Transfer Learning-Based Tomato Disease Prediction using Leaf Images

Abstract

Tomato plants are among the most important of all the fundamental crops, valued highly by both farmers and consumers. The biggest concern for tomato farmers is that the plant is becoming infected with unidentified diseases. Farmers must recognize the type of disease to treat the plants appropriately when numerous diseases could hinder growth. The goal of this project is to create three distinct transfer learning models that can accurately identify the type of sickness. The most effective algorithm for this purpose is also finalized. For this purpose, a dataset of more than 10,000 images is collected from Kaggle. This collection contains pictures of tomato plant leaves in both a healthy and a diseased state. Three sorts of data must be created: training, testing, and validation. The smallest portion is utilized to validate the model, while the second largest portion is used to test the model for its final parameters. After separating, the photographs are preprocessed. In this study, image scaling is the only preprocessing technique used. The transfer learning models that were developed utilizing three different transfer learning strategies are then tested and trained using the preprocessed and divided images. In this experiment, the VGG-16, AlexNet, and hybrid algorithms were applied. Testing, validation, and training outcomes are recorded. But next, to enhance readability and facilitate analysis, the data is converted by being plotted onto a graph. Throughout the training, the hybrid model showed significantly greater performance than the other two models. The models created using the VGG-16 and AlexNet algorithms displayed a lot of oscillations, even when the margin shrank during validation, which could lead to inconsistent performance. The model can be used in the future to help farmers and other agriculturists accurately identify a problem with their plant and search for a workable remedy by integrating it into the backend processor of a website or software application uses pre-trained convolutional weights to optimize the model.