CLA: A self-supervised contrastive learning method for leaf disease identification with domain adaptation

Abstract

Plant leaf diseases cause a decrease in crop yield and degrade the quality, which presents the urgent need for leaf disease identification. Recently, deep learning technologies, especially computer vision, have emerged as a powerful tool in plant leaf disease identification. However, existing methods invariably rely on large-scale labeled data for model training. Although self-supervised learning which uses large-scale unlabeled data for pre-training provides a useful scheme, these unlabeled data are messy, e.g., images with different shooting angles and backgrounds. In this regard, such data results in distribution-shifted training datasets, which degrades model performance. To address these problems, we propose a self-supervised Contrastive learning method for Leaf disease identification with domain Adaptation (CLA), including pre-training with forwarding and fine-tuning with domain adaptation stage. Specifically, CLA utilises large-scale yet messy unlabeled data to train the encoder and obtains their visual representations in the pre-training stage. Based on small labeled data, CLA trains the domain adaptation layer (DAL) and the classifier in the fine-tuning stage. Due to the DAL, CLA can align labeled data and unlabeled data in the fine-tuning stage and generates more general visual representations, which improves the domain adaptation ability of CLA. Experiments are conducted to evaluate the performance of CLA and demonstrate that CLA outperforms the comparison methods by a large margin on both domain adaptation and accuracy performance, with the highest accuracy of 90.52%. To better understand our proposed method, additional experiments are also conducted to explore the influencing factors of CLA.