Online Identification Method of Tea Diseases in Complex Natural Environments

An intelligent Internet-of-Things (IoT) hardware system in the field tea plantations was built, comprising collection of tea images by HD zoom cameras in a cluster structure and deployment of the detection model by cluster-head edge computing nodes. Data was sent to customer premise equipment through edge nodes and gateways and then to cloud platforms, which provided a hardware platform for identifying remote tea disease online. Field-placed cameras were used as the main acquisition means to study various diseases on Yashixiang, a typical variety of Chaozhou Dancong tea, in different seasons and weather conditions and shooting angles in a natural year period with complex backgrounds. In turn, we constructed a natural environment high-quality dataset covering major diseases e.g., tea anthracnose, tea leaf blight, tea grey blight, Pseudocercospora theae, etc. and explored the feasibility of deep learning algorithms for automatic identification of Chaozhou Dancong tea diseases. Results showed that the recognition rate of Swim Transformer reached 94% in complex natural environments. This paper demonstrated the effectiveness of the dataset and the feasibility of deep learning algorithms applied to the automatic identification of diseases of Chaozhou Dancong tea, laying a foundation for the practical application of the technology in complex natural environments.

Chaozhou City is the representative tea area of Guangdong Province, one of the four major production areas of Chinese Oolong tea, an important inheritance and birthplace of Chaozhou Kung Fu Tea culture, is also the origin of Fenghuang-Dancong Oolong Tea. There are more than 100 Guangdong provincial tea industrial parks, specialized towns and villages have been built in Chaozhou. Tea planting areas reached 15300 hectares, the production value of tea at the beginning of the year exceeded 6.4 billion CNY, driving employment of more than 500000 people. However, because of Chaozhou's subtropical marine monsoon climate, with its hot and humid climate characteristics, tea trees are prone to bacterial infection. Tea diseases can cause discoloration, deformation, withering, shedding and thus affect the growth and development of tea trees, that could lead to cloudy tea soup, bitter and astringent taste of tea, and some of the diseases are contagious, which seriously affect tea yield and quality and cause economic losses to tea farmers [1], [2]. For various reasons such as the variety of tea diseases and different control measures for different diseases, high similarity of certain diseases, different seasons and different regions with different manifestations at the same time [3], [4], [5], [6], [7], [8], it is difficult for common tea farmers to identify them. Coupled with the difficulty for plant protection experts to carry out comprehensive on-site guidance, it is of greater practical significance to study the online automatic identification method of tea diseases [9], [10], [11].
There are two types of approaches for the identification and classification of crop diseases, classical machine learning and deep learning network-based models. Sun et al. [12] proposed an algorithm combining SLIC (simple linear iterative clustering) with SVM (support vector machine) to test 261 disease images with an accuracy of 96.8%. Billah et al. [13] proposed a tea disease diagnosis system based on adaptive neuro-fuzzy inference system using wavelet transform to extract color tea image features, and obtained an accuracy of 95.7% on the authors' database with only 45 images as the training set and 30 images as the test set. The advantages of the two classical machine learning algorithms mentioned above are that they do not need to be based on a large number of images, and the model structure level is small, the amount of calculation is also small, and the calculation time is short; Their disadvantage is that they cannot obtain higher level semantic and depth features from the original image, their anti-interference performance and algorithm generalization are poor. [14]. In recent years, the deep learning technology, which can extract deeper image features, has been widely studied [15], [16], [17], [18], [19], [20]. Hu [21] utilized GAN technique to expand a tea leaf database with only 120 disease images, and then used VGG to achieve disease classification achieving an average accuracy of 90%. Mu et al. [22] proposed an image identification algorithm for tea leaf disease based on SENet and depth-separable convolutional capsule network, achieved an accuracy of 94.20% with the training database comes from 15000 images of tea leaves hand-picked in the Tea Creek Valley of Tai'an City, Shandong Province. Zhang et al. [23] used an efficient net training method to identify 2816 cucumber pest images with a maximum precision of 96.00%. Wang et al. [24] used Multi -scale Res Net to identify 19517 images of three plant pests and diseases in Plant Village, AI challenge data set with the highest accuracy of 95.95%. Transformer models is a popular state-of-the art deep learning method in recent years, which was originally proposed for the