The core process of the method described in this invention includes: First, using a drone equipped with a high-resolution camera and a Raspberry Pi embedded unit, video stream data of the canopy layer of Camellia oleifera forest is collected along a preset route. Then, based on the improved CO_YOLOv11 model, the collected disease images are detected in real time and filtered with high confidence, thereby accurately identifying the types of diseases such as soft rot, Fiori anthracnose, and leaf spot. Finally, the confirmed disease images and their associated coordinate data are stored in the cloud to support subsequent data analysis.

[0038] A disease detection system for camellia trees based on video stream data and the YOLO algorithm includes:

[0039] Drones use high-definition cameras to collect real-time video stream data of crops to be inspected.

[0040] The control unit receives video stream data collected by the drone, processes the data, and uploads it to the cloud server in real time.

[0041] A cloud server receives and stores video stream data, and inputs the received video stream data into an optimized and improved YOLOv11 model loaded on it to obtain the crop disease detection results.

[0042] Furthermore, it also includes a communication module that provides network access for drones.

[0043] Furthermore, the control unit is a Raspberry Pi. The data processing generally includes compressing video data, etc.

[0044] This invention employs a hardware architecture with a Raspberry Pi as the core control unit, integrating a 4G communication module, and collaborating with a drone to achieve remote data interaction and intelligent operation in complex outdoor environments. This module features rich peripheral interfaces, providing stable hardware support for multi-device connections. A USB connection to the communication module, coupled with a SIM card, provides network access, ensuring real-time data transmission during high-altitude drone flight. Simultaneously, a high-speed communication link is established with the drone via an E-port coaxial cable. This link supports bidirectional data transmission, satisfying both low-latency control command delivery and stable transmission of large-capacity video data. The execution unit utilizes a DJI Matrice 4T drone, which possesses characteristics suitable for outdoor operations, such as long endurance and anti-interference capabilities, and can adapt to the complex terrain of camellia oil forests. During operation, the drone uses a high-definition camera to capture real-time images of the camellia oil forest canopy, and the data is transmitted to the Raspberry Pi via the E-port interface. After receiving the data, the Raspberry Pi compresses the images to reduce bandwidth consumption, and then uploads the processed data to the cloud server in real-time via the communication module.

[0045] The camellia oleifera disease detection method integrating UAV technology and YOLO algorithm provided by this invention has the advantages of high detection accuracy, strong real-time response, and accurate spatial positioning. It can significantly improve the efficiency and accuracy of camellia oleifera disease prevention and control, provide a basis for optimizing subsequent planting and management strategies, and has important application value and promotion prospects in the fields of smart agriculture and forestry disease prevention and control. Attached Figure Description

[0046] Figure 1 A flowchart illustrating the disease detection process for camellia oleifera trees under complex planting conditions;

[0047] Figure 2 Example diagram of setting up a disease detection equipment for camellia trees;

[0048] Figure 3 An example diagram illustrating the culture and inoculation of pathogenic bacteria in Camellia oleifera.

[0049] Figure 4 Example diagram of diseases affecting camellia trees;

[0050] Figure 5 Example diagram of CO_YOLOv11 model;

[0051] Figure 6 Example image of CO_YOLOv11 detection results. Detailed Implementation

[0052] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0053] like Figure 1 As shown, a method for detecting diseases in camellia trees based on drones and the YOLO algorithm includes the following steps:

[0054] Step S1: Setting up disease detection equipment

[0055] Step S1.1: This plan addresses the practical needs of ecological monitoring and precise disease control in camellia oleifera forests, specifically by establishing the following structure: Figure 2 As shown, it mainly includes Raspberry Pi, drones, cloud servers, 4G communication modules, and E-port coaxial cables.

