Lab-to-Field Quantization Robustness

Deploying deep learning models for plant disease recognition on resource-constrained agricultural edge devices requires balancing computational efficiency with robustness to real-world environmental degradation. In this work, we present a systematic lab-to-field evaluation of quantization strategies for plant disease classification, examining the trade-off between model efficiency and robustness under realistic agricultural field conditions. Using a ResNet-18 model trained on the 38-class PlantVillage dataset, we compare four configurations: FP32, Dynamic INT8, Post-Training Quantization (PTQ INT8), and Quantization-Aware Training (QAT INT8). We evaluate model size, single-sample CPU inference latency, throughput, and clean classification accuracy, together with robustness under eight simulated field corruptions, including canopy shadowing, solar contrast changes, soil dust splatter, lens dust, out-of-focus blur, and CMOS sensor noise. Full INT8 quantization through PTQ and QAT reduces model size by approximately 3.96× and increases single-thread CPU throughput by approximately 1.8× without reducing clean accuracy. Furthermore, QAT improves robustness to several structural field corruptions, including soil dust splatter and lens blur, suggesting that quantization-aware training can provide a regularization benefit beyond computational efficiency. However, all evaluated models exhibit severe degradation under CMOS sensor noise, indicating that quantization alone cannot address all field deployment challenges. Our findings highlight the importance of combining efficient INT8 models with noise-aware preprocessing or denoising strategies for reliable agricultural edge AI deployment.