Efficient Multi-Class Image-Based Rosemary Variety Verification and Classification Model Using Deep Learning: A Scientific Investigational Study

Efficient Multi-Class Image-Based Rosemary Variety Verification and Classification Model Using Deep Learning: A Scientific Investigational Study | 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 Efficient Multi-Class Image-Based Rosemary Variety Verification and Classification Model Using Deep Learning: A Scientific Investigational Study Tsega Asresa, Melaku Bayih, Shashi Kant Gupta, Eman Abdullah Aldakheel, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6432850/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 Artificial intelligence (AI) has a subfield called computer vision that allows systems and computers to extract replacement data from digital photos and videos. It is used in many fields, including agriculture, health care, education, self-driving cars, and daily living. In Ethiopia, rosemary is a well-known aromatic and therapeutic plant. It is an evergreen herb that belongs to the shrub family and it is widely used in Ethiopia with three varieties: WG rosemary I, WG rosemary II, WG rosemary III. Botanists, researchers, herbal industries, pharmacists and domain experts are facing challenges to classify appropriate varieties. And there is a lack of research and technology for identifying and classifying those varieties in Ethiopia. To address this gap, the proposed study employs supervised machine learning and multi class image classification. Specially, this study is conducted using a convolutional neural network (CNN) employing a SoftMax activation function in the last layer is used to develop the classification models. In this study, five cutting-edge models: convolutional neural network, Inception V3 and exception have been selected. After a comprehensive review of the best-performing models. The 80/20 percentage split was used to evaluate the model, and classification metrics were used to evaluate and compare the models. The pre-trained Inception V3 model outperforms well, achieving training and validation accuracy of 98.8% and 97.7%, respectively. Biological sciences/Computational biology and bioinformatics/Image processing Biological sciences/Plant sciences Artificial intelligence convolutional neural network Computer vision Varity verification Rosemary Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction Agriculture is a cornerstone of Ethiopia's economy, contributing significantly to the country's GDP. Rosemary, known locally as "Yetebes Ketel" (meaning a leaf used for roasting) and "Azmerino," is a popular aromatic and medicinal plant in Ethiopia [1]. This perennial shrub, if properly managed, can yield economically for 4 to 7 years [2]. Despite its widespread cultivation in various ecological zones, the genetic variability of Ethiopian rosemary has not been thoroughly studied, leading to limited knowledge of the agronomic and chemical characteristics within the local germplasm. This knowledge gap hinders the full exploitation of rosemary's potential in the country. Therefore, the current study aims to evaluate the variability among Ethiopian rosemary germplasms for morpho-agronomic and key quality traits. The study focuses on three identified varieties—WG rosemary I, WG rosemary II, and WG rosemary III—verified by researchers at the Wondo Genet Agricultural Research Center [3]. This research is crucial for the development of new varieties and the effective utilization of rosemary's potential in Ethiopia [3]. Researchers are currently developing a Rosemary Variety Identification and Verification model to assist experts in the field. Multi-class image classification is essential for identifying and verifying rosemary varieties using deep learning techniques. Previous research has addressed rosemary identification and classification, but questions remain regarding the specific varieties of the rosemary plant. Automating this process will facilitate the determination of plant parts used for oil production, medicine, cosmetics, and perfumery. The proposed study aims to create an automatic classification model for aromatic and medicinal plants in the agricultural sector, leveraging state-of-the-art deep learning. This model will utilize multi-class image classification to effectively identify and verify rosemary varieties. 2. Literature review Various researches have looked into various aspects of plant classification, with the majority of them focusing on standard image classification using machine learning models. However, there is a lack of research to classify rosemary Varity classification and verification model using the current state of the art Deep learning. Extensive researches have been done in plant classification [4]. However, the researchers did not address the rosemary Varity classification and verification model. According to this study, the researcher tried to classify the plant images using six deep Learning algorithms. Deep learning approaches, specifically multi-class image classification [5]. are being used to develop a model for identifying and verifying rosemary varieties. This model aims to automate the classification of aromatic and medicinal plants, particularly for the production of oil, medicine, cosmetics, and perfumery. This research is significant for the agriculture sector and aims to create an automatic classification model using deep learning for identifying and verifying rosemary varieties [6]. The study [7] demonstrated the plant image classification using the Deep Learning models. However, the researcher did not answer the Question the rosemary Varity classification and verification model using the current state of the art models. According to the study [8] et. al the researcher uses the convolutional neural network for efficient identification of rosemary pollen in the case of Spain. The paper is significant for the proposed study. However, the researcher is failed to identify the rosemary Varity verification and classification model using the current state of the art algorithm. The proposed study will alleviate the challenges of botanists to identify the three types of rosemary by employing the Deep learning methods. The study [9] Numerous techniques for classifying plants based on their size, shape, color, and texture have been developed and published in the literature. This field is covered by a few works. Using a probabilistic neural network (PNN) classifier, the researcher was combined fuzzy local binary pattern and fuzzy color histogram with a dataset of 2448. However, the proposed model extracts the leaf feature by using the max-pooling layer. 3. Methods and Methodology In order to automatically extract several distinctive properties of rosemary leaves, this study used bespoke CNN, the foundation of a deep learning algorithm. By including batch normalization at the fully connected layer of CNN, adjusting weights, and training hyperparameters, the study also used the transfer learning technique. In this work, we used the underside of the leaves to help the model extract the features in a unique way and correctly identify the rosemary variety. In this study the researcher used experimental research design with Qualitative and quantitative research Approach. The present study employed six distinct techniques, namely data collecting, data annotation (labelling), picture preprocessing, feature extraction and training, testing (model evaluation), and categorization of Rosemary components. A high-level system architecture for the investigation is Acquisition of the image and annotation (data labelling) are included in the input image phase. Training and feature extraction are tasks included in the Model Building process. The sections that follow one another provide a detailed description of each. A . Image Acquisition Data