UAV-based Remote Sensing of Bee Nesting Aggregations with Computer Vision for Object Detection

Abstract

O_LIPollinating insects are in decline globally, threatening pollination services and driving a growing interest in pollinator monitoring and conservation. However, the implementation of conservation programs for these insects is often hindered by labor-intensive monitoring methods and in turn insufficient data to assess population trends. C_LIO_LIWe detail a method for surveying and censusing ground nesting bee aggregations, pairing automated UAV image capture with a custom trained computer vision-based object detection workflow using the YOLOv5m architecture. To highlight the ease of application and accuracy of the workflow, we surveyed a roughly 65m2 portion of a large Colletes inaequalis nesting aggregation. We compared the efficiency and performance of our model to manual counts of a technician. C_LIO_LIOur model detected the location of 1,094 nests, representing 88% of the nests present in our test dataset, and a true positive rate of 97%. Adjusting for error, our model estimated a total of 1,250 nests across the study site, comparable to the total estimated from a manual count of 1,259 nests. Our model detected nests 20 times faster than the manual counts while mapping the aggregation with millimeter accuracy. Spatial analyses show that bee nest density was heterogenous, with dense spatially clustered regions comprised of upwards of 60 nests per m2. C_LIO_LISynthesis and applications: Our novel application of UAV imagery and object detection models for mapping and censusing a ground nesting bee aggregation represents a rapid, cost-effective solution for overcoming limitations in traditional manual methods. Our workflow generates essential data with the high throughput required to help inform the conservation decisions needed to stem global bee declines. C_LI