A YOLO-Based Pest Detection System for Precision Agriculture

TLDR

Early pest detection is essential for designing effective crop defense strategies in precision agriculture. The study proposes a data‑driven pest detection system for hazelnut orchards, collecting a custom outdoor dataset and training a YOLO‑based CNN to achieve approximately 94.5 % average precision. The system focuses on true bugs, trains a YOLO‑based CNN on the custom dataset, and evaluates performance using data augmentation and depth information. Deployment on a NVIDIA Jetson Xavier reaches about 50 fps, enabling real‑time on‑board processing for robotic platforms.

Abstract

In this work, inspired by the needs of the H2020 European project PANTHEON for the precision farming of hazelnut orchards, we propose a data-driven pest detection system. Indeed, the early detection of pests represents an essential step towards the design of effective crop defense strategies in Precision Agriculture (PA) settings. Among the possible pests, we focus on true bugs as they can heavily compromise hazelnut production. To this aim, we collect a custom dataset in a realistic outdoor environment and train a You Only Look Once (YOLO)-based Convolutional Neural Network (CNN), achieving ≈ 94.5% average precision on a holdout dataset. We extensively evaluate the detector performance by also analyzing the influence of data augmentation techniques and of depth information. We finally deploy it on a NVIDIA Jetson Xavier on which it reaches ≈ 50 fps, enabling online processing on-board of any robotic platform.