field of NLP. In 2021, Alexey et al. [25] applied the Transformer technology to the CV field and introduced the Vision Transformer network, and achieved very outstanding results. Unlike classical CNN networks that use convolution to extract local texture features of the recognition target, Vision Transformer focuses not only on local texture extraction, but also on the inter-regional position of the image. In the same year, Liu et al. [26] from Microsoft Research Institute proposed the Swim Transformer network, which uses Hierarchical feature maps similar to those used in convolutional neural networks, allowing the network to extract target features from more dimensions than the Vision Transformer network. Compared to the Vision Transformer network, Swim Transformer uses Hierarchical feature maps similar to those used in convolutional neural networks, allowing the network to extract target features from more dimensions. Swim Transformer Network uses Windows Multi-Head Self-Attention (W-MSA) to drastically reduce the amount of computation and enhance the connection between different blocks with the Shifted Windows Multi-Head Self-Attention (SW-MSA) module. The method won the championship in several image recognition and segmentation tasks upon its introduction, and received the honorary title of ICCV 2021 best paper. Deep learning is outstanding in the application of target detection and classification, but requires a large amount of training data [27].
All the above methods are based on the construction of high-quality data sets and achieve better results in classification identification by extracting image features. However, at present, disease identification has been carried out more in major crops such as maize and rice, and there are relatively few studies on the construction of tea disease data sets and their identification methods, especially for Chaozhou Dancong Tea. There are no relevant reports on tea disease data sets and automatic identification methods. In addition, tea diseases are affected by factors such as climate, environment and varietal differences in the cultivation areas where they are located, and the data sets are not universally applicable, as Mohanty et al. [28] applied the deep learning model trained on the PlantVillage data set to another data set of the same type of plant, and the accuracy dropped to less than 50%. In summary, the above methods are mainly aimed at the classification and recognition of tea diseases in a single piece of tea in an ideal environment in the laboratory, while for real scenario applications, the recognition of tea diseases in natural environments with complex backgrounds. There are still relatively few research results, and the robustness of the above methods still needs to be verified.
Based on the above-mentioned methods, and facing the practical application scenario of complex natural environment, we took Yashixiang Fenghuang Dancong Tea as the research object, and build an intelligent IoT hardware system based on visual recognition in the outdoor tea plantation, using the deployed cameras as the main means of tea image acquisition, and study the construction of a data set of tea major diseases with complex background in natural environment reflecting different seasons, weather, light, etc., and explore the feasibility of automatic identification methods of Chaozhou Dancong Tea diseases based on deep learning algorithms, which provides a solution for realizing the practical application of automatic tea disease identification in complex natural environments.
The main work of this paper is: a) A set of remote tea disease automatic recognition system based on machine vision of the Internet of Things was constructed to provide a reference scheme for scientific tea planting. This system constructs an automatic remote picture acquisition and classification path through the network pin-top camera-routermobile Wi-Fi-cloud server-identification terminal. The highlights of this path are: a. Relying on the full coverage of cellular network 4G signal, the use of mobile Wi-Fi to solve the problem of remote mountainous areas without broadband signal; b. Use the cloud server to lay the hardware foundation for multiple access terminals. b) The typical flavor of Chaozhou single cluster tea, "Yashixiang", was selected as the research object. The deployed camera was used as the main means of tea image collection, and more than 10000 tea disease images were collected under different seasons, weather and shooting angles in natural environment. The tea images were screened, cleaned and labeled together with experts from Chaozhou Tea Scientific Research Center. The high quality tea disease data set close to the actual application scenario was constructed to fill the blank of single cluster tea disease database. Swim Transformer algorithm with Transfer Learning is used to classify the tea diseases of Chaozhou Dancong Tea taken in complex natural environment, and a good classification effect is obtained, which verifies the scientific nature of the data set constructed in this paper, and lays a foundation for further automatic disease recognition of Teochew single cluster tea for practical scenarios.

