Framework-Dependent Quantization Stability in Agricultural Edge AI Under Resource-Constrained Deployment

Reliable edge deployment in agricultural computer vision is often hindered by accuracy degradation during model compression. This study presents a systematic audit of a 14,154-image wheat disease dataset and identifies 11.6% cross-split data leakage using perceptual hashing (pHash) and MD5 verification techniques. Three modern image classification architectures—ConvNeXt-Tiny, ResNet50, and MobileNetV3-Large—were evaluated under multiple quantization and deployment backends. Results reveal substantial performance variability across optimization strategies. CPU-based dynamic quantization using ONNX Runtime caused severe degradation in MobileNetV3-Large accuracy (85.53% to 31.04%), whereas entropy-calibrated static quantization with TensorRT restored performance to 82.54%. These findings indicate that quantization instability is strongly influenced by deployment backend selection and calibration methodology rather than being an inherent limitation of a given architecture. While ConvNeXt-Tiny achieved the highest predictive accuracy overall, MobileNetV3-Large demonstrated strong deployment efficiency when optimized with TensorRT, achieving 54.5 FPS on constrained edge hardware. This work contributes a reproducible benchmark for evaluating quantized agricultural computer vision models and highlights the importance of data integrity auditing and deployment-aware optimization in edge AI systems.