Detection And Grading of Rust Disease Severities from Wheat Images Using Deep Learning Techniques

Detection And Grading of Rust Disease Severities from Wheat Images Using Deep Learning Techniques | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Detection And Grading of Rust Disease Severities from Wheat Images Using Deep Learning Techniques Jibril Muhammadnur, Million Meshesha, Jemal Abate, Araba Aman This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5796873/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Wheat production, a cornerstone of food security in Ethiopia, is heavily impacted by stripe rust disease, which leads to significant economic losses. Traditional methods for detecting and classifying disease severity are labor-intensive, error-prone, and costly. This study introduces a novel convolutional neural network (CNN)-based model, WRNet, designed for the detection and severity classification of wheat yellow rust disease, along with treatment recommendations. Utilizing 20,000 annotated images collected from Ethiopia, the model applies advanced preprocessing techniques such as noise removal and segmentation using bilateral filtering and k-means algorithms. The WRNet model achieved superior performance with 99.11% training accuracy, 99.04% validation accuracy, and 99% testing accuracy, surpassing pre-trained models such as InceptionV3, InceptionResNetV2, and MobileNetV2. Additionally, the system provides fungicide dosage recommendations tailored to severity levels, ensuring effective disease management. A user-friendly prototype interface developed using Flask enables domain experts to classify disease severity and receive treatment recommendations, offering a scalable solution for precision agriculture in Ethiopia and beyond. Physical sciences/Energy science and technology Physical sciences/Engineering Physical sciences/Mathematics and computing Wheat yellow rust deep learning CNN WRNet disease severity classification precision agriculture image preprocessing treatment recommendation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Introduction The agriculture sector plays a crucial role in the process of economic growth of the country. Farmers consult specialists to get their opinions on the probability of crop diseases occurring and suggestions for controlling the diseases to increase crop yield production. Early disease detection and classification are particularly easy in small-scale farming. Due to the difficulty of monitoring, early disease detection, and classification in large-scale farms, serious disease occurrences and pest expansion cannot be controlled organically. In general, visual-based crop disease detection is expensive, inefficient, inaccurate, and difficult (Haile, 2019 ). Wheat is a food crop that delivers the most nutrition worldwide (Tiwari and Shoran, 2022 ). Wheat is one of the most significant crops for Ethiopia's food security, it is grown on 2.1 million hectares of land and yields 6.7 million tons yearly. Ethiopia's wheat production has a huge potential to grow, but it is currently constrained by the prevalence of rust disease stresses, yield gaps brought on by the slow implementation of new technologies, the massive price and limited availability of inputs, the lack of public and private investments, and the country's insufficient infrastructure and marketing systems (Tadesse et al., 2022 ). The degree of production losses and quality degradation may vary in wheat cultivation fields depending on the severity of the disease, and as a result, economic losses may differ (Hayit et al., 2021 ). Manual optical examination of the plant is a time-consuming and tedious method for diagnosing plant disorders, and certain diseases are not present in the plants. There are comprehensive methods that can identify and categorize the rust disease of wheat. Deep Learning is the approaches that have been applied in the agriculture sector and have greatly increased in popularity (Kabir et al., 2020 ). It is a class of machine learning models inspired by the structure and functioning of the human brain composed of interconnected nodes, often referred to as neurons organized into layers. Deep learning models have the advantage of automatically extracting features from raw data, a process known as feature learning (Ganesh and Kumar, 2018 ). Deep learning in agriculture is not only helping farmers to detect and classify diseases at early stages but also to predict treatment dosage based on severity classification and shifts to precise cultivation for higher and better-quality crop yield. The SoftMax classifier was used in this study to categorize the severity levels of the rust disease and treatment recommendations. The classifier is easy to use and has good computational performance. For the development of automatic diagnostics recommendations, the complete process of knowledge acquisition and engineering for disease management has been applied and a new knowledge base has been applied knowledge with deep learning techniques. Therefore, the objective of this research is to design and develop a deep learning-based model for the wheat yellow rust diseases detection and severity classification and treatment recommendations. The focus of a design and development step can be on front-end analysis, planning, production, and/or evaluation. We have developed a wheat yellow rust disease severity classification and treatment recommendation architecture using the deep learning method. In this study, the image was collected and then image preprocessing was performed. The preprocessed data is used for training and testing the classification model, based on which specified severities class classification is done. The knowledge base is also constructed for treatment recommendations. Related Works Several researches were conducted to automate the detection of wheat rust diseases through the application of deep learning, image processing, and machine learning techniques. One specific area of research is the identification of wheat diseases using deep learning (DL) approaches, which has seen an increase in studies over the years. Aboneh et al. (2021) proposed a deep-learning approach for detecting various wheat diseases. Their method utilized four popular deep learning architectures: Inceptionv3, ResNet50, and VGG16/19. They evaluated the computation cost and CPU run time using a Jetson Nano Board accelerator, a low-power embedded graphics processing unit that enables multiple neural networks to run simultaneously and supports a computer vision algorithm for image classification. The authors compared the performance of these architectures on a dataset of 1500 images representing three classes of wheat diseases. They conducted data preprocessing, standardization, formatting, removal, and rescaling. The dataset was a combination of images from wheat farms in Ethiopia, gathered in collaboration with the Bishoftu Agricultural Research Institute, and a publicly available online repository. The experimental results showed that the VGG19 model achieved 99.38% accuracy in classifying wheat diseases. However, the study did not include severity classification and treatment recommendations. Azadbakht et al. (2019) proposed a machine learning (ML)-based method for detecting wheat leaf rust disease. They constructed their dataset using canopy scale data at various leaf area index (LAI) levels. Their framework identified the severity of wheat leaf rust at the canopy scale using four ML techniques: Random Forests Regression (RFR), ν-support vector regression (ν-SVR), Gaussian process regression (GPR), and boosted regression trees (BRT). Hyperspectral signature data collected in the field using a spectroradiometer served as inputs for the models. The ML inversion models all performed well at high and medium LAI levels, achieving up to 99% accuracy with ν-support vector regression (ν-SVR), which outperformed other ML methods. The results indicated that the performance of the ML methods improved as the LAI value increased. However, the experiments focused solely on leaf rust disease severity and did not include yellow rust of wheat and also treatment recommendations. Xu et al. (2017) conducted a study on Automatic Wheat Leaf Rust Detection and Grading Diagnosis via an Embedded Image Processing System. The study focused on detecting color images, which involved the processes of image capturing, image processing with detection, and disease classification. The method used for each process varied based on the characteristics of different types of diseases. The research employed Sobel edge detection based on exploration. Another significant finding was the grading of infected regions based on the percentage of scab area. The system captured images using a web camera connected to an ARM microcontroller through USB, and the results were displayed on a dedicated LCD. The proposed system achieved an accuracy rate of up to 96.2% for recognition. However, the study did not include wheat yellow rust scaling classification and treatment recommendation. Assefa (2021) conducted research on the detection and grading of stem