Design and Research of an Online Diagnosis Platform for Tomato Seedling Facilities Production Diseases

doi:10.21203/rs.3.rs-3121099/v1

Download PDF

Research Article

Design and Research of an Online Diagnosis Platform for Tomato Seedling Facilities Production Diseases

https://doi.org/10.21203/rs.3.rs-3121099/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The facility-based production method is an important stage in the development of modern agriculture, lifting natural light and temperature restrictions and helping to improve agricultural production efficiency. To address the problems of difficulty and low accuracy in detecting pests and diseases in the dense production environment of tomato facilities, an online diagnosis platform for tomato plant diseases based on deep learning and cluster fusion was proposed by collecting images of eight major prevalent pests and diseases during the growing period of tomatoes in a facility-based environment. The diagnostic platform consists of three main parts: pest and disease information detection, clustering and decision making of detection results, and platform diagnostic display. Firstly, based on the YOLOv5 algorithm, the key information of the disease was extracted by adding attention module (CBAM), multi-scale feature fusion was performed using weighted bi-directional feature pyramid network (BiFPN), and the overall construction was designed to be compressed and lightweight; Secondly, a K-means clustering algorithm is used to fuse with the deep learning results to output pest identification decision values to further improve the accuracy of identification applications; Finally, a web-side pest detection platform was designed and developed using Python, including the front-end, back-end, and database of the system to realize online diagnosis and interaction of tomato plant diseases and pests. The experiment shows that the algorithm detects tomato plant diseases and insect pests with mAP of 92.7% ([email protected]), weights of 12.8M, inference time of 33.6ms. Compared with the current mainstream single-stage detection series algorithms, the improved algorithm model has achieved better performance; The accuracy rate of the platform diagnosis output disease information of tomato images is 91.2%, and the accuracy rate of video disease information is 95.2%.

Pest and disease detection

YOLO

Diagnosis Platform

K-means clustering

Facility production base

Tomatoes are affected by different pests and diseases at different stages of growth, which is the main reason for reduced yield. Accurate identification and diagnosis of pests and diseases in tomato seedlings during the growing period, as well as early detection and treatment, will not only provide a healthy growing environment for tomatoes but will also effectively increase tomato yields to a large extent [1].In particular, the high relative temperature and humidity, poor lighting conditions, and poor circulation in the facility base provide excellent environmental conditions for the rapid spread of pathogens, greatly increasing the chances of pathogenic infestation and leading to a disaster. The most serious diseases of tomato seedlings during the growing period are Bacterial Spot [2], Early Blight [3], Late Blight [4], and Leaf Mold [5], which have a greater than 50% chance of developing [6, 7]. The disease mainly affects the leaves, starting with individual seedlings and then spreading rapidly in all directions with these plants as the center, infecting neighboring plants. The main pest species are Aphid [8], Helicoverpa Armigera [9], Spider Mite [10] and White Fly [11], all of which are highly reproductive, fast-growing and widespread when the environment is suitable, and in addition to direct damage, can also spread directly or promote secondary infection of the disease [12]. According to statistics, approximately 15% of global tomato production is affected by pests and diseases each year, with average yield reductions in severe regions capable of reaching 40–80% [13, 14]. Careful control of pests and diseases is a key task to reduce losses and increase crop yields. Once a pest or disease has invaded a field, it must be detected in time for farmers to treat it and prevent it from spreading [15]. Therefore, it is necessary to select pests and diseases that cause serious damage to tomatoes as research objects, to collect and collate relevant information, to achieve accurate identification and detection, and to provide a theoretical basis for targeted early warning and prevention.

Traditional detection methods no longer meet the needs of research and production in terms of identification efficiency, accuracy, and application scenarios. With the continuous development of the Internet, the application of information technology has provided new methods and ideas for the identification of crop pests and diseases. The successful application of deep learning in other fields has attracted the attention of many agricultural scholars and applied it to the agricultural field [16, 17]. Deep learning (DL) methods, especially those based on convolutional neural networks (CNN), are widely used in object detection and classification in the agricultural field, demonstrating excellent performance and classification in applications such as plant pest detection and plant identification [18].

