<p>Weed classification is a significant challenge in precision agriculture, particularly in resource-constrained agricultural settings with limited access to rich annotated datasets. Deep learning models are computationally intensive in this setting due to their heavy reliance on richly annotated data. Few-shot learning (FSL) provides a better option by allowing accurate classification from a few labelled instances. This paper introduces the hybrid attention-based prototypical network (HAPN), a state-of-the-art FSL model that enhances feature representation through an attention mechanism and adaptive prototype estimation. The proposed approach significantly enhances feature embeddings, providing better generalization in low-data settings. The study evaluates HAPN against baseline FSL models on the real-world weed benchmark, the Deep Weeds dataset, and establishes its superiority. Experimental evaluations demonstrate that HAPN substantially enhances classification accuracy, providing a viable solution to real-time weed classification in resource-constrained agricultural settings. Integrating attention within HAPN is a simple yet effective modification that enhances performance without significantly increasing computational complexity.</p>