Content uploaded by Zsófia Varga-Szilay
Author content
All content in this area was uploaded by Zsófia Varga-Szilay on Oct 19, 2023
Content may be subject to copyright.
CONCLUSION –Our results justify the application of computer vision-based monitoring in studying insect behaviour
and highlight them as a potentially trusted and non-destructive tool in pollinator research.
Zsófia Varga-Szilay1*, Gábor Pozsgai2
1Doctoral School of Biology, Institute of Biology, ELTE Eötvös Loránd University, Budapest, Hungary
2cE3c –Centre for Ecology, Evolution and Environmental Changes/Azorean Biodiversity Group, CHANGE –Global Change
and Sustainability Institute, Departamento de Ciências e Engenharia do Ambiente, Universidade dos Açores, Açores, Portugal
*zsofia@vargaszilay.hu, https://www.vargaszilay.hu
Flower visitation through the lens: exploring bumblebees’ behaviour
with computer vision-based application
INTRODUCTION –To understand the processes behind pollinator declines, besides community monitoring, we also have to understand the
behaviour of pollinators and the plant-pollinator interactions. In addition to traditional trapping and visual observation methods, developing fast and
reliable automated methods which are capable of both monitoring species occurrences and flower visitation events would be highly advantageous.
The buff-tailed bumblebee (Bombus terrestris)is common in Europe, and it is thus a good model species to examine whether flower visitation
behaviour can be explored with computer vision-based methods.
METHODS –The study was conducted in urban and seminatural areas of Terceira in the Azores, Portugal, during July and August 2022. We
recorded the bumblebees on pink-headed persicaria (Persicaria capitata) and red clover (Trifolium pratense) patches in five-minute-long slots at
5K resolution (30 frame per sec). For the automated bumblebee detection in the recorded videos we created computer vision models based on
YOLOv5 (You Only Look Once) deep learning algorithm, with custom datasets. To assess the similarity of the colour of a pixel to a predefined
‘optimal flower colour’(OFC), we calculated the Euclidean distance between the measured and optimal value for each pixel in the image. To
estimate how much time bumblebees spent on flowers (‘handling time’) and away from flowers (‘travelling time’), we recorded this color
similarity from areas where insects were detected and compared their standardised values between the two focal plant species.
Acknowledgements –The conference participation was supported by the Royal Entomological Society and the Talent Support Council of Eötvös Loránd University.
pollinators –behavioural ecology –plant-insect interactions –video monitoring –artificial intelligence –computer vision
RESULTS –We found that bumblebees spent significantly more time on red clover than on pink-headed persicaria (t = -253.061, p < 0.001),
which suggests that the optimum duration of stay on larger-headed Trifolium may be longer than that on the smaller-headed Persicaria.
The figures are examples of the
movement of bumblebees within a
patch of Persicaria (left) and a
Trifolium (right). Each red dot
represents a detected bumblebee. X
and Y axis units are pixel-coordinates.
Scan the QR codes in the bottom left
corner and watch the videos of
automated bumblebee detection.
The histograms show the deviation
distribution of patch colour from OFC
(green, OFC = 0.0 on the X axis) and
the deviation distribution of the colour
under bounding boxes enclosing
detected bumblebees (orange).
OBJECTIVE –Exploring and understanding the food-gathering behaviour of natural population of bumblebees in field conditions.