Current mainstream target detection algorithms can be divided into two types. One is a two-stage target detection based on the candidate region method, which requires a proposal (a pre-selected box that may contain the object to be detected) followed by fine-grained object detection. Such as RCNN (Regions with CNN features), Fast RCNN, and Faster RCNN, etc. Deng et al. [19] proposed a multiple pest detection technology based on federated learning (FL) and improved fast regional convolutional neural network (R-CNN). The improved R-CNN has an average accuracy of 90.27% for multiple pest detection in orchards, and the detection time of each image is only 0.05 s, realizing the accurate identification of small pests and diseases in complex environments. Jiao et al. [20] proposed an anchor-free area proposal network (AFRPN) and combined it with Fast R-CNN to detect 24 types of pests in an end-to-end manner. The mAP and recall of the improved model are 7.5% and 15.3% higher than that of Faster R-CNN, the running time can reach 0.07 seconds per image, which meets real-time detection. Zhang et al. [21] proposed an improved Faster RCNN algorithm to detect tomato diseased leaves, using ResNet101 instead of VGG16 for feature extraction and k-means clustering algorithm for clustering bounding boxes, the accuracy rate increased by 2.71%, which can effectively detect and recognize tomato diseases. Xie et al. [22] proposed Faster DR-IACNN algorithm on the self-built grape leaf disease dataset (GLDD). Inception-v1 module, InceptionResNet v2 module and SE module are introduced, mAP is 81.1%, and detection speed reached 15.01 FPS. One is a single-stage target detection based on regression methods, which requires only one input to the network to predict all bounding boxes, and extracts feature directly from the network to predict the classification and location of targets. Such as Single Shot MultiBox Detector (SSD), You Only Look Once (YOLO), etc. Wang et al. [23] proposed a Deep Block Attention SSD (DBA_SSD) method for plant leaf disease identification by combining an improved VGG network and a channel attention mechanism, achieving 92.2% accuracy on the plant Village dataset. Sun et al. [24] built a new apple leaf disease detection model based on the Mobile AppleNet SSD algorithm using the MEAN module and the Inception module, which could achieve 83.12% mAP and 12.53 FPS in a complex background. Liu and Wang [25] constructed a dataset of tomato pests and diseases in a real natural environment, used image pyramids to optimize the feature layer of the YOLO V3 algorithm, realized multi-scale feature detection, and could accurately, quickly detect the location and type of tomato pests and diseases. Wang et al. [26] achieved early real-time detection of tomato pests and diseases with an F1 value of 94.77% and an AP value of 91.81%, with a false detection rate of only 2.1%, based on YOLOv3 with fused expanded convolution and convolution factor decomposition. Liu et al. [27] proposed a tomato pest identification algorithm based on the improved YOLOv4 fusion triple attention mechanism (YOLOv4-TAM) by introducing the focal loss function and the K-means + + clustering algorithm, with an average identification accuracy of 95.2%. Qi et al. [28] added the Squeeze and Excite (SE) module to the YOLOv5 model for the detection of tomato virus diseases in natural backgrounds, with an accuracy rate of 91.07%. Chen et al. [29] integrated the involute bottleneck module and SE module on the basis of the original YOLOv5 network algorithm. The detection accuracy of the algorithm for powdery mildew and anthracnose is 86.5% and 86.8%, respectively.

As mentioned above, deep learning-based target detection algorithms can better extract features from images and show good performance in the identification of agricultural objects, but relatively little research has been done on the identification of tomato pests and diseases in the growth period. In this paper, we have created a tomato growing period pest and disease dataset consisting of 2547 images by collecting images of tomato pests and diseases in a facility environment, using the Plant plant public dataset and Internet crawler technology for additions (The dataset has been shared at https://drive.google.com/file/d/1V1cRPVwtqrJuBJifGyQj0z2b0N0Dh_OS/view?usp=sharing). It can provide data to support research on tomato pests and diseases during the growing period. In order to further improve the efficiency and identification accuracy of tomato pests and diseases, the YOLOv5 algorithm was improved by introducing attention, weighted bi-directional feature pyramid network (BiFPN), and C3Ghost modules as a means to achieve high accuracy and algorithm lightweight. Based on the identification of the improved algorithm and the analysis of the results of the k-means clustering algorithm, an online diagnosis platform for pests and diseases was designed and developed, capable of real-time detection of images or videos, using the k-means clustering algorithm to select the detection results when detecting videos, and outputting the optimal identification results and confidence levels to the platform interface. The platform was applied at a tomato production facility and the experimental results met the need for rapid and accurate detection of pests and diseases during the growing period, which is of great significance for tomato pest and disease control research and the development of smart agriculture.

2.1 Data acquisition

Eight tomato pests were selected for this study: Aphid (AP), Bacterial Spot (BS), Early Blight (EB), Helicoverpa Armigera (HA), Late Blight (LB), Leaf Mold (LF), Spider Mite (SM), and White Fly (WF). The data was collected from the facility production base in the Luolong District of Luoyang City, the scene is shown in Fig. 1, the facility was well-lit and photographed with a Sony DSC-H300 camera, and a total of 1978 images were acquired with 640 x 480 pixels. The dataset was also supplemented by a plant public dataset [30] and a web crawler technique [31], with a total of 569 images at 256 × 256 pixels.

These images contain different light intensities and object sizes, ensuring a diversity of data. Through screening and collation, a total of 2388 valid multi-scene and multi-scale pest and disease images were obtained. Figure 2 shows an example of some of the pest and disease images in the dataset.

2.2 Data expansion and enhancement

To avoid overfitting problems due to the small number of datasets, the pests in the dataset were expanded using CycleGAN [32], using contrast enhancement, rotation and flipping to expand the pests. CycleGAN is trained iteratively against each other to achieve the effect of expanding the dataset of this paper. The coordinates and categories of the dataset were manually annotated using the lightweight image annotation tool LabelImg software to obtain the corresponding label files for the dataset images, and the annotated dataset was divided into a training set and a test set. The distribution of the training and test sets and the labels for each category are shown in Table 1.

Table 1

Distribution of the number of pest and disease images
Category	Train/enhance	Test/enhance	Sum	Label
AP	259/777	45/135	1216	0
BS	249/747	45/135	1176	1
EB	253/759	45/135	1192	2
HA	256/768	45/135	1204	3
LB	250/750	45/135	1180	4
LM	250/750	45/135	1180	5
SM	251/753	45/135	1184	6
WF	260/780	45/135	1220	7
Sum	8112	1440	9552

The training data is enhanced by the use of a data loader during the data input process. The data loader uses a variety of data enhancement methods, including color transformation, random rotation and panning, Mosaic [33], Mixup [34], and a random combination of these methods stitched together to enrich the target and background. The enhancement process is shown in Fig. 3.

