Abstract: Agriculture plays a critical role in global food security, yet crop diseases continue to cause significant eco- nomic losses worldwide. Traditional deep learning models, while achieving high accuracy in disease detection, face substantial limitations when deployed for real-time, in-field applications due to their computational complexity and large memory requirements. This research addresses these challenges by developing a lightweight, quantized model specifi- cally designed for edge device deployment. We propose applying model compression techniques through quantization on MobileNetV2, a lightweight neural network architecture, to create an efficient model suitable for resource-constrained environments. Our methodology involves comprehensive comparison of Post-Training Quantization (PTQ) and Dynamic Range Quantization (DRQ) techniques applied to rice leaf disease classification. The results demonstrate a significant reduction in model size from approximately 9 MB to 2.5 MB while maintaining acceptable accuracy levels. The DRQ model achieved 92.23% accuracy with an F1-score of 0.9212, compared to the original model’s 94% accuracy, repre- senting a minimal 1.77% accuracy trade-off for a 72% size reduction. These findings highlight the practical viability of quantized models for automated disease detection systems in precision agriculture, enabling real-time deployment on smartphones and embedded devices for farmers in remote locations.

Keywords: Crop disease classification, edge computing, model quantization, MobileNetV2, precision agriculture, deep learning compression.

Downloads: | DOI: 10.17148/IJIREEICE.2025.131006