Effectiveness of Random Search in Enhancing CNN Performance for Rice Plant Disease Classification

Problems in the classification of plant diseases are often caused by difficulties in determining optimal hyperparameters, which can affect the accuracy and generalization ability of the model. This study aims to explore the influence of hyperparameter variations, especially the number of units, dropout rate, and learning rate, and compare the effectiveness of manual optimization methods with Random Search. Experiments are conducted on MobileNetV2-based CNN models using the Rice Plant Disease (RPD) dataset, which consists of 3698 images of rice leaf diseases classified into 10 categories, including fungal, bacterial, and virus infections. Random Search is applied to find optimal combinations. The results showed that settings with low dropouts and small learning rates, such as 0.001, resulted in higher accuracy. Random Search shows superior performance with an accuracy of 97.84% and significantly reduces validation losses, especially with longer training durations. These findings underscore the importance of the proper selection of hyperparameters and the effectiveness of Random Search in finding the optimal configuration for the CNN model.