rust in wheat to address the severity levels of stem rust according to CIMMYT grading guidelines. The study aimed to utilize a deep convolutional neural network for this purpose, consisting of four components: preprocessing, segmentation, feature extraction, and classification. Adaptive Thresholding was employed for segmentation, while a Gabor filter was proposed for texture feature enhancement in the image to detect and select important disease features. For classification, a convolutional neural network and a 12-way Softmax were utilized for grading into specific classes (Resistant (TR), Moderately resistant (MR), and Susceptible (S)). To address overfitting, data augmentation and dropout techniques were applied. The study evaluated three state-of-the-art CNN architectures, including AlexNet, VGGNet, and GoogLeNet, with SRNet achieving a training accuracy of 92% and a testing accuracy of 92.01%. The proposed system was implemented using Keras (with TensorFlow as a backend) in Python and trained and tested using an image dataset collected and prepared with the assistance of pathologists from Haramaya University, divided into 70/30 for training and testing, respectively. The model achieved an accuracy of 92.02% for training and 92.01% for testing in detecting and grading wheat stem rust. It's important to note that the study focused on stem rust disease and did not include yellow rust and treatment recommendations. Zhang et al. (2019) proposed a deep learning-based method for the automated detection of yellow rust disease from high-resolution hyperspectral UAV images. They utilized multiple Inception-ResNet layers for deep feature extraction, optimized to determine the most suitable depth and width. The model incorporated spatial and spectral information for yellow rust detection, resulting in improved accuracy. The study confirmed the potential of the proposed deep-learning architecture for crop rust disease detection, despite increased complexity and slightly lower accuracy compared to leaf-scale reports by other researchers. In a carefully planned field experiment, UAVs captured hyperspectral imagery of healthy and rust-infected wheat plots on five distinct dates throughout a crop cycle. After deep feature extraction using Inception-ResNet blocks, the feature maps were converted into a three-class classifier for rust, healthy, and others using an average pooling layer and a fully connected layer. The study did not include a classification of yellow rust severities or treatment recommendations. Automatic detection of yellow rust in wheat using image processing and machine learning approach done by (Tsiyon, 2020 ) a proposed system that integrates components, such as preprocessing (Image resizing, Histogram equalization, and Noise removal), segmentation, feature extraction, and classification. They perform implementation using Keras (with Tensor Flow as a backend) in Python and they take samples of wheat leaf from South Gonder Zone from a specific place, Gobgob Kebele. They take the total number of images taken was 3322 containing wheat leaf. To build the detection models for the prediction of wheat yellow rust, they investigate Softmax and SVM classifiers. An overall accuracy of 98.04% is achieved by using the SVM classifier; on the other hand, the Softmax classifier achieves an accuracy of 97.74%. However, this research is done on yellow rust detection, which did not include severities classification and treatment recommendation. To the best of researcher knowledge, there is no research works done that focus on automatically detecting, severities assessment, and treatment recommendation of yellow rust of wheat using deep learning. Materials and Methods Design science research is selected as general approaches to design and develop wheat yellow rust diseases detection and severity classification and treatment recommendations. By facilitating the change from the current situation to the expected state, this strategy attempts to develop and evaluate the artifacts necessary to solve the problems that have been identified. The recognition of plant diseases passes through a series of steps/procedures that would be applied to separate plants, in which each wheat yellow rust disease must be categorized into one of predefined classes based on observed attributes or features. DSR is selected for the following reasons: first, this methodology has a strong emphasis on developing artifacts and their usage by the specialist to solve real-world problems. Second, DSR, which focuses on real-world problem solving, contains prescriptive or solution-oriented knowledge that is used for designing complex and relevant field problems by using the results of scientific reasoning (such as prediction). Finally, within the realm of information systems (IS), DSR aims to acquire knowledge and comprehension of a problem and develop a solution for the demands of the business and its environment (Yimenu, 2021 ). In this study, a CNN architecture is used. Because deep learning via CNN didn't need a different algorithm for feature extraction like machine learning methods. It uses the learned characteristics to automatically learn features and classify them into the best-fitting group or class. The deep-learning workflow for solving real-world problems includes three processing steps: data analysis and preprocessing, DL model development and training, and validation and interpretation. However, feature extraction in the DL model is automated rather than manual, in contrast to ML modeling (Sarker, 2021 ). CNN architecture is designed using convolutional layers, pooling layers, fully connected layers, and dropout layers. These operations are then applied to all subsequent convolutional layers. The images are passed through the convolutional layer, Rectified Linear Unit layer, batch normalization layer, and max pooling layer for feature extraction. Finally, the fully connected layer is used for disease classification. The input layer of the CNN model can directly receive the image as input. A color image has three color channels. Input layer The image can be directly fed into the CNN model as input. The number of color channels in a color image is three, while in a gray-scale image, it is one. In this study, 950 x 250 x 3 (height, width, and number of color channels) was used respectively, because it is the most recommended one by various researchers, to use all of the features of the image in their full resolution (Fasnacht et al., 2020 ). Convolutional Layer : The convolution operation requires four parameters: the number of filters, filter size, stride size, and amount of zero-padding. In this study, we have tested different numbers of convolution layers and filters, such as 32, 64, 128, 256, 512, and 1024. We selected the ones that achieved higher accuracy. To ensure better detection, we used filter sizes of 5 x 5, 3 x 3, and 2 x 2 at a single layer. A 2 x 2 filter size can detect a feature that cannot be detected by a 3 x 3 filter size. We used stride sizes of 2 (2, 2) and 1 (1, 1) to cover the entire image and to avoid overlapping. We have used "the same padding" parameter in our convolutional operations, which ensures that the output is of the same size as the input. The amount of zero-padding is used to adjust the output size. Padding is a technique used to resolve issues in normal convolutional operations, such as the reduction of the height and width of the output feature map. A typical padding method is zero-padding. We have performed the convolution operation repeatedly before downsampling the input image using the pooling operation. Using multiple convolution layers before applying a pooling layer allows the model to develop more complex features before the destructive pooling operation is carried out. Experimental Result To accomplish the experiments, we have implemented our designed system with the required parameters. Four experiments conducted. The first experiment conducted with the creating CNN model from scratch and others three experiments were conducted with pre-trained deep learning models such as MobileNetV2, InceptionV3, and InceptionResnetV2. The experiments are performed with the same corpus size. The effect of applying different preprocessing techniques such as noise removal and segmentation, and after data augmentation, and dropout applied are evaluated and compared with CNN. In addition, the test results were presented and compared with other researchers' work. Experiment 1: Creating a Model from Scratch Training from Scratch is a method of training a CNN network, where the researcher decides all the hyperparameters such as the optimizer, learning rate, number of layers, epochs, filters per layer, activation function, and fully connected layers. Our proposed 'WRNet' model was evaluated through a series of experiments in which different noise removal techniques were applied, including denoising, bilateral filter, K-means segmentation, and data augmentation. Hereunder we present experiments conducted using segmentation and filtering methods. The experiment conducted by applying bilateral filtering and k-means segmentation