Research and Publications

Introduction

This paper presents neural network–based image classification models designed to automatically identify species found in Aotearoa New Zealand, leveraging a large, diverse dataset sourced from iNaturalist. Trained across nearly 15,000 species spanning Animalia, Plantae, Fungi, and several smaller kingdoms, the models achieve over 76% accuracy and provide calibrated confidence estimates, enabling reliable offline species identification on both web and mobile platforms.

Problems addressed

• New Zealand’s biodiversity is unique, yet manual species identification remains time‑consuming and inaccessible for many users.

• Existing tools often require internet connectivity, limiting their usefulness in remote environments where species are observed.

• Large‑scale species classification is challenging due to high class imbalance, fine‑grained distinctions, and limited labeled data for many taxa.

• Many ML models lack interpretable outputs, making it difficult to understand why a prediction was made.

Methods overview

The authors develop deep neural network models trained on 14,991 species from iNaturalist, covering organisms observed both in the wild and in captivity. The models output calibrated class probabilities using temperature scaling, enabling users to gauge prediction confidence. Input attribution techniques highlight image regions influencing the model’s decisions, improving interpretability. The resulting models are optimized for deployment on mobile devices, supporting offline species identification across Aotearoa.

Key outcomes

• Delivers a large‑scale species identification model tailored to New Zealand’s biodiversity.

• Achieves 76%+ accuracy across nearly 15,000 species.

• Provides calibrated confidence scores for more trustworthy predictions.

• Includes input attribution for interpretable model outputs.

• Supports offline use on mobile devices and is available as open‑source software.

Experimental findings

The trained models achieve over 76% accuracy across all species, demonstrating strong performance despite the large number of classes and fine‑grained distinctions. Calibration improves the reliability of probability estimates, and attribution methods provide meaningful insights into model behavior. The models are released publicly as downloadable files and integrated into open‑source web and mobile applications for real‑world use.