was collected using high-resolution cameras, specifically the TECHNO SPARK with 16 MP and SAMSUNG A30S with 25 MP, provided by the Agricultural Research Center at Wondo Genet. To ensure the integrity of the data, attention was paid to the direction and intensity of light when capturing images of plant leaves. Photographs of the backsides of the leaves were taken to accurately extract vein features. The focal length varied based on the width of the leaves to maintain consistency in the images. The dataset comprises 3284 photos of leaves from medicinal plants, with an equal number of leaf photos for each plant classified into the WG_rosemary I, WG_rosemary II, and WG_rosemary III classes. These rosemary types represent the varieties available at the Wondo Genet Agricultural Research Center and have been verified by researchers. In total, 3,482 rosemary leaf images were collected from the research center, the researcher collected leaf images of WG-rosemary I, WG-rosemary II, WG-rosemary III 2220, 935,327 respectively. According to Ethiopian Institute of Agricultural research center three are three different types of Rosemary Varity WG- Rosemary I, WG- Rosemary II, WG- Rosemary III. Each varity are unique characteristics are Described Figure 1. As illustrated in Figure 2, the distribution of rosemary classes shows that WG_rosemary I has a significantly higher number of samples compared to WG_rosemary II and WG_rosemary III. Therefore, it is necessary to balance the dataset using techniques for imbalanced learning. B. Image Preprocessing The input image was reduced to 150 by 150 in order to shorten the computing time required for training. Normalization in the context of the system architecture refers to image pixel scaling. The integer values of an image pixel range from 1 to 255. Processing a big integer number in CNN can interfere with or slow down learning. As a result, the image's pixels were normalized, or scaled, from 0 to 1. The histogram of the Digital image is given with the intensity level 0 to L-1 h(r k )=n k…………………………………………………………………………………………. ………………………….( 1 ) where r k is the kth intensity level and n k is the number of pixels in the image having that intensity level. We can also normalize the histogram by dividing it by the total number of pixels in the image. For an N x N image, we have the following definition of a normalized histogram function: P(rk)=nk/N2………………………………………………. ………………………(2) As shown in Figure 3, a plot illustrating the intensity transformation across the red, green, and blue channels of the images has been demonstrated. As evident from the graph, the gray level transformation between the image color channels is approximately around 250. As indicated in Figure 4, various experiments have been conducted, which aimed at transitioning the intensity of the original image to grayscale. This adjustment was made to accommodate the model's sensitivity within the dual-color spectrum. C. Feature extraction Adjusted CNN models such as Inception_V3, MobileNet, and VGG16 are utilized to extract informative leaf features, subsequently fed into the classifier. The CNN feature learning layer manages these feature extractions, leveraging several techniques. This includes employing the Rectified Linear Unit (ReLu) activation function, which accelerates and enhances learning by mitigating the vanishing gradient problem. Additionally, the Max-pooling technique is employed to reduce the spatial dimensions of the convolved image features. Furthermore, a filtering kernel is utilized, sliding over image pixels and computing dot products to generate diverse image features. Through this feature extraction process, the deep learning model autonomously extracts essential attributes of plant leaves. This structure adheres to the convolutional neural network framework. As presented in Figure 5, the feature learning phase for rosemary variety classification is executed in this stage. Throughout the training phase of the convolutional neural network, convolution and pooling operations are employed to extract pertinent features from the input. As outlined in Figure 6, input images sized at 150x150x3 are passed into the convolution layer to reduce dimensionality. Within the convolutional neural network, features are autonomously extracted via the max pooling layer. Subsequently, the images are transformed into a 1D vector upon reaching the fully connected layer. Ultimately, the proposed Deep Learning-based rosemary Variety Verification model is presented in Figure 7 The researcher collected rosemary leaf images, conducted preprocessing of the images, and subsequently partitioned the image data into training and testing sets. As described in Figure 6 the proposed model design consists of three Phases data collection, data preprocessing, and data splitting. During the data collection phase, the researcher labeled the RGB images in folder and read the images using the panda's data frame. And in the preprocessing phase, the images are resized, rescaled, and vectorization. Finally, the data is split into training and testing datasets. After data splitting the model is developed using the training data and evaluated using the testing data. D. Model Evaluation Methods Accuracy is the proportion of samples that are correctly classified. More precisely it is the sum of the number of true positive and true negative and divided by the number of samples in the dataset. The model is evaluated using cutting edge evaluation metrics Accuracy, pression, recall, confusion matrix and F1 score …………………………………………………….. (3) It describes the proportion of correct prediction among all classes and labels. On the other hand, it …………………………………… ………………………. (4) The proportion of samples of a certain class of labels that have been predicted by the model as belonging to the class or labels. On the other hand, it is the proportion of true positive among true samples. ……………………………..………… …………….................(5) 4. Experimental Result and discussion 4. 1 Analysis of Results Utilizing Convolutional Neural Networks In this section, we employed a conventional Convolutional Neural Network (CNN) to present experimental findings. The dataset was divided, with eighty percent allocated for training and twenty percent for testing purposes. Across the initial three convolutional layers, the researcher utilized 16, 32, and 64 filters respectively. This study utilized a dataset comprising 3482 color images, meticulously labeled according to their respective variations. Recent advancements in deep convolutional neural networks have demonstrated remarkable efficacy in image classification tasks. Thus, all experiments herein were conducted with hyperparameters configured based on a fundamental three-layer CNN model. Specifically, a set of 150 × 150 x 3 down-sampled RGB images was fed into the model. The researcher employed a 2x2 max pooling parameter and a 3x3 kernel size to construct a bespoke CNN model. Multiple experiments were conducted utilizing the convolutional neural network model. The dataset was stratified into three groups by the researcher, with ratios of 70/30, 80/20, and 90/10 allocated for training and testing respectively. Table 1: Analysis of Results Utilizing Convolutional Neural Networks Optimizer Training Accuracy Validation accuracy Training Loss Validation loss SGD 99.6% 92.6% 0.01% 0.024% Adam 99.5% 99.2% 0.012% 0.023% Adagrad 99.5% 99.3% 0.01% 0.02% Ada delta 81% 82.3% 0.43% 0.54 RMS Prop 99% 99.2% 0.04% 0.024% As we presented the above Table 1 The result that is obtained from the experiment of the proposed model by using different optimizers is presented in the following Table 1 by using the classification accuracy, validation accuracy, hamming loss, validation loss metrics in the form of training data, validation data, and test data samples. The proposed model is