algorithms. This proposed CNN model experiment has been tested by splitting the whole dataset into 70% for training and 30% for testing. The following figure shows the training process of the model with 10 number of epochs. Both training accuracy and validation accuracy are increasing with the number of epochs. The result recorded from training accuracy was 99.11% and validation accuracy of 99.04%. The trained model evaluated by three performance measurements metrics and the result recorded as follows. Throughout the curve in Fig. 1 , the training accuracy is always higher than the validation accuracy. However, the gap between the two becomes smaller over time. The validation loss remains relatively small during the entire curve. Additionally, both the training and validation loss decrease significantly. After training completed, the model predicts the output for each class of severity level. Table 1: Confusion Matrix of model with filter and segmentation The above table shows the graphical representation of the confusion matrix for test images for S(Susceptible), MS, MR, R disease, and healthy level. As shown above for S, MS, and all images (218) are correctly classified in their classes, while no images are incorrectly classified. For MR, the model correctly classified 214 images as MR, with only 4 images incorrectly classified as healthy. Additionally, 215 images were correctly classified as healthy, with only 3 images incorrectly classified as MR. This indicates that the model has effectively learned from the training data and can generalize well to the validation and testing datasets. So, the model constructed after applying k-means segmentation with bilateral filtering is selected and proposed for comparison with pre-trained models. Experiment 2: InceptionV3 model This experiment performed with pre-trained InceptionV3 by applying bilateral filtering and k-means segmentation algorithms. The following figure shows the training process of the model with 10 number of epochs. The pre-trained model achieved a training accuracy of 95.72%, a validation accuracy of 95.12%, and a testing accuracy of 87.03%. These classification accuracies were attained through the utilization of image bilateral filtering, k-means segmentation after data augmentation, and dropout application. It has been observed that the InceptionV3 Model has identified overfitting, which is indicated by a significant difference between the training and validation loss. The training loss is relatively small and remains stable throughout the curve, while the validation loss is unstable and shows fluctuations. This suggests that the model is not generalizing well to new data. After training and testing the Inceptionv3 model, the prediction result for one image produces the above result with a percentage of confidence. Experiment 3: InceptionResnetv2 model by applying filtering and Segmentation The third experiment uses InceptionResnetv2 to measure the performance of training accuracy and testing accuracy of model at epochs 40. InceptionResnetv2 achieves the accuracy of training, validation and testing of 97.47%, 97.28%, and 83% respectively after applying image bilateral filtering and k-means segmentation algorithms. Experiment 4: With MobileNetV2 After applying filtering and Segmentation The training accuracy consistently surpasses the validation accuracy, as evidenced by the curve. Notably, the difference between the two accuracies diminishes compared to the previous curve. Throughout the curve, the validation loss remains relatively small. Meanwhile, the training loss exhibits a steep decline from 0.14, while the validation loss fluctuates up and down and peaks at 0.13. After training and testing the MobileNetV2 model, the prediction result for one image produces the above result with a percentage of confidence. Experimental Discussion From the above experimental result conducted on both the WRNet model and pre-trained models such as InceptionV3, InceptionResnetV2, MobileNetV2 pre-trained model, the researcher trained with epochs of 50. For each experiment, we used a total of 20,000 images for training, validation, and testing the model. Table 2 Comparison of models Models Training Accuracy % Validation Accuracy % Testing Accuracy% Testing Loss% InceptionV3 95.72 95.12 87.03 0.37 InceptionResnetV2 97.47 97.28 83 0.27 MobileNetV2 99.02% 99.02% 89.01% 0.39 WRNet model 99.11% 99.04% 99.00% 0.004 From our comparison, we have found that the proposed model with a k-means segmentation algorithm and bilateral filtering is very effective. The training of the WRNet model increases the testing accuracy from 83–99% with no over-fitting. And in the case of InceptionV3, InceptionResNetV2, and MobileNetv2, model the testing accuracy increases from 83%, 87.03–89.01% respectively. The validation and testing accuracy performances of our WRNet model in all cases are maximum. At the same time, the amount of time it takes to train and test the dataset showed a better performance. In addition, the chance of overfitting is rigorously minimal in the case of using bilateral filtering and k-means segmentation image training. We are subjected to adding different regularization techniques to reduce overfitting in pre-trained model cases. In all cases, bilateral filtering and k-means segmentation method has given the most effective result compared to those other filtering and segmentation. Analyzing the created models, the comparison showed that the proposed WRNet model trained with bilateral filtering and k-means segmentation images is superior to all other models experimented with. Therefore, we propose this model for Wheat Rust disease detection and severity classification. Treatment Recommendation. To effectively control the spread of yellow rust, it is suggested to apply fungicide solutions such as propiconazole 250 EC (Tilt), tebuconazole 250 EC (Folicur), or triadimefon 250 WP (Bayleton) (Erdado, 2019 ). For all of the severity Propiconazole 0.65 li/ha with 250 li/ha, Opera Max 1 li/ha with 250 li/ha, Tilt 250EC 0.5-1 li/ha with 250 li/ha, and Progress 250EC 0.5 li/ha with 250 li/ha. But the main difference is the number of times the fungicide is applied. Based on this the percentage of dosage pesticides for each classification of wheat yellow rust contains for R one time, for MR two times per week, for MS 2–3 times per week, and for S 3–4 times per week is important (Erdado, 2019 ). Prototype Deployment After conducting the study, the researcher developed a user interface for the proposed research work. This user interface is designed for domain experts or end-users to classify wheat yellow rust severity diseases and receive treatment recommendations. The user interface is created using Python and the Flask framework, which is used for developing web applications in Python. The system can detect five types of wheat yellow rust disease severity, including healthy plants. This enables the user to perform a wheat yellow rust disease severity classification service. To access the system, the end-user can launch the web application by typing the URL ( http://127.0.0.1:5000 ). into any web browser and accessing it through port 5000. When the user submits an HTTP request to the webserver, the server responds with an HTTP response. The prototype's initial user interface includes a button for uploading images. The user interface depicted in Fig. 36 above is the first page that is opened when the user enters the URL of the wheat yellow rust disease classification and treatment recommendation system. In this prototype flask server is run at http://localhost:5000 (in the browser) to allow the user to import the images that need to detect severity level and recommendation. The user imports the wheat yellow rust image by using the ‘choose file button’ that is displayed in the user interface. The home page of the developed model after importing the images is shown in Fig. 37 below. After uploading images and clicking on the "predict" button, the user will receive a classification of the wheat yellow rust severity. A sample result is depicted in Fig. 19 below. Conclusion Wheat is a major crop worldwide, including in Ethiopia. However, wheat production and quality are affected by stripe rust (Puccinia striiformis f. sp. tritici), which is the most common fungal disease, especially in cool climates. Traditional monitoring and examination methods are employed after disease progression reaches a higher stage, and are time-consuming, labor-intensive, error-prone, and very expensive. In this study, we constructed an optimal model using a CNN model that detects and classifies the severity of disease and provides treatment recommendations for wheat yellow rust. In this study, 20,000 yellow rust leaf images were collected from the Kulumsa Research Center in the Arsi zone of Asela town. After collecting the required images, the researcher applied different image preprocessing techniques, such as image resizing, noise removal, filtering, and segmentation. To conduct this study, training from scratch (i.e., the researcher settings different hyper-parameters) and transfer learning methods