Automated disease classification in (Selected) agricultural crops using transfer learning

• Published in: Automatika Vol. 61; no. 2; pp. 260 - 272
• Main Authors: Rangarajan Aravind , Krishnaswamy, Raja, Purushothaman
• Format: Journal Article Paper
• Language: English
• Published: Ljubljana Taylor & Francis 02.04.2020; Taylor & Francis Ltd; KoREMA - Hrvatsko društvo za komunikacije,računarstvo, elektroniku, mjerenja i automatiku; Taylor & Francis Group
• Subjects: Accuracy; Artificial neural networks; automated classification; Automation; convolutional neural networks; Crop disease diagnosis; Crop diseases; Crops; Image acquisition; Image classification; Image transmission; Local area networks; Machine learning; Medical imaging; Model accuracy; Personal computers; Reduction; smartphone; Smartphones; Tobacco; transfer learning; Viruses
• ISSN: 0005-1144; 1848-3380
• DOI: 10.1080/00051144.2020.1728911

The biotic stress of agricultural crops is a major concern across the globe. Especially, its major effects are felt in economically poor countries where advanced facilities for diagnosis of a disease is limited as well as lack of awareness among the farmers. A recent revolution in smartphone technology and deep learning techniques have created an opportunity for automated classification of disease. In this study images acquired through smartphone are transmitted to a personal computer via a wireless Local Area Network (LAN) for classification of ten different diseases using transfer learning in four major agricultural crops which are least explored. Six pre-trained Convolutional Neural Network (CNN) have been used namely AlexNet, Visual Geometry Group 16 (VGG16), VGG19, GoogLeNet, ResNet101 and DenseNet201 with its corresponding results explored. GoogLeNet resulted in the best validation accuracy of 97.3%. The misclassification was mainly due to Tobacco Mosaic Virus (TMV) and two-spotted spider mite. In test conditions, images were classified in real-time and prediction scores have been evaluated for each disease class. It depicted a reduction in accuracy in all models with VGG16 resulting in the best accuracy of 90%. Various factors contributing to the reduction in accuracy and future scope for improvement have been elucidated.