outperforming in Adam optimizer in iteration per epoch. Whereas the other optimizers adagrad, adadelta, and rmsprop is fluctuating in iteration per epoch. As illustrated in Figure 8 a first the training and validation loss goes different epochs. After some epochs the validation loss became down and scaled up after some epochs. Finally, the training and the validation loss become equal. Therefore, there is no overfitting and under-fitting in the proposed CNN model. The model works well in unseen data As illustrated in Figure 8 b the training and validation accuracy is equal after some epochs. In addition, the gap between the training and validation accuracy is wide. This shows there is some overfitting. Table 2: Result using Different Activation function Activation function Learning rate Training accuracy Validation accuracy Training Loss Validation loss Softmax 0.001 99.5% 99.2% 0.012% 0.02 Sigmoid 0.001 88% 87% 0.02% 0.0276 As we described in Table 2 using the sigmoid activation function in the last fully connected layer and the output layer of the multi-class image classification is better than soft max which is preferred to multi-class classification. The main secret using the softmax activation function in multi-class image classification is softmax computes the probability between n samples. The result that is obtained from the experiment of the proposed model by using different activation functions is presented in the following Table 2 by using the classification accuracy, validation accuracy, As we observed from the above Table 3 the Confusion matrix of the CNN classifier as we mentioned the model is correctly classified the samples of each label in the test data. The model predicts 86% WG_rosemary I, 66% WG_rosemary II, 67 % WG_rosemary III. Table 4: Precision, Recall and F1 Score Convolutional Neural Networks Labels Precision Recall F1 Score WG_rosmary I 0.97 1.00 0.99 WG_rosmary III 0.97 0.99 0.99 WG_rosmary III 0.99 0.97 0.99 As we describe in the above Table 4 the precision and recall of the convolutional neural network. The classifier is performing well in all class labels. 4.2 Result Analysis Utilizing Inception V3 Network Google and GoogleNet developed effective deep learning architecture known as Inception V3, named after a popular internet meme. Despite comprising numerous layers and neurons, computation is primarily executed by a single layer. When considering deeper networks, it's crucial to acknowledge the increased risk of overfitting. In response, the Inception model incorporates sparsely connected filters across multiple layers within a single layer to mitigate this issue. Trained on the ImageNet dataset, consisting of approximately a million images, Inception V3 accepts 299 × 299 x 3 images as input and outputs 1000 classes. This architectural design is grounded in rigorous mathematical principles. Table 5: Training accuracy, validation accuracy and Loss Training Accuracy Validation Accuracy Training Loss Validation Loss 99.8% 98.7% 0.001% 0.05% According to Table 5 the accuracy of the model by utilizing the inception v3 model. The model performs well 99.8% in tranning and 98.7% in validation. As illustrated in Figure 9 a first the training and validation loss goes different epochs. After some epochs the validation loss became down and scaled up after some epochs. Finally, the training and the validation loss become equal. Therefore, there is no overfitting and under-fitting in the proposed Inception V3 model. The model works well in unseen data As illustrated in Figure 9 b the training and validation accuracy is equal after some epochs. In addition, the gap between the training and validation accuracy is wide. This shows there is some overfitting. As we observed from the above Table 6 the confusion matrix of the Inception V3 classifier as we mentioned the model is correctly classified the samples of each label in the test data. The model predicts 92% WG_rosmary I, 70% WG_rosmary II, 55 % WG_rosmary III. Table7: Precision, Recall and F1 Score Inception V3 model Labels Precision Recall F1 Score WG_rosmary I 0.98 1.00 0.99 WG_rosmary III 0.98 0.99 0.99 WG_rosmary III 0.98 0.97 0.99 As we describe in the above Table 7 the precision and recall of the Inception V3 The classifier is performing well in all class labels. 5. Conclusion Throughout history, Rosemary plants have played a significant role in various domains including medicine, cuisine, decoration, perfumery, and oil production, exerting considerable influence on the economic, social, cultural, and environmental landscapes of the local communities. Despite its extensive use, identifying the different varieties of rosemary poses a challenge for botanists, researchers, and domain experts, necessitating advanced algorithmic solutions. This study aimed to address these challenge by developing an optimized model for verifying rosemary varieties. To achieve this goal, researchers gathered leaf images representing three classes of rosemary varieties, amassing a total of 3482 images sourced from the Wondo Genet Agricultural Research Center (WARC). Deep learning methodologies were applied to classify three distinct types of rosemary (WG_rosemary I, WG_rosemary II, and WG_rosemary III) utilizing data validated by the WARC for their aromatic and therapeutic attributes. Employing an experimental research approach, the dataset was meticulously prepared for training and testing to evaluate the classification model. A convolutional neural network (CNN) was employed, dynamically extracting features from the resized images of aromatic medicinal plants. A fundamental CNN architecture was utilized for both training and prediction, leveraging preprocessed image samples. Following training, the model exhibited a high accuracy rate of 98.8%, albeit encountering a relatively lower validation accuracy of 97.7%. Researchers conjectured that augmenting the number of CNN layers could potentially enhance validation accuracy, thereby elevating the quality of the obtained results. Declarations Funding Statement This research was funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2024R409), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia Acknowledgment The authors would like to acknowledge the support of Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2024R409), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Conflict of Interest The authors have declared that they have no conflict of interest. References B. M. P. M. Tigist German, Rosemary Production and utilization, Addis Ababa, Ethiopia, December 2016. B. D. F. eemnet Mengesha, "Variability in Ethiopian Rosemary (Rosmarinus officinalis L.) Collections for Agronomic and," in Biennial Conference of the Ethiopian HorticulturalScience society (EHSS) Science society (EHSS) Science society (EHSS) , Shshemene, Ethiopi, 2018. T. G. W. G. B. M. Dejene Tadesse Banjaw, "Rosemary (Rosmarinus officinalis L.) Variety Verification Trial at," Medicinal & Aromatic Plants, 2016. L.-M. G.-B. G.-M. Z.-P. C. Cheng-Li Zhou, "A comprehensive comparison on current deep learning approaches for plant image classification," Journal of Physics, 2021. S. H. N. K. R. P. Ganbayar Batchuluun, "Deep Learning-Based Plant-Image Classification Using a Small Traning Dataset," MDPI, 2022. B.-D. M. S. M. Z. Rashad, "Plants Images Classification Based on Textural Features using Combined Classifier," International Journal of Computer Science & Information Technology (IJCSIT) Vol 3, No 4,, August 2011. Y. H. ,. Z. Ying Chen, "Plant image recognition with deep learning: A review," Computers and Electronics in Agriculture, 2023. os´e Miguel Valiente, "Automatic pollen recognition using convolutional neural networks: The case of the main pollens present in Spanish citrus and rosemary honey," Journal of Food Composition and Analysis, 2023. M. M. S. Yeni Herdiyeni, "Combination of Morphological, Local Binary Pattern Variance and Color Moments Features for Indonesian," Bogor Agricultural University, 2016. "KEY FINDINGS OF THE 2014/2015 (2007 E.C.) AGRICULTURAL SAMPLE SURVEYS," Adis Ababa, September, 2015. Tables Tables 3 and 6 are available in the Supplementary Files section. 