Accurate Multilevel Classification for Wildlife Images.

Area of Science:

• Computer Science
• Machine Learning
• Biodiversity Informatics

Background:

• Traditional classification models infer single classes, which can be inefficient for hierarchical data.
• Human knowledge is structured taxonomically, suggesting hierarchical approaches may be more effective.
• Classifying wild animals and plants presents challenges due to the vast number of species and their natural taxonomic relationships.

Purpose of the Study:

• To explore and evaluate different methods for multilevel classification.
• To apply these methods to the specific task of classifying wild animal and plant species.
• To identify the optimal model architecture and data processing techniques for hierarchical species classification.

Main Methods:

• An exhaustive study of various multilevel classification techniques was performed.
• Different convolutional neural network (CNN) backbones, data configurations, and ensembling methods were investigated.
• The study focused on using an EfficientNetB5 backbone with a 300x300 px input size and a Multiscale Crop data augmentation strategy.

Main Results:

• The optimal setup for tree-structured datasets involved an EfficientNetB5 backbone and a multilevel classifier.
• A Multiscale Crop data augmentation process was crucial for achieving high performance.
• The best-performing model achieved 62% top-1 accuracy and 88% top-5 accuracy on the species classification task.

Conclusions:

• Multilevel classification offers a more intuitive and potentially more effective approach for hierarchical data like species taxonomy.
• The combination of EfficientNetB5, specific input size, and Multiscale Crop augmentation yields strong results for species classification.
• While ensemble methods could further boost accuracy, their computational cost currently limits practical application.