Self-supervised learning advanced plant disease image classification with SimCLR

Abstract

Supervised learning will be a bottleneck for developing plant disease identification since it relies on learning from massive amounts of carefully labeled images, which is costly and time-consuming. On the contrary, self-supervised learning has succeeded in various image classification tasks; however, it has not been applied broadly in the plant disease analysis process. This work, therefore, studies the effectiveness of self-supervised learning using contrastive pre-training with SimCLR for plant disease image classification. We investigated unsupervised pre-training scenarios on unlabeled plant images across multiple architectures, including supervised fine-tuning on labeled samples. In addition, we explored the label efficiency of the self-supervised approach, acquired by fine-tuning the models on various fractions of labeled images. Our results demonstrated that the performance of self-supervised learning on plant disease became comparable to that of the supervised training approach.