First results from applying novel technologies to study plant–pollinator interactions in restored sand grasslands

Monitoring plant–pollinator interactions is essential for combating the loss of biodiversity and the associated ecosystem service of pollination. Data gaps in this area partly result from the labor-intensive nature of sampling and the lack of direct observation of plant–pollinator interactions, but are most notably due to the challenging identification of highly diverse pollinating insects. To fill this gap the SEPPI project (https://seppi-pollinate.weebly.com/) aims to test automated, image-based methods using machine learning and compare them to traditional sampling in terms of accuracy, feasibility and cost. The project also seeks to evaluate the effectiveness of the new method across different biogeographical regions and various ecological gradients, such as land use, restoration and altitude. In the Pannonian ecoregion, sand grassland was selected to represent reference, control, and restored grassland. In addition to plant composition surveys, four orders of pollinating insects (mainly: bees, wasps, flies, butterflies) were captured and images were collected by automated time-laps cameras (camera-AI method) to compare the two approaches. The initial results indicate that 54 automated sampling sessions produced 6,446 pollinator images, encompassing 17 plant species (effort: avg. 2 hours per flower). In contrast, traditional netting sampling detected 1,932 plant-pollinator interactions across 31 plant species (effort: 30 minutes per transect). Specifically, traditional sampling recorded 687 individuals from 51 pollinator species in the restored areas, 670 individuals from 49 species in the control areas, and 606 individuals from 20 species in the reference areas. To assess and enhance the automated sampling, we are currently annotating the images collected during the first year and will repeat the sampling in the subsequent years. Our efforts also focus on improving the sampling protocols, sharing insights gained during the project, and advancing the camera-AI method to better recognize and identify pollinators from images, thereby enhancing its accuracy and applicability in diverse environments. These findings highlight the potential of automated, image-based methods to complement traditional sampling by providing extensive and detailed insights into plant-pollinator interactions, paving the way for more efficient and scalable biodiversity monitoring efforts in restored sand grasslands.