Tomato leaf disease classification by exploiting transfer learning and feature concatenation
Mehdhar S. A. M. Al-Gaashani, Fengjun Shang, Mohammed Saleh Ali Muthanna, Mashael Khayyat, Ahmed A. Abd El‐Latif
Abstract
Abstract Tomato is one of the most important vegetables worldwide. It is considered a mainstay of many countries’ economies. However, tomato crops are vulnerable to many diseases that lead to reducing or destroying production, and for this reason, early and accurate diagnosis of tomato diseases is very urgent. For this reason, many deep learning models have been developed to automate tomato leaf disease classification. Deep learning is far superior to traditional machine learning with loads of data, but traditional machine learning may outperform deep learning for limited training data. The authors propose a tomato leaf disease classification method by exploiting transfer learning and features concatenation. The authors extract features using pre‐trained kernels (weights) from MobileNetV2 and NASNetMobile; then, they concatenate and reduce the dimensionality of these features using kernel principal component analysis. Following that, they feed these features into a conventional learning algorithm. The experimental results confirm the effectiveness of concatenated features for boosting the performance of classifiers. The authors have evaluated the three most popular traditional machine learning classifiers, random forest, support vector machine, and multinomial logistic regression; among them, multinomial logistic regression achieved the best performance with an average accuracy of 97%.