II. TEA LEAF DISEASE REMOTE ONLINE IDENTIFICATION HARDWARE SYSTEM
A visual IoT remote monitoring system was constructed and deployed to achieve remote online identification of tea diseases in complex natural environments. The system has a three-layer architecture of sensing, transmission and processing, consisting of HD cameras, gateways, edge computing nodes, CPEs (Customer Premise Equipment), and OneNET cloud platform and some operating terminals. The platform framework is shown in Fig. 1.
The perceptual layer adopts a split-cluster structure, and each cluster has multiple cameras to capture images of tea. A cluster head is an edge computing node that can serve deep learning models or preprocess images. The network layer consists of gateways and CPE devices in each cluster (see in Fig. 2), gateways aggregate each cluster's information into CPE nodes with mobile Wi-Fi capabilities, and CPE nodes transmit information to the cloud platform over the Internet. The cloud platform can store and process data for tea disease identification, thus providing decision support for tea disease management. Terminal devices such as PC, tablet or cell phone can realize camera control and real-time query of disease identification results by accessing the cloud server.
The parameters of each key device are as follows: a) The camera is made by Hikvision. The model is DS-2DC4423IW-D(C). The camera is installed with a height of 2 meters and a horizontal tilt Angle of 45°. The monitoring range of each shooting is about 1.5 square meters. The camera has the function of 360°head rotation, and one picture is sampled every 90°rotation. b) The gateway adopts the Universal TL-WDR5620 gigabit wireless router, which can satisfy the signal transmission of multiple cameras and the access function of current wireless devices. c) The signal transmission from the router to the cloud server adopts FLASH FISH mobile Wi-Fi. Because the tea seedling garden is located in a remote area and there is no wired broadband signal, the signal transmission from the router to the cloud server is realized by using the cellular signal of mobile portable Wi-Fi. d) OneNET cloud platform of China Mobile is selected as the cloud server, which can provide complete intelligent hardware access and cloud storage functions to meet the requirements of this project. Since the disease of tea seedlings usually presents "regional", the monitoring of tea seedling diseases, due to the realization cost, does not need to achieve full coverage of camera shooting, only needs to be done when the planting environment (temperature, humidity, ventilation conditions, planting altitude), planting conditions (irrigation, fertilization, weeding, shading, planting density) and planting varieties of tea seedlings have obvious changes. Add terminal detection nodes (including cameras, edge computing nodes and mobile Wi-Fi) for sampling and detection.
In addition, since tea diseases are not mutational and the characteristics of diseases need to change significantly after several days, the requirements for continuous monitoring and real-time performance are low. Basically, one frame of picture can be taken every hour for detection. The current resolution of pictures taken by webcams is 1280 × 720 and occupies 224 KB. It is calculated that each camera needs 24 MB of cloud storage space per day. We can dynamically plan the use of cloud storage space according to the number of sampling nodes.

III. TEA LEAF DISEASE DATA SET CONSTRUCTION
The construction of a high-quality tea disease database is the basis for conducting research on tea disease identification methods. We took "Yashixiang", a typical variety of Chaozhou Dancong Tea, as the research object, and focused on the feasibility of establishing a multi-angle and multi-size tea disease database to fully reflect the tea under different time, season and light environment.
In order to study the impact of different seasons, climate and weather changes on tea, this paper took a natural year as a cycle, took pictures of tea diseases at different times, in different weather and from different shooting angles, and successfully collected tens of thousands of tea photos. These pictures are composed of nearly 10 different varieties of tea. 'Yashixiang' is one of the varieties popular in the local area, and the planting scale is relatively large. Therefore, our project team chooses 'Yashixiang' as our experimental object. The three team members who participated in the work of classifying and labeling tea pictures were trained by experts from Chaozhou Tea Science Research Center and were familiar with the characteristics of tea diseases of Chaozhou Dancong Tea to ensure the correct classification of tea diseases. Through screening, cleaning and classification, a total of 1377 tea leaf images were selected as the sample set and divided into 5 categories, including 166 healthy tea leaves, 582 tea leaves infected with Tea anthracnose, 126 with Pseudocercospora theae, 384 with Tea leaf blight and 119 with Tea grey blight. Photo samples of (a) healthy tea leaves, (b) Tea leaf blight (c) Tea anthracnose (d) Pseudocercospora theae (e) Tea grey blight can be seen in Fig. 2.

