EMBEDDED RESIDUAL NEURAL NETWORKS FOR REAL-WORLD PLANT DISEASE IDENTIFICATION IN DIGITAL AGRICULTURE

This study addresses the challenge of real-time plant disease identification on resource-constrained embedded platforms, a critical need for improving agricultural productivity. Using the NVIDIA Jetson Orin Developer Kit and the PlantVillage dataset, the research evaluates Residual Neural Networks (ResNets), focusing on ResNet-50, ResNet-101, and ResNet-152. The study highlights the balance between model depth, batch size, accuracy, and computational efficiency. ResNet-101, optimized with a batch size of 64, achieved 90.62% accuracy and an average identification time of 17.6 milliseconds, emerging as the most effective configuration. These findings demonstrate the feasibility of deploying deep learning models on embedded devices and provide insights into optimizing architectures for real-time agricultural applications.