Deep learning based assessment of disease severity for early blight in tomato crop

Abstract

Assessment of disease severity is one of the major challenges which helps in the prediction of yield quantitatively and to decide the control factors that improve the yield of any crop. Hence a perfect system is essential to measure the severity level of the disease in order to improve its productivity. An intelligent state of the art technique i.e., deep learning plays an inevitable role in most of the real-time applications including smart farming. Tomato crops are frequently affected by a dangerous fungal disease i.e., early blight, resulting in 100% production loss to farmers. In this work, an identification of early blight disease in tomato leaves is performed by a recently invented paper microscope named Foldscope. Further, a deep Residual Network101 (ResNet101) of Convolutional Neural Network (CNN) architecture is used to measure the severity level of early blight disease in tomato leaves. The dataset in the model is trained by using an open database i.e., PlantVillage dataset for mild, moderate, and severely diseased leaves along with healthy tomato leaves. The results of ResNet101 architecture is compared with other pre-trained CNN such as Visual Geometry Group16 (VGG16), VGG19, GoogLeNet, AlexNet, and ResNet50. Among these networks, the deep ResNet101 architecture has achieved the highest accuracy of 94.6%. Finally, a case study has been conducted based on the estimated severity levels and the required fungicide treatment is also prescribed.