Towards Computer Vision and Deep Learning Facilitated Pollination Monitoring for Agriculture

Abstract

Globally, pollinators affect 35% of agricultural land and play a key role in food production. Consequently, monitoring is useful to understand the contribution insects make towards crop pollination. Traditional sampling techniques used in insect monitoring have several drawbacks, including that they are labour intensive and potentially unreliable. Some of these drawbacks may be overcome using computer vision and deep learning-based approaches to automate pollination monitoring. In this paper, we present a pipeline for computer vision-based pollination monitoring and propose a novel algorithm, Polytrack, that tracks multiple insects simultaneously in complex agricultural environments. Our algorithm uses deep learning and fore-ground/background segmentation to detect and track in-sects. We achieved precision and recall rates of 0.975 and 0.972 respectively when monitoring honeybees foraging in our test sites within the polytunnels of an industrial straw-berry farm. Polytrack includes a flower identification module to automate collection of insect-flower interaction data, and a low-resolution processing mode that reduces computational demands placed on the processor to bring the software towards the requirements of low-powered monitoring hardware.