Enhancing Plant Identification: Exploring Deep Learning for Herbarium Specimen Classification

Herbarium specimen is a preserved plant specimen collected from the wild or cultivated sources that is carefully mounted, dried, and stored for scientific purposes. It is necessary to have an automated identification system that can classify the species to their respective categories appropriately. However, creating such a classifier is difficult due to the visual similarity of leaves as well as high intra and less inter-class variations. The problem becomes more challenging when the leaves are dry. The proposed method aims to enhance the performance of herbarium specimen classification by emphasizing on the fine variations in the visual traits among distinct species. In order to perform accurate classification of different species of the flowering plant family ’Melastomataceae’, an approach is proposed that aims to improve the performance of herbarium specimen classification by learning the fine variations in the visual traits among distinct species. The proposed approach comprises of pre-processing, undersampling of majority while augmentation of minority classes, segmentation and classification using CNN to recognize the visual patterns associated with different plant species. The evaluation and comparison of proposed method with existing techniques on a publicly available dataset demonstrates its better performance in terms of Accuracy and F1-Score.