A Scalable Distillation and Metric-Learning Pipeline for Adaptive Weed Classification on Edge Devices

Smart agriculture increasingly relies on automated weed detection to reduce inputs and labor. Deploying deep learning on edge devices is difficult due to limited compute and evolving weed classes. We propose a pipeline that combines partial fine-tuning of an EfficientNet-B7 teacher, embedding level distillation into lightweight students (MobileNetV3, ShuffleNetV2, EfficientNet-B0), and semi-hard triplet metric learning. The system learns 2048D embeddings and is evaluated with N-way/K-shot episodes to mimic a few-label condition. Dynamic INT8 quantization enables CPU-only inference with minimal accuracy loss. The approach adapts rapidly to novel species with few labels while meeting real-time edge constraints, supporting sustainable herbicide management in practice.