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-6432850","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":450888098,"identity":"ee522953-d5e5-44a7-bbf1-47407cf81e0f","order_by":0,"name":"Tsega Asresa","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/klEQVRIiWNgGAWjYNCDDxDKAK8iHmQO4wyEFtzaULQw8xCjxZ7/8OEXPxjs7M3b2x9+tm2rS2xgb94mwVDxB7ctEmlplj0MyYlzzpwxls5tO5zYwHOsTILhDB6HSfCYGfAwMCdISOQwSOduO5DYIJFjJsHYhkcL//lvhn8Y6u0l5J8//m25Degw+TdALf/weT+H+TEPw2HGGRIMZtKM25iBtvAAtTTg0XIjzYxZxuB44gyeHDPL3n+Hjdt40ootEo4Z49TC3n/48cc3FdX2EuzHH9/4caZOtp/98MYbH2rkcGoBAjYJlEhgAxEJ+DQAI/ADfvlRMApGwSgY8QAArZ5Kq5H/1MQAAAAASUVORK5CYII=","orcid":"","institution":"Wolaita Sodo University","correspondingAuthor":true,"prefix":"","firstName":"Tsega","middleName":"","lastName":"Asresa","suffix":""},{"id":450888099,"identity":"f2c106ea-6a7e-4f1e-8e7e-a2f54d53c8bd","order_by":1,"name":"Melaku Bayih","email":"","orcid":"","institution":"Wolaita Sodo University","correspondingAuthor":false,"prefix":"","firstName":"Melaku","middleName":"","lastName":"Bayih","suffix":""},{"id":450888100,"identity":"eedb0a23-6dec-4d0b-85d7-32ad6bcf0545","order_by":2,"name":"Shashi Kant Gupta","email":"","orcid":"","institution":"Chitkara University Institute of Engineering and Technology. Chitkara University","correspondingAuthor":false,"prefix":"","firstName":"Shashi","middleName":"Kant","lastName":"Gupta","suffix":""},{"id":450888101,"identity":"fb6c69d4-26f2-4680-84f6-e911e11f09e9","order_by":3,"name":"Eman Abdullah Aldakheel","email":"","orcid":"","institution":"Princess Nourah Bint Abdulrahman University","correspondingAuthor":false,"prefix":"","firstName":"Eman","middleName":"Abdullah","lastName":"Aldakheel","suffix":""},{"id":450888102,"identity":"c284da0a-5eb9-4757-be28-5c820984893f","order_by":4,"name":"Diaa Salama AbdElminaam","email":"","orcid":"","institution":"Benha University","correspondingAuthor":false,"prefix":"","firstName":"Diaa","middleName":"Salama","lastName":"AbdElminaam","suffix":""}],"badges":[],"createdAt":"2025-04-12 07:23:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6432850/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6432850/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82165864,"identity":"e9e933a3-acc1-4917-a84d-a827fb030552","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":248856,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVarity of Rosemary\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/142fcf5c8bd619bbfb136243.png"},{"id":82165863,"identity":"c3bfe739-22c5-4b76-b08b-d4fa2ee8f0de","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":15545,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCategory Distribution between rosemary classes\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/34843d8ed0becff01a9d2b4c.png"},{"id":82165891,"identity":"05d90ba5-a30e-4ac0-b4f6-459b7b7be6d9","added_by":"auto","created_at":"2025-05-07 09:10:33","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":28710,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIntensity Transformation between RGR image color channels\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/4e4948281905fe520a74c10b.png"},{"id":82165868,"identity":"95806d66-c9bf-4024-973a-eb449e17bae9","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":276238,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConversion to gray scale image\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/0b0df65dbe5085030db131a6.png"},{"id":82165867,"identity":"286bcea0-0f03-44d9-8ac9-7f60e4d55a60","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":162568,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSimple leaf feature extraction\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/5e1fce9189666c0622e22215.png"},{"id":82165865,"identity":"608c469a-edcc-4917-a942-dfcaeaef283a","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":321774,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature Extraction\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/9515a1656feb85844c84b211.png"},{"id":82166928,"identity":"5357271f-fd34-4457-ac94-4841e2c007e9","added_by":"auto","created_at":"2025-05-07 09:18:05","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":78754,"visible":true,"origin":"","legend":"\u003cp\u003eProposed Rosemary Varity classifier Model Design\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/56cac4418622f201b896ea13.png"},{"id":82165870,"identity":"e39e40fe-e22e-4dcd-bd85-a4781252b471","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":48489,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining and validation accuracy\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"image8.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/48bb8cb4a283058bac0859f0.png"},{"id":82165866,"identity":"91b1114a-f542-4e24-bffb-43e4d35cdc31","added_by":"auto","created_at":"2025-05-07 09:10:05","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":49919,"visible":true,"origin":"","legend":"\u003cp\u003eModel accuracy and loss\u003c/p\u003e","description":"","filename":"image9.png","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/508fe3c1b4697de4a9c3a5b5.png"},{"id":82222921,"identity":"b6a92d8f-ad25-4f10-9246-6736e6c65423","added_by":"auto","created_at":"2025-05-08 03:01:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1695299,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6432850/v1/5b339d67-9722-464b-9d9c-789019a71158.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Efficient Multi-Class Image-Based Rosemary Variety Verification and Classification Model Using Deep Learning: A Scientific Investigational Study","fulltext":[{"header":"1.\tIntroduction ","content":"\u003cp\u003eAgriculture is a cornerstone of Ethiopia\u0026apos;s economy, contributing significantly to the country\u0026apos;s GDP. Rosemary, known locally as \u0026quot;Yetebes Ketel\u0026quot; (meaning a leaf used for roasting) and \u0026quot;Azmerino,\u0026quot; is a popular aromatic and medicinal plant in Ethiopia\u0026nbsp;[1]. This perennial shrub, if properly managed, can yield economically for 4 to 7 years\u0026nbsp;[2]. Despite its widespread cultivation in various ecological zones, the genetic variability of Ethiopian rosemary has not been thoroughly studied, leading to limited knowledge of the agronomic and chemical characteristics within the local germplasm. This knowledge gap hinders the full exploitation of rosemary\u0026apos;s potential in the country. Therefore, the current study aims to evaluate the variability among Ethiopian rosemary germplasms for morpho-agronomic and key quality traits. The study focuses on three identified varieties\u0026mdash;WG rosemary I, WG rosemary II, and WG rosemary III\u0026mdash;verified by researchers at the Wondo Genet Agricultural Research Center\u0026nbsp;[3]. This research is crucial for the development of new varieties and the effective utilization of rosemary\u0026apos;s potential in Ethiopia\u0026nbsp;[3].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eResearchers are currently developing a Rosemary Variety Identification and Verification model to assist experts in the field. Multi-class image classification is essential for identifying and verifying rosemary varieties using deep learning techniques. Previous research has addressed rosemary identification and classification, but questions remain regarding the specific varieties of the rosemary plant. Automating this process will facilitate the determination of plant parts used for oil production, medicine, cosmetics, and perfumery. The proposed study aims to create an automatic classification model for aromatic and medicinal plants in the agricultural sector, leveraging state-of-the-art deep learning. This model will utilize multi-class image classification to effectively identify and verify rosemary varieties.