Rough Grouping Enhances YOLO's Pollinator Classification and Detection from Small Datasets

Abstract

This study addresses the challenges in pollinator monitoring by proposing an effective data structure for automated systems, with a focus on the use of machine learning to handle underrepresented groups in small datasets. By experimenting with grouping the top three pollinators (bee, butterfly, hoverfly) and non-pollinators in datasets of fewer than 300 samples, the research aims to enhance classification and detection accuracy. During 4-hour filming sessions, 181 images of insects larger than 1 cm were captured and classified into three grouping methods: "Pollinator/ Non-pollinator", "Bee/Butterfly/Hoverfly/Ant", and "Bumblebee/Honeybee/Butterfly/Hoverfly/ Ant". YOLO V8 models were trained, validated, and tested with these datasets based on different class grouping methods. The study found that the "Pollinator/Non-pollinator" YOLOv8 model performed best across all metrics, suggesting it is more reliable for categorizing groups and detecting target objects, especially with smaller, imbalanced datasets. This finding aligns with the trend in machine learning that providing more training opportunities for individual classes improves accuracy. Therefore, using broader categorization methods can enhance the reliability and accuracy of automated monitoring systems when training data is insufficient.