A comprehensive comparison on current deep learning approaches for plant image classification

• Published in: Journal of physics. Conference series Vol. 1873; no. 1; p. 12002
• Main Authors: Zhou, Cheng-Li, Ge, Lin-Mei, Guo, Yan-Bu, Zhou, Dong-Ming, Cun, Yu-Peng
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.04.2021
• Subjects: Accuracy; Biodiversity; Classification; Datasets; Deep learning; Image classification; Machine learning; Model testing; Physics; Taxonomy
• ISSN: 1742-6588; 1742-6596
• DOI: 10.1088/1742-6596/1873/1/012002

Abstract Plant identification and classification play a key role in understanding, protecting and conserving biodiversity. Traditional plant taxonomy needs long time intensive training and experience, which limited others to identify plant categories. With the development of automated image-based classification, machine learning (ML) is becoming a popular tool. Image classification, especially plant images taxonomy, has achieved great improvement in these years by deep learning (DL) methods. In this study, we first reviewed current deep learning applications in the field of plant image classification, and then we tested six deep learning methods in four public plant image datasets. In order to test the classification power of DL methods at cultivar level, we prepared a Camellia sasanqua Thunb. dataset, which is called Camellia@clab, for assessing classification performance of the six DL methods. These DL models’ classification performance all exceeded 70% in the four public plant image datasets, and LeNet and DenseNet had stable good performance, with median prediction accuracy of the LeNet was over 87.29% and that of DenseNet was over 93.8% in the four public datasets at species level. At cultivar level, the lowest median prediction accuracy of those DL methods decreased to 62%, but LeNet and DenseNet still performed very well. The prediction accuracy of LeNet and DenseNet was 82.3% and 100% in the Camellia@clab dataset, respectively. DenseNet model showed a stable best classification performance among the five datasets. To our knowledge, this is the first study that provides a comprehensive review and comparison on applying current DL methods to plant image classification. This study will provide guidance for DL applications in plant image classification, and point out the protentional DL research direction for modeling improvement.