EfficientNetV2M for Image Classification of Tomato Leaf Deseases

Abstract

Favorable climatic conditions make tomato plants (Solanum Lycopersicon) a widely cultivated horticultural crop in Indonesia. However, the increase in tomato production is often accompanied by a decrease in both the quantity and quality of the plants, which can be caused by a variety of factors such as bacteria, fungi, viruses, and insects like Late Blight and Two-Spotted Spider Mite diseases that attack the tomato leaves. To help farmers identify leaf diseases that have similar characteristics, this study employs image processing with the Convolutional Neural Network (CNN) algorithm and transfer learning models. Specifically, the study uses the EfficientNetV2M transfer learning architecture which has superior parameter efficiency and training speed compared to other transfer learning models. Additionally, this study conducts four experimental scenarios on preprocessing, including green channel + CLAHE, green channel + Gaussian Blur, CLAHE without green channel, and Gaussian Blur without green channel. The dataset used in this study includes 5,176 images with three labels: Tomato Healthy, Tomato Late Blight, and Tomato Two-Spotted Spider Mite. These images were used to train and produce models, which were then tested using a different dataset from the trained dataset. The testing dataset included 30 image samples divided into three labels. Based on the test results of the four models with different scenarios, the best model was found to be the one with the green channel preprocessing scenario + CLAHE, which was able to precisely predict all 30 image samples with high accuracy.