Abstract:
Tomatoes, an indispensable and consumable crop worldwide, exhibit variations in size based on
fertilization methods. Leaf disease is the primary factor that influences the amount and standard
of tomato yield. Consequently, it is imperative to effectively identify and classify these diseases.
Various conditions can impact tomato production, and identifying these conditions beforehand
would minimize the adverse effects of diseases on tomato plants and enhance crop yield. There
are various factors that limit tomato yield, some of which are brought on by fungi, viruses, and
bacteria. The visual classification of tomato leaf diseases is poor due to the multiplicity of
illnesses. This research aims to develop a precise and effective model for classifying tomato leaf
diseases using Convolutional Neural Networks (CNNs), which have proven highly successful for
this purpose. The proposed methodology includes several stages: preprocessing, feature
extraction, image enhancement, and segmentation techniques. To get the data ready for analysis,
preprocessing included normalizing the photos and eliminating noise. In order to classify
diseases, feature extraction concentrated on identifying important image properties. To increase
image quality, image enhancement techniques were used. In order to separate sick from healthy
zones and enable more precise analysis, segmentation techniques were applied. Ten classifier
labels are employed for this purpose, specifically Tomato Early Blight, Tomato Spider Mites,
Tomato Bacterial Spot, Tomato Late Blight, Tomato Target Spot, Tomato Septoria Leaf Spot,
Tomato Yellow Leaf Curl Virus, Healthy Leaf, Tomato Leaf Mold, and Tomato Mosaic Virus. The
experiments carried out to evaluate the proposed CNN model involved the utilization of two
distinct categories of datasets: the RGB dataset and an augmented image dataset (which
integrates RGB images and employs a range of augmentation techniques). Various dataset ratios
were employed during the model development process, including 90/10, 80/20, and 70/30. After a
comparison study of the results, the 90/10 data ratios were selected. The accuracy experienced a
substantial enhancement when the model underwent training on the RGB image dataset,
attaining an accuracy rate of 96.60% after 90 training epochs and a learning rate of 0.0001.
Ultimately, the performance was further improved by employing the augmented image dataset,
resulting in an accuracy of 99.60% after 70 training epochs and a learning rate of 0.000