\u0026nbsp;\u003c/p\u003e"},{"header":"2. Literature review ","content":"\u003cp\u003eVarious researches have looked into various aspects of plant classification, with the majority of them focusing on standard image classification using machine learning models. However, there is a lack of research to classify rosemary Varity classification and verification model using the current state of the art Deep learning. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eExtensive researches have been done in plant classification\u0026nbsp;[4]. However, the researchers did not address the rosemary Varity classification and verification model. According to this study, the researcher tried to classify the plant images using six deep Learning algorithms. Deep learning approaches, specifically multi-class image classification\u0026nbsp;[5]. are being used to develop a model for identifying and verifying rosemary varieties. This model aims to automate the classification of aromatic and medicinal plants, particularly for the production of oil, medicine, cosmetics, and perfumery. This research is significant for the agriculture sector and aims to create an automatic classification model using deep learning for identifying and verifying rosemary varieties\u0026nbsp;[6].\u003c/p\u003e\n\u003cp\u003eThe study [7] demonstrated the plant image classification using the Deep Learning models. However, the researcher did not answer the Question the rosemary Varity classification and verification model using the current state of the art models.\u003c/p\u003e\n\u003cp\u003eAccording to the study\u0026nbsp;[8] et. al the researcher \u0026nbsp; uses the convolutional neural network for efficient identification of rosemary pollen in the case of Spain. The paper is significant for the proposed study. However, the researcher is failed to identify the rosemary Varity verification and classification model using the current state of the art algorithm. \u0026nbsp; The proposed study will alleviate the challenges of botanists to identify the three types of rosemary by employing the Deep learning methods.\u003c/p\u003e\n\u003cp\u003eThe study [9] Numerous techniques for classifying plants based on their size, shape, color, and texture have been developed and published in the literature. This field is covered by a few works. Using a probabilistic neural network (PNN) classifier, the researcher was combined fuzzy local binary pattern and fuzzy color histogram with a dataset of 2448. However, the proposed model extracts the leaf feature by using the max-pooling layer.\u003c/p\u003e"},{"header":"3.\tMethods and Methodology ","content":"\u003cp\u003eIn order to automatically extract several distinctive properties of rosemary leaves, this study used bespoke CNN, the foundation of a deep learning algorithm. By including batch normalization at the fully connected layer of CNN, adjusting weights, and training hyperparameters, the study also used the transfer learning technique. In this work, we used the underside of the leaves to help the model extract the features in a unique way and correctly identify the rosemary variety.\u003c/p\u003e\n\u003cp\u003eIn this study the researcher used experimental research design with Qualitative and quantitative research Approach. The present study employed six distinct techniques, namely data collecting, data annotation (labelling), picture preprocessing, feature extraction and training, testing (model evaluation), and categorization of Rosemary components. A high-level system architecture for the investigation is Acquisition of the image and annotation (data labelling) are included in the input image phase. Training and feature extraction are tasks included in the Model Building process. \u0026nbsp;The sections that follow one another provide a detailed description of each.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA\u003c/strong\u003e.\u003cstrong\u003e\u0026nbsp;Image Acquisition\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData was collected using high-resolution cameras, specifically the TECHNO SPARK with 16 MP and SAMSUNG A30S with 25 MP, provided by the Agricultural Research Center at Wondo Genet. To ensure the integrity of the data, attention was paid to the direction and intensity of light when capturing images of plant leaves. Photographs of the backsides of the leaves were taken to accurately extract vein features. The focal length varied based on the width of the leaves to maintain consistency in the images. The dataset comprises 3284 photos of leaves from medicinal plants, with an equal number of leaf photos for each plant classified into the WG_rosemary I, WG_rosemary II, and WG_rosemary III classes. These rosemary types represent the varieties available at the Wondo Genet Agricultural Research Center and have been verified by researchers. In total, 3,482 rosemary leaf images were collected from the research center, the researcher collected leaf images of WG-rosemary I, WG-rosemary II, WG-rosemary III 2220, 935,327 respectively.\u003c/p\u003e\n\u003cp\u003eAccording to Ethiopian Institute of Agricultural research center three are three different types of Rosemary Varity \u0026nbsp; \u0026nbsp; \u0026nbsp;WG- Rosemary I, WG- Rosemary II, WG- Rosemary III. Each varity are unique characteristics are Described Figure 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs illustrated in Figure 2, the distribution of rosemary classes shows that WG_rosemary I has a significantly higher number of samples compared to WG_rosemary II and WG_rosemary III. Therefore, it is necessary to balance the dataset using techniques for imbalanced learning.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB. Image Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe input image was reduced to 150 by 150 in order to shorten the computing time required for training. \u0026nbsp;Normalization in the context of the system architecture refers to image pixel scaling. The integer values of an image pixel range from 1 to 255. Processing a big integer number in CNN can interfere with or slow down learning. As a result, the image\u0026apos;s pixels were normalized, or scaled, from 0 to 1.\u003c/p\u003e\n\u003cp\u003eThe histogram of the Digital image is given with the intensity level 0 to L-1\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;h(r\u003csub\u003ek\u003c/sub\u003e)=n\u003csub\u003ek\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;. \u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;.(\u003c/sub\u003e1\u003csub\u003e)\u003c/sub\u003e\u003c/p\u003e\n\u003cp\u003ewhere \u003cem\u003er\u003csub\u003ek\u003c/sub\u003e\u003c/em\u003e is the kth intensity level and \u003cem\u003en\u003csub\u003ek\u003c/sub\u003e\u003c/em\u003e is the number of pixels in the image having that intensity level. We can also normalize the histogram by dividing it by the total number of pixels in the image. For an N x N image, we have the following definition of a normalized histogram function:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eP(rk)=nk/N2\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;. \u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;(2)\u003c/p\u003e\n\u003cp\u003eAs shown in Figure 3, a plot illustrating the intensity transformation across the red, green, and blue channels of the images has been demonstrated. As evident from the graph, the gray level transformation between the image color channels is approximately around 250.