Deep learning based mapping of bee-friendly trees through remote sensing: A novel approach to enhance pollinator conservation

The global decline in wild bee populations poses significant risks to ecosystem stability given bees' essential role as pollinators. Conserving bee-friendly habitats is critical for the promotion of wild bees and prevention of further losses, which requires a good understanding of the bee's ecology. This study explores the relationship between nesting sites of the ground-nesting bee Andrena vaga and the distribution of Salix trees, an essential pollen source for this and other bee species, within the city of Braunschweig, Germany. Our approach integrates multi-temporal PlanetScope imagery, a tabular transformer deep learning model, and a LiDAR-derived 3D tree model to automate the mapping of Salix trees. The mapping achieved an F1 score of 0.73 (precision: 0.69, recall: 0.78). Field surveys were conducted, documenting Andrena vaga nest aggregations and aggregation sizes. On average, the nearest Salix tree was located approximately 150 m from an aggregation, while the nearest five trees were within 300 m. Literature-guided estimates of the required Salix density for a given aggregation size showed that, on average, the theoretically necessary number of trees was found within 300 m, though for one aggregation the distance exceeded 1000 m. While overall the number of Salix trees around nest aggregations seems to increase with aggregation size, the relationship did not prove to be statistically significant. Our study illustrates the potential of remote sensed based mapping of tree species to enhance our understanding of floral resource availability in pollinator habitats, thereby supporting informed conservation of essential resources for bees and other insects. It also highlights how advances in remote sensing can play an important role in ecological research and habitat conservation.