Integrating computer vision, machine learning and technologies for automated monitoring of pollinators

Abstract

This dissertation investigates how accessible technologies, particularly smartphones combined with deep learning, can be applied to automated pollinator monitoring. Using time-lapse photography, a large annotated image dataset of flower visitors was created (Chapter 2). Citizen science image repositories were then used to train lightweight YOLO object detection models that classify eight major arthropod groups known to visit flowers in Europe (Chapter 3). Finally, these models were evaluated on out-of-distribution field images, revealing performance declines for small, blurred insects and mimicry cases, providing a framework for realistic assessment of model generalisation (Chapter 4). The findings contribute to bridging the gap between computer vision research and scalable pollinator monitoring.