Hyperparameter Optimization on CNN Using Hyperband on Tomato Leaf Disease Classification

Abstract

Convolutional Neural Network (CNN) has been successfully applied to image classification, one of which is plant or leaf disease. However, choosing the optimal architecture and hyperparameters is a challenge in its implementation. The purpose of this study was to optimize the Convolutional Neural Network (CNN) hyperparameter on the classification of tomato leaf diseases. In this research, optimization of hyperparameter Convolutional Neural Network (CNN) using Hyperband on Tomato Leaf Disease Detection dataset. The dataset consists of 10,000 training data and 1,000 testing data with 10 classes. In the training data, the distribution of the dataset is 80% for training data and 20% for data validation. This study uses the Keras-Tuner library which aims to optimize two hyperparameters, namely the number of dense neurons and the learning rate. The best hyperparameter value resulting from hyperparameter optimization is 128 for the number of dense neurons and 0.001 for the learning rate. The proposed method succeeded in achieving an accuracy value of 95.690% in the training phase and 88.50% in the validation phase. These results were obtained from model training of 50 epochs. In addition, the model testing obtained an accuracy value of 88.60%. Therefore, hyperparameter optimization on Convolutional Neural Network (CNN) using Hyperband can be an alternative in choosing the optimal architecture and hyperparameters.