A proposed deep learning architecture for detecting dragon fruit disease

Abstract

Diseases have remained as one of the major issues affecting dragon fruit production due to its susceptibility to infections. The framework of this research work is the use of a deep learning technique for an assessment of disease infections in dragon fruits with data sourced from a local nursery. The dataset comprises images annotated with four target attributes. A dataset of 2000 high-resolution images, captured under various lighting conditions from local nursery and farms. After Augmentation and Labeling 4,827 images was annotated with four target attributes. New-germination (1300), Germination-white (1203), Heavy red germination (1200), and Wilting (1124). Five models: Xception, VGG19, InceptionV3, MobileNetV2 and one custom CNN model was used for the classification of diseases. Among those, the highest accuracy level was recorded at 93.49 % by using MobileNetV2, while the other three models: Xception, VGG19, InceptionV3 achieved 91.01%, 76.80% and 86.04%, respectively. The result of the complex metric in MobileNetV2 outperforms all other methods significantly while the good performance of the custom CNN which achieved an impressive 91.82% proves that it is an eminently suitable solution designed specifically to act on this task. This implementation helps farmers be able to identify diseases from their crops early enough to reduce crop loss and at the same time help reduce the use of chemicals. Future research could consider several directions for improving the use of deep learning in the identification of dragon fruit diseases. Exploring the possibility of using techniques, in which several deep learning models are employed simultaneously, may prove beneficial to increase the general reliability of the results