such as InceptionV3, InceptionResnetV2, and MobileNetV2 were used. The developed model was evaluated using a confusion matrix, accuracy, recall, precision, f1-score, and support. Among all the models, the proposed model trained from scratch performed better, with 99.11% training accuracy, 99.04% validation accuracy, and 99% testing accuracy. Additionally, the researcher creates knowledge modeling by utilizing a decision tree to represent the concepts involved in the classification and treatment of wheat yellow rust disease severity. Next, a rule-based representation technique is used to represent the knowledge. With the help of this developed model, a treatment plan can be quickly suggested, and the disease can be detected early. Declarations Author Contribution J.M. and M.M. conceived and designed the study. J.M. conducted the experiments and implemented the proposed model. M.M. and J.A. collected and preprocessed the data. J.A. wrote the main manuscript text and conducted the comparative analysis. A.A. developed the user interface and assisted with the treatment recommendation module. J.A. and A.A. prepared Figures 1–12 and Tables 1–2. All authors reviewed and approved the final manuscript. Data Availability The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request. References Erdado, F. Integrated Management of Stripe Rust (Puccinia striiformis f.sp. tritici) OF Bread Wheat. Unpublished , November , 74. (2019). Fasnacht, L., Renard, P. & Brunner, P. Robust input layer for neural networks for hyperspectral classification of data with missing bands. Appl. Comput. Geosci. 8 (August), 100034. https://doi.org/10.1016/j.acags.2020.100034 (2020). Ganesh, B. & Kumar, C. Deep learning Techniques in Image processing. National Conference On Emerging Trends in Computing Technologies (NCETCT) , 18 , 1–5. (2018). Haile, E. Plant Disease Detection and Classification Using Artificial Neural Network. Addis Ababa Univ. Repository . 01 , 83 (2019). Hayit, T., Erbay, H., Varçın, F., Hayit, F. & Akci, N. Determination of the severity level of yellow rust disease in wheat by using convolutional neural networks. J. Plant. Pathol. 103 (3), 923–934. https://doi.org/10.1007/s42161-021-00886-2 (2021). Kabir, M. M., Ohi, A. Q. & Mridha, M. F. A Multi-Plant Disease Diagnosis Method using Convolutional Neural Network. Res. Gate , 1–13. (2020). Sarker, I. H. Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions. SN Comput. Sci. 2 (6), 1–20. https://doi.org/10.1007/s42979-021-00815-1 (2021). Tadesse, W., Zegeye, H., Debele, T. & Kassa, D. Wheat Production and Breeding in Ethiopia: Retrospect and Prospects. Crop Breed. Genet. Genomics . 1–22. https://doi.org/10.20900/cbgg20220003 (2022). Tiwari, V. & Shoran, J. Growth and Production of Wheat. EOLSS (Encyclopedia Life Support Systems) . I , 1–9 (2022). Tsiyon, W. Automatic Detection of Yellow rust in Wheat using Image Processing and Machine Learning Approach. DSpace Institution’s Institutional Repository . 01 , 1–90 (2020). Yimenu, D. F. Image Based Sorghum Leaf Disease Classification Using Deep Learning Approach. Debre Berhan Univ. Repository , 1–138. (2021). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5796873","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":399977975,"identity":"1b8c01e7-768d-4a10-a271-8a3f0ccc6cc4","order_by":0,"name":"Jibril Muhammadnur","email":"","orcid":"","institution":"Haramaya University","correspondingAuthor":false,"prefix":"","firstName":"Jibril","middleName":"","lastName":"Muhammadnur","suffix":""},{"id":399977978,"identity":"858f34c1-8b48-4a9e-b63c-4b05994d43fc","order_by":1,"name":"Million Meshesha","email":"","orcid":"","institution":"Addis Ababa University","correspondingAuthor":false,"prefix":"","firstName":"Million","middleName":"","lastName":"Meshesha","suffix":""},{"id":399977980,"identity":"a55c30fc-3d7b-4128-a64a-961eba136a25","order_by":2,"name":"Jemal Abate","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7klEQVRIiWNgGAWjYDACCRDBA8TszAdAXBlitDA2gLUwsyWAuDxEagEBZh4DmIX4Af/s5ucPPsjYRPM383x+daPGgoeB/fDRDXgtuXPMsHEGT1rujMO826xzjgEdxpOWdgOvNTcSDJt5eA7nNgC1GOewAbVI8Jjh1SJ/I/1j8x+e/7nzD/M8M875R4QWgxs5hs0MPAdyNxzmYX6c20aEFsM7Zwpn9vAk5248zGbGnNsnwcNGyC9yt9s3fPjZY5c773jz48853+rk+NkPH8PvfRBg7AFTbOCUwEZQORj8AJPMH4hTPQpGwSgYBSMNAADSD0onNBE4rwAAAABJRU5ErkJggg==","orcid":"","institution":"Haramaya University","correspondingAuthor":true,"prefix":"","firstName":"Jemal","middleName":"","lastName":"Abate","suffix":""},{"id":399977982,"identity":"859cfd7d-e084-437b-85fb-d85b2834dbc2","order_by":3,"name":"Araba Aman","email":"","orcid":"","institution":"Haramaya University","correspondingAuthor":false,"prefix":"","firstName":"Araba","middleName":"","lastName":"Aman","suffix":""}],"badges":[],"createdAt":"2025-01-09 13:23:31","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5796873/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5796873/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":73520165,"identity":"6bf5da43-fd64-44bc-9450-cdee6c0a33fd","added_by":"auto","created_at":"2025-01-10 18:12:46","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":50275,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eModel Training and Validation Accuracy and Loss\u003c/em\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/d0ff0d2a7988e4d5c930d705.png"},{"id":73520412,"identity":"26a2c8d5-4b35-4a5f-89ee-ca9cf806bdda","added_by":"auto","created_at":"2025-01-10 18:20:46","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":742195,"visible":true,"origin":"","legend":"\u003cp\u003eWRNet Model prediction result.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/83ade22b25a9f5a4ae32e9bc.png"},{"id":73520181,"identity":"b643ea2c-c780-400b-be21-de4cb8e57ee9","added_by":"auto","created_at":"2025-01-10 18:12:46","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":246518,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eInceptionV3 model Training and Validation Training and Loss\u003c/em\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/ee8201634b63bd7b17ab97f5.png"},{"id":73520174,"identity":"53a20e74-b90c-4be4-965e-27523a8eac84","added_by":"auto","created_at":"2025-01-10 18:12:46","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":554649,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eInceptionV3 model prediction\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/d7e1aff75c6ca550bc2c098a.png"},{"id":73520179,"identity":"32f8414c-00d6-4532-8a5e-d10443433f8a","added_by":"auto","created_at":"2025-01-10 18:12:46","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":295086,"visible":true,"origin":"","legend":"\u003cp\u003eInceptionResnetv2 Model Training and Validation Accuracy and Loss\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/8447f7d03945df802540fba7.png"},{"id":73520176,"identity":"06a1f3cb-83ba-46da-923f-1399404f6ba6","added_by":"auto","created_at":"2025-01-10 18:12:46","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":464929,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eInceptionResnetV2 model prediction result\u003c/em\u003e\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/3c0974e94810b19b545bbd34.png"},{"id":73520184,"identity":"24cd5de3-70a6-491e-81b6-c64cbe09b796","added_by":"auto","created_at":"2025-01-10 18:12:47","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":160481,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eTraining and Validation Accuracy and Losses MobileNetV2 model\u003c/em\u003e\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/e4beb8f364ac25ff1539c911.png"},{"id":73520868,"identity":"074b85f6-996a-4c1a-a7ba-aa180fd05fee","added_by":"auto","created_at":"2025-01-10 18:28:46","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":632411,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eMobileNetV2 model prediction result\u003c/em\u003e\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/b6f5517ac689c332bc886432.png"},{"id":73520408,"identity":"934e7eef-cef9-407c-983e-2eba520bb4ee","added_by":"auto","created_at":"2025-01-10 18:20:46","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":164287,"visible":true,"origin":"","legend":"\u003cp\u003eTreatment recommended example\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/35c28a792f477975183a9d53.png"},{"id":73520187,"identity":"9be54e04-cc83-4152-94e8-f10b48d072f0","added_by":"auto","created_at":"2025-01-10 18:12:47","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":88463,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eUser interface\u003c/em\u003e\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/03ebc469157c9937927f3f60.png"},{"id":73520183,"identity":"8a7be4db-6420-411a-bb88-6b9e9a607932","added_by":"auto","created_at":"2025-01-10 18:12:46","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":166560,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eHomepage after image uploaded\u003c/em\u003e\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/e1071f7820a8e2025687d30e.png"},{"id":73520414,"identity":"58e99e9e-2907-4970-bdf1-6471c00210c3","added_by":"auto","created_at":"2025-01-10 18:20:46","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":301357,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eResults