Barnyard grass, a pernicious weed thriving in rice fields, poses a significant challenge to agricultural productivity. Detection of barnyard grass before the four-leaf stage is critical for effective control measures. However, due to their striking visual similarity, separating them from rice seedlings at early growth stages is daunting using traditional visible light imaging models. To explore the feasibility of hyperspectral identification of barnyard grass and rice in the seedling stage, we have pioneered the DeepBGS hyperspectral feature parsing framework. This approach harnesses the power of deep convolutional networks to automate the extraction of pertinent information. Initially, a sliding window-based technique is employed to transform the one-dimensional spectral band sequence into a more interpretable two-dimensional matrix. Subsequently, a deep convolutional feature extraction module, ensembled with a bilayer LSTM module, is deployed to capture both global and local correlations inherent within hyperspectral bands. The efficacy of DeepBGS was underscored by its unparalleled performance in discriminating barnyard grass from rice during the critical 2-3 leaf stage, achieving a 98.18% accuracy rate. Notably, this surpasses the capabilities of other models that rely on amalgamations of machine learning algorithms and feature dimensionality reduction methods. By seamlessly integrating deep convolutional networks, DeepBGS independently extracts salient features, indicating that hyperspectral imaging technology can be used to effectively identify barnyard grass in the early stages, and pave the way for the development of advanced early detection systems.