[0056] This implementation scheme employs a hardware architecture with a Raspberry Pi as the core control unit, integrating a 4G communication module, and collaborating with a drone to achieve remote data interaction and intelligent operation in complex outdoor environments. This module features rich peripheral interfaces, providing stable hardware support for multi-device connections. The Raspberry Pi connects to the 4G communication module via USB; inserting a SIM card provides network access, ensuring real-time data transmission during high-altitude drone flight. Simultaneously, a high-speed communication link is established with the drone via an E-port coaxial cable. This link supports bidirectional data transmission, satisfying both low-latency control command delivery and stable transmission of large-capacity video data. The execution unit uses a DJI Matrice 4T drone, which possesses characteristics suitable for outdoor operations such as long endurance and anti-interference capabilities, adapting to the complex terrain of camellia oil forests. During operation, the drone uses a high-definition camera to capture real-time images of the camellia oil forest canopy, and the data is transmitted to the Raspberry Pi via the E-port interface and E-port coaxial cable. After receiving the data, the Raspberry Pi compresses the images to reduce transmission bandwidth consumption, and then uploads the processed data to the cloud server in real-time via the 4G communication module.

[0057] Step S1.2: The core of cloud server configuration is establishing the environment for video stream reception and the YOLO algorithm. Specifically, this includes deploying and configuring a MediaMTX streaming media server on the cloud server, opening RTSP / RTMP ports, and receiving real-time video streams from drones. Simultaneously, configuring the YOLOv11 target detection environment involves installing Python, PyTorch, and YOLO dependencies, and loading a disease detection-specific model, namely the improved YOLOv11 model (denoted as CO_YOLOv11 model).

[0058] Step S2: Disease Data Collection

[0059] Step S2.1: Due to the scarcity of research data on leaf diseases in Camellia oleifera, obtaining high-quality image datasets is quite difficult. In order to meet the needs of deep learning model training for large-scale, high-quality labeled data, this study uses artificially induced diseases to construct the dataset.

[0060] The specific steps are as follows:

[0061] First, pathogenic fungi were isolated and extracted from leaf samples of diseased Camellia oleifera trees in their natural environment. The diseased leaf parts were disinfected with 75% alcohol and rinsed with sterile water before being inoculated into a culture medium. The samples were then placed in a constant temperature incubator for cultivation. Mycelia at the colony edges were picked for preliminary isolation. The obtained fungi were then purified and cultured. Simultaneously, the cultivation location, time, and fungal species were recorded on the culture dishes using a marker. Figure 3 As shown in (a), Fiorio anthrax bacillus (LKY2) was cultured at the Zhejiang Academy of Forestry Sciences on July 23, 2025; after the colony stabilized, its species classification was accurately identified by DNA sequence alignment technology.

[0062] Secondly, the cultured pathogenic fungi were inoculated onto healthy camellia leaves using artificial inoculation techniques to induce disease. Healthy camellia leaves were selected, and the wound inoculation method was employed. First, the leaf surface was wiped clean with alcohol, then the leaf surface was punctured with a sterile needle, and the cultured fungal inoculum was inoculated into the damaged area of ​​the leaf. Figure 3 As shown in (b); then wrap and seal with a sterile moisturizing film, as shown in Figure (b). Figure 3 As shown in (c); after 2 weeks of cultivation, the moisturizing film was removed, and the disease status of the leaves was observed.

[0063] Step S2.2: Record the entire development process of disease symptoms using a camera system, collecting typical image data at different stages of the disease, resulting in a total of 2580 images. For example... Figure 4 The image shows three different types of diseases collected, among which... Figure 4 (a) in the text represents soft rot. Figure 4 (b) in the image represents Fiori anthrax. Figure 4 (c) in the image represents leaf spot disease. After the images were taken, they were categorized and organized according to the disease type. The dataset was then manually annotated using the professional image annotation tool LabelImg, with annotation information including the disease type and the corresponding bounding box coordinates. Through these steps, a complete and accurate image dataset of camellia leaf diseases was constructed, providing data support for subsequent research on intelligent disease identification and prevention.