\u003c/p\u003e\n\u003cp\u003eAs indicated in Figure 4, various experiments have been conducted, which aimed at transitioning the intensity of the original image to grayscale. This adjustment was made to accommodate the model\u0026apos;s sensitivity within the dual-color spectrum.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC. Feature extraction\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Adjusted CNN models such as Inception_V3, MobileNet, and VGG16 are utilized to extract informative leaf features, subsequently fed into the classifier. The CNN feature learning layer manages these feature extractions, leveraging several techniques. This includes employing the Rectified Linear Unit (ReLu) activation function, which accelerates and enhances learning by mitigating the vanishing gradient problem. Additionally, the Max-pooling technique is employed to reduce the spatial dimensions of the convolved image features. Furthermore, a filtering kernel is utilized, sliding over image pixels and computing dot products to generate diverse image features. Through this feature extraction process, the deep learning model autonomously extracts essential attributes of plant leaves. This structure adheres to the convolutional neural network framework.\u003c/p\u003e\n\u003cp\u003eAs presented in Figure 5, the feature learning phase for rosemary variety classification is executed in this stage. Throughout the training phase of the convolutional neural network, convolution and pooling operations are employed to extract pertinent features from the input.\u003c/p\u003e\n\u003cp\u003eAs outlined in Figure 6, input images sized at 150x150x3 are passed into the convolution layer to reduce dimensionality. Within the convolutional neural network, features are autonomously extracted via the max pooling layer. Subsequently, the images are transformed into a 1D vector upon reaching the fully connected layer.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eUltimately, the proposed Deep Learning-based rosemary Variety Verification model is presented in Figure 7 The researcher collected rosemary leaf images, conducted preprocessing of the images, and subsequently partitioned the image data into training and testing sets. As described in Figure 6 the proposed model design consists of three Phases data collection, data preprocessing, and data splitting. During the data collection phase, the researcher labeled the RGB images in folder and read the images using the panda\u0026apos;s data frame. And in the preprocessing phase, the images are resized, rescaled, and vectorization. Finally, the data is split into training and testing datasets. After data splitting the model is developed using the training data and evaluated using the testing data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eD. Model Evaluation Methods\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccuracy is the proportion of samples that are correctly classified. More precisely it is the sum of the number of true positive and true negative and divided by the number of samples in the dataset. The model is evaluated using cutting edge evaluation metrics Accuracy, pression, recall, confusion matrix \u0026nbsp;and F1 score\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"183\" height=\"27\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;.. (3)\u003c/p\u003e\n\u003cp\u003eIt describes the proportion of correct prediction among all classes and labels. On the other hand, it\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" align=\"\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 576px;\"\u003e\n \u003cp\u003e\u003cimg width=\"146\" height=\"27\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u0026nbsp;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip; \u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;. (4)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe proportion of samples of a certain class of labels that have been predicted by the model as belonging to the class or labels. On the other hand, it is the proportion of true positive among true samples.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 28px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 576px;\"\u003e\n \u003cp\u003e\u003cimg width=\"136\" height=\"27\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u0026nbsp;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;..\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip; \u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;.................(5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"4. Experimental Result and discussion ","content":"\u003cp\u003e\u003cstrong\u003e4. 1 Analysis of Results Utilizing Convolutional Neural Networks\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this section, we employed a conventional Convolutional Neural Network (CNN) to present experimental findings. The dataset was divided, with eighty percent allocated for training and twenty percent for testing purposes. Across the initial three convolutional layers, the researcher utilized 16, 32, and 64 filters respectively. This study utilized a dataset comprising 3482 color images, meticulously labeled according to their respective variations. Recent advancements in deep convolutional neural networks have demonstrated remarkable efficacy in image classification tasks. Thus, all experiments herein were conducted with hyperparameters configured based on a fundamental three-layer CNN model. Specifically, a set of 150 \u0026times; 150 x 3 down-sampled RGB images was fed into the model. The researcher employed a 2x2 max pooling parameter and a 3x3 kernel size to construct a bespoke CNN model. Multiple experiments were conducted utilizing the convolutional neural network model. The dataset was stratified into three groups by the researcher, with ratios of 70/30, 80/20, and 90/10 allocated for training and testing respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1: Analysis of Results Utilizing Convolutional Neural Networks\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"642\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 97px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOptimizer\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 114px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Loss\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation loss\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 97px;\"\u003e\n \u003cp\u003eSGD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e99.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e92.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 114px;\"\u003e\n \u003cp\u003e0.01%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e0.024%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 97px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdam\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.5%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.2%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 114px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.012%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.023%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 97px;\"\u003e\n \u003cp\u003eAdagrad\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e99.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e99.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 114px;\"\u003e\n \u003cp\u003e0.01%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e0.02%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 97px;\"\u003e\n \u003cp\u003eAda delta\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e81%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e82.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 114px;\"\u003e\n \u003cp\u003e0.43%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 97px;\"\u003e\n \u003cp\u003eRMS Prop\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e99%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 156px;\"\u003e\n \u003cp\u003e99.