Accurate detection and identification of insects from camera trap images with deep learning

• DOI: 10.1371/journal.pstr.0000051
• Publication Date: 2023-03-15T17:35:37Z
• Authors: Kim Bjerge; Jamie Alison; Mads Dyrmann; Carsten Eie Frigaard; Hjalte M. R. Mann; Toke Thomas Høye
• Keywords: Identification; Citizen Science

Reported insect declines have dramatically increased the global demand for standardized monitoring data. Image-based can generate such data cost-efficiently and non-invasively. However, extracting ecological from images is more challenging insects than vertebrates because of their small size great diversity. Deep learning facilitates fast accurate detection identification, but lack training coveted deep models a major obstacle application. We present large annotated image dataset functionally important taxa. The primary consists 29,960 representing nine taxa including bees, hoverflies, butterflies beetles across two million recorded with ten time-lapse cameras mounted over flowers during summer 2019. was extracted using an iterative approach: First, preliminary model identified candidate insects. Second, were manually screened by users online citizen science platform. Finally, all annotations quality checked experts. used to train compare performance selected You Only Look Once (YOLO) algorithms. show that these detect classify in complex scenes unprecedented accuracy. best performing YOLOv5 consistently identifies dominant species play roles pollination pest control Europe. reached average precision 92.7% recall 93.8% classification species. Importantly, when presented uncommon or unclear not seen training, our detects 80% individuals usually interprets them as closely related This useful property (1) rare which are absent, (2) new correctly identify those future. Our camera system, framework promising results non-destructive Furthermore, resulting quantify phenology, abundance, foraging behaviour flower-visiting Above all, this represents critical first benchmark future development evaluation identification.