Animal pollination is involved in the reproduction of 90 % of flowering plants. Approximately 70% of crops at global scale rely on pollinators, and growing concerns about insect decline highlight the need for effective monitoring of their activity. Traditional monitoring methods are often time-consuming and destructive. Technological advances now allow the development of passive techniques, such as computer vision and acoustic recording, combined with machine learning. These methods offer improved spatial and temporal coverage for biodiversity monitoring. Passive acoustic monitoring is particularly promising for tracking pollinators but remains underutilized and often relies on outdated machine learning approaches. Recently, deep learning methods—originally designed for image analysis—have begun to be applied to spectrograms of acoustic monitoring of various taxa, including flying insects. In this study, we propose a method for quantifying pollinator activity in sunflower fields based on the automatic detection of wingbeat sounds. We tested both a random forest and a deep learning algorithm using a new open-access software tool for acoustic biodiversity monitoring, TadariDeep. Our results show that deep learning outperforms random forest algorithms in classifying pollinator flight sounds. Comparisons with a standard visual observation protocol confirm the validity of the acoustic approach. Moreover, acoustic monitoring provides a more continuous and accurate assessment of pollinator activity than visual methods. Therefore, combining passive acoustic monitoring with deep learning presents a reliable way to assess pollinator activity at broad spatial and temporal scales. Nonetheless, further refinement is needed to improve species-level identification.
Crochard et al. (Fri,) studied this question.