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 114px;\"\u003e\n \u003cp\u003e0.04%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 118px;\"\u003e\n \u003cp\u003e0.024%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAs we presented the above Table 1 The result that is obtained from the experiment of the proposed model by using different optimizers is presented in the following Table 1 by using the classification accuracy, validation accuracy, hamming loss, validation loss metrics in the form of training data, validation data, and test data samples. The proposed model is outperforming in Adam optimizer in iteration per epoch. Whereas the other optimizers adagrad, adadelta, and rmsprop is fluctuating in iteration per epoch.\u003c/p\u003e\n\u003cp\u003e\u003cspan id=\"_Toc108032636\"\u003eAs illustrated in Figure \u003cem\u003e8 a first\u003c/em\u003e the training and validation loss goes different epochs. After some epochs the validation loss became down and scaled up after some epochs. Finally, the training and the validation loss become equal. Therefore, there is no overfitting and under-fitting in the proposed CNN model. The model works well in unseen data As illustrated in \u003cem\u003eFigure 8 b\u003c/em\u003e\u003c/span\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003ethe training and validation accuracy is equal after some epochs. In addition, the gap between the training and validation accuracy is wide. This shows there is some overfitting.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2: Result using Different Activation function\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"642\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eActivation function\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eLearning rate\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Loss\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 126px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation loss\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSoftmax\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.001\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.5%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.2%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.012%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 126px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.02\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eSigmoid \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 103px;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003e87%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 96px;\"\u003e\n \u003cp\u003e0.02%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 126px;\"\u003e\n \u003cp\u003e0.0276\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAs we described in Table 2 using the sigmoid activation function in the last fully connected layer and the output layer of the multi-class image classification is better than soft max which is preferred to multi-class classification. The main secret using the softmax \u0026nbsp;activation function in multi-class image classification is softmax \u0026nbsp;computes the probability between n samples. The result that is obtained from the experiment of the proposed model by using different activation functions is presented in the following Table 2 by using the classification accuracy, validation accuracy,\u003c/p\u003e\n\n\u003cp\u003eAs we observed from the above Table 3 the Confusion matrix of the CNN classifier as we mentioned the model is correctly classified the samples of each label in the test data. The model predicts 86% WG_rosemary I, 66% WG_rosemary II, 67 % WG_rosemary III.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4: Precision, Recall and F1 Score Convolutional Neural Networks\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Labels\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1 Score\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eWG_rosmary I\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eWG_rosmary III\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eWG_rosmary III\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp id=\"_Toc109378926\"\u003eAs we describe in the above\u003cem\u003e\u0026nbsp;\u003c/em\u003eTable 4 the precision and recall of the convolutional neural network. The classifier is performing well in all class labels.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Result Analysis\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eUtilizing Inception V3 Network\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGoogle and GoogleNet developed effective deep learning architecture known as Inception V3, named after a popular internet meme. Despite comprising numerous layers and neurons, computation is primarily executed by a single layer. When considering deeper networks, it\u0026apos;s crucial to acknowledge the increased risk of overfitting. In response, the Inception model incorporates sparsely connected filters across multiple layers within a single layer to mitigate this issue. Trained on the ImageNet dataset, consisting of approximately a million images, Inception V3 accepts 299 \u0026times; 299 x 3 images as input and outputs 1000 classes. This architectural design is grounded in rigorous mathematical principles.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5: Training accuracy, validation accuracy and Loss\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 158px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation Accuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Loss\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 127px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation Loss\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 158px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.8%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 163px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.7%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.001%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 127px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.05%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAccording to Table 5 the accuracy of the model by utilizing the inception v3 model. The model performs well 99.8% in tranning and 98.7% in validation.\u003c/p\u003e\n\u003cp\u003eAs illustrated in Figure \u003cem\u003e9\u003cstrong\u003e\u0026nbsp;a\u003c/strong\u003e first\u003c/em\u003e the training and validation loss goes different epochs. After some epochs the validation loss became down and scaled up after some epochs. Finally, the training and the validation loss become equal. Therefore, there is no overfitting and under-fitting in the proposed Inception V3 model. The model works well in unseen data As illustrated in \u003cem\u003eFigure 9 \u003cstrong\u003eb\u003c/strong\u003e\u003c/em\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003ethe training and validation accuracy is equal after some epochs. In addition, the gap between the training and validation accuracy is wide. This shows there is some overfitting.\u003c/p\u003e\n\u003cp\u003eAs we observed from the above Table 6 the confusion matrix of the Inception V3 classifier as we mentioned the model is correctly classified the samples of each label in the test data. The model predicts 92% WG_rosmary I, 70% WG_rosmary II, 55 % WG_rosmary III.