Modified-residual network for maize stalk rots diseases classification

Abstract

In this article, image classification of maize stalk rots diseases was carried out. The experiment used primary data taken from maize plantations in Bangkalan, Madura. The data consists of three classes: healthy, anthracnose, and gibberella. For deep learning experiments, we augmented the primary data. The total data was 2,211 images. An investigation is composed of three sections. First, we used five different Convolutional Neural Network (CNN) architectures, and second, the best CNN will be modified. Finally, it performed varied layer types from the second section. The parameters used were epoch 10, learning rate 3.10e-4, and minibatch-size 64. The distribution of training, validation, and testing data were 40:40:20. The result shows the best performance for the first section is ResNet18. Next step, we modify ResNet18 into six different architectures. From the second section, the best results were ResNet18 and modified-ResNet, but modified-ResNet has less number of parameters. The third section’s results showed accuracy, precision, and recall were 99.55%, 99.53%, and 99.73%, respectively. The modified-ResNet architecture is suitable for classifying maize stalk rots diseases.