Comparative and edge-hybrid modeling of EfficientNetV2 and MobileNetV2 for multi-class crop disease classification with statistical validation

Abstract

Crop disease classification is critical for global food security, yet deploying accurate deep learning models on resource-constrained edge devices remains challenging. This study systematically compares EfficientNetV2 and MobileNetV2 while proposing an edge-optimised hybrid architecture integrating both with vision transformers (ViT). Evaluated on PlantVillage and field-collected images, MobileNetV2 demonstrated superior edge compatibility with 99.0% accuracy, 0.0938 s/image inference speed, minimal resources (30.38 MB size), and statistical superiority (z-test p=0.0071). The hybrid model combines MobileNetV2’s texture analysis and EfficientNetV2’s multiscale detection through a dual-branch architecture enhanced with SE blocks, ViT (16×16 patches), and attention-guided fusion. It achieved 99.5% test accuracy with real-time performance (0.15 s/image) and 97.97% field accuracy via Android deployment. Statistical validation confirmed robustness: Kruskal-Wallis H=597.40 (p<0.05), near-perfect AUC (0.999998), and minimal confidence variance (0.000010). Ablation studies verified architectural efficacy (98.68% accuracy with SE/gating modules). This work advances precision agriculture through a scalable framework unifying hybrid deep learning with edge-compatible deployment.