In order to achieve a facility-based pest detection application, an algorithm based on the YOLOv5 is proposed for the detection of eight pests and diseases during the growing period of tomatoes. By clustering the identified information using the k-mean clustering algorithm, the frequency of pest and disease occurrence and the confidence level are statistically analyzed to achieve rapid data clustering and fusion and online pest and disease diagnosis and decision making. Design and develop a pest and disease detection webpage to introduce information on various types of pests and disease and prevention methods, and output of the algorithm identification results on the platform after analysis by clustering.

Figure 4 shows the architecture of the online pest and disease diagnosis process. After the model is trained and saved, it is connected to the online diagnosis platform, where users upload images or videos of tomato pests and diseases, and the platform will identify and output prediction results.

3.1 Identification algorithm model

The pest and disease diagnostic platform needs to meet the real-time needs of users. Accurate and real-time detection is a prerequisite for easy use of the webpage, and YOLOv5 was chosen as the detection model, taking into account factors such as real-time pest and disease detection, identification accuracy, and future deployment. YOLOv5 is a more mature single-stage target detection model in the YOLO [35] family, capable of locating and classifying targets by directly regressing the relative positions of candidate frames. There are three main components: the backbone network (Backbone), the neck network (Neck), and the detection (Head) network. The backbone network consists of convolution module (Conv), C3 module, and a modified spatial pyramid module (SPPF) for extracting features from the input image and passing them to the Neck. Neck uses the structure of FPN and PAN, where FPN uses a high resolution of low-level features and semantic information of high-level features in top-down delivery of semantic information, and PAN is bottom-up delivery of localization information, making it easier to propagate low-level information to the top level and enhancing the ability of network feature fusion. Head is responsible for classifying the features extracted after compression and fusion, generating detection frames, and classifying them into the appropriate coordinates, categories, and confidence levels.

A tomato pest detection algorithm is proposed based on the YOLOv5s algorithm. In Bankbone, the CBAM attention mechanism has been introduced to help the network locate the location of pests and diseases more accurately, thus reducing the interference of target background information. In Neck, the use of the BiFPN structure instead of the original PANet structure improves the detection of small target pests by increasing the weights to adjust the contribution at each scale of the algorithm. The C3Ghost module is used instead of the C3 module to reduce the number of parameters, achieve a lighter model, and balance identification accuracy with detection speed. Figure 5 shows the structure of the improved algorithm.

3.1.1 Improvement details

Improvement 1: The attention mechanism can focus on important information with high weights and ignore irrelevant information with low weights, and can continuously adjust the weights so that important information is selected in different situations [36]. There are problems of multiple targets, small size and complex backgrounds in the homemade pest and disease dataset. A CBAM module [37] was added to the original feature extraction network part, which consists of a channel attention module and a spatial attention module. It enables the network to better extract feature information of pests and diseases and improve the characterization capability.

Improvement 2: The neck of the original YOLOv5s simply adds up different scale pest features in the feature fusion process, the method does not make full use of feature information between different scales, and the importance of feature information varies at different scales, which can lead to lower detection accuracy of the network. To solve this problem, the BiFPN network structure [38], a combination of Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) structures, was introduced for enhancing the depth of information mining and further improving the feature extraction capability of the algorithm.

Improvement 3: During the extraction of pest and disease features, the standard convolution extracts feature maps that are very similar on many feature channels, with some duplication. Therefore, the desired feature maps can be obtained without the need for a full convolution operation. To reduce the number of operations, the Ghost module [39] is used, and the C3Ghost module is used to replace the C3 module in the network in a two-step operation: first, the ordinary convolution calculation is used to obtain a feature map with a smaller number of channels; then, more feature maps are obtained using linear operations, and the output results are stitched with the constant mapping in the channel direction to obtain the output feature map.

3.2 Cluster analysis model

As an unsupervised clustering algorithm in data mining, K-means is simple and effective [40]. This algorithm is a distance-based clustering algorithm, which uses the distance between samples as the sample similarity measure, and its basic process is as follows:

(1) Select k objects from the known model data as the initial cluster center, and the i-th mean value is recorded as u_i;

(2) Classify the categories by calculating the distance d from each cluster object to the cluster center;

$$d=\sqrt {\sum\limits_{{i=1}}^{n} {{{(xi - uj)}^2}} }$$

where j = 1,2…,k, i = 1,2…,n, d is the distance from the i-th point to the j-th mean, and x_i is the i-th data value;

(3) Calculate each cluster center again;

$$uk=\frac{{\sum\limits_{{i=1}}^{{{n_k}}} {{X_i}} }}{{{n_k}}}$$

where n_k denotes the number of objects in the kth cluster;

(4) Repeat steps (2) and (3) until the clustering centres no longer change or the maximum number of iterations is reached, then stop.

To perform unsupervised k-means clustering on the given data in this paper, it is first necessary to determine the optimal number of clusters, using the Elbow Method, by plotting the intra-cluster sum of squares (WCSS) against the number of clusters and selecting the number of clusters where WCSS decreases at a slower rate and an inflection point occurs.

