IUCNN

IUCNN predicts IUCN Red List (RL) conservation status for Not Evaluated and Data Deficient species using neural networks trained on geographic occurrence records and environmental and trait data.

Key Features:

• Deep Learning Integration: Implements neural networks and deep learning models to predict species extinction risk, demonstrated on datasets including ~14,000 orchid species.
• R-package Implementation: Provided as an R package for model training and prediction using R data structures and workflows.
• Ordinal RL Handling: Incorporates regression-based methods to respect the ordinal nature of IUCN Red List categories.
• Class Imbalance Strategies: Includes strategies to manage class imbalance within training data to improve predictive performance.
• Uncertainty Quantification: Implements a Bayesian approach to quantify uncertainty in model predictions.
• Target Accuracy Thresholds: Limits predictions to species meeting target accuracy thresholds to enhance reliability of assigned statuses.
• Efficiency and Performance: Optimized to deliver rapid extinction risk assessments on large datasets.

Scientific Applications:

• Large-scale ecological and biogeographic analyses: Enables assessment of conservation status across broad taxonomic and geographic scales.
• Estimating threatened species composition: Estimates the number and identity of threatened species across custom geographic or taxonomic datasets.
• Addressing assessment biases: Helps mitigate taxonomic and geographic biases in traditional IUCN Red List assessments by enabling rapid evaluation of understudied species.

Methodology:

Trains neural networks using geographic species occurrences, environmental data, and existing IUCN Red List assessments; employs regression-based methods for ordinal RL categories; applies strategies for class imbalance; uses a Bayesian approach for uncertainty quantification; and applies target accuracy thresholds to limit predictions to confident cases.