Optimizing Lightweight Plant Disease Diagnostics: A PSO-Tuned MobileNetV2 Approach Across Multi-Domain Datasets

Abstract

Plant diseases affect more than 80% of global crops annually, causing up to 40% yield loss and threatening global food security. Traditional detection methods depend heavily on manual inspection, which delays intervention and limits scalability. While deep learning has transformed automated diagnostics, the trade-off between model complexity and edge-device deployability remains a critical challenge. This paper proposes a hybrid AI framework combining the efficiency of MobileNetV2 with the global search capabilities of Particle Swarm Optimization (PSO) for hyperparameter tuning. The primary contribution is a rigorous multi-domain evaluation spanning three dataset types-controlled laboratory images (PlantVillage), real-world noisy field images (PaddyDoc), and a mixed-domain hybrid dataset (PlantDoc)-a setting that existing literature largely overlooks by testing on a single domain. PSO tuning raised accuracy from 95.28% to 96.45% on the lab dataset, from 68% to 71% on the field dataset, and from 50.40% to 52.80% on the hybrid dataset, while requiring only a one-time $\mathbf{1 4 0}$-minute offline run with no impact on inference latency. 8-bit Integer (INT8) quantization further reduced model size from 14 MB to approximately 3.8 MB with under 0.3% accuracy loss, enabling sub-50 ms inference on entry-level smartphones and demonstrating the framework's practical readiness for edge deployment in agricultural settings.