3.3 Interface design module

The platform adopts B/S (Browser/Server) system structure and is divided into three parts: front-end, back-end and database. The development mainly uses Python language and the front end uses html combined with js, bootstrap and css to display the page and interact with the back end. The back end uses the SSM framework, the SSM framework is Spring MVC, Spring and Mybatis framework integration, using Spring MVC is responsible for request forwarding and view management, Spring to achieve business object management, Mybatis as the data object persistence engine. MVC is composed of Model (JavaBean), View (Html), and Controller (Servlet). Model is a module used to carry data and perform calculations on requests submitted by users. View provides a user interface for the user and interacts directly with the user. The Controller is used to forward the user request to the corresponding Model for processing and provide the corresponding response to the user with the calculation result of the processing Model. Use the server session to cache the currently logged-in user, upload the file by copying the file to the tomcat path, and then use the url to access it, and the database uses mysql to store the database.

4.1 Experimental environment

To ensure the reliability of the experimental data, all algorithms used in this paper were tested under the same experimental conditions. The hardware and software details are shown in Table 2.

The model was trained using stochastic gradient descent (SGD) for optimization, with the learning scheduler set to Poly, the momentum parameter set to 0.937, the weight decay set to 0.0005, the learning rate set to 0.001, the epochs set to 270, and the batch size set to 16. The aspect ratio of the image subject was set to 1:1, and for consistency, part of the image was cropped from the original image was cropped out of the original image for consistency.

Table 2

Hardware and software environment
Item	Type
CUDA CPU	11.7 Intel(R) Core(TM) i5-10100F CPU @ 2.9GHz
GPU	NVIDIA GeForce GT 1030
Operating system	Windows 10
Deep learning frame	Pytorch
Python	3.9

4.2 Algorithm performance analysis

In order to objectively assess the model performance, precision (Pr), recall (Re), average precision mean ([email protected]), operational parameters, and algorithm weights are used as evaluation metrics for comparison experiments.

4.2.1 Ablation experiments

Ablation research observes the impact on performance by removing or adding some features of the algorithm [41]. To verify the effectiveness of the introduced module in improving the performance of the algorithm, four algorithms were built to perform ablation experiments on the dataset of this paper, all in the same training environment.

Table 3

Comparison of ablation experiment results
Algorithm	CBAM	BiFPN	C3Ghost	Pr(%)	Re(%)	[email protected](%)	Weights(M)	Parameter
YOLOv5s	×	×	×	90.3	87.5	89.6	14.3	7031701
YOLOv5s-CB	✓	×	×	90.2	88.6	90.4	14.7	7228407
YOLOv5s-BF	×	✓	×	91.5	85.9	90.8	14.3	7031701
YOLOv5s-CG	×	×	✓	90.6	86.7	91.1	12.4	6069413
YOLOv5s-CBC	✓	✓	✓	90.6	88.9	92.7	12.8	6266128

Among them, YOLOv5s-CB indicates the introduction of the CBAM module into the Backbone tail. YOLOv5s-BF indicates the introduction of the Bi-directional Feature Pyramid Network (BiFPN) structure into the Neck. YOLOv5s-CG indicates the replacement of the C3 module in the Neck with the C3Ghost module. YOLOv5s-CBC is the algorithm proposed in this paper.It can be seen from Table 3 that compared with YOLOv5s, the precision rate decreased by 0.1% after the fusion of the CBAM module, while the recall rate and mAP increased by 1.1% and 0.8% respectively, indicating that after adding the attention mechanism, the model is more focused on the target of pests and diseases area, the detection ability of the model is increased, but the algorithm weight is increased by 0.4M. After integrating the BiFPN module, the recall rate decreased by 1.6%, and the precision rate and mAP increased by 1.2% and 1.2%, indicating that the feature fusion ability of the network is improved after adding BiFPN, and the feature consistency of pests and diseases at different scales is ensured. After integrating the C3Ghost module, the recall rate is reduced by 0.8%, the precision rate and mAP are increased by 0.3% and 1.5%, and the weight is reduced by 1.9M. It shows that the weight and operation parameters of the algorithm are reduced after adding C3Ghost, but the detection accuracy of the algorithm is not affected. Combining the above three not only improves the performance of the algorithm, but also reduces the weight of the algorithm to a certain extent, balancing the increase in the weight of the network after the attention mechanism is added, and the final trained algorithm can ensure high accuracy while maintaining a fast detection speed.

Figure 6 shows the change curve of mAP with epoch. The mAP curve fluctuates greatly at the beginning of training from 0 to 100 times, indicating that the convergence speed of the model training is fast in the early stage, which meets the requirements of model training. After 100 times, it is stable and the change is small, indicating that the model is well-trained and there is no overfitting. After 240 times, the curve basically tends to be stable, indicating that the training of the pest detection model is basically completed at this time. The best results of the improved model mAP were obtained when the curves of the five models became stable.

4.2.2 Comparison of different detection models

In order to verify the performance of the improved algorithm, the proposed algorithm was tested against single-stage target detection algorithms such as YOLOv3 tiny, YOLOv4-tiny, YOLOv5n and YOLOv7. The specific results are shown in Table 4.

