Recognizability bias in citizen science photographs

Area of Science:

• Ecology
• Computer Science
• Biodiversity Informatics

Background:

• Citizen science and automated methods rely on image recognition for vast observational data.
• Image recognition models require extensive data, creating a feedback loop with data collection methods.

Purpose of the Study:

• To investigate the 'recognizability bias' in image data used for species recognition.
• To understand how this bias affects training data and model performance across different taxa.

Main Methods:

• Trained species recognition models on diverse datasets.
• Analyzed image data prevalence in relation to species recognizability by humans and algorithms.

Main Results:

• Found evidence of a 'recognizability bias' across multiple taxa.
• Easily identifiable species were disproportionately represented in image datasets.
• Bias was not explained by image quality, biological traits, or other data collection metrics.

Conclusions:

• The 'recognizability bias' can limit the effectiveness of training data for under-reported species.
• This bias has significant implications for the performance of future AI models in ecological research.
• Addressing this bias is crucial for accurate biodiversity monitoring and conservation.