Image-Based Automated Species Identification: Can Virtual Data Augmentation Overcome Problems of Insufficient...

Area of Science:

• Biodiversity Informatics
• Computational Biology
• Machine Learning Applications

Background:

• Automated species identification is crucial but challenging, particularly for rare species with limited sampling data.
• Scarcity of infraspecific data hinders the performance of machine learning models in distinguishing between closely related species.
• Existing methods struggle with low or exaggerated interspecific morphological variation.

Purpose of the Study:

• To evaluate the effectiveness of a data augmentation strategy for enhancing automated visual species identification.
• To address the challenge of limited infraspecific sampling data in machine learning for species identification.
• To improve the accuracy of automated species identification, especially for rare and under-sampled taxa.

Main Methods:

• Employed a stepwise data augmentation approach including image rotation, visual augmentation, and generative adversarial networks (GANs).
• Extracted descriptive feature vectors from VGG-16 convolutional neural network bottleneck features, reduced dimensionality using Global Average Pooling and Principal Component Analysis (PCA).
• Utilized synthetic oversampling in feature space to augment limited training data.

Main Results:

• The proposed data augmentation approach significantly outperformed a deep learning baseline using non-augmented data.
• The augmentation method also showed superior performance compared to a traditional 2D morphometric approach (Procrustes analysis).
• The study demonstrated improved identification accuracy across diverse datasets including beetles, bees, and butterflies.

Conclusions:

• Data augmentation is an effective strategy to overcome the limitations of scarce training data in automated species identification.
• The combined approach of deep learning, GANs, and synthetic oversampling offers a robust solution for identifying rare species.
• This methodology holds significant promise for advancing biodiversity research and conservation efforts through improved automated identification tools.