A. HEALTHY TEA LEAVES
Healthy tea leaves ( Fig. 2(a)) are free of mottling and curling throughout the body, and the leaves are full and healthy in tissue. To enhance the robustness of the system, the tea image datasets are all taken from the natural environment, adding tea images from different shooting angles, lighting and seasons.

B. TEA LEAF BLIGHT
Tea leaves infected with Tea leaf blight are as shown in Fig. 2), the disease peak incidence in June and between late August and early September; mainly affects mature leaves, but also new tips, fruit or branches; mostly in adult, old or young leaves of the leaf tip or leaf edge to produce round to irregular water-soaked spots. The diseased leaves start as yellow-green or yellow-brown, and later become brown, with undulating brown, gray interspersed with clouds. Diseased leaves often fall prematurely, the new tips die, the tree becomes weak.

C. TEA ANTHRACNOSE
The typical symptoms of Tea anthracnose is shown in Fig. 2(c). The disease is most likely to occur in seasons of high temperature and humidity such as May to August, when the leaves with low polyphenol content or thin and soft leaf structure are more susceptible to the disease. Mainly affects adult leaves; first in the leaf margin or leaf tip to form spots, lesions leaf color is light brown or yellow-brown, and finally gray-white, its scattered black dots, spots without whorls.

D. PSEUDOCERCOSPORA THEAE
Tea leaves with Pseudocercospora theae are as shown in Fig. 2(d), the disease is a low temperature and high humidity type disease, spring and autumn rainy season can occur. When the disease on the leaves produce small round spots, and then expand into a gray-white middle sunken round spots, with dark brown or purple-brown raised line at the edge, the central reddish-brown, late spots scattered black dots in the middle. When the humidity is high, the leaves are covered with gray moldy material.

E. TEA GREY BLIGHT
The typical disease of Tea grey blight is shown in Fig. 2(e), the leaf disease is divided into two types: verticillate and cloudy. The whorl type affects the old and adult leaves, invading from the leaf tip, leaf edge or wound, the disease part is yellowish green small spots at first, then gradually expanded to round or even irregular brown spots, later the central spot becomes gray-white and gray-brown concentric circular whorl, in wet conditions along the whorl appears thick ink small grain point, the edge with brown elevated line, the junction of disease and healthy part is obvious; the back of the spot is gray-brown, whorl is not significant, slightly black spots, the border is not obvious. The cloud pattern occurs on the young leaves, from the leaf tip or leaf edge inward expansion for gray-brown cloud-like spots, no obvious lines, the junction of the healthy part of the disease is not obvious, the later part of the disease has a small number of black dots, black dots occasionally aggregated into a black velvet-like. The spots are often interconnected into brown spots, causing the whole buds and leaves to scorch when serious.

IV. SWIM TRANSFORMER TEA DISEASE CLASSIFICATION ALGORITHM BASED ON TRANSFER LEARNING
Commonly used image classification techniques can be divided into classical machine learning techniques and deep learning techniques. deep learning is gradually becoming mainstream due to the advantages of strong learning ability, good adaptability, sustainable optimization, and good portability of the model. Swim Transformer [26] is currently a more advanced deep learning model and has won the championship in several image processing competitions. In this paper, we try to use the Swim Transformer network model with the above local tea disease database in Chaozhou as the data set and introduce Transfer Learning to improve the generalization performance of the model and achieve automatic classification of tea leaf diseases.

