An efficient and consistent framework for multi-rank taxonomic identification in wildlife images

Area of Science:

• Biodiversity research
• Computational biology
• Machine learning for ecology

Background:

• Accurate taxonomic classification is crucial for biodiversity research but current image-based methods lack taxonomic consistency.
• Field studies face computational and network limitations on edge devices, hindering reliable biodiversity monitoring.
• Existing models often fail to ensure hierarchical consistency in species identification.

Purpose of the Study:

• To develop a reliable and computationally efficient method for hierarchical taxonomic classification.
• To address the challenge of taxonomic inconsistency in image-based classification models.
• To enable accurate biodiversity monitoring on resource-constrained edge devices.

Main Methods:

• Proposed TaxonomyNet, an ensemble detection model with six independent heads for taxonomic classification.
• Introduced the Weighted Agreement Loss (WAL) metric to enforce structural coherence and consistency in predictions.
• Trained and evaluated the model on a dataset of 50 Australian animal species.

Main Results:

• TaxonomyNet achieved high detection performance across all taxonomic ranks (mAP: 90.7-99.75%).
• The WAL metric improved species-level accuracy by up to 3.87% compared to baseline and foundation models.
• Demonstrated superior computational efficiency, reducing processing delay by 22 minutes for 1500 samples.

Conclusions:

• TaxonomyNet offers a practical and extensible solution for reliable hierarchical classification in biodiversity monitoring.
• The WAL metric effectively enforces taxonomic consistency, enhancing classification accuracy and scientific reliability.
• The model's efficiency makes it suitable for deployment on edge devices in real-world ecological studies.