Towards ML Methods for Biodiversity: A Novel Wild Bee Dataset and Evaluations of XAI Methods for ML-Assisted Rare Species Annotations

AI can help address biodiversity loss, but requires specialized datasets: The paper highlights that while deep learning offers promise for monitoring declining insect populations, the field lacks benchmark datasets for insects. The authors created a new wild bee dataset to help fill this gap, demonstrating that developing AI solutions for conservation requires domain-specific training data rather than relying on generic models.

Expert knowledge remains essential for rare species identification: Annotating training data for rare species requires biologist expertise, which is a significant bottleneck. Rather than replacing experts, the paper suggests AI should be designed to assist them—a key principle for responsible AI in high-stakes domains like conservation.

Explainability methods need evaluation for real-world use: The paper investigates which XAI (Explainable AI) methods best support biologists in annotation tasks, recognizing that simply having an accurate model isn't enough. For responsible AI deployment in conservation, the model's decision-making process must be interpretable to domain experts who need to verify and trust its recommendations.

Fine-grained classification remains technically challenging: Wild bee classification is difficult due to high visual similarity within species and significant variation within species (high intraclass variability). This technical challenge underscores that responsible AI requires acknowledging where models may struggle and designing appropriate human oversight accordingly.