Deep Learning with Taxonomic Loss for Plant Identification.

Area of Science:

• Computer Science
• Botany
• Machine Learning

Background:

• Plant identification is crucial for biodiversity assessment and requires fine-grained classification of family, genus, and species.
• Deep learning models often struggle with the hierarchical nature of taxonomic labels.

Purpose of the Study:

• To introduce a novel taxonomic loss function for deep learning-based plant identification.
• To evaluate the effectiveness of the proposed loss function against traditional methods.

Main Methods:

• A taxonomic loss function was developed to incorporate hierarchical relationships into the deep learning objective.
• Various neural networks were trained using both the proposed taxonomic loss and cross-entropy loss on PlantCLEF 2015 and 2017 datasets.

Main Results:

• The taxonomic loss function demonstrated significant performance improvements over cross-entropy loss.
• On the PlantCLEF 2017 dataset, the SENet-154 model with taxonomic loss achieved higher accuracies at family (84.07%), genus (79.97%), and species (73.61%) levels.

Conclusions:

• The proposed taxonomic loss is effective and easy to implement for fine-grained plant identification.
• This approach enhances deep learning models by better utilizing hierarchical label structures.