Maximizing citizen scientists' contribution to automated species recognition

Area of Science:

• Biodiversity informatics
• Computational biology
• Citizen science

Background:

• Artificial intelligence (AI) tools are increasingly used for species identification from images.
• Citizen science data is crucial for addressing biodiversity knowledge gaps and taxonomic bias.
• Current AI models are often trained on biased data, over-representing common species.

Purpose of the Study:

• To investigate the impact of training data quantity on species recognition model performance across different taxa.
• To determine optimal data collection strategies for improving AI models in biodiversity research.
• To address the taxonomic bias in citizen science data.

Main Methods:

• Utilized a large citizen science dataset from Norway with independently collected images and identifications.
• Analyzed the influence of varying amounts of training data on species recognition model accuracy.
• Evaluated model performance for different taxonomic groups.

Main Results:

• Increasing training data generally improves recognition models, especially for under-represented taxa.
• Significant deviations from the 'more is better' paradigm were observed.
• Strategic data collection yields greater improvements than simply increasing data volume.

Conclusions:

• Focused data collection, prioritizing under-represented species, is more effective than a general increase in data.
• Improved AI models can enhance the representativeness of biodiversity data.
• This approach can help mitigate taxonomic bias in citizen science.