Table 4

Comprehensive performance of the network under different models
Algorithm	Pr(%)	Re(%)	mAP(%)@0.5	Parammeters	Inference time(ms)	Weights(M)
YOLOv3-tiny	89.4	83.9	87.5	8682862	38.6	17.4
YOLOv4-tiny	89.6	84.8	89.2	9129509	66	18.6
YOLOv5s	90.3	87.5	89.6	7031701	43.4	14.3
YOLOv5n	90.5	87.9	90.8	20881221	83.5	42.1
YOLOv7	88.6	82.8	88.9	37236527	135.1	72.8
YOLOv5-CBC	90.6	88.9	92.7	6266128	33.6	12.8

Compared to other algorithms, the improved algorithm has 0.1-2% higher accuracy, 1.0%-6.1% higher recall, 1.9%-3.8% higher mAP, and the lowest Parameters and Weights. mAP is 3.1% higher compared to YOLOv5s, and Parameters and algorithm Weights are 10.9% and 1.5M lower, respectively. The improved algorithm has a significant advantage over other models in terms of detection accuracy and speed while reducing the algorithm parameters.

Figure 7 is a comparison curve of precision and recall. It can be seen from the figure that both show a rapid growth trend within 100 epochs. After 240 epochs of training, the network has stabilized on each evaluation index. The YOLOv5s algorithm is stable at 90.3% and 87.5%, and the YOLOv5s-CBC is stable at 90.6% and 88.9%, both of which have improved.

4.2.3 Visual comparison of results

To verify the practical effectiveness of this improved algorithm, an application test was carried out in a tomato facility production base and the results are shown in Fig. 8. Some of the white flies in the first column are missed by the original algorithm and are successfully identified by the improved algorithm. In the second column the original algorithm incorrectly identifies late blight as helicoverpa and lacks identification of late blight, while the improved algorithm identifies the disease but incorrectly identifies late blight as early blight. In the third column, the improved algorithm accurately identifies the right half of the early blight that the original algorithm failed to identify. In the fourth column, the original algorithm misses spider mite and misidentifies the background, but the improved algorithm identifies it accurately. The original algorithm in the fifth column misdetects the background as a helicoverpa, whereas the improved algorithm does not make the error. The results show that the improved algorithm is able to identify pests and diseases missed by the original algorithm, as well as increasing the detection of small targets and complex background targets.

4.3 Analysis of clustering results

Five groups of videos of 10s duration were selected for detection, 1s is 12 frames, a total of 120 images can be detected, the detection results are shown in Table 5.

Table 5 Detection data

Number/ Category	AP	BS	EB	HA	LB	LM	SM	WF
1	98	0	0	3	2	2	3	12
2	0	7	92	0	9	5	1	4
3	10	2	4	2	0	0	6	95
4	3	9	9	0	13	81	0	0
5	2	0	0	109	4	0	3	2

Before clustering, the data needs to be preprocessed, and the value under each label is regarded as a dimension of the vector. Eight kinds of diseases and insect pests correspond to labels 1-8 respectively, and multiple one-dimensional arrays are connected together, and they are Converted to a single-column two-dimensional array for use in calculations. The optimal number of clusters is first determined using the Elbow Method based on the data obtained from the algorithm identification, plotting the intra-cluster sum of squares (WCSS) versus the number of clusters. The WCSS is the sum of the squares of the distances between each point and the center of mass of its assigned cluster, and as the number of clusters increases, the WSS usually decreases, followed by a 'twist'. The WCSS-Number of clusters obtained by the Elbow Method is shown in Figure 9(a), with the five data sets clustering best at k=2.

Figure 1(b) shows the data rendering after using k-means clustering. The results obtained by clustering the five groups of data are AP (0.94), EB (0.80), WF (0.86), LM (0.83), and HA (0.95), respectively corresponding to the highest red scatter points in the figure, indicating that k-means clustering can accurately output the results predicted by the algorithm. Analysis of several experiments revealed that k-means clustering was able to successfully output the pest species with the highest number of occurrences and high confidence levels, in line with the expected results.

4.4 Platform realisation

Based on the data detected by the algorithm and the results of the cluster analysis, a tomato pest detection webpage was designed and developed, with some additional features added to make this webpage even better. Figure 10 shows the detection web interface. (a) is the login interface, where the user enters an account and password before entering. (b) is the main query interface, where the user can search for various types of tomato pests and diseases by directly entering the name of the pest or by clicking on the three images below. (c) is the rough query interface, where you can search for pests and diseases based on their characteristics or by clicking on the images directly. (d) is the details interface, which includes the form of the pest, the plant affected, the characteristics of the pest and the prevention and treatment methods. (e) is the detection interface, where images or videos of the pests to be detected are uploaded by clicking on the camera button. (f) is a display of the detection results, showing the optimal identification results and confidence levels calculated by the k-means clustering algorithm, and the corresponding control method and related agent. By integrating the modules, the final platform for tomato pest and disease detection was integrated.

In this study, the YOLOv5 algorithm is improved according to the characteristics of various pests and diseases, using the CBAM module to suppress irrelevant features, the BiFPN module to fuse multi-scale information and the Ghost module to streamline the network. By using an efficient and simplified network structure, the improved method enhances the detection capability of the algorithm and reduces the algorithm parameters and weights, reducing the hardware equipment requirements and detection time of the identification procedure and improving the overall performance of the pest and disease image detection task during the tomato growing period. Designed and developed an online detection platform for pests and diseases. After logging in, users can query information about pests and diseases and upload images or videos to detect unknown pests and diseases. By adding video detection to improve the identification accuracy, the platform uses the k-means clustering algorithm to analyze the identification results and outputs the optimal identification results and confidence to the platform interface.

To verify the feasibility of the platform, 125 photographs were taken at the tomato facility production site using a Sony DSC-H300 camera, as well as 10s of video of different angles of each image at 640 x 480 pixels, uploaded and tested, and the platform showed accuracy as shown in Table 6.

Table 6 Platform experiment results

	YOLOv5s	YOLOv5s-CBC
Image/Video	109/112	114/119
accuracy rate (%)	87.2/89.6	91.2/95.2

It is found that when performing image detection, the original algorithm correctly detected 109 images with an accuracy rate of 87.2% and the improved algorithm correctly detected 114 images with an accuracy rate of 91.2%, an improvement of 4 percentage points. When performing video detection, the original algorithm correctly detected 112 images with an accuracy of 89.6% after using the k-means clustering algorithm to analyse the results output, and the improved algorithm correctly identified 119 images with an accuracy of 95.2%, an improvement of 5.6 percentage points. It is proved that the improved algorithm YOLOv5s-CBC combined with k-means clustering algorithm in this paper can achieve good results and improve the detection ability of pests and diseases.

Figure 11 shows the comparison of the experiment results, after changing the separate image detection to video detection, both can improve the identification accuracy, indicating that video detection can effectively avoid the false detection and leakage problems caused by the interference of light, background, and shooting angle. Compared to other deep learning target detection algorithms, the improved algorithm is able to identify the type and location of pests on the image in a timely and effective manner, while maintaining accuracy with a small weight and inference time. The method can replace traditional manual identification and meet the task requirements of tomato disease identification in real-life situations during the growing period. A comprehensive analysis of the misjudgments and omissions that still exist in the identification process is related to the dispersed image acquisition, the complex background environment, and the limited performance of the computer. Compared to other studies that used data from the same study area, the data from this experiment were selected from 3 sections, with differences in the data itself, but the results are more adaptable. The following two aspects could be investigated in the future: firstly, in terms of tomato pest and disease datasets, although the homemade dataset used in this study includes pests and backgrounds of different complexity, it still lacks images of different species and scenes collected. Future datasets with more extensive and diverse multi-species backgrounds should be used, and the functionality of this diagnostic platform should be continuously enriched. Secondly, on the algorithm side, while continuously validating and improving it, incorporating the latest benchmark networks and modules of the day to develop a more robust algorithm, and improving the clustering algorithm by adding weight to the detection results in the hope of arriving at a more accurate output.

By collecting images of eight major prevalent pests and diseases during the tomato growth period in the facility environment, a data set was created to solve the application problem of target detection in the detection of tomato pests and diseases. Based on the improved YOLOv5 algorithm and k-means clustering algorithm, we designed and developed an online diagnosis platform for pests and diseases, through the identification of disease information, clustering and decision making of detection results, and display of the diagnosis platform, to achieve online diagnosis and interaction of tomato plant diseases during the growing period. And the application of the online diagnosis platform of diseases and insect pests was carried out in the tomato facility production base. The following conclusions are drawn from the experiments described in this paper:

(1) An image database of eight pests and diseases during the growing period of tomato has been established (The dataset has been shared at https://drive.google.com/file/d/1V1cRPVwtqrJuBJifGyQj0z2b0N0Dh_OS/view?usp=sharing), which can provide data support for the detection of pests and diseases during the growing period of tomato. The ablation experiments demonstrate that the CBAM module improves feature extraction for multi-scale pests and diseases, the BiFPN module improves the network's ability to detect small-scale targets by increasing weights, and the Ghost module effectively reduces algorithm weights and operational parameters while ensuring accuracy.

(2) The mAP, precision, and recall of the improved algorithm are 92.7%, 90.6%, and 88.9%, respectively. Compared with YOLOv3-tiny, YOLOv4-tiny, YOLOv5s, YOLOV5n, and YOLOv7, mAP increased by 5.2%, 3.5%, 3.1%, 1.9%, and 3.8%, respectively, and weights decreased by 4.6M, 5.8M, 1.5M, 29.3M, and 60M, respectively. The improved algorithm is proved to have higher detection accuracy and smaller weights, meeting the requirements of accuracy and lightness of the algorithm for tomato pest detection in the growing period.

(3) Based on the improved algorithm and k-means clustering algorithm, a web platform for online diagnosis of tomato pests and diseases in the growing period is designed and developed through the SSM framework to achieve real-time monitoring of tomato pests and diseases or videos. The accuracy of the image diagnosis output is 91.2% and the video diagnosis output is 95.2%, which are 4% and 5.6% higher than the original algorithm, respectively, and can provide a reference for tomato pest and disease prediction and control during the growing period.

Acknowledgments

Appreciations are given to the editors and reviewer of the Journal Plant Method.

Authors contributions

XJ contributed to the conception of the study; XZ performed the data analyses and wrote the manuscript; JJ contributed significantly to the analysis and manuscript preparation; ML helped perform the analysis with constructive discussions; XX performed the experiment; BZ designed the method and created the model; All authors reviewed the manuscript.

Funding

This study was supported by the National Key Research and Development Program of China Project (No. 2021YFD2000700), the Natural Science Foundation of Henan Province (No. 202300410124), Talent Support Plan in Henan Province (No. ZYQNBJRC2021-04), and Special projects for industrial foundation reconstruction and high-quality development of manufacturing industry in MIIT.

Availability of data and materials

For relevant data and codes, please contact the corresponding author of this manuscript.

Conflicts of interest

The authors declare no competing interests.

Ethics approval and consent to participate

Not applicable.

