Distilled Ensembles and Quantized Prototypical Networks for Adaptive, Edge-Based Weed Classification

Precision agriculture increasingly relies on automated weed detection to optimize herbicide usage and reduce environmental impact. This paper tackles edge-device constraints and domain shifts by combining knowledge distillation, few-shot meta-learning, and quantization. An EfficientNet-B7 teacher network provides high-quality features, which are distilled into an ensemble of three lightweight student networks, thereby inheriting strong representational power with minimal overhead. A weighted ensemble merges their outputs into a single embedding, enabling rapid adaptation to new weed species using only a handful of annotated examples. Dynamic quantization further reduces the model footprint, making it practical for resource-constrained devices. Experimental results on varied weed datasets confirm robust performance under real-world conditions. By uniting teacher–student distillation, ensemble embeddings, few-shot adaptation, and quantization, the methodology advances precision agriculture across diverse field scenarios. Additionally, our approach surpasses existing classification benchmarks. This synergy fosters advanced weed classification on constrained hardware deployments ensuring more sustainable farmland management practices.