Plant disease classification using lightweight deep learning architectures

Abstract

Detecting early-stage pathogens of plant diseases is a challenging task due to the low pathogen density and the subtlety or absence of visible symptoms. If not identified at an early stage, plant diseases can progress rapidly, making them significantly harder to control. Therefore, real-time and in-situ monitoring of plants is essential for tracking diseases and enabling early intervention. Within the scope of smart agriculture applications, crop health can be assessed and productivity can be maintained through mobile devices. In this study, the performance of lightweight neural network architectures on various plant disease datasets was compared, targeting devices with limited memory resources such as mobile devices and embedded systems, which we consider potential candidates for edge deployment based on the measured accuracy-efficiency trade-offs. The architectures evaluated include SqueezeNet, MobileViT, EdgeNeXt, GhostNet, EfficientNet, and LeViT. The plant datasets used in the experiments were Plant Disease, Cassava, and Turmeric. Each model was evaluated using Accuracy, F1-Score, Precision, and Recall metrics. Among the three datasets, the best performance was achieved on the Turmeric dataset using the GhostNet architecture, with scores of 0.9978 (Accuracy), 0.9976 (F1-Score), 0.9976 (Precision), and 0.9976 (Recall), respectively.