Agricultural Crop Disease Detection Using Convolutional Neural Networks

Abstract

Mitigating food insecurity in Niger is of paramount importance. According to some researchers, agricultural yield is declining. The permanent annual food shortage was estimated to be between 25 and 50%. Loss of cereal agricultural production is caused by pests and plant diseases. Knowing that ordinary techniques became inefficient. Automatic detection of plant diseases using artificial intelligence becomes the best solution. Several methods applying Convolutional Neural Networks (CNN) have been used in the recent literature. However, these networks are significantly affected by the vanishing gradient problem. Thus, to overcome this challenge, some CNN architectures have been proposed in recent literature. In this study, some properties of these CNN architectures, such as Densely Connected Convolutional Network (DenseNet) and AlexNet, were combined to propose a new efficient architecture called AlexNetDense. Some works prove that not all connections in DenseNet play positive roles for small datasets. And, reducing the connections between layers can improve the efficiency of the network model. Based on AlexNet, the proposed architecture adds some connections not only between consecutive layers. The proposed method was evaluated on PlantVillage. It produced average accuracy, F-measure, and MCC rates of 99.30, 99.12, and 98.90% respectively. Average accuracy, F-measure, and MCC rates of 99.94, 99.94, and 99.88% for this model with Millet Leaf Dataset were obtained. The result showed that the proposed model achieved a higher performance compared to most of the state-of-the-art models, such as LeNet-5 (LeNet), AlexNet, Visual Geometry Group (VGGNet), ResNet, DenseNet, and EfficientNet.