Deep Learning-based Classification of Tomato Leaf Diseases

Abstract

Background: In the Indian economy agriculture plays important role. Many of the crops are damaged due to diseases, therefore plant leaf disease detection at early stage is important. Tomatoes are the second most consumed vegetable in Indian households, with a rank second largest producer and consumption in world. Tomatoes are of economically important crop in India with large scale cultivation. Yet, unfavourable environmental factors tend to cause numerous diseases due to bacteria, fungi and viruses that infect different plant parts. These diseases cause the yield to be lower, leading to heavy economic losses for the farmers. For effective disease management, it is important to detect the disease in early stage using the advanced techniques for maximising the yielding of crop. Methods: In the proposed approach, tomato leaf images were first enhanced using CLAHE to improve local contrast, after that data augmentation to increase dataset diversity. Four transfer learning models based on pretrained CNN architecture: VGG16, VGG19, ResNet50 and MobileNetV2, were used as fixed feature extractors, where the pre-trained layers were fixed and newly added classification layers were trained for tomato leaf disease classification. Although transfer learning has been widely applied for tomato leaf disease classification, comparative investigations on the effect of CLAHE-enhanced images across different pre-trained architectures remain limited. This study shows performance of these four CNN architectures under identical preprocessing and training conditions to analyze the influence of CLAHE-based contrast enhancement on disease feature representation and classification performance. Result: The outcomes of experimental results shows strong classification performance on the plant village dataset for detection of tomato leaf disease, achieving best accuracies of 93.75%, 95.75%, 86.38% and 98.54% using the VGG16, VGG19, ResNet50 and MobileNetV2 models respectively.