A. SWIM TRANSFORMER NETWORK MODEL
Swim Transformer network architecture is shown in Fig. 3. Swim Transformer input the picture into the Patch Partition module for block. Every pixel adjacent to 4 × 4 is a Patch. The RGB three-channel picture is 4 × 4 × 3 = 48, and flatten in the channel direction. channel data for each pixel is transformed from 48 to C by Linear Embedding layer, and then connected to the Swim Transformer Block. The feature maps of different sizes were constructed by four stages. Except for a Linear Embedding layer in Stage1, the remaining three stages were subsampled by a Patch Merging layer first. As can be seen from the blue part on the left of Fig. 3, Swim Transformer uses Hierarchical feature maps similar to those in convolution neural networks to achieve different resolutions, which helps the network cope with different use environments and improve its robustness.
As shown in Fig. 3, two structures are used in the Swim Transformer Block, which differ only in that one uses the Windows Multi-Head Self-Attention (W-MSA) structure and the other uses the Shifted Windows Multi-Head Self-Attention (SW-MSA) structure. In order to solve the problem of high parameter computation caused by the VIT global attention mechanism, Swim Transformer proposed a windowbased multihead attention mechanism (W-MSA), which divides the image into multiple windows and interacts only within the windows, which can greatly reduce the computation, especially for large resolution images. Expression (1) is the computation of MSA module and expression (2) is the computation of WMSA module.
(MSA) = 4hwC 2 + 2(hw) 2 C ( 1 ) where h represents the height of the feature map, w represents the width of the feature map, C represents the depth of the feature map, and M represents the size of each window.
In order to solve the problem of the lack of interconnection between the windows of the picture when the WMSA module is introduced, Swim Transformer proposes SW-MSA in order to introduce the interconnection between the windows while keeping the computational complexity low, the SW-MSA windows are offset by 0.5M pixels from the upper left corner to the right and the bottom respectively. In each Swim Transformer block, a W-MSA structure is used before a SW-MSA structure, and these two structures are used in pairs, thus achieving the problem of information exchange between different windows with low computational cost.
To sum up, Swim transformer uses a hierarchical construction method similar to CNN network to improve network expression ability. A multi-window attention mechanism is proposed, which divides the image into multiple Windows and only interacts within the Windows, thus greatly reducing the complexity of computation. The sliding window multiattention mechanism can be used to make up the problem of insufficient information interaction between different Windows caused by window division. It better integrates the strengths of CNN network and transformer network, with excellent performance and wide application range.

B. TRANSFER LEARNING
Yosinski et al. [30] published a paper in NIPS, which studied the transfer ability (or generalization) of individual layer features in deep learning and proposed that usually the first layer is not particularly related to the specific image data set, while the last layer of the network is closely related to the selected data set and its task goals; the shallow layer feature is called general features and the last layer is called specific features, and the use of Transfer Learning can effectively improve the generalization performance of the model. Therefore, in order to improve the model recognition accuracy this paper introduces Transfer Learning, which disguisedly increases the size of the database. The process of Transfer Learning is to use the model parameters trained on a large database, and all the shallow layer parameters are directly migrated, and the final fully connected layer is deleted and retrained.
The specific steps are as follows. a) Download the same network structure with a pre-trained model that has been trained on ImageNet (the size of this database is about 1.2 million samples with 1000 classifications). b) Create a new neural network model, i.e., the target model, and then load all the weight parameters of the pre-trained model into the target model. c) Remove the output layer of the target model with 1000 classification loaded with the pre-trained model and remap it to the 4 classification output layer. d) For comparison, the hyperparameter settings are modified to be consistent with experiments in this paper.

V. EXPERIMENTAL RESULTS AND ANALYSIS
The experiment is divided into the following four parts: The first part introduces the experimental environment, including the hardware used in the experiment, the platform, the parameter setting of the experiment process, and the way of data preprocessing; The second part applies the latest and classic deep learning algorithm to the database of this paper, and selects the most appropriate network model to apply to this project based on the experimental results. The third part, through the analysis of the experimental process, further understand a performance of the model in this database. In the fourth part, the fuzzy matrix is used to analyze the causes of network errors and lay a foundation for further improving the accuracy.