Consent to publish

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Jia Z, Zhang Y, Wang H,Liang D. Identification Method of Tomato Disease Period Based on Res2Net and Bilinear Attention Mechanism.Transactions of the Chinese Society for Agricultural Machinery,2022,53(7):259-266.https://doi.org/10.6041/j.issn.1000-1298.2022.07.027
Osdaghi E, Jones JB, Sharma A, Goss EM, Abrahamian P, Newberry EA, Potnis N, Carvalho R, Choudhary M, Paret ML. A centenary for bacterial spot of tomato and pepper. Molecular Plant Pathology 2021, 22(12):1500.https://doi.org/10.1111/mpp.13125
Hu Z, Zhang Y. Effect of Dimensionality Reduction and Noise Reduction on Hyperspectral Recognition During Incubation Period of Tomato Early Blight. Spectroscopy and Spectral Analysis. 2023, 43(3):744-752.https://doi.org/10.3964/j.issn.1000-0593(2023)03-0744-09
Wu KT, Gevens AJ, Silva EM. Exploring grower strategies and needs for enhancing organic disease management of tomato late blight. Renewable Agriculture and Food Systems 2022, 37(5):382-398.https://doi.org/10.1017/s1742170521000594
Zhao T, Pei T, Jiang J, Yang H, Zhang H, Li J, Xu X. Understanding the mechanisms of resistance to tomato leaf mold: A review. Horticultural Plant Journal 2022.https://doi.org/10.1016/j.hpj.2022.04.008
Massumi H, Shaabanian M, Pour AH, Heydarnejad J, Rahimian H. Incidence of viruses infecting tomato and their natural hosts in the southeast and central regions of Iran. Plant disease 2009, 93(1):67-72.https://doi.org/10.1094/pdis-93-1-0067
Mansour A, Al-Banna L, Alsmairat N. Disease management of organic tomato under greenhouse conditions in the Jordan Valley. Crop Protection 2014, 60:48-55. https://doi.org/10.1016/j.cropro.2014.03.001
Mantelin S, Bhattarai KK, Kaloshian I. Ethylene contributes to potato aphid susceptibility in a compatible tomato host. New Phytologist 2009, 183(2):444-456.https://doi.org/10.1111/j.1469-8137.2009.02870.x
Reddy GV, Tangtrakulwanich K. Module of integrated insect pest management on tomato with growers’ participation. 2014.https://doi.org/10.5539/jas.v6n5p10
Obopile M, Munthali D, Matilo B. Farmers’ knowledge, perceptions and management of vegetable pests and diseases in Botswana. Crop Protection 2008, 27(8):1220-1224. https://doi.org/10.1016/j.cropro.2008.03.003
Saleem M, ul Hasan M, Sagheer M, Atiq M. Determination of insecticide resistance in Bemisia tabaci (Hemiptera: Aleyrodidae) populations from Punjab, Pakistan. International Journal of Tropical Insect Science 2021, 41:1799-1808.https://doi.org/10.1007/s42690-020-00393-4
Mateos Fernández R, Petek M, Gerasymenko I, Juteršek M, Baebler Š, Kallam K, Moreno Giménez E, Gondolf J, Nordmann A, Gruden K. Insect pest management in the age of synthetic biology. Plant Biotechnology Journal 2022, 20(1):25-36.https://doi.org/10.1111/pbi.13685
Arunnehru J, Vidhyasagar B, Anwar Basha H. Plant Leaf Diseases Recognition Using Convolutional Neural Network and Transfer Learning. In: International Conference on Communication, Computing and Electronics Systems: 2020. Springer: 221-229.https://doi.org/10.1007/978-981-15-2612-1_21
Singh R, Ao NT, Kangjam V, Rajesha G, Banik S. Plant growth promoting microbial consortia against late blight disease of tomato under natural epiphytotic conditions. Indian Phytopathology 2022, 75(2):527-539.https://doi.org/10.1007/s42360-022-00464-1
Li Y, Wang H, Dang LM, Sadeghi-Niaraki A, Moon H. Crop pest recognition in natural scenes using convolutional neural networks. Computers and Electronics in Agriculture 2020, 169:105174.https://doi.org/10.1016/j.compag.2019.105174
Liu J, Wang X. Plant diseases and pests detection based on deep learning: a review. Plant Methods 2021, 17(1).https://doi.org/10.1186/s13007-021-00722-9
Kamilaris A, Prenafeta-Boldú FX. Deep learning in agriculture: A survey. Computers and electronics in agriculture 2018, 147:70-90. https://doi.org/10.1016/j.compag.2018.02.016
Lu J, Tan L, Jiang H. Review on convolutional neural network (CNN) applied to plant leaf disease classification. Agriculture 2021, 11(8):707.https://doi.org/10.3390/agriculture11080707
Deng F, Mao W, Zeng Z, Zeng H, Wei B. Multiple Diseases and Pests Detection Based on Federated Learning and Improved Faster R-CNN. IEEE Transactions on Instrumentation and Measurement 2022, 71:1-11.https://doi.org/10.1109/tim.2022.3201937
Jiao L, Dong S, Zhang S, Xie C, Wang H. AF-RCNN: An anchor-free convolutional neural network for multi-categories agricultural pest detection. Computers and Electronics in Agriculture 2020, 174:105522.https://doi.org/10.1016/j.compag.2020.105522
Zhang Y, Song C, Zhang D. Deep learning-based object detection improvement for tomato disease. IEEE Access 2020, 8:56607-56614.https://doi.org/10.1109/access.2020.2982456