[0064] Step S2.3: Randomly divide all labeled camellia leaf disease image datasets into training set, validation set and test set according to a sample size ratio of 7:2:1.

[0065] Step S3 involves constructing a disease detection model for camellia leaves, specifically including:

[0066] Step S3.1: This study uses the CO_YOLOv11 model to detect leaf diseases in Camellia oleifera. CO_YOLOv11 is an improved version of YOLOv11, derived from the existing YOLOv11 model. Its model structure is as follows: Figure 5 As shown ( Figure 5 In this diagram, SPPF represents the fast spatial pyramid pooling module; C3k2 is an improved convolutional feature extraction module introduced in YOLOv11. To address the complex and diverse characteristics of tea oleifera leaf disease features under complex environments, and their varying scales, the original YOLOv11 network structure was specifically improved.

[0067] In the feature extraction backbone network, we removed the original C2PSA module and introduced the CBAM and Biformer modules, which were combined into the C2B module. The C2B module was then integrated into the YOLOv11 model.

[0068] Traditional CBAM modules, through a series of channel attention modules and spatial attention modules, can effectively enhance the model's attention to the target object. The relevant formulas are as follows:

[0069]

[0070] The formula is as follows:

[0071]

[0072]

[0073] The parameters are explained below:

[0074] F represents the input feature map. Where C, H, and W represent the number of channels, height, and width of the image, respectively;

[0075] F' represents the feature map output after processing by the channel attention module;

[0076] F” represents the feature map output after processing by the spatial attention module;

[0077] This indicates element-wise multiplication;

[0078] M C ( ) represents a one-dimensional convolution of the channel attention module. ;

[0079] M S ( ) represents a two-dimensional convolution of the spatial attention module. ;

[0080] ( ) represents the Sigmoid activation function;

[0081] MLP ( ) represents a multilayer perceptron;

[0082] This represents a 7×7 convolution operation;

[0083] AvgPool and MaxPool represent global average pooling and max pooling operations, respectively.

[0084] This step allows us to fully capture the key information within the features.

[0085] BiFormer, through its dynamic and sparse two-layer routing attention mechanism, can efficiently compute global context information and model long-distance spatial relationships, as shown in the following formula:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] The parameters are explained below:

[0092] X is the input feature map, which is the same as feature map F. The feature map X is divided into S×S non-overlapping regions using the Unfold() function, such that each region contains H×W×S regions. 2 The feature vector is then transformed using the Reshape() function. ;

[0093] W q W k W v These are learnable weight matrices, used to weight the input features X respectively. r The mapping is a query (Q), key (K), value (V) tensor;

[0094] To query the degree of association between each key; Q r ,K r Let Q and K be the average values ​​of the token dimension in each region, and we obtain the region-level aggregated representation;

[0095] To For each row, select the k most relevant regions; topkIndex() represents the operation of retrieving the top-k indices;

[0096] K g V g The key-value tensors are aggregated. `gather()` represents the tensor aggregation operation based on indices, O represents the final output, `Attention()` represents the attention calculation, and `LCE()` represents the local context enhancement module, which compensates for local information that may be lost due to sparse attention. Global information can be captured through BiForme.

[0097] The input features of the C2B module first pass through a simplified and efficient BiFormer routing attention branch, thereby capturing the global semantic context and long-distance spatial interactions in the image with extremely low computational overhead, making up for the global perception shortcomings of CBAM. At the same time, the features are input in parallel into a CBAM structure branch, which performs fine calibration on the channel dimension relationship of the feature map through channel attention, and then focuses on the most salient local detail regions through spatial attention, enhancing the representation of the subtle texture of lesions. Finally, the outputs of the two branches are integrated through an adaptive weighting, fundamentally improving the model's perception and discrimination capabilities in complex scenes.

[0098] Meanwhile, the adaptive spatial feature fusion mechanism ASFF is used in the detection head section. This mechanism is located between the Neck output and the decoupling head and aims to solve the consistency problem between multi-scale feature layers.

