2025/10/30 by Pau Enric Serra, Albert Solé‐Ribalta, Arancha Lana +3 · 1 voice · 1 citation
Agricultural and Biological Sciences · Environmental Science · Psychology · #Plant and animal studies #Species Distribution and Climate Change #Animal and Plant Science Education
paper · pdf · doi:10.1111/2041-210x.70165
Abstract Pollinator declines pose a significant threat to ecosystem services, making effective monitoring methods critical for conservation efforts. Current research on pollination interactions remains constrained by traditional methods such as direct observations, which have limited spatial and temporal coverage and are inherently biased toward diurnal interactions. Moreover, the presence of human observers can alter pollinator behaviour, and field notebooks serve as the only permanent records, restricting data accessibility and reproducibility. To overcome these challenges, we developed an Automated Camera System (ACS) that integrates Raspberry Pi hardware with YOLOv5, a deep learning‐based object detection model, to detect pollinators in plant communities. The system was trained on video recordings of plant–pollinator interactions collected over 4 years across four different islands. Detected pollinators were subsequently classified to the species level by an entomologist, who also determined which individuals had landed on floral structures. We also conducted 13 field campaigns across six study sites over two spring seasons to monitor all flowering plants using both direct observations and multiple ACS units. This approach allowed us to compare the plant–pollinator data acquired by each method. The resulting datasets derived from ACS and direct observations shared similar characteristics (e.g. species richness and Shannon diversity ) and core interactions. However, ACS recorded higher pollinator visitation rates per individual and species than direct observations, likely due to the absence of human observers. Additionally, ACS detected a greater proportion of interactions, particularly low‐frequency ones, which significantly influenced network metrics and lead to higher connectance and specialisation and lower robustness . Despite these advantages, ACS had difficulties in detecting small‐bodied pollinators (<5 mm), highlighting an area for future refinement. This is the first study to introduce an open‐source tool for automatically detecting both nocturnal and diurnal plant–pollinator interactions across natural plant communities and to cross‐validate it with direct observations. We foresee that the proposed system has broad applications in research and conservation, providing valuable insights into pollination networks across diverse species and ecosystems.