Resource-Efficient MobileNetV2 Model for Multiclass Plant Disease Prediction Using Real-Time Data in Smart Farming

Abstract: Agriculture remains a cornerstone of Namibia’s economy, yet small-scale crop farmers continue to face significant productivity losses due to late or inaccurate diagnosis of plant diseases. Tomato, a major crop in the country’s semi-arid regions, is highly susceptible to fungal and bacterial infections that spread rapidly under local climatic conditions. Manual inspection is labour-intensive, subjective, and ineffective for large-scale monitoring. In the literature, many studies have used high-quality datasetsto train deep learning models. However, these datasets are not real-time and rarely reflect Namibia’s specific atmospheric and climatic conditions. To address this challenge, this study uses a blended dataset combining the Plant Village Tomato Leaf Dataset from Kaggle with real-time images collected from small farms in Namibia. The study further investigates a resourceefficient and reliable deep learning model, namely MobileNetV2, for multiclass classification of plant diseases. The proposed framework using the MobileNetV2 model is benchmarked against the VGG16 and ResNet50 models, both trained and fine-tuned on the blended dataset. The models are compared in terms of the overall prediction accuracy from the multiclass confusion matrix and their computational cost. The results indicate that the proposed multiclass classification model based on the MobileNetV2 architecture has achieved the best performance near to 90% accuracy, compared to VGG16 (88.33%) and ResNet50 (58.02%), while incurring minimal computational cost. The model achieved fast predictions with reasonable accuracy, enabling mobile deployment to monitor crop health in the field. The results show that MobileNetV2 offers a low-cost way to assesstomato crop health and support farmers in Namibia using digital technologies.