Esra Kavalcı Yılmaz
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Esra Kavalcı Yılmaz has a total of 9 co-authors. Here are the top co-authors by number of shared publications:
- Kemal Adem
Sivas Cumhuriyet Üniversitesi
3 shared publications - Halit Bakır
Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Sivas University of Science and Technology, Sivas 58000, Türkiye
2 shared publications - Fatih Ölmez
Sivas Bilim ve Teknoloji Universitesi
1 shared publication - Hatice Aktas Aydın
Sivas University of Science and Technology
1 shared publication Kemal Adem
1 shared publication- Kübra Çelik
Sivas Bilim ve Teknoloji Universitesi
1 shared publication
Latest Publications
- 22 April 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 199
BiKAN-ViT: Enhancing vision transformers via spline-based patch embedding and nonlinear token mixing
- 22 April 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 199
BiKAN-ViT: Enhancing vision transformers via spline-based patch embedding and nonlinear token mixing
- Top Cited Papers
- 6 February 2026
- Published by Springer Nature in Applied Fruit Science
- Vol. 68 (1) , 64
A Hybrid Approach Based on ViT and Capsule Networks for Classification of Agricultural Plant Diseases
- 6 February 2026
- Published by Springer Nature in Applied Fruit Science
- Vol. 68 (1) , 64
Top Cited PapersA Hybrid Approach Based on ViT and Capsule Networks for Classification of Agricultural Plant Diseases
- Open Access
- 31 January 2025
- Published by International Journal of Informatics Technologies in Bilişim Teknolojileri Dergisi
- Vol. 18 (1) , 45-61
Advanced Android Malware Detection: Merging Deep Learning and XGBoost Techniques
The increasing importance of Android devices in our lives brings with it the need to secure personal information stored on these devices, such as contact details, documents, location data, and browser data. These devices are often targeted by attacks and malware designed to steal this data. In response, this work takes a novel approach to Android malware detection by integrating deep learning with traditional machine learning algorithms. An extensive experimental study was conducted using the DroidCollector network traffic analysis dataset. Eight different deep learning methods are analysed for malware classification. In the first phase, experiments were conducted on both original and stabilised datasets and the most effective methods were identified. In the second phase, the best performing deep learning methods were combined with XGBoost for classification. This hybrid approach increased classification success by 3-4%. The highest F1 and accuracy values obtained after 150 epochs of training with BiLSTM+XGBoost were 95.12% and 99.33% respectively. These results highlight the superiority of combining deep learning and traditional machine learning techniques over individual models and significantly improve classification accuracy. This integrated method provides a very important strategy for developing high-performance models for various applications.
- 31 January 2025
- Published by International Journal of Informatics Technologies in Bilişim Teknolojileri Dergisi
- Vol. 18 (1) , 45-61
Open AccessAdvanced Android Malware Detection: Merging Deep Learning and XGBoost Techniques
The increasing importance of Android devices in our lives brings with it the need to secure personal information stored on these devices, such as contact details, documents, location data, and browser data. These devices are often targeted by attacks and malware designed to steal this data. In response, this work takes a novel approach to Android malware detection by integrating deep learning with traditional machine learning algorithms. An extensive experimental study was conducted using the DroidCollector network traffic analysis dataset. Eight different deep learning methods are analysed for malware classification. In the first phase, experiments were conducted on both original and stabilised datasets and the most effective methods were identified. In the second phase, the best performing deep learning methods were combined with XGBoost for classification. This hybrid approach increased classification success by 3-4%. The highest F1 and accuracy values obtained after 150 epochs of training with BiLSTM+XGBoost were 95.12% and 99.33% respectively. These results highlight the superiority of combining deep learning and traditional machine learning techniques over individual models and significantly improve classification accuracy. This integrated method provides a very important strategy for developing high-performance models for various applications.
- Open Access
- 30 December 2024
- Published by Cumhuriyet University in Cumhuriyet Science Journal
- Vol. 45 (4) , 701-706
Classification of Grapevine Leaf Types with Vision Transformer Architecture
Viticulture plays an important role in agriculture. Farmers prefer grapevine cultivation because not only its fruit but also its leaves are used in various fields. Both the use and trade of grapevine leaves within the country is an important source of income. Grapevine leaves, which are grown in almost all countries and used as edible, vary in terms of species. Determining and cultivating the species according to their suitability in terms of productivity is important. In this study, artificial intelligence methods were used to classify grapevine leaf species. The dataset consisting of five different classes, including 100 grapevine leaf images for each class, totalling 500 images, was classified using ViT, VGG19 and MobileNet methods. When the methods used in this study to help increase productivity in production are evaluated, ViT method has the best accuracy rate with 94%.