of the developed System\u003c/em\u003e\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/9578fbea3252a4bc0d048f86.png"},{"id":85370617,"identity":"8ed0a406-f3d8-4cf2-9161-004684b4feff","added_by":"auto","created_at":"2025-06-25 07:25:41","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5438747,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5796873/v1/4e04c432-307a-4130-a93e-40f6b966469b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Detection And Grading of Rust Disease Severities from Wheat Images Using Deep Learning Techniques","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe agriculture sector plays a crucial role in the process of economic growth of the country. Farmers consult specialists to get their opinions on the probability of crop diseases occurring and suggestions for controlling the diseases to increase crop yield production. Early disease detection and classification are particularly easy in small-scale farming. Due to the difficulty of monitoring, early disease detection, and classification in large-scale farms, serious disease occurrences and pest expansion cannot be controlled organically. In general, visual-based crop disease detection is expensive, inefficient, inaccurate, and difficult (Haile, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWheat is a food crop that delivers the most nutrition worldwide (Tiwari and Shoran, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Wheat is one of the most significant crops for Ethiopia's food security, it is grown on 2.1\u0026nbsp;million hectares of land and yields 6.7\u0026nbsp;million tons yearly. Ethiopia's wheat production has a huge potential to grow, but it is currently constrained by the prevalence of rust disease stresses, yield gaps brought on by the slow implementation of new technologies, the massive price and limited availability of inputs, the lack of public and private investments, and the country's insufficient infrastructure and marketing systems (Tadesse et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The degree of production losses and quality degradation may vary in wheat cultivation fields depending on the severity of the disease, and as a result, economic losses may differ (Hayit et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Manual optical examination of the plant is a time-consuming and tedious method for diagnosing plant disorders, and certain diseases are not present in the plants.\u003c/p\u003e \u003cp\u003eThere are comprehensive methods that can identify and categorize the rust disease of wheat. Deep Learning is the approaches that have been applied in the agriculture sector and have greatly increased in popularity (Kabir et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). It is a class of machine learning models inspired by the structure and functioning of the human brain composed of interconnected nodes, often referred to as neurons organized into layers. Deep learning models have the advantage of automatically extracting features from raw data, a process known as feature learning (Ganesh and Kumar, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDeep learning in agriculture is not only helping farmers to detect and classify diseases at early stages but also to predict treatment dosage based on severity classification and shifts to precise cultivation for higher and better-quality crop yield. The SoftMax classifier was used in this study to categorize the severity levels of the rust disease and treatment recommendations. The classifier is easy to use and has good computational performance. For the development of automatic diagnostics recommendations, the complete process of knowledge acquisition and engineering for disease management has been applied and a new knowledge base has been applied knowledge with deep learning techniques. Therefore, the objective of this research is to design and develop a deep learning-based model for the wheat yellow rust diseases detection and severity classification and treatment recommendations. The focus of a design and development step can be on front-end analysis, planning, production, and/or evaluation. We have developed a wheat yellow rust disease severity classification and treatment recommendation architecture using the deep learning method. In this study, the image was collected and then image preprocessing was performed. The preprocessed data is used for training and testing the classification model, based on which specified severities class classification is done. The knowledge base is also constructed for treatment recommendations.\u003c/p\u003e"},{"header":"Related Works","content":"\u003cp\u003eSeveral researches were conducted to automate the detection of wheat rust diseases through the application of deep learning, image processing, and machine learning techniques. One specific area of research is the identification of wheat diseases using deep learning (DL) approaches, which has seen an increase in studies over the years.\u003c/p\u003e \u003cp\u003eAboneh et al. (2021) proposed a deep-learning approach for detecting various wheat diseases. Their method utilized four popular deep learning architectures: Inceptionv3, ResNet50, and VGG16/19. They evaluated the computation cost and CPU run time using a Jetson Nano Board accelerator, a low-power embedded graphics processing unit that enables multiple neural networks to run simultaneously and supports a computer vision algorithm for image classification. The authors compared the performance of these architectures on a dataset of 1500 images representing three classes of wheat diseases. They conducted data preprocessing, standardization, formatting, removal, and rescaling. The dataset was a combination of images from wheat farms in Ethiopia, gathered in collaboration with the Bishoftu Agricultural Research Institute, and a publicly available online repository. The experimental results showed that the VGG19 model achieved 99.38% accuracy in classifying wheat diseases. However, the study did not include severity classification and treatment recommendations.\u003c/p\u003e \u003cp\u003eAzadbakht et al. (2019) proposed a machine learning (ML)-based method for detecting wheat leaf rust disease. They constructed their dataset using canopy scale data at various leaf area index (LAI) levels. Their framework identified the severity of wheat leaf rust at the canopy scale using four ML techniques: Random Forests Regression (RFR), ν-support vector regression (ν-SVR), Gaussian process regression (GPR), and boosted regression trees (BRT). Hyperspectral signature data collected in the field using a spectroradiometer served as inputs for the models. The ML inversion models all performed well at high and medium LAI levels, achieving up to 99% accuracy with ν-support vector regression (ν-SVR), which outperformed other ML methods. The results indicated that the performance of the ML methods improved as the LAI value increased. However, the experiments focused solely on leaf rust disease severity and did not include yellow rust of wheat and also treatment recommendations.\u003c/p\u003e \u003cp\u003eXu et al. (2017) conducted a study on Automatic Wheat Leaf Rust Detection and Grading Diagnosis via an Embedded Image Processing System. The study focused on detecting color images, which involved the processes of image capturing, image processing with detection, and disease classification. The method used for each process varied based on the characteristics of different types of diseases. The research employed Sobel edge detection based on exploration. Another significant finding was the grading of infected regions based on the percentage of scab area. The system captured images using a web camera connected to an ARM microcontroller through USB, and the results were displayed on a dedicated LCD. The proposed system achieved an accuracy rate of up to 96.2% for recognition. However, the study did not include wheat yellow rust scaling classification and treatment recommendation.\u003c/p\u003e \u003cp\u003eAssefa (2021) conducted research on the detection and grading of stem rust in wheat to address the severity levels of stem rust according to CIMMYT grading guidelines. The study aimed to utilize a deep convolutional neural network for this purpose, consisting of four components: preprocessing, segmentation, feature extraction, and classification. Adaptive Thresholding was employed for segmentation, while a Gabor filter was proposed for texture feature enhancement in the image to detect and select important disease features. For classification, a convolutional neural network and a 12-way Softmax were utilized for grading into specific classes (Resistant (TR), Moderately resistant (MR), and Susceptible (S)). To address overfitting, data augmentation and dropout techniques were applied. The study evaluated three state-of-the-art CNN architectures, including AlexNet, VGGNet, and GoogLeNet, with SRNet achieving a training accuracy of 92% and a testing accuracy of 92.01%. The proposed system was implemented using Keras (with TensorFlow as a backend) in Python and trained and tested using an image dataset collected and prepared with the assistance of pathologists from Haramaya University, divided into 70/30 for training and testing, respectively. The model achieved an accuracy of 92.02% for training and 92.01% for testing in detecting and grading wheat stem rust. It's important to note that the study focused on stem rust disease and did not include yellow rust and treatment recommendations.