A Machine Learning Application to Camera-Traps: Robust Species Interactions Datasets for Analysis of Mutualistic...

Area of Science:

• Ecology
• Biodiversity Research
• Computational Biology

Background:

• Ecological interaction data is crucial for understanding biodiversity and ecosystem stability.
• Camera traps combined with computer vision (CV) offer a powerful method for documenting plant-animal interactions.
• Current methods for creating ecological interaction databases are often labor-intensive and lack standardization.

Purpose of the Study:

• To present a detailed methodology for creating robust ecological interaction databases using CV-enhanced tools.
• To highlight potential pitfalls and limitations of CV models in ecological contexts, particularly for specific species.
• To extend current methodologies to behavioral studies using video-based image recognition.

Main Methods:

• Development of a streamlined methodology for ecological database creation using CV-enhanced tools.
• Evaluation of CV model performance in estimating plant-animal interaction frequency (PIE).
• Application of complex network analysis tools and comparison with existing camera trap standards.

Main Results:

• CV may miss up to 10% of pairwise interactions, with variation depending on species and context.
• Information loss from CV is minimal compared to the vast data acquired, especially for community-level analyses.
• Community-level estimates of PIE and interaction strengths remained largely unaffected by CV data loss.

Conclusions:

• The proposed methodology efficiently supports the creation of ecological interaction databases.
• CV significantly enhances the capacity for large-scale ecological data collection, proving indispensable for community-level research.
• Guidelines are provided for collecting reliable data while acknowledging and addressing CV's limitations in capturing unbiased interaction data.