Consistency regularization based semi-supervised plant disease recognition

Plant diseases in agriculture are one of the most important factors affecting the quality and efficiency of production. Autonomous early detection of diseases in plants allows the necessary pesticide work to be carried out on time. Such an autonomous system can minimize production losses and improve product quality. This study highlights the data collection and labeling challenges in the field of plant diseases and proposes to solve the plant disease recognition problem with semi-supervised learning. To this end, we first evaluate three open datasets in the literature and settle on one. Then, numerous experiments are conducted to show that the supervised learning solution to the plant disease recognition problem is ineffective with limited labels. With these experiments, the effects of factors are measured, including limited labeled data, dataset balance, batch size, different fine-tuning strategies, and different levels of input augmentation in supervised learning. Finally, semi-supervised learning experiments showed an improvement in accuracy of up to 5.52 % over supervised learning.