Detection and Characterisation of Tomato Plant Diseases Using Deep Learning

Leong, Jia Shean (2021) Detection and Characterisation of Tomato Plant Diseases Using Deep Learning. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Leong Jia Shean.pdf Restricted to Registered users only Download (1MB)

Abstract

Plant disease has been affecting tomatoes crop yield and traditional disease identification methods are inconsistent and inefficient (Nagaraju and Chawla, 2020). The trend of tomato plant disease identification is utilising machine vision to classify the diseases based on the leave images. There are many attempts in assessing plant diseases with deep learning, but it is found that over 85% of models display inefficiency in the subject. Various researches employ popular deep learning models like ANN, RNN, and CNN in image classification. Chapter 2 discusses previous works reported in the literature about these models in the context of plant disease identification, their limitations and possible improvements. In ANN, MLP, RBF and PSO are studied, whereby PSO outperforms other models in terms of accuracy and computational power (Singh, 2019). In RNN, LSTM and PC-RNN are studied, whereby LSTM has 89% of accuracy while PC-RNN generally performs better (Chandra and Sharma, 2017). In CNN, AlexNet, VGG16 and ResNet are studied, whereby ResNet 18 and 50 can have the highest accuracy of 99% (Maeda-Gutiérrez et al., 2020). However, the deeper ResNet is more computationally costly. Most researches on tomato plant disease detection and classification use CNN models. This is because CNN requires no data pre-processing, has a better convergence rate and training performance. The state-of-the-art CNN architectures are AlexNet, VGG16, and ResNet. Through reviewing previous studies, it is concluded that the best models of plant disease identification and classification are these three models and applications of them to tomato plant diseases are to be investigated further. Chapter 3 outlines the research methodology of the research. The techniques used in the training model are transfer learning, data augmentations and cloud computing. These techniques are to solve the limitations as described in section 2.4 and also to work under constrained computing resources in the Covid-19 pandemic. AlexNet is being replaced by MoibleNetV2 because of the accessibility to the model and the model to transfer learning. The three models, VGG16, ResNet 50 and MobileNetV2 have experimented under different training-to-testing ratios and absence and presence of data augmentation to investigate how their accuracy, loss and training time will be affected. Chapter 4 shows the experiment results and discussions of them. There is no significance found in loss and accuracy for different training-to-testing ratio and presence and absence of data augmentation. After 10 training epochs, the highest accuracy acquired is 90.19% for VGG16, 68.08% for ResNet 50 and 90.84% for MobileNetV2. The accuracy obtained is comparable to a similar study that also uses 10 training epochs. In that study, the accuracy obtained is 90.4% for VGG16 and 80% for ResNet 50. The models trained in the experiment under crucial limitations of limited computational resources. In future works, it is recommended to have a consistent model training environment, model fine-tuning, more training epochs and more evaluation metrics to produce models with more reliable results.