Work

Insectopedia: An On-Device Mobile Pipeline for Fine-Grained Agricultural Pest Identification

Year: 2026
Type: conference-paper
Abstract: Insectopedia is a machine learning-based mobile application for fine-grained agricultural pest identification that runs entirely on-device, requiring no internet connectivity. The application integrates a three stage computer vision pipeline, YOLO26 for detection, RepViT-SAM for segmentation, and group-specific ResNet18 for classification, exported to ONNX and deployed natively on consumer mobile hardware via Flutter. Achieving 86.09% species-level accuracy across top 20 Ontario agricultural pest species and 9 taxonomic groups on a held-out test set of 1,524 field images, the system demonstrates that a purpose-built lightweight pipeline can substantially outperform general-purpose cloud vision models while operating entirely of-fline. In a zero-shot benchmark against five state-of-the-art vision language models including Claude Sonnet 4.6, GPT-5.1, GPT-5.4, Gemini 3.1 Flash-Lite, and Claude Haiku 4.5, Insectopedia outperforms all evaluated models by at least 25 percentage points in species-level accuracy. The application further incorporates a human-in-the-loop confirmation mechanism and local scan logging to support continuous field data collection for future model improvement.
Language: English
Subfield: Plant Science
Open Access status: hybrid
OpenAlex ID: W7204246049
FWCI: 9.042
Cites: 7
Cited by: 2
Related to: 10