Classification of Corn Leaf Disease Using the Optimized DenseNet-169 Model

Abstract

Corn is the second main commodity after rice in Indonesia. Meanwhile, in its cultivation, there are main obstacles, namely pests and diseases. Diseases in plants usually occur in different parts, such as roots, leaves, and stems. However, the leaves are the most common parts to detect the disease because of the differences in size, shape, and color of the leaves. This makes it a major challenge to identify and classify diseases. Classification of leaf diseases of corn is one way to increase the accuracy of diagnosis by utilizing the symptoms and signs found on the leaves of corn plants. This paper presents one of the Deep Convolutional Neural Network (CNN) models, namely DenseNet-169 optimized. Applied models trained with an open dataset from the plant village dataset and primary data obtained from four districts in Madura, Indonesia. Because the amount of primary data is not much, data augmentation is carried out, namely rotate range 90⁰, flip horizontal, flip vertical, brightness random 0.6 to 2.0, zoom range 0.65 to 0.95. To evaluate the model's performance, different optimization parameters were included, namely, Stochastic Gradient Descent (SGD) optimization compared to Adam optimization. The implemented model achieves 62.3%, 75.66%, 98.08% and 99.32% accuracy of corn leaf disease classification for the original primary dataset, the augmented primary dataset, the original secondary dataset and the augmented secondary dataset for the SGD optimizer. As for the Adam optimizer, this model produces a classification accuracy of corn leaf disease of 67.78%, 83.5%, 99% and 99.32% with the same conditions. The accuracy results show that the DenseNet-169 model with Adam optimizer is more hopeful and can significantly affect the efficient recognition of diseases. This makes it possible to have the potential to detect disease in real-time farming systems.