\u003c/p\u003e \u003cp\u003eZhang et al. (2019) proposed a deep learning-based method for the automated detection of yellow rust disease from high-resolution hyperspectral UAV images. They utilized multiple Inception-ResNet layers for deep feature extraction, optimized to determine the most suitable depth and width. The model incorporated spatial and spectral information for yellow rust detection, resulting in improved accuracy. The study confirmed the potential of the proposed deep-learning architecture for crop rust disease detection, despite increased complexity and slightly lower accuracy compared to leaf-scale reports by other researchers. In a carefully planned field experiment, UAVs captured hyperspectral imagery of healthy and rust-infected wheat plots on five distinct dates throughout a crop cycle. After deep feature extraction using Inception-ResNet blocks, the feature maps were converted into a three-class classifier for rust, healthy, and others using an average pooling layer and a fully connected layer. The study did not include a classification of yellow rust severities or treatment recommendations.\u003c/p\u003e \u003cp\u003eAutomatic detection of yellow rust in wheat using image processing and machine learning approach done by (Tsiyon, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) a proposed system that integrates components, such as preprocessing (Image resizing, Histogram equalization, and Noise removal), segmentation, feature extraction, and classification. They perform implementation using Keras (with Tensor Flow as a backend) in Python and they take samples of wheat leaf from South Gonder Zone from a specific place, Gobgob Kebele. They take the total number of images taken was 3322 containing wheat leaf. To build the detection models for the prediction of wheat yellow rust, they investigate Softmax and SVM classifiers. An overall accuracy of 98.04% is achieved by using the SVM classifier; on the other hand, the Softmax classifier achieves an accuracy of 97.74%. However, this research is done on yellow rust detection, which did not include severities classification and treatment recommendation. To the best of researcher knowledge, there is no research works done that focus on automatically detecting, severities assessment, and treatment recommendation of yellow rust of wheat using deep learning.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eDesign science research is selected as general approaches to design and develop wheat yellow rust diseases detection and severity classification and treatment recommendations. By facilitating the change from the current situation to the expected state, this strategy attempts to develop and evaluate the artifacts necessary to solve the problems that have been identified. The recognition of plant diseases passes through a series of steps/procedures that would be applied to separate plants, in which each wheat yellow rust disease must be categorized into one of predefined classes based on observed attributes or features.\u003c/p\u003e \u003cp\u003eDSR is selected for the following reasons: first, this methodology has a strong emphasis on developing artifacts and their usage by the specialist to solve real-world problems. Second, DSR, which focuses on real-world problem solving, contains prescriptive or solution-oriented knowledge that is used for designing complex and relevant field problems by using the results of scientific reasoning (such as prediction). Finally, within the realm of information systems (IS), DSR aims to acquire knowledge and comprehension of a problem and develop a solution for the demands of the business and its environment (Yimenu, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn this study, a CNN architecture is used. Because deep learning via CNN didn't need a different algorithm for feature extraction like machine learning methods. It uses the learned characteristics to automatically learn features and classify them into the best-fitting group or class. The deep-learning workflow for solving real-world problems includes three processing steps: data analysis and preprocessing, DL model development and training, and validation and interpretation. However, feature extraction in the DL model is automated rather than manual, in contrast to ML modeling (Sarker, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). CNN architecture is designed using convolutional layers, pooling layers, fully connected layers, and dropout layers. These operations are then applied to all subsequent convolutional layers. The images are passed through the convolutional layer, Rectified Linear Unit layer, batch normalization layer, and max pooling layer for feature extraction. Finally, the fully connected layer is used for disease classification. The input layer of the CNN model can directly receive the image as input. A color image has three color channels.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInput layer\u003c/strong\u003e \u003cp\u003eThe image can be directly fed into the CNN model as input. The number of color channels in a color image is three, while in a gray-scale image, it is one. In this study, 950 x 250 x 3 (height, width, and number of color channels) was used respectively, because it is the most recommended one by various researchers, to use all of the features of the image in their full resolution (Fasnacht et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eConvolutional Layer\u003c/b\u003e: The convolution operation requires four parameters: the number of filters, filter size, stride size, and amount of zero-padding. In this study, we have tested different numbers of convolution layers and filters, such as 32, 64, 128, 256, 512, and 1024. We selected the ones that achieved higher accuracy. To ensure better detection, we used filter sizes of 5 x 5, 3 x 3, and 2 x 2 at a single layer. A 2 x 2 filter size can detect a feature that cannot be detected by a 3 x 3 filter size. We used stride sizes of 2 (2, 2) and 1 (1, 1) to cover the entire image and to avoid overlapping.\u003c/p\u003e \u003cp\u003eWe have used \"the same padding\" parameter in our convolutional operations, which ensures that the output is of the same size as the input. The amount of zero-padding is used to adjust the output size. Padding is a technique used to resolve issues in normal convolutional operations, such as the reduction of the height and width of the output feature map. A typical padding method is zero-padding. We have performed the convolution operation repeatedly before downsampling the input image using the pooling operation. Using multiple convolution layers before applying a pooling layer allows the model to develop more complex features before the destructive pooling operation is carried out.\u003c/p\u003e"},{"header":"Experimental Result","content":"\u003cp\u003eTo accomplish the experiments, we have implemented our designed system with the required parameters. Four experiments conducted. The first experiment conducted with the creating CNN model from scratch and others three experiments were conducted with pre-trained deep learning models such as MobileNetV2, InceptionV3, and InceptionResnetV2. The experiments are performed with the same corpus size. The effect of applying different preprocessing techniques such as noise removal and segmentation, and after data augmentation, and dropout applied are evaluated and compared with CNN. In addition, the test results were presented and compared with other researchers\u0026apos; work.\u003c/p\u003e\n\u003ch3\u003eExperiment 1: Creating a Model from Scratch\u003c/h3\u003e\n\u003cp\u003eTraining from Scratch is a method of training a CNN network, where the researcher decides all the hyperparameters such as the optimizer, learning rate, number of layers, epochs, filters per layer, activation function, and fully connected layers. Our proposed \u0026apos;WRNet\u0026apos; model was evaluated through a series of experiments in which different noise removal techniques were applied, including denoising, bilateral filter, K-means segmentation, and data augmentation. Hereunder we present experiments conducted using segmentation and filtering methods. The experiment conducted by applying bilateral filtering and k-means segmentation algorithms. This proposed CNN model experiment has been tested by splitting the whole dataset into 70% for training and 30% for testing. The following figure shows the training process of the model with 10 number of epochs. Both training accuracy and validation accuracy are increasing with the number of epochs. The result recorded from training accuracy was 99.11% and validation accuracy of 99.04%. The trained model evaluated by three performance measurements metrics and the result recorded as follows.