Automatic species identification from images for Aotearoa.

Area of Science:

• Ecology
• Computer Science
• Bioinformatics

Background:

• Accurate species identification is crucial for conservation and education.
• New Zealand's unique biodiversity presents an opportunity for automated identification systems.
• Machine learning offers a powerful approach for image classification tasks.

Purpose of the Study:

• To develop and evaluate neural network-based image classification models for New Zealand species.
• To enable accurate, offline species identification on mobile devices.
• To provide open-source tools for biodiversity research and public engagement.

Main Methods:

• Utilized a dataset of 14,991 species from iNaturalist, covering Animalia, Plantae, Fungi, and other kingdoms.
• Trained neural network models for image classification.
• Calibrated model confidence using temperature scaling and employed input attribution for interpretability.

Main Results:

• Achieved over 76% classification accuracy across all species.
• Generated calibrated class probability estimates for confidence assessment.
• Demonstrated the utility of input attribution for understanding model predictions.

Conclusions:

• The developed models provide accurate and reliable species identification for New Zealand organisms.
• The open-source, offline-capable applications support conservation, education, and research.
• This work facilitates biodiversity monitoring and public engagement with Aotearoa's unique species.