Automated Detection and Monitoring of Ground‐Nesting Bee Nests Using Drone Imagery and Deep Learning

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Bees that nest in the soil in self‐excavated burrows comprise the majority of wild bee species and provide important pollination and soil functions, yet many species are threatened. Conservation efforts for ground‐nesting bees are often hindered by limited knowledge of their nesting habitat requirements, in part because nests are difficult to locate and efficient methods for monitoring nesting sites are lacking. Automated detection and monitoring of soil mounds (tumuli) produced by ground‐nesting bees, indicating nest presence, could provide new insights into bee nesting biology and population dynamics while also providing crucial data to support conservation and management. Image‐based methods, such as the automated acquisition of high‐resolution aerial imagery using drones combined with modern computer vision techniques, offer a promising path toward scalable systems for detecting and monitoring bee nest tumuli across large areas. Here, we evaluate the feasibility of integrating drone‐based image acquisition with deep learning to detect tumuli representing bee nests and to distinguish them from other soil surface deposits, such as earthworm casts. We demonstrate this approach on a 120 m 2 area of a densely populated nesting aggregation of Lasioglossum malachurum on bare soil containing numerous earthworm casts. Our model reliably detected bee nest tumuli, achieving an F 1 score of 0.90 (precision: 0.89, recall: 0.91). Misclassifications mainly arose from atypically shaped tumuli (e.g., new and incomplete or damaged), and from cases where tumuli overlapped, but no earthworm casts were confused for bee nest tumuli. This pilot study represents a step toward more efficient monitoring of ground‐nesting bees and demonstrates the potential of this approach under specific conditions. Future work could evaluate its applicability across additional habitats and species and explore alternative methods, such as image segmentation, which may be better suited for cases with less distinct tumuli or extensive overlap among soil mounds.
Funding Information
  • Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (185273)

This publication has 26 references indexed in Scilit: