On the application of image augmentation for plant disease detection : A systematic literature review

Agriculture significantly influences the global economy, especially in developing countries, but plant diseases can drastically reduce crop yields and economic gains if not detected early. To combat this, the agriculture sector must adopt innovative technologies like image augmentation for early disease detection, though it faces challenges such as limited datasets for model training. Conventional Image augmentation techniques and Generative Adversarial Networks (GANs) have been used to generate more image data in prior studies. Several reviews and surveys have been conducted that provide overviews of datasets, models, and GAN architectures used in Plant Disease Detection (PDD) but do not investigate the comparative use of GANs and the basic data augmentation approaches. In this paper, we conducted a tertiary Systematic Literature Review (SLR) to summarize the current state in the field of automatic plant disease detection. 49 secondary reviews, including over 555 unique primary studies, are considered. This review shows that GANs are increasingly becoming the go-to image augmentation technique in PDD. Particularly, Deep Convolutional GAN (DCGAN) is found to be the most used image augmentation technique in PDD. Interestingly, the basic image augmentation techniques, image flipping and image rotation were found to be very popular among researchers. Additionally, the review reveals that convolutional neural network (CNN) models, especially VGG models, have the best performance in the field of plant disease detection. It was also revealed that insufficient data continues to be a huge challenge, as insufficient dataset limits the representativeness, universality and generalizability of the models. Again, most datasets are private, while open-source datasets are often too small or modified under laboratory conditions; which makes them impractical. Other findings extracted from the review are (a) the issues of the unbalanced dataset, (b) the questionable effectiveness of GANs due to challenges in generating realistic images of plant diseases as well as challenges in practical applications, and (c) the lack of farm-field practicality of the models. It is concluded that plant disease image augmentation and disease detection models perform well on certain datasets, but more research on real-life data, and therefore open-source real in-field datasets, are needed.

Bibliographic Details

Main Authors:	Antwi, Kwame, Bennin, Kwabena Ebo, Pobi Asiedu, Derek Kwaku, Tekinerdogan, Bedir
Format:	Article/Letter to editor biblioteca
Language:	English
Subjects:	Convolutional neural network, Data augmentation, Deep learning, Generative adversarial networks, Machine learning,
Online Access:	https://research.wur.nl/en/publications/on-the-application-of-image-augmentation-for-plant-disease-detect
Tags:	Add Tag No Tags, Be the first to tag this record!