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable7: Precision, Recall and F1 Score Inception V3 model\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 180px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Labels\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1 Score\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 180px;\"\u003e\n \u003cp\u003eWG_rosmary I\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 180px;\"\u003e\n \u003cp\u003eWG_rosmary III\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 180px;\"\u003e\n \u003cp\u003eWG_rosmary III\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 78px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eAs we describe in the above\u003cem\u003e\u0026nbsp;\u003c/em\u003eTable 7 the precision and recall of the Inception V3 The classifier is performing well in all class labels.\u003c/p\u003e"},{"header":"5. Conclusion ","content":"\u003cp\u003eThroughout history, Rosemary plants have played a significant role in various domains including medicine, cuisine, decoration, perfumery, and oil production, exerting considerable influence on the economic, social, cultural, and environmental landscapes of the local communities. Despite its extensive use, identifying the different varieties of rosemary poses a challenge for botanists, researchers, and domain experts, necessitating advanced algorithmic solutions. This study aimed to address these challenge by developing an optimized model for verifying rosemary varieties. To achieve this goal, researchers gathered leaf images representing three classes of rosemary varieties, amassing a total of 3482 images sourced from the Wondo Genet Agricultural Research Center (WARC). Deep learning methodologies were applied to classify three distinct types of rosemary (WG_rosemary I, WG_rosemary II, and WG_rosemary III) utilizing data validated by the WARC for their aromatic and therapeutic attributes. Employing an experimental research approach, the dataset was meticulously prepared for training and testing to evaluate the classification model. A convolutional neural network (CNN) was employed, dynamically extracting features from the resized images of aromatic medicinal plants. A fundamental CNN architecture was utilized for both training and prediction, leveraging preprocessed image samples. Following training, the model exhibited a high accuracy rate of 98.8%, albeit encountering a relatively lower validation accuracy of 97.7%. Researchers conjectured that augmenting the number of CNN layers could potentially enhance validation accuracy, thereby elevating the quality of the obtained results.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eFunding Statement\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2024R409), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgment\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to acknowledge the support of Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2024R409), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have declared that they have no conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eB. M. P. M. Tigist German, Rosemary Production and utilization, Addis Ababa, Ethiopia, December 2016. \u003c/li\u003e\n\u003cli\u003eB. D. F. eemnet Mengesha, \u0026quot;Variability in Ethiopian Rosemary (Rosmarinus officinalis L.) Collections for Agronomic and,\u0026quot; in \u003cem\u003eBiennial Conference of the Ethiopian HorticulturalScience society (EHSS) Science society (EHSS) Science society (EHSS)\u003c/em\u003e, Shshemene, Ethiopi, 2018. \u003c/li\u003e\n\u003cli\u003eT. G. W. G. B. M. Dejene Tadesse Banjaw, \u0026quot;Rosemary (Rosmarinus officinalis L.) Variety Verification Trial at,\u0026quot; \u003cem\u003eMedicinal \u0026amp; Aromatic Plants, \u003c/em\u003e2016. \u003c/li\u003e\n\u003cli\u003eL.-M. G.-B. G.-M. Z.-P. C. Cheng-Li Zhou, \u0026quot;A comprehensive comparison on current deep learning approaches for plant image classification,\u0026quot; \u003cem\u003eJournal of Physics, \u003c/em\u003e2021. \u003c/li\u003e\n\u003cli\u003eS. H. N. K. R. P. Ganbayar Batchuluun, \u0026quot;Deep Learning-Based Plant-Image Classification Using a Small Traning Dataset,\u0026quot; \u003cem\u003eMDPI, \u003c/em\u003e2022. \u003c/li\u003e\n\u003cli\u003eB.-D. M. S. M. Z. Rashad, \u0026quot;Plants Images Classification Based on Textural Features using Combined Classifier,\u0026quot; \u003cem\u003eInternational Journal of Computer Science \u0026amp; Information Technology (IJCSIT) Vol 3, No 4,, \u003c/em\u003eAugust 2011. \u003c/li\u003e\n\u003cli\u003eY. H. ,. Z. Ying Chen, \u0026quot;Plant image recognition with deep learning: A review,\u0026quot; \u003cem\u003eComputers and Electronics in Agriculture, \u003c/em\u003e2023. \u003c/li\u003e\n\u003cli\u003eos\u0026acute;e Miguel Valiente, \u0026quot;Automatic pollen recognition using convolutional neural networks: The case of the main pollens present in Spanish citrus and rosemary honey,\u0026quot; \u003cem\u003eJournal of Food Composition and Analysis, \u003c/em\u003e2023. \u003c/li\u003e\n\u003cli\u003eM. M. S. Yeni Herdiyeni, \u0026quot;Combination of Morphological, Local Binary Pattern Variance and Color Moments Features for Indonesian,\u0026quot; Bogor Agricultural University, 2016. \u003c/li\u003e\n\u003cli\u003e\u0026quot;KEY FINDINGS OF THE 2014/2015 (2007 E.C.) AGRICULTURAL SAMPLE SURVEYS,\u0026quot; Adis Ababa, September, 2015.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 3 and 6 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"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":"Artificial intelligence, convolutional neural network, Computer vision, Varity verification, Rosemary","lastPublishedDoi":"10.21203/rs.3.rs-6432850/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6432850/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Artificial intelligence (AI) has a subfield called computer vision that allows systems and computers to extract replacement data from digital photos and videos. It is used in many fields, including agriculture, health care, education, self-driving cars, and daily living. In Ethiopia, rosemary is a well-known aromatic and therapeutic plant. It is an evergreen herb that belongs to the shrub family and it is widely used in Ethiopia with three varieties: WG rosemary I, WG rosemary II, WG rosemary III. Botanists, researchers, herbal industries, pharmacists and domain experts are facing challenges to classify appropriate varieties. And there is a lack of research and technology for identifying and classifying those varieties in Ethiopia. To address this gap, the proposed study employs supervised machine learning and multi class image classification. Specially, this study is conducted using a convolutional neural network (CNN) employing a SoftMax activation function in the last layer is used to develop the classification models. In this study, five cutting-edge models: convolutional neural network, Inception V3 and exception have been selected. After a comprehensive review of the best-performing models. The 80/20 percentage split was used to evaluate the model, and classification metrics were used to evaluate and compare the models. The pre-trained Inception V3 model outperforms well, achieving training and validation accuracy of 98.8% and 97.7%, respectively.","manuscriptTitle":"Efficient Multi-Class Image-Based Rosemary Variety Verification and Classification Model Using Deep Learning: A Scientific Investigational Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-07 09:10:00","doi":"10.21203/rs.3.rs-6432850/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":"6c72a644-5e8b-407a-a19b-8bfd3e91f392","owner":[],"postedDate":"May 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":47971439,"name":"Biological sciences/Computational biology and bioinformatics/Image processing"},{"id":47971440,"name":"Biological sciences/Plant sciences"}],"tags":[],"updatedAt":"2025-05-08T02:53:36+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-07 09:10:00","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6432850","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6432850","identity":"rs-6432850","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}