<p>Polymetallic nodules, rich in critical metals, are a vital resource for deep-sea mining and sustainable energy applications. Traditional detection techniques, including manual surveys and early image processing methods, struggle with the low visibility, non-uniform textures, and high pressure of deep-sea environments. While deep learning models have improved detection accuracy, they often lack a lightweight design, robust feature attention, and scalability for real-time deployment. To address these limitations, we propose NodYOLO-GAM, a novel deep learning architecture that integrates a Global Attention Mechanism (GAM) into a YOLO-based backbone for the non-destructive detection of polymetallic nodules. The proposed model enhances both spatial and channel-wise feature extraction, supported by Multi-Scale Feature Fusion (MSF) using upsampling and SPPF (Spatial Pyramid Pooling Fusion) modules. It is trained and evaluated on our newly introduced NoD-Sea dataset, comprising annotated low-resolution RGB underwater images. NodYOLO-GAM achieves a mAP@50 of 69.8%, outperforming state-of-the-art models in precision and speed. The novelty of this study lies in its lightweight attention-enhanced architecture tailored for real-time AUV deployment, and in the creation of the NoD-Sea dataset for benchmarking. This work contributes to scalable, sustainable deep-sea exploration, with potential relevance to ocean conservation initiatives.</p>