A machine learning application to camera-traps: robust species interactions datasets for analysis of mutualistic networks

Abstract

Recording and quantifying ecological interactions is vital for understanding biodiversity, ecosystem stability, and resilience. Camera traps have become a key tool for documenting plant-animal interactions, especially when combined with computer vision (CV) technology to handle large datasets. However, creating comprehensive ecological interaction databases remains challenging due to labor-intensive processes and a lack of standardization. While CV aids in data processing, it has limitations, including information loss, which can impact subsequent analyses. This study presents a detailed methodology to streamline the creation of robust ecological interaction databases using CV-enhanced tools. It highlights potential pitfalls in applying CV models across different contexts, particularly for specific plant and animal species. The approach aligns with existing camera trap standards and incorporates complex network analysis tools. It also addresses a gap in ecological research by extending the methodology to behavioral studies using video-based image recognition, as most current studies rely on still images. The study evaluates CVs performance in estimating species interaction frequency (PIE) and its ecological implications, with examples of plant-frugivores interactions for seed dispersal. Results show that up to 10% of pairwise interactions may be missed with CV, with information loss varying among focal species and individual plants. This poses challenges for individual-based approaches, where unbiased data collection requires extra caution. However, the loss is minimal compared to the vast data CV enables researchers to gather. For community-level approaches, only three out of 344 unique pairwise interactions were missed, and overall estimates of both PIEs and interaction strengths remained largely unaffected. The methodology provides a valuable resource for ecologists seeking to document ecological interactions efficiently. It offers guidelines for collecting reliable data while addressing CVs limitations in capturing unbiased species interaction data. Despite its constraints, CV significantly enhances the ability to gather large-scale interaction data, particularly at the community level, making it an indispensable tool for ecological research.