[0099] Specifically, ASFF unifies the feature maps from Neck at three different scales (P3, P4, P5) to the same spatial size through upsampling or downsampling operations. Then, it generates adaptive weight matrices for both spatial and channel dimensions using learnable parameters, recalibrating and fusing features from different levels. For each spatial location (i,j) and feature level l, its adaptive weights... , , After normalization using the Softmax function, the final fused feature map can be represented as follows:

[0100]

[0101] Among them, F out (i,j) is the fused feature value output by the ASFF module at position (i,j) in row i and column j. , , These represent the element values ​​at position (i,j) in the output feature maps of layers P3, P4, and P5 of the backbone network. , , To output feature maps (after resizing to be consistent) from layers P3, P4, and P5 of the backbone network at position (i,j) , , Adaptive fusion weights, A larger value indicates that the location is more dependent on the detailed features of layer P3. The larger the value, the more dependent the location is on the features of layer P4. The larger the value, the more the position depends on the higher-level semantic features of the P5 layer.

[0102] The fused feature map contains rich semantic and detailed information across scales, and is then fed into YOLO's decoupled detection head for bounding box regression and class prediction, respectively. This design not only effectively alleviates semantic conflicts between multi-scale features, but also significantly enhances the model's sensitivity and accuracy in perceiving and locating small target lesions against complex leaf backgrounds.

[0103] Step S3.2: Train the camellia tree disease detection algorithm model. Specifically, firstly, preprocess the images in the training and validation sets from Step S2.3, uniformly adjusting all images to a fixed size (640×640), and performing pixel value normalization. Data augmentation operations such as random horizontal flipping, brightness and contrast adjustment, and random cropping are only performed on the training set, while the validation set only undergoes size adjustment and normalization without additional augmentation. Then, the SGD optimizer is used, with an initial learning rate of 0.01, momentum of 0.937, and weight decay coefficient of 0.0005. The batch size is set to 16, and the maximum training epochs are set to 600. In each training epoch, the model's precision, recall, and mean average precision (mAP) are evaluated on the validation set. An early stopping strategy is employed: if the validation set mAP does not improve for 50 consecutive epochs, training is automatically terminated. The training hardware environment consisted of an Intel(R) Xeon(R) CPU E5-2699 v4 and an NVIDIA GeForce RTX 3090 GPU, while the software environment used Torch-2.5.1 + CUDA 12.4 and Python 3.8 as the programming language.

[0104] To verify the performance of the CO_YOLOv11 model in camellia oleifera disease detection, precision, recall, and mean precision were selected as evaluation metrics, and compared with other YOLO series algorithms. The experimental results are shown in Table 1. Precision refers to the proportion of samples predicted as positive by the model that are actually positive. A higher value indicates a lower false positive rate for disease-affected areas or ecological anomalies. Recall refers to the proportion of truly positive samples successfully identified by the model. A higher value indicates a stronger ability to detect disease-affected areas and ecological anomalies. Mean precision is the weighted average of the mean precision for each category, calculated using the PR curve, comprehensively measuring the model's accuracy and recall. A higher value indicates a more balanced recognition effect of the model on multiple categories in complex camellia oleifera forest environments, and superior overall performance.

[0105] As shown in Table 1, the improved CO_YOLOv11 performed best in most detection accuracy metrics: its recall (0.737), mean precision (mAP0.5, 0.772), and mean precision (mAP0.5:0.95, 0.698) were significantly higher than other comparative models. Although YOLOv9t's precision (0.91) was slightly higher than CO_YOLOv11's (0.89), CO_YOLOv11 outperformed YOLOv9 by approximately 2 percentage points in all other key metrics mentioned above. Therefore, this model improves the accuracy and reduces false negatives in the detection of diseases in camellia oleifera trees.

