Global food security relies heavily on rice production, which is continuously threatened by crop diseases. Although deep learning (DL) models have demonstrated strong diagnostic performance, their high computational and memory demands limit deployment on low-power edge devices for in field applications. This study presents a systematic empirical evaluation of model quantization techniques for enabling efficient edge deployment of agricultural AI systems. Three widely used convolutional neural network (CNN) architectures InceptionV3, ResNet50 and MobileNetV2 were analyzed under four optimization schemes: full-precision (FP32), Quantization-Aware Training (QAT), Post-Training Static Quantization (PTQ-Static) and Post-Training Dynamic Quantization (PTQ-Dynamic). Experiments were conducted on the public Paddy Doctor dataset, which contains rice leaf images captured under diverse real-world field conditions. In addition, a conceptual framework for an integrated precision agriculture system is proposed, combining the optimized vision-based disease classifier with a real-time IoT sensor network for monitoring key soil and environmental parameters. An energy-efficient Lightweight Secure Data Transmission (LSDT) scheme using AES-128 and ECC ensures secure, end-to-end data transmission for the constrained IoT nodes. Model interpretability was enhanced using explainable AI (XAI) through Gradient-weighted Class Activation Mapping (Grad-CAM).The QAT-optimized InceptionV3 model achieved the best overall performance, reaching a classification accuracy of 97.68%, exceeding its FP32 baseline.This had been alongside 75 percent model size reduction and twofold inference latency which proved to be highly suitable in edge deployment.The findings prove Quantization-Aware Training to be a strong and stable technique used to develop model-forceful agricultural vision systems. The work offers realistic points of reference and core issues to bake in the implementation of edge AI systems in precision agriculture.