Weeds significantly reduce agricultural yields. This study focuses on using artificial intelligence, specifically deep learning, for early-stage weed detection on edge devices, which have limited computational power. We propose a three-part strategy to address this challenge: employing lightweight architectures to reduce the model size and computational demand; using transfer learning to overcome the limitations of small data sets; and applying Bayesian optimization to fine-tune model parameters. Our results show that MobileNetV2 and EfficientNetV2B0 models achieve high accuracy (95.09% and 95.79% respectively) with MobileNetV2 being nearly as accurate but much smaller in size (13.96 MB compared to 23 MB). This demonstrates MobileNetV2's suitability for computationally constrained edge devices.