EfficientNetB0-Based Lightweight AI System for Real-Time Diagnosis of Lettuce Diseases in Resource-Constrained Agricultural Settings

Lettuce (Lactuca sativa) is critical for Libyan food security and smallholder income, but production is severely affected by fungal, bacterial, and viral diseases. Conventional diagnosis relies on subjective, time-consuming expert inspection, often unavailable in resource limited rural areas. This paper presents a lightweight Artificial Intelligence (AI) system for real time lettuce disease diagnosis, tailored to the Libyan agricultural context. Methodology: A dataset of 4,619 RGB images was constructed from public sources and field captures in northwestern Libya (2024-2025). Five deep learning architectures (Custom CNN, VGG16, ResNet50, AlexNet, EfficientNetB0) were evaluated using macro F1 score (primary fairness metric), per class recall, and CPU inference latency on a Lenovo Yoga 370 (no GPU). Lighting variability was addressed via histogram equalization and CLAHE as part of preprocessing. Key results: EfficientNetB0 achieved 90.0% test accuracy and a macro F1 score of 0.78, with 100% recall for Downy Mildew (11/11 cases) – critical for humid greenhouse systems. Inference latency was 158 ms on CPU only hardware, well below the 500 ms real time threshold. A lightweight fallback model (Custom CNN) runs in 28 ms with 89.7% accuracy. Deployment: The system is integrated into a bilingual (Arabic-English) web prototype providing agronomically validated, Libya compliant treatment recommendations. Results show that efficient deep learning can bridge the gap between AI research and practical field deployment in low resource regions.