Optimised MobileNet for very lightweight and accurate plant leaf disease detection

Abstract

Abstract The development of accurate and efficient plant disease classification systems is vital for addressing the challenges of climate change and the growing global demand for food. This study presents $$\hbox {V}^2$$ PlantNet, a novel lightweight multi-class classification model based on a modified MobileNet architecture, designed to detect plant leaf diseases across a diverse range of crop types. $$\hbox {V}^2$$ PlantNet employs depthwise separable convolutions to significantly reduce model complexity without compromising accuracy. The architecture integrates Batch Normalization (BN) and Rectified Linear Unit (ReLU) activation after each convolutional layer, while a multi-stage design enhances feature extraction and overall performance. Despite its compact size, comprising only 389,286 parameters and requiring just 1.46 MB of memory, $$\hbox {V}^2$$ PlantNet achieved up to 99% training accuracy, with validation and test accuracies of 97% and 98%, respectively. Across most classes, precision, recall, and F1-scores ranged from 0.97 to 1.0, demonstrating consistent and robust generalization across diverse plant species. These architectural innovations enable $$\hbox {V}^2$$ PlantNet to outperform larger models such as ResNet-50 and Inception V3 in terms of computational efficiency, owing to its smaller model size (1.46 MB), reduced parameter count (389,286), and faster inference time (0.676 s), offering a scalable solution for real-time plant disease detection in precision agriculture.

Code and data availability

The paper's Data Availability statement explicitly links the PlantVillage dataset used to train and evaluate PlantNet, hosted publicly on Kaggle. No author code or trained model checkpoint is deposited.