Enhancing crop productivity and sustainability through disease identification in maize leaves: Exploiting a large dataset with an advanced vision transformer model

Abstract

The timely identification of diseases in maize leaf offers several benefits such as increased crop productivity, reduced reliance on harmful chemicals, and improved production of healthy crops, resulting in enhanced economic returns. Computer-aided systems (CAD) play a crucial role in agriculture by enabling timely and efficient disease identification in plant leaves. Deep learning-based CAD systems facilitate accurate and rapid diagnosis of maize leaf diseases. In this research, we introduce an advanced vision transformer model that achieves exceptional accuracy and inference speed in detecting diseases in maize leaves. To begin with, we adapt the Multi-axis vision transformer (MaxViT) model to a 4-class maize dataset, creating a lightweight structure that offers improved accuracy and inference speed. Furthermore, we enhance the accuracy by replacing the conventional convolutional structure in the MaxViT architecture's Stem with a Squeeze-and-Excitation (SE) block. In addition, to boost accuracy further, we employ the Global Response Normalization (GRN)-based MLP from the ConvNexTv2 architecture instead of the MLP in the MaxViT architecture. Notably, we combine the PlantVillage, PlantDoc, and CD&S datasets from the literature, resulting in the creation of the most extensive dataset available. This dataset is then divided into three sets: training, validation, and testing, enabling the evaluation of the generalization abilities of the deep learning models. Our study goes beyond previous research by offering a comprehensive comparison of the performance of over 28 CNN models and more than 36 vision transformer models on the newly created dataset. By achieving a remarkable accuracy rate of 99.24% and a high inference speed, the proposed method outperforms all existing deep learning models in the literature. Therefore, it has been demonstrated that this advanced vision transformer model, based on MaxViT, is exceedingly effective for practical applications in agriculture.