A. INTRODUCTION TO THE EXPERIMENTAL ENVIRONMENT
The operating system of the experimental platform is Windows 10, the CPU is Intel(R) Core(TM) i7-10700 CPU @ 2.90GHz, the GPU is NVIDIA GeForce RTX 3080 10GB, and the CUDA version number is 11.6. The model training development environment is configured using Anaconda, and the programming language is Python 3.7. and the deep learning framework is Pytorch 1.9.
In order to ensure the same comparison conditions, the deep learning algorithm adopted in this paper tries to keep the same parameter Settings. The main parameter Settings are as follows: Adam network optimization algorithm is adopted for network parameter optimization in the deep learning process, and the learning rate in training is set to 0.0001. The model training process adopts the batch training method, the Batch Size is set to 24, and the model processes all the training and test images for one iteration (Epoch), and the total number of experiments in this paper is 100. A total of 1377 pieces of 'Yashixiang' data set were divided into 1104 pieces of training data set and 273 pieces of test data set in an 8:2 ratio.
In order to enhance the diversity and comprehensiveness of the data and thus improve the generalization ability of the model, the training images are randomly range cropped, scaled and rotated in the experiments. To improve the gradient disappearance and gradient explosion problems, data normalization processing, weight initialization, and through BN 17 (Batch Normalization) are used to achieve accelerated convergence of the network and improve the accuracy.

B. COMPARISON OF CLASSIFICATION EFFECTS
As shown in Fig. 4, the experimental results are optimal with Swim Transformer, which has the relatively fastest learning rate and classification accuracy. Its final recognition accuracy reaches 94%, Resnet network's final recognition accuracy is about 90%, DenseNet is about 91%, and Vision Transformer's is about 85%. Swim Transformer uses hierarchical construction method similar to CNN network to improve network expression ability. On the other hand, it has multi-window based attention mechanism, which can be said to be a combination of advantages of CNN network and Transformer network. In the database constructed in this paper, Swim Transformer has the best effect. Therefore, it is selected as the backbone network of this paper.

C. ANALYSIS OF TRAINING AND TEST RESULTS
Based on the experimental results in 5.2, it can be seen that Swim Transformer has good recognition effect under the data constructed in this paper, so Swim Transformer is selected as the network model for tea disease classification in this subject.
Experimental results are shown in Fig. 5(a)-(d) The Swim Transform model was used to classify the tea leaf disease data set, and 100 Epochs were trained with and without Transfer Learning respectively. The experimental results are shown in Fig. 5(a)-(d). Fig. 5(a) shows the training error graph, which shows that the starting value of the training error is high and decreases slowly when migration learning is not used; while when migration learning is introduced, the training error of the first 20 epochs shows rapid convergence, and the training error is basically stable at 0.7 by the 60th Epoch; Fig. 5(b) shows the training accuracy graph, which shows that the training accuracy increases slowly when migration learning is not used, and the training accuracy is only 74.5% even by the 100th Epoch; while when migration learning is introduced, the training accuracy is only 74.5% by the first 20 Epochs. When migration learning is introduced, the training accuracy rises slowly and is only 74.5% even at the 100th Epoch, while the first 20 Epochs make the training accuracy reach 96% quickly and stabilizes at 98% after the 40th Epoch. Fig. 5(c) shows that without using Transfer Learning, the validation set error decreases steadily from the first Epoch until the 80th Epoch when the training error stabilizes at 0.8; while when Transfer Learning is introduced, the training error is smoother and stabilizes in a small range of fluctuations around 0.6 after the 60th Epoch. Fig. 5(d) shows the graph of validation accuracy. This result shows that without adding migration learning, the training progress slowly improves until the 80th epoch, when the validation accuracy only stabilizes at 75%; after adding migration learning, the validation accuracy of this model quickly stabilizes at 93% in the first 20 epochs of training accuracy.

