Analysis of Deep Learning Algorithms for Detection and Classification of Tomato Leaf Diseases

Abstract

The utilization of deep learning models has gained significant popularity for identifying and categorizing Tomato Leaf Disease, owing to the intricacy of the task and the proficiency required for precise identification. However, the current models employed for this function are typically judged based on their ability to distinguish between healthy and diseased leaves. This model evaluation approach is limiting. The present models are trained and validated on a small dataset that covers only 2000- 5000 images, referred to as Plant Village. This limited dataset size is deemed unreliable, leading to inefficiency and a lack of robustness in the models. Consequently, it makes agriculture-based applications challenging, particularly when spotting new visual effects of Tomato Leaf Disease in smaller datasets. Additionally, the lack of multiclass classification/detection and the absence of damaged regions’ localization in the existing studies constitute a grave concern. Such a deficiency in localization can result in incorrect diagnoses and treatments, leading to crop losses, therefore highlighting the necessity of comprehensive evaluation and validation of these models. This study’s primary focus is on addressing the limitations of a small dataset, and it proposes using a more extensive multi-source dataset of around 40000−45000 images. This study performs data preprocessing to balance the dataset’s classes using data augmentation techniques. Afterward, the final augmented dataset consisted of 60000 images. In concomitance, it notably conducted to an augmented discriminative aptitude, substantiated by a resultant classification efficacy of 99.97%. Concurrently, this methodical intervention correlated with a salutary diminution in the temporal overhead implicated in the training of the expansive cohort encompassing 40,000 distinct tomato leaf images.