\u003c/p\u003e\n\u003cp\u003eThroughout the curve in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, the training accuracy is always higher than the validation accuracy. However, the gap between the two becomes smaller over time. The validation loss remains relatively small during the entire curve. Additionally, both the training and validation loss decrease significantly. After training completed, the model predicts the output for each class of severity level.\u003c/p\u003e\n\u003cp\u003eTable 1: Confusion Matrix of model with filter and segmentation\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cimg src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003eThe above table shows the graphical representation of the confusion matrix for test images for S(Susceptible), MS, MR, R disease, and healthy level. As shown above for S, MS, and all images (218) are correctly classified in their classes, while no images are incorrectly classified. For MR, the model correctly classified 214 images as MR, with only 4 images incorrectly classified as healthy. Additionally, 215 images were correctly classified as healthy, with only 3 images incorrectly classified as MR. This indicates that the model has effectively learned from the training data and can generalize well to the validation and testing datasets. So, the model constructed after applying k-means segmentation with bilateral filtering is selected and proposed for comparison with pre-trained models.\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eExperiment 2: InceptionV3 model\u003c/h3\u003e\n\u003cp\u003eThis experiment performed with pre-trained InceptionV3 by applying bilateral filtering and k-means segmentation algorithms. The following figure shows the training process of the model with 10 number of epochs. The pre-trained model achieved a training accuracy of 95.72%, a validation accuracy of 95.12%, and a testing accuracy of 87.03%. These classification accuracies were attained through the utilization of image bilateral filtering, k-means segmentation after data augmentation, and dropout application.\u003c/p\u003e\n\u003cp\u003eIt has been observed that the InceptionV3 Model has identified overfitting, which is indicated by a significant difference between the training and validation loss. The training loss is relatively small and remains stable throughout the curve, while the validation loss is unstable and shows fluctuations. This suggests that the model is not generalizing well to new data.\u003c/p\u003e\n\u003cp\u003eAfter training and testing the Inceptionv3 model, the prediction result for one image produces the above result with a percentage of confidence.\u003c/p\u003e\n\u003ch3\u003eExperiment 3: InceptionResnetv2 model by applying filtering and Segmentation\u003c/h3\u003e\n\u003cp\u003eThe third experiment uses InceptionResnetv2 to measure the performance of training accuracy and testing accuracy of model at epochs 40. InceptionResnetv2 achieves the accuracy of training, validation and testing of 97.47%, 97.28%, and 83% respectively after applying image bilateral filtering and k-means segmentation algorithms.\u003c/p\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003eExperiment 4: With MobileNetV2 After applying filtering and Segmentation\u003c/h2\u003e\n \u003cp\u003eThe training accuracy consistently surpasses the validation accuracy, as evidenced by the curve. Notably, the difference between the two accuracies diminishes compared to the previous curve. Throughout the curve, the validation loss remains relatively small. Meanwhile, the training loss exhibits a steep decline from 0.14, while the validation loss fluctuates up and down and peaks at 0.13.\u003c/p\u003e\n \u003cp\u003eAfter training and testing the MobileNetV2 model, the prediction result for one image produces the above result with a percentage of confidence.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Experimental Discussion","content":"\u003cp\u003eFrom the above experimental result conducted on both the WRNet model and pre-trained models such as InceptionV3, InceptionResnetV2, MobileNetV2 pre-trained model, the researcher trained with epochs of 50. For each experiment, we used a total of 20,000 images for training, validation, and testing the model.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of models\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModels\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining Accuracy %\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation Accuracy %\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTesting Accuracy%\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTesting Loss%\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInceptionV3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e95.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e87.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInceptionResnetV2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e97.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e97.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMobileNetV2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e99.02%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e99.02%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e89.01%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWRNet model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e99.11%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e99.04%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99.00%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFrom our comparison, we have found that the proposed model with a k-means segmentation algorithm and bilateral filtering is very effective. The training of the WRNet model increases the testing accuracy from 83\u0026ndash;99% with no over-fitting. And in the case of InceptionV3, InceptionResNetV2, and MobileNetv2, model the testing accuracy increases from 83%, 87.03\u0026ndash;89.01% respectively. The validation and testing accuracy performances of our WRNet model in all cases are maximum. At the same time, the amount of time it takes to train and test the dataset showed a better performance. In addition, the chance of overfitting is rigorously minimal in the case of using bilateral filtering and k-means segmentation image training. We are subjected to adding different regularization techniques to reduce overfitting in pre-trained model cases. In all cases, bilateral filtering and k-means segmentation method has given the most effective result compared to those other filtering and segmentation. Analyzing the created models, the comparison showed that the proposed WRNet model trained with bilateral filtering and k-means segmentation images is superior to all other models experimented with. Therefore, we propose this model for Wheat Rust disease detection and severity classification.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTreatment Recommendation.\u003c/b\u003e \u003c/p\u003e \u003cp\u003eTo effectively control the spread of yellow rust, it is suggested to apply fungicide solutions such as propiconazole 250 EC (Tilt), tebuconazole 250 EC (Folicur), or triadimefon 250 WP (Bayleton) (Erdado, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). For all of the severity Propiconazole 0.65 li/ha with 250 li/ha, Opera Max 1 li/ha with 250 li/ha, Tilt 250EC 0.5-1 li/ha with 250 li/ha, and Progress 250EC 0.5 li/ha with 250 li/ha. But the main difference is the number of times the fungicide is applied. Based on this the percentage of dosage pesticides for each classification of wheat yellow rust contains for R one time, for MR two times per week, for MS 2\u0026ndash;3 times per week, and for S 3\u0026ndash;4 times per week is important (Erdado, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003ePrototype Deployment\u003c/h3\u003e\n\u003cp\u003eAfter conducting the study, the researcher developed a user interface for the proposed research work. This user interface is designed for domain experts or end-users to classify wheat yellow rust severity diseases and receive treatment recommendations. The user interface is created using Python and the Flask framework, which is used for developing web applications in Python. The system can detect five types of wheat yellow rust disease severity, including healthy plants. This enables the user to perform a wheat yellow rust disease severity classification service. To access the system, the end-user can launch the web application by typing the URL (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://127.0.0.1:5000\u003c/span\u003e\u003cspan address=\"http://127.0.0.1:5000\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). into any web browser and accessing it through port 5000. When the user submits an HTTP request to the webserver, the server responds with an HTTP response. The prototype's initial user interface includes a button for uploading images.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe user interface depicted in Fig.