Xie X, Ma Y, Liu B, He J, Li S, Wang H. A deep-learning-based real-time detector for grape leaf diseases using improved convolutional neural networks. Frontiers in plant science 2020, 11:751.https://doi.org/10.3389/fpls.2020.00751
Wang J, Yu L, Yang J, Dong H. DBA_SSD: A novel end-to-end object detection algorithm applied to plant disease detection. Information 2021, 12(11):474.https://doi.org/10.3390/info12110474
Sun H, Xu H, Liu B, He D, He J, Zhang H, Geng N. MEAN-SSD: A novel real-time detector for apple leaf diseases using improved light-weight convolutional neural networks. Computers and Electronics in Agriculture 2021, 189:106379.https://doi.org/10.1016/j.compag.2021.106379
Liu J, Wang X. Tomato diseases and pests detection based on improved Yolo V3 convolutional neural network. Frontiers in plant science 2020, 11:898. https://doi.org/10.3389/fpls.2020.00898
Wang X, Liu J, Zhu X. Early real-time detection algorithm of tomato diseases and pests in the natural environment. Plant Methods 2021, 17(1).https://doi.org/10.1186/s13007-021-00745-2
Liu J, Wang X, Liu G. Tomato pests recognition algorithm based on improved YOLOv4. Frontiers in Plant Science 2022:1894.https://doi.org/10.3389/fpls.2022.814681
Qi J, Liu X, Liu K, Xu F, Guo H, Tian X, Li M, Bao Z, Li Y. An improved YOLOv5 model based on visual attention mechanism: Application to recognition of tomato virus disease. Computers and Electronics in Agriculture 2022, 194:106780.https://doi.org/10.1016/j.compag.2022.106780
Chen Z, Wu R, Lin Y, Li C, Chen S, Yuan Z, Chen S, Zou X. Plant disease recognition model based on improved YOLOv5. Agronomy 2022, 12(2):365. https://doi.org/10.3390/agronomy12020365
Hughes D, Salathé M. An open access repository of images on plant health to enable the development of mobile disease diagnostics. arXiv preprint arXiv:151108060 2015. https://arxiv.org/abs/1511.08060
Shi Z, Shi M, Lin W. The implementation of crawling news page based on incremental web crawler. In: 2016 4th Intl Conf on Applied Computing and Information Technology/3rd Intl Conf on Computational Science/Intelligence and Applied Informatics/1st Intl Conf on Big Data, Cloud Computing, Data Science & Engineering (ACIT-CSII-BCD): 2016. IEEE: 348-351. https://doi.org/10.1109/ACIT-CSII-BCD.2016.073
Zhu J-Y, Park T, Isola P, Efros AA. Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision: 2017. 2223-2232. https://arxiv.org/abs/1703.10593
Bochkovskiy A, Wang C-Y, Liao H-YM. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:200410934 2020. https://arxiv.org/abs/2004.10934
Zhang H, Cisse M, Dauphin YN, Lopez-Paz D. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:171009412 2017. https://arxiv.org/abs/1710.09412
Redmon J, Divvala S, Girshick R, Farhadi A. You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition: 2016. 779-788.https://arxiv.org/abs/1506.02640
Hao S, Lee D-H, Zhao D. Sequence to sequence learning with attention mechanism for short-term passenger flow prediction in large-scale metro system. Transportation Research Part C: Emerging Technologies 2019, 107:287-300.https://doi.org/10.1016/j.trc.2019.08.005
Woo S, Park J, Lee J-Y, Kweon IS. Cbam: Convolutional block attention module. In: Proceedings of the European conference on computer vision (ECCV): 2018. 3-19. https://arxiv.org/abs/1807.06521
Tan M, Pang R, Le QV. Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition: 2020. 10781-10790. https://arxiv.org/abs/1911.09070
Han K, Wang Y, Tian Q, Guo J, Xu C, Xu C. Ghostnet: More features from cheap operations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition: 2020. 1580-1589.https://arxiv.org/abs/1911.11907
Liu H, Chen J, Dy J, Fu Y. Transforming Complex Problems into K-means Solutions. IEEE Transactions on Pattern Analysis and Machine Intelligence 2023. https://doi.org/10.1109/tpami.2023.3237667
Zhang Y, Ma B, Hu Y, Li C, Li Y. Accurate cotton diseases and pests detection in complex background based on an improved YOLOX model. Computers and Electronics in Agriculture 2022, 203.https://doi.org/10.1016/j.compag.2022.107484

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Design and Research of an Online Diagnosis Platform for Tomato Seedling Facilities Production Diseases

Status:

Version 1

Abstract

Figures

1. Introduction

2. Construction of pest and disease datasets

2.1 Data acquisition

2.2 Data expansion and enhancement

3. Detection platform design

3.1 Identification algorithm model

3.1.1 Improvement details

3.2 Cluster analysis model

3.3 Interface design module

4. Results and analysis

4.1 Experimental environment

4.2 Algorithm performance analysis

4.2.1 Ablation experiments

4.2.2 Comparison of different detection models

4.2.3 Visual comparison of results

4.3 Analysis of clustering results

4.4 Platform realisation

5. Discussion

6. conclusion

Declarations

References

Additional Declarations

Status:

Version 1