[0106] Table 1 Comparative Experimental Results

[0107]

[0108] Step S4: System testing and saving of test results

[0109] Step S4.1: Model Detection: The UAV, equipped with a high-resolution camera, collects real-time video stream data of camellia trees along a preset flight path and uploads the data to a cloud server. Diseases in the camellia video stream are detected based on the CO_YOLO11 model, and the confidence level of the disease target frame images is filtered to retain high-confidence disease images, generating disease type data.

[0110] Step S4.2: Data Archiving: Store the disease type data in a cloud database. The detection results are as follows: Figure 6 The results shown are for the test set. This provides complete and reliable basic data support for subsequent work such as analyzing the occurrence patterns of camellia tree diseases, optimizing prevention and control strategies, and formulating precise pesticide application plans.

[0111] Finally, it should be noted that the above implementation is only used to illustrate this solution and is not intended to limit the scope of protection of this solution. Even though the method has been described in detail based on the above implementation, those skilled in the art should understand that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this solution should be included within the scope of protection of this solution.

Claims

1. A method for detecting diseases in camellia trees based on unmanned aerial vehicles (UAVs) and the YOLO algorithm, characterized in that, include: The drone collects video stream data of the camellia trees to be detected in real time, and inputs the obtained video stream data into the optimized and improved YOLOv11 model to obtain the disease detection results of the camellia trees. The improved YOLOv11 model replaces the C2PSA module in the backbone network of the original YOLOv11 model with the C2B module, and introduces an adaptive spatial feature fusion module between the neck network and the head network of the original YOLOv11 model.

2. The method for detecting diseases in camellia trees based on UAVs and the YOLO algorithm according to claim 1, characterized in that, The C2B module consists of parallel CBAM module branches and BiFormer module branches.

3. The method for detecting diseases in camellia trees based on UAVs and the YOLO algorithm according to claim 1, characterized in that, The adaptive spatial feature fusion module is located between the neck network output and the head network input in the original YOLOv11 model structure.

4. The method for detecting diseases in camellia trees based on UAVs and the YOLO algorithm according to claim 1, characterized in that, The dataset used to optimize the improved YOLOv11 model is a tea oil leaf disease dataset constructed using artificially induced disease methods.

5. The method for detecting diseases in camellia trees based on UAVs and the YOLO algorithm according to claim 4, characterized in that, The method for constructing the camellia leaf disease dataset is as follows: (1) Pathogenic fungi were isolated and extracted from samples of diseased Camellia oleifera leaves in the natural environment and classified to obtain pathogenic fungi whose species were determined; (2) The pathogenic fungi obtained by culture were inoculated onto the leaves of healthy Camellia oleifera trees using artificial inoculation techniques to induce disease. (3) Collect typical image data of different types of diseases and different disease stages of diseased camellia trees; (4) The typical image data obtained are labeled according to the disease category. The labeling information includes the disease category and the corresponding bounding box coordinates, and finally the tea tree leaf disease dataset is obtained.

6. The method for detecting diseases in camellia trees based on UAVs and the YOLO algorithm according to claim 1, characterized in that, The obtained video stream data is input into the optimized and improved YOLOv11 model to obtain disease target frame images labeled with disease categories and corresponding bounding box information. The confidence of the disease target frame images is filtered, and high-confidence disease images are retained to generate crop disease detection results.

7. A detection system for implementing the camellia oleifera tree disease detection method based on UAV and YOLO algorithm as described in any one of claims 1 to 6, characterized in that, include: Drones use high-definition cameras to collect real-time video stream data of the camellia trees to be inspected. The control unit receives video stream data collected by the drone, processes the data, and uploads it to the cloud server in real time. The cloud server receives and stores video stream data, and inputs the received video stream data into the optimized and improved YOLOv11 model loaded on it to obtain the detection results of the camellia tree diseases.

8. The detection system according to claim 7, characterized in that, It also includes a communication module that provides network access for the drone; the control unit is a Raspberry Pi.

Images