D. ANALYSIS OF FUZZY MATRIX OF CLASSIFICATION
The confusion matrix of classification results after the introduction of Transfer Learning is shown in Fig. 6, and the results of precision, recall and specificity derived from the confusion matrix are shown in Table 1. Table 1 shows that the precision of Tea anthracnose and tea grey blight is relatively low, with values of 0.905 and 0.875, respectively, and the recall of tea grey blight is the lowest, with a value of  0.609, which indicates that some leaves with tea grey blight were incorrectly classified into other types; the specificity of Tea anthracnose is the lowest, with a value of 0.924, which indicates that the classification model misidentified other types of diseases as Tea anthracnose. As can be seen from Fig. 6, among the 126 anthracnose images in the validation set, one image was labeled as Tea leaf blight, accounting for 0.79%; two were labeled as Pseudocercospora theae, accounting for 1.59%, and nine were labeled as tea grey blight, accounting for 7.14%. Out of the total 16 misclassifications, the number of misclassifications for Tea anthracnose was 12, accounting for 75% of the misclassifications. There are two main reasons for this result, one is that Tea anthracnose is the disease with the highest incidence in local tea gardens. In this experimental data, there are a total of 1377 pictures of tea pests and diseases, among which there are 582 pictures of Tea anthracnose, whose sample size is significantly more than other classifications.
In future research, we will gradually increase the types of disease samples, balance the number of images in each type, and optimize the database to improve the recognition rate of the model; Second, when viewing the misclassified images, we found that multiple diseases coexisted in a particular tea image, and the similarity between Tea anthracnose and tea grey blight was high, which affected the classification accuracy. In conclusion, according to the experimental results, the classification accuracy of Swim Transform model reaches 94% on average, indicating that the introduction of Transfer Learning can effectively improve the accuracy of training and recognition, which is suitable for the classification of this data set.

VI. DISCUSSIONS
Compared with the laboratory environment, the camera based online identification method in the natural environment has many problems, such as complex background of tea images, large light changes, easy overlap between tea leaves, and difficult hardware system deployment. In particular, the characteristics of different tea diseases in different tea producing areas and varieties are also quite different. Therefore, the identification of tea diseases has its particularity. It is necessary to build a single type of tea disease database in a certain producing area.
This paper builds a hardware platform of intelligent Internet of Things in the field tea garden, and attempts to conduct online data collection for a period of one year under various natural environment conditions. Practice has proved that the network system is stable and reliable. This study took the typical single cluster tea variety "Yashixiang" in Chaozhou as the research object, constructed a data set of five diseases, and used the Swim Transformer depth learning algorithm with migration learning to identify diseases, with the recognition rate reaching 94%, which shows that the data set constructed in this paper is scientific and effective, and further proves that it is feasible to apply the Swim Transformer depth learning algorithm based on migration learning to the disease identification of single cluster tea in Chaozhou, It provides a valuable reference for further research on real-time online automatic identification of tea diseases.
However, there are still many challenges: a) The natural environment is relatively complex, with many interfering factors. In the natural environment there are leaves, branches, mud, stones, overlapping tea leaves, weeds, flying butterfly insects, uneven lighting, light spots, and many other unforeseen interference, affecting the accuracy of tea disease identification. b) The real-time outdoor online identification system and the reliability in the face of bad weather conditions is also a challenge. Our subsequent research will be carried out in the following aspects: a) The database will be expanded, more pictures of tea diseases will be taken and attempts will be made to include pictures of other varieties of tea diseases in the data to further enhance the depth, breadth and generalization ability of the database. b) Optimize the network architecture to further improve the recognition rate of tea diseases. Compress the network scale and further improve the recognition speed. Complete the deployment of edge computing nodes and improve the real-time performance of the system. c) Multiple factors will be fused for classification. Tea diseases are closely related to season, external environment (temperature, humidity, ventilation) and other factors, and environmental factors are subsequently integrated into the tea disease recognition model to improve the recognition rate.