- 30 December 2024
- Published by Cumhuriyet University in Cumhuriyet Science Journal
- Vol. 45 (4) , 701-706
Open AccessClassification of Grapevine Leaf Types with Vision Transformer Architecture
Viticulture plays an important role in agriculture. Farmers prefer grapevine cultivation because not only its fruit but also its leaves are used in various fields. Both the use and trade of grapevine leaves within the country is an important source of income. Grapevine leaves, which are grown in almost all countries and used as edible, vary in terms of species. Determining and cultivating the species according to their suitability in terms of productivity is important. In this study, artificial intelligence methods were used to classify grapevine leaf species. The dataset consisting of five different classes, including 100 grapevine leaf images for each class, totalling 500 images, was classified using ViT, VGG19 and MobileNet methods. When the methods used in this study to help increase productivity in production are evaluated, ViT method has the best accuracy rate with 94%.
- Open Access
- 4 March 2024
- Published by Uluslararasi Muhendislik Arastirma ve Gelistirme Dergisi in Uluslararası Muhendislik Arastirma ve Gelistirme Dergisi
- Vol. 16 (2) , 659-667
A Comparative Analysis of Deep Learning Parameters for Enhanced Detection of Yellow Rust in Wheat
Wheat, one of the most important food sources in human history, is one of the most important cereal crops produced and consumed in our country. However, if diseases such as yellowpas, which is one of the risk factors in wheat production, cannot be detected in a timely and accurate manner, situations such as decreased production may be encountered. For this reason, it is more advantageous to use decision support systems based on deep learning in the detection and classification of diseases in agricultural products instead of experts who perform the processes in a longer time and have a higher error rate. In this study, the effects of the number of layers, activation function and optimization algorithm variables on the classification of deep learning models used for the classification of yellow rust disease in wheat were examined. As a result of the study, the highest success value was obtained with 97.36% accuracy when using a 5-layer CNN model using Leaky ReLU activation function and Nadam optimization algorithm.
- 4 March 2024
- Published by Uluslararasi Muhendislik Arastirma ve Gelistirme Dergisi in Uluslararası Muhendislik Arastirma ve Gelistirme Dergisi
- Vol. 16 (2) , 659-667
Open AccessA Comparative Analysis of Deep Learning Parameters for Enhanced Detection of Yellow Rust in Wheat
Wheat, one of the most important food sources in human history, is one of the most important cereal crops produced and consumed in our country. However, if diseases such as yellowpas, which is one of the risk factors in wheat production, cannot be detected in a timely and accurate manner, situations such as decreased production may be encountered. For this reason, it is more advantageous to use decision support systems based on deep learning in the detection and classification of diseases in agricultural products instead of experts who perform the processes in a longer time and have a higher error rate. In this study, the effects of the number of layers, activation function and optimization algorithm variables on the classification of deep learning models used for the classification of yellow rust disease in wheat were examined. As a result of the study, the highest success value was obtained with 97.36% accuracy when using a 5-layer CNN model using Leaky ReLU activation function and Nadam optimization algorithm.
- Open Access
- 29 February 2024
- Published by Politeknik Dergisi in Journal of Polytechnic
- Vol. 27 (1) , 343-353
Hyperparameter Tunning and Feature Selection Methods for Malware Detection
Smartphones have started to take an essential place in every aspect of our lives with the developing technology. All kinds of transactions, from daily routine work to business meetings, payments, and personal transactions, started to be done via smartphones. Therefore, there is a significant amount of very important user information stored in these devices which makes them a target for malware developers. For these reasons, machine learning (ML) methods have been used to detect malicious software on android devices quickly and reliably. In this study, a machine learning-based Android malware detection system has been developed, optimized, and tested. To this end, firstly, the data in the dataset has been balanced with 3 different methods namely SMOTE, SMOTETomek and ClusterCentroids. Afterward, the obtained results have been tried to be optimized by using different feature selection approaches including mRMR, Mutual Information, Select From Model, and Select k Best. Finally, the most two successful methods from the five tested ML algorithms (i.e. RF, SVM, LR, XGBoost, and ETC) have been tuned using GridSearch, Random Search, and Bayesian Optimization algorithms in order to investigate the effects of hyperparameter tuning on the performance of ML algorithms.
- 29 February 2024
- Published by Politeknik Dergisi in Journal of Polytechnic
- Vol. 27 (1) , 343-353
Open AccessHyperparameter Tunning and Feature Selection Methods for Malware Detection
Smartphones have started to take an essential place in every aspect of our lives with the developing technology. All kinds of transactions, from daily routine work to business meetings, payments, and personal transactions, started to be done via smartphones. Therefore, there is a significant amount of very important user information stored in these devices which makes them a target for malware developers. For these reasons, machine learning (ML) methods have been used to detect malicious software on android devices quickly and reliably. In this study, a machine learning-based Android malware detection system has been developed, optimized, and tested. To this end, firstly, the data in the dataset has been balanced with 3 different methods namely SMOTE, SMOTETomek and ClusterCentroids. Afterward, the obtained results have been tried to be optimized by using different feature selection approaches including mRMR, Mutual Information, Select From Model, and Select k Best. Finally, the most two successful methods from the five tested ML algorithms (i.e. RF, SVM, LR, XGBoost, and ETC) have been tuned using GridSearch, Random Search, and Bayesian Optimization algorithms in order to investigate the effects of hyperparameter tuning on the performance of ML algorithms.