A Comparative Analysis of Lightweight Deep Learning Models for Real-Time Apple Orchard Monitoring

Abstract: Apple cultivation represents a significant global investment; profit losses result when growers are unable to detect leaf disease at early stages. Deep learning has accelerated disease diagnosis, although a deployment gap for mobile devices remains. The majority of models are not efficient enough to run on the standard smartphones that farmers actually own. For this reason, this research evaluates the practical suitability of three deep learning models: EfficientNet-B0, MobileNetV2 and VGG16. To determine the actual trade-offs between the accuracy and power drain of a model, this research trained each model on a dataset containing apple scab, black rot, and cedar apple rust. The highest validation accuracy achieved by EfficientNet-B0 was 97.8% and F1-score of 0.977. MobileNetV2, however, was more feasible for edge deployment owing to its far fewer parameters. This study provides a scalable, real-time apple orchard monitoring solution by evaluating the on-device performance of these models.

Keywords: Apple leaf disease detection, Deep learning, EfficientNet-B0, MobileNetV2, Lightweight model, Apple scab, Black rot, Precision agriculture.