\u0026nbsp;36 above is the first page that is opened when the user enters the URL of the wheat yellow rust disease classification and treatment recommendation system. In this prototype flask server is run at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://localhost:5000\u003c/span\u003e\u003cspan address=\"http://localhost:5000\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (in the browser) to allow the user to import the images that need to detect severity level and recommendation. The user imports the wheat yellow rust image by using the \u0026lsquo;choose file button\u0026rsquo; that is displayed in the user interface. The home page of the developed model after importing the images is shown in Fig.\u0026nbsp;37 below.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAfter uploading images and clicking on the \"predict\" button, the user will receive a classification of the wheat yellow rust severity. A sample result is depicted in Fig.\u0026nbsp;19 below.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWheat is a major crop worldwide, including in Ethiopia. However, wheat production and quality are affected by stripe rust (Puccinia striiformis f. sp. tritici), which is the most common fungal disease, especially in cool climates. Traditional monitoring and examination methods are employed after disease progression reaches a higher stage, and are time-consuming, labor-intensive, error-prone, and very expensive. In this study, we constructed an optimal model using a CNN model that detects and classifies the severity of disease and provides treatment recommendations for wheat yellow rust. In this study, 20,000 yellow rust leaf images were collected from the Kulumsa Research Center in the Arsi zone of Asela town. After collecting the required images, the researcher applied different image preprocessing techniques, such as image resizing, noise removal, filtering, and segmentation. To conduct this study, training from scratch (i.e., the researcher settings different hyper-parameters) and transfer learning methods such as InceptionV3, InceptionResnetV2, and MobileNetV2 were used. The developed model was evaluated using a confusion matrix, accuracy, recall, precision, f1-score, and support.\u003c/p\u003e \u003cp\u003eAmong all the models, the proposed model trained from scratch performed better, with 99.11% training accuracy, 99.04% validation accuracy, and 99% testing accuracy. Additionally, the researcher creates knowledge modeling by utilizing a decision tree to represent the concepts involved in the classification and treatment of wheat yellow rust disease severity. Next, a rule-based representation technique is used to represent the knowledge. With the help of this developed model, a treatment plan can be quickly suggested, and the disease can be detected early.\u003c/p\u003e "},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJ.M. and M.M. conceived and designed the study. J.M. conducted the experiments and implemented the proposed model. M.M. and J.A. collected and preprocessed the data. J.A. wrote the main manuscript text and conducted the comparative analysis. A.A. developed the user interface and assisted with the treatment recommendation module. J.A. and A.A. prepared Figures 1\u0026ndash;12 and Tables 1\u0026ndash;2. All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eErdado, F. Integrated Management of Stripe Rust (Puccinia striiformis f.sp. tritici) OF Bread Wheat. \u003cem\u003eUnpublished\u003c/em\u003e, \u003cem\u003eNovember\u003c/em\u003e, 74. (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFasnacht, L., Renard, P. \u0026amp; Brunner, P. Robust input layer for neural networks for hyperspectral classification of data with missing bands. \u003cem\u003eAppl. Comput. Geosci.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e (August), 100034. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.acags.2020.100034\u003c/span\u003e\u003cspan address=\"10.1016/j.acags.2020.100034\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGanesh, B. \u0026amp; Kumar, C. Deep learning Techniques in Image processing. \u003cem\u003eNational Conference On Emerging Trends in Computing Technologies (NCETCT)\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e, 1\u0026ndash;5. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaile, E. Plant Disease Detection and Classification Using Artificial Neural Network. \u003cem\u003eAddis Ababa Univ. Repository\u003c/em\u003e. \u003cb\u003e01\u003c/b\u003e, 83 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHayit, T., Erbay, H., Var\u0026ccedil;ın, F., Hayit, F. \u0026amp; Akci, N. Determination of the severity level of yellow rust disease in wheat by using convolutional neural networks. \u003cem\u003eJ. Plant. Pathol.\u003c/em\u003e \u003cb\u003e103\u003c/b\u003e (3), 923\u0026ndash;934. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s42161-021-00886-2\u003c/span\u003e\u003cspan address=\"10.1007/s42161-021-00886-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKabir, M. M., Ohi, A. Q. \u0026amp; Mridha, M. F. A Multi-Plant Disease Diagnosis Method using Convolutional Neural Network. \u003cem\u003eRes. Gate\u003c/em\u003e, 1\u0026ndash;13. (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarker, I. H. Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions. \u003cem\u003eSN Comput. Sci.\u003c/em\u003e \u003cb\u003e2\u003c/b\u003e (6), 1\u0026ndash;20. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s42979-021-00815-1\u003c/span\u003e\u003cspan address=\"10.1007/s42979-021-00815-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTadesse, W., Zegeye, H., Debele, T. \u0026amp; Kassa, D. Wheat Production and Breeding in Ethiopia: Retrospect and Prospects. \u003cem\u003eCrop Breed. Genet. Genomics\u003c/em\u003e. 1\u0026ndash;22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.20900/cbgg20220003\u003c/span\u003e\u003cspan address=\"10.20900/cbgg20220003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTiwari, V. \u0026amp; Shoran, J. Growth and Production of Wheat. \u003cem\u003eEOLSS (Encyclopedia Life Support Systems)\u003c/em\u003e. \u003cb\u003eI\u003c/b\u003e, 1\u0026ndash;9 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTsiyon, W. Automatic Detection of Yellow rust in Wheat using Image Processing and Machine Learning Approach. \u003cem\u003eDSpace Institution\u0026rsquo;s Institutional Repository\u003c/em\u003e. \u003cb\u003e01\u003c/b\u003e, 1\u0026ndash;90 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYimenu, D. F. Image Based Sorghum Leaf Disease Classification Using Deep Learning Approach. \u003cem\u003eDebre Berhan Univ. Repository\u003c/em\u003e, 1\u0026ndash;138. (2021).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Wheat yellow rust, deep learning, CNN, WRNet, disease severity classification, precision agriculture, image preprocessing, treatment recommendation","lastPublishedDoi":"10.21203/rs.3.rs-5796873/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5796873/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWheat production, a cornerstone of food security in Ethiopia, is heavily impacted by stripe rust disease, which leads to significant economic losses. Traditional methods for detecting and classifying disease severity are labor-intensive, error-prone, and costly. This study introduces a novel convolutional neural network (CNN)-based model, WRNet, designed for the detection and severity classification of wheat yellow rust disease, along with treatment recommendations. Utilizing 20,000 annotated images collected from Ethiopia, the model applies advanced preprocessing techniques such as noise removal and segmentation using bilateral filtering and k-means algorithms. The WRNet model achieved superior performance with 99.11% training accuracy, 99.04% validation accuracy, and 99% testing accuracy, surpassing pre-trained models such as InceptionV3, InceptionResNetV2, and MobileNetV2. Additionally, the system provides fungicide dosage recommendations tailored to severity levels, ensuring effective disease management. A user-friendly prototype interface developed using Flask enables domain experts to classify disease severity and receive treatment recommendations, offering a scalable solution for precision agriculture in Ethiopia and beyond.\u003c/p\u003e","manuscriptTitle":"Detection And Grading of Rust Disease Severities from Wheat Images Using Deep Learning Techniques","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-01-10 18:12:41","doi":"10.21203/rs.3.rs-5796873/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e2682988-4be3-4303-b241-9c62033e811e","owner":[],"postedDate":"January 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":42625758,"name":"Physical sciences/Energy science and technology"},{"id":42625759,"name":"Physical sciences/Engineering"},{"id":42625760,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2025-06-25T07:25:15+00:00","versionOfRecord":[],"versionCreatedAt":"2025-01-10 18:12:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5796873","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5796873","identity":"rs-5796873","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}