Modification of the YOLOv8 Model for Inference of Disease Detection and Nutritional Deficiency in Corn Plants on Edge Computing Devices

Abstract

Early detection of diseases and nutritional deficiencies in plants is crucial in supporting precision agriculture, particularly in corn cultivation. This study proposes modifying the YOLOv8 architecture with patch-based image segmentation techniques to improve edge computing devices' detection accuracy and inference efficiency. Four model configurations were evaluated: YOLOv8 baseline, YOLOv8 with patches, modified YOLOv8, and the proposed model, which is modified YOLOv8 with patches. Experimental results show that the proposed model achieves the best detection performance, with a mean mAP@0.5 (mAP@0.5) of 87%, precision of 0.98, recall of 0.77, and the highest F1-Score of 0.89. Additionally, this model demonstrates superior inference efficiency, with an execution time of 1.9 ms/image on an NVIDIA T4 GPU, 21 seconds/image on a Raspberry Pi, and 121.1 ms/image on a Jetson Nano. Memory consumption is also the lowest, at 308.62 MB on Raspberry Pi and 278.82 MB on Jetson Nano. These results indicate that modifying YOLOv8 with patch-based preprocessing improves detection accuracy and enables practical deployment on edge devices with limited computational resources. This makes the model highly suitable